Commun. Math. Phys. 207, 1 – 42 (1999)
Communications in
Mathematical Physics
© Springer-Verlag 1999
On the Discrete Spectrum of the Nonstationary Schrödinger Equation and Multipole Lumps of the Kadomtsev–Petviashvili I Equation Javier Villarroel1 , Mark J. Ablowitz2 1 Universidad de Salamanca, Dept. de Matematicos puras y aplicadas, 37008 Salamanca, Spain 2 Department of Applied Mathematics, University of Colorado, Boulder, CO 80309-0526, USA
Received: 30 September 1998 / Accepted: 30 March 1999
Abstract: The discrete spectrum of the nonstationary Schrödinger equation and localized solutions of the Kadomtsev–Petviashvili-I (KPI) equation are studied via the inverse scattering transform. It is shown that there exist infinitely many real and rationally decaying potentials which correspond to a discrete spectrum whose related eigenfunctions have multiple poles in the spectral parameter. An index or winding number is asssociated with each of these solutions. The resulting localized solutions of KPI behave as a collection of individual humps with nonuniform dynamics. 0. Introduction In this paper we study the discrete spectrum of the nonstationary Schrödinger equation (cf. Eq. (1.1) below). It is well known that this operator can be used to linearize the Kadomtsev–Petviashvili-I (KPI) equation (cf. Eq. (1.5) below) via the inverse scattering transform (IST) (see e.g. [1]). Both of these equations have wide application in physics. The Schrödinger equation is central in the study of quantum mechanics and the KP equation is a basic nonlinear equation that models long two dimensional waves with weak nonlinearity and slow transverse variation. The KP equation arises in water waves, plasma physics, as well as having important mathematical implications. Both the continuous and discrete spectrum of the stationary Schrödinger operator have been extensively investigated. Far less is known in the nonstationary case. With regard to the discrete spectrum, corresponding to decaying potentials, it is known [2] that there exist complex conjugate pairs of simple eigenvalues of the nonstationary Schrödinger operator which, via IST, yield rational slowly decaying (they decay as O(1/r 2 ) where r 2 = x 2 + y 2 ) soliton solutions, usually referred to as lumps, of KPI. These lump solutions have been heavily studied since they were first found by direct methods [3]. Here we show that there exist infinitely many real, rationally decaying potentials of the nonstationary Schrödinger operator and corresponding solutions to KPI equations. As functions of the spectral parameter, the analogue of the “energy” in the stationary case,
2
J. Villarroel, M. J. Ablowitz
the related eigenfunctions (or wavefunctions) have, in general, multiple poles. Thus the potentials correspond to a discrete spectrum associated with poles of the eigenfunction of multiplicity m ≥ 1. This is in contrast to the stationary case where all real localized potentials correspond to simple poles. We have found that to characterize the discrete spectrum of this operator it does not suffice to require that the potential vanishes at infinity with a certain decay rate (like O( r12 )). Additional restrictive conditions must be given which are expressed in terms of new integer valued quantities. We discuss the characterization of the potential in terms of the pole structure of the eigenfunction and this additional number, which we refer to as the charge which has an implicit interpretation as an index or winding number (see Sect. 3; result V, part iii, where using Green’s theorem we prove that they are winding numbers). We note, as with the ordinary lump solutions, the potentials found here are not absolutely integrable, decaying as O(1/r 2 ), where r 2 = x 2 + y 2 . This slow decay is the underlying reason for the appearance of the charge. We also note that these results are relevant to, and many of the methods can be carried over to, other completely integrable partial differential equations in 2 + 1-dimensions such as the Davey-Stewartson equations. The resulting class of potentials is related to analytic and rationally decaying solutions in the KPI equation when the appropriate temporal dependence is inserted. These solutions of KPI are termed multipole solutions when m > 1. They can be seen as a collection of individual humps that generically evolve with nonuniform dynamics and display a behaviour that is in contrast to the uniform motion of standard solitary waves; they interact in a nontrivial manner (see also ref. [6]). Direct methods for finding rational solutions of KPI and connection with Calogero-Moser systems have also been considered in [10], though the full manifold of physically relevant solutions (real and nonsingular) was not obtained. In a subsequent paper we shall show how the problem of classifying the various infinite families of real nonsingular localized solutions to KPI, can be found by direct methods (cf. ref. [11]). The results discussed here were announced in ref. [4] (see also ref. [9]). The purpose of the present paper is to derive and elaborate on these results. Several examples are considered. The framework of the paper is as follows: in Sect. 1 we review the the known results of IST for the case of eigenfunctions with only simple poles. New results regarding the discrete spectrum are presented in Sects. 2–3 with relevant proofs (the proofs in these two sections can be skipped by a reader wishing to only understand the basic phenomena). Section 4 employs the results of the previous sections to obtain concrete examples of reflectionless potentials. Section 5 discusses how some of these results can be obtained by coalescence of poles. Examples of corresponding decaying solutions of the KPI equation are given in Sect. 6 where we also study the trajectories and dynamics of the humps that each solution contains. Galilean interactions that correspond to the above lump trajectories are also derived. 1. Some Previous Results The nonstationary Schrödinger equation is iψy + ψxx + uψ = 0. We introduce the complex spectral parameter k via ψ = µ(x, y, k)eikx−ik that µ solves (i∂y + ∂xx + 2ik∂x + u)µ = 0.
(1.1) 2y
and find (1.2)
Discrete Spectrum of Nonstationary Schrödinger and Kadomtsev–Petviashvili Equation
3
In this paper we restrict ourselves to the study of the class u(x, y) of potentials that are real, nonsingular and decay rationally at infinity. u(x, y) is reconstructed from solutions of (1.2) from (see [1] for this and other facts below) u(x, y) = −2i
∂ µ1 , ∂x
(1.3)
where µ = 1 + µ1 /k + O( k12 ) as k → ∞. The compatibility of (1.2) and (∂t + 4∂xxx + 12ik∂xx − 12k 2 ∂x + 6u∂x + 3ux + 6iku − 3i(∂y ∂x−1 u))µ = 0 (1.4) yields the KPI equation: (ut + 6uux + uxxx )x = 3uyy .
(1.5)
The knowledge of the Schrödinger potential u(x, y) and corresponding eigenfunction µ(x, y, k) allows one to obtain solutions to KPI by letting certain constants (the scattering data) evolve in time t in a precise way as required through compatibilty with (1.4). Similarly, setting t = constant in the relevant KPI solution u(x, y, t) one obtains a Schrödinger potential for which (1.1,1.2) is solvable. If the potential is regular and has suitable decay (1.2) can be converted into the integral equation Gµ(x, y, k) = 1, (1.6) where the operator G(k) acts as follows: Z ∞ Z ∞ dx 0 dy 0 G(x − x 0 , y − y 0 , k)(uf )(x 0 , y 0 ). (1.7) Gf (x, y) = f (x, y, k) − −∞
−∞
The Green’s function of the problem has different forms for Im k > 0, Im k < 0. Indeed one has that G(x, y, k) is given on the upper-lower half planes by Z ∞ Z ∞ 1 dm dn exp [i(mx + iny)] /((m2 + 2mk + n) (1.8) G± (x, y, k) = (2π)2 −∞ −∞ Z ∞ h i 1 exp i(mx − (m2 + 2km)y) (θ (y)θ (∓m) − θ (−y)θ (±m))dm. = − 2π i −∞ This function satisfies Z ∞ h i 1 sign(m) exp i(mx − (m2 + 2km)y) , Im k = 0, G+ − G− = 2πi −∞ 1 ∂G(x, y, k) = −i(x − 2ky)G(x, y, k) − sign (Im k), Im k 6 = 0, ∂k 2π i ¯ ¯ = G(−x, −y, k), G(x, y, k)
(1.9.a) (1.9.b) (1.9.c)
where bar denotes complex conjugate. We consider solutions µ to (1.6) that correspond to regular and decaying potentials u(x, y) tending to 1 as |k| → ∞. Generically as functions of the complex parameter k, µ has a discontinuity across the real axis, proportional to a scattering function F (k), which in analogy to the stationary problem is called the “reflection” coefficient. The Fredholmlike nature of (1.6) implies that at some points kα (the eigenvalues), eigenfunctions or
4
J. Villarroel, M. J. Ablowitz
solutions to its homogeneous version, Gkα 8(x, y) = 0, may exist. Here Gkα ≡ G(k = kα ). Equivalently these eigenfunctions correspond to solutions of (1.2) at k = kα : (i∂y + ∂xx + 2ikα ∂x + u)8α = 0
(1.20 )
vanishing as x 2 + y 2 → ∞. At these eigenvalues µ(k) has poles with the residue given by the eigenfunction 8. Solutions corresponding to the discrete spectrum of the nonstationary Schrödinger equation correspond, by definition, to having a vanishing reflection coefficient F (k) and are taken to be purely meromorphic. We first review the results of [2] corresponding to pure discrete spectrum. In [2] it is assumed that the discrete spectrum consists of single poles (here we take an even number), i.e. µ(k) is written: µ(k) = 1 +
2N X α=1
8α , (k − kα )
(1.10)
where the residues 8α remain to be found. The potential is obtained from u(x, y) = −2i
2N X
∂x 8α .
(1.11)
α=1
In [2] the following fundamental equation is found that relates the residues 8α with the spectral function µ ν(kα ) = −i(fα + γα )8α , α = 1, . . . 2N,
(1.12)
f (k) = x − 2ky; fα ≡ f (kα ) = x − 2kα y, ν(kα ) ≡ lim (µ(k) − singular part of µ at kα )
(1.13)
where
k→kα
and γα are “normalization” constants (in KPI they are functions of time). In addition, a constraint, conveniently normalized now for the purposes described later in this paper, (1.14a) Q(kα ) = 1 was obtained where 1 sign (Im kα ) Q(kα ) ≡ 2πi
Z Z u8α dxdy
(1.14b)
(integration is over the plane). Later we show that this constraint is a certain integer; i.e. an index or winding number. From (1.12) the following linear algebraic system of equations ensues: 1+
2N X β=1 β6=α
8β = −i(fα + γα )8α , α = 1, . . . 2N. kα − kβ
(1.15)
Discrete Spectrum of Nonstationary Schrödinger and Kadomtsev–Petviashvili Equation
5
After solving this system, the corresponding “reflectionless” potentials for (1.1) are obtained as 2N ∂2 ∂ X 8α = 2 2 log F, (1.16) u(x, y) = −2i ∂x ∂x α=1
where F is the determinant of the matrix Bαβ = (fα + γα )δαβ − i(1 − δαβ )
1 kα − kβ
, α, β = 1, 2 . . . 2N.
(1.17)
Solutions to KPI (lumps) are given by (1.16) letting the constants in (1.12) “evolve” as γα (t) = γα (t = 0) + 12kα2 t.
(1.18)
Real nonsingular potentials are obtained by taking kα , γα to satisfy kj +N = k¯j , γj +N = γ¯j , j = 1, . . . N, where the bar denotes complex conjugate. The simplest real nonsingular solution of KPI corresponds to taking N=1, with k1 = a + ib, F = z2 + 4b2 y 02 +
1 , 4b2
(1.19a)
where z = x 0 − 2ay 0 , x 0 = x − 12(a 2 + b2 )t − x0 , y 0 = y − 12at − y0 ,
(1.19b)
γI a − γR b γI , y0 = , b 2b where a, b, γR , γI are arbitrary real constants. This is the well known one lump solution. Later we will say that the solution has index or charge one. The foregoing discussion summarizes what had been considered (previous to ref. [4]) regarding the discrete spectrum of the operator (1.6) and the associated lump type solutions which correspond to simple poles kα with the condition Q(kα ) = 1, ∀α enforced. The KPI solutions are localized structures that move uniformly and do not suffer a phase shift upon interaction. An important discovery was that there exist nonsingular localized decaying solutions u(x, y, t) to KPI away from the above class of lump solutions. One such new ∂2 configuration is u(x, y, t) = 2 ∂x 2 log F , where x0 =
γI − δI )2 F = (z2 − 4b2 y 02 + δR − 24bt)2 + (2y 0 (1 + 2bz) + b 1 1 1 + 2 (z − )2 + 4b2 y 02 + 2 , b 2b 4b
(1.20)
where z, x 0 , y 0 , are defined above in Eq. (1.19b) and a, b, γR , γI , δR , δI are arbitrary real constants. This particular solution of KPI was constructed in [5], but without reference to the underlying scattering problem. It went almost unnoticed until recently (cf. [6]), where it was rederived, again purely in terms of KP solutions and related dynamics. The above solution, although sharing some of the properties of solitary waves, is not encompassed
6
J. Villarroel, M. J. Ablowitz
by the aforementioned class of lump solutions (1.16)–(1.18). We shall show that this solution is really part of the discrete spectrum corresponding to poles of higher order and a higher order index, in this case index two. Thus the previous theory regarding the discrete spectrum of the Schrödinger operator, based upon assuming simple poles, was not complete. Additionally there is the further interesting underlying characterization property of an index and its meaning with regard to inverse scattering. 2. Solutions and the Spectrum of the Nonstationary Schrödinger Operator In what follows we assume that u(x, y) is real (the extension to more general potentials follows along similar lines), and suitably decaying such that the relevant integral equations are of Fredholm type. We leave the justification of this to later study. All results obtained are consistent with this assumption. Our results may then be summarized as follows: i) If k = k1 is an eigenvalue, then its complex conjugate, k¯1 , is an eigenvalue. The number of linearly independent homogeneous solutions to (1.6) at k1 and k¯1 are equal (Result I below). ii) Eigenfunctions corresponding to eigenvalues k1 , k2 are orthogonal (with respect to an appropriate pairing) if k1 6 = k¯2 (Result II below). Otherwise their scalar product equals the “charge" Q (Result IV, below, see Eq. 2.11.2). iii) The charge Q can take on other natural values besides Q = 1. If Q 6= 1, then Eq. (1.12) is no longer valid. Instead a new equation holds; this is discussed in Sect. 3. iv) Q has a topological interpretation as Q = winding number H, where H is related to the residue 8 of the eigenfunction µ satisfying Eq. (1.6) by 8 =i
1 ∂ log H + O( 2 ), r 2 ≡ x 2 + y 2 → ∞, ∂f r
where f ≡ x − 2k1 y and the derivative is taken along a closed contour at infinity: 0∞ that surrounds the origin once. In fact we find H = f Q + O(1/r) as r → ∞. v) Solutions of the nonstationary Schrödinger operator (Eq. (1.2),(1.6)) can have higher order poles. vi) Characterization of the class of reflectionless potentials requires giving the position and multiplicity of the poles. Additional determining data are the multiplicity and the associated indices Qj of the eigenfunction satisfying Eq. (1.6). These meromorphic wave functions correspond to real, nonsingular and decaying potentials of the form ∂2 u(x, y) = 2 ∂x 2 log F (x, y, γ , δ, . . . ), where F is a certain polynomial and γ , δ, . . . constants. They also correspond to localized, real, nonsingular, decaying solutions of the KPI equation of the form ∂2 log F (x, y, γ (t), δ(t) . . . ). (2.1) ∂x 2 The dynamical evolution of the maxima of u(x, y, t) is nontrivial as opposed to the situation for standard lumps or solitons. Equations (1.16–17) and (1.20) are members of this class. u(x, y, t) = 2
Discrete Spectrum of Nonstationary Schrödinger and Kadomtsev–Petviashvili Equation
7
We now elaborate on the above statements. As previously mentioned, we denote Gkα ≡ G(k = kα ). We call homogeneous solutions of the operator Gk1 , ω(x, y) which satisfy Z Gk1 ω ≡ ω(x, y) −
dx
0
Z
dy 0 G(x − x 0 , y − y 0 , k1 )u(x 0 , y 0 )ω(x 0 , y 0 ) = 0. (2.2.1)
The span (vector space) of all such functions is denoted Ker Gk1 . Similarly we refer to homogeneous adjoint solutions as ρ(x, y). Homogeneous solutions to the adjoint operator Gk†1 satisfy Z Gk†1 ρ
≡ ρ(x, y) −
dx
0
Z
¯ 0 − x, y 0 − y, k1 )u(x, y)ρ(x 0 , y 0 ) = 0. dy 0 G(x
(2.2.2)
We reiterate that certain properties of these operators, e.g. compactness and decay of their solutions, is assumed in this discussion. Result I. At any point k1 the spaces KerGk†1 , KerGk1 , KerGk¯1 satisfy i) KerGk†1 = u(x, y)KerGk¯1 ,
(2.3.1)
i.e. ω1¯ (x, y) is a homogeneous solution at k¯1 if ρ(x, y) ≡ uω1¯ (x, y) is an adjoint function at k1 . (Note, the potential is assumed to be real.) In particular DimKerGk†1 = DimKerGk¯1 . ii) DimKerGk¯1 = DimKerGk1 .
(2.3.2)
Thus, at any point k1 the number of linearly independent homogeneous solutions to Gk1 ω1 = 0, Gk†1 ρ = 0 and Gk¯1 ω1¯ = 0 is the same. Proof. i) Using (1.9.c) we have for the adjoint equation Z ρ=
dx 0 dy 0 u(x, y)G(x − x 0 , y − y 0 , k¯1 )ρ(x 0 , y 0 ).
Hence if a solution ρ to this equation exists, letting ρ = ζ u implies that ζ satisfies Gk¯1 ζ = 0. Thus , ρ ∈ Ker Gk†1 ⇒ ζ ∈ Ker Gk¯1 . The proof holds also in the opposite direction and i) follows. ii) Fredholm’s first theorem gives that the number of linearly independent solutions to Gk1 8 = 0, and Gk†1 ρ = 0 is the same. Thus, Dim Ker Gk1 = Dim KerGk†1 . These two facts together imply the statement. u t
8
J. Villarroel, M. J. Ablowitz
Result II. The system of eigenfunctions satisfy the orthogonality relationship ¯ < ωα , ωβ >= 0, β 6 = α, where ωα denotes an eigenfunction at k = kα , where Z 1 f¯x gdxdy < f, g >≡ π
(2.4)
(2.5)
is the relevant bilinear form. Note that we find it convenient to define kα¯ ≡ k¯α . Hence ωα¯ denotes an eigenfunction at k¯α . Proof. (1.20 ) implies that Z 0 = dxdy ω¯ β (i∂y + ∂xx + u + 2ikα ∂x )ωα Z Z ¯ ¯ = dxdyωα {(−i∂y + ∂xx + u − 2i kβ ∂x )ω¯ β } − 2i(kα − kβ ) dxdyωα ∂x ω¯ β Z ⇒ (kα − k¯β ) dxdyωα ∂x ω¯ β = 0. (Actually we can extend this result, see Appendix I). u t We assume that eigenfunctions µ(k) corresponding to “pure” multipole lumps have the following meromorphic representation: µ(k) = 1 +
n X
M
{
α=1
α X 8α 9r,α }, + (k − kα ) (k − kα )r
(2.6)
r=2
where 8α , 9r,α depend only on x, y. Around any pole k1 we have the Laurent expansion M
µ(k) = ν(k) +
X 9r 81 + (k − k1 ) (k − k1 )r r=2
M
= ν˜ (k) + ν(k1 ) +
X 81 9r + , (k − k1 ) (k − k1 )r r=2
where ν(kα ) ≡ limk→kα µ(k) − singular part of µ(k) at kα , and ν˜ (k) is zero at kα and tends to 0 as k → ∞. Note that ν(kα ) = 1 +
n X β=1,β6=α
Mβ
X 8β 9r,β { }. + (kα − kβ ) (kα − kβ )r
(2.7)
r=2
The next result establishes the connection between the pole structure of µ(k) and the discrete spectrum.
Discrete Spectrum of Nonstationary Schrödinger and Kadomtsev–Petviashvili Equation
9
Result III. The (negative) Laurent coefficients of µ(k) around kα :{ν(kα ), 8α , 9r,α }, r = 2 . . . M, M ≥ 2 satisfy the integral equations ∂G 9M,α }k=kα = . . . ∂k ∂G 1 ∂ M−1 G 9M,α }k=kα = 0 92,α + . . . + = {G8α + ∂k (M − 1)! ∂k M−1
{G9M,α }k=kα = {G9M−1,α +
{Gν +
1 ∂ 2G 1 ∂M G ∂G 92,α + . . . + 9M,α }k=kα = 1. 8α + 2 ∂k 2 ∂k M! ∂k M
(2.8)
Proof. Take the limit k → kα in Eq. (1.6) and set the coefficients of 1/(k − kα )j , j = 0, 1, . . . M equal to zero. u t Result III implies that the poles of µ(k) are eigenvalues, i.e. they are points of the discrete spectrum, while 9M,α are eigenfunctions at the eigenvalue kα . By Result I we have that the discrete spectrum is an even dimensional set {kj , k¯j }j =1...N . As a consequence, pure meromorphic functions, µ have the structure (2.6), where the poles occur in pairs (kj , kj +N ) ≡ (kj , k¯j ) and hence without loss of generality the position can be arranged such that Im (kj ) > 0; j = 1, 2 . . . N. We find it convenient to define j¯ = j + N. Suppose µ corresponds to the discrete spectrum of the Schrödinger operator and has 2N poles {kj , k¯j }j =1...N with multiplicities {Mj , Mj¯ }j =1...N , and tends to 1 as k → ∞. Such an eigenfunction is given by N X { µ(k) = 1 + j =1
M
M¯
j j X X 8j¯ 9r,j¯ 9r,j 8j + + + }, r (k − kj ) (k − k¯j ) (k − kj ) (k − k¯j )r r=2 r=2
(2.60 )
where we have used kj +N = k¯j in (2.6) j = 1 . . . N. Fundamental objects in our development are certain quantities that we call indices (winding numbers) or charges. At any pole kα , α = 1, . . . 2N we define the index of the pole as Z Z 1 u8α dxdy, (2.9) sign (Im kα ) Qα ≡ 2πi where 8α is the residue of µ at the pole. We also introduce the secondary indices Z Z 1 sign (Im kα ) u9r,α dxdy, 2 ≤ r ≤ Mα . (2.10) Qr,α ≡ 2πi Note. To stress the importance of the charges Q1,α we drop the subscript and we write Qα ≡ Q1,α . Result IV. Let µ(k) be given by Eq. (2.60 ) with discrete spectrum {kj , k¯j }N j =1 . Then the indices Qα satisfy the following properties: PN PN (2.11.1) i) j =1 Qj = j =1 Qj¯ , j = 1 . . . N, ii)
Qα = sign (Im kα ) < 8α¯ , 8α >,
(2.11.2)
10
J. Villarroel, M. J. Ablowitz
Qr,α = sign (Im kα ) < 8α¯ , 9r,α > iii) iv)
¯ α = Qα¯ , Q ¯ r,j = Q ¯ , Q r,j QM,α = 0,
(2.11.3) (2.11.4)
where M ≥ 2 is the multiplicity of the pole. Proof of these facts are given in Appendix I and II.
3. Equations Determining the Eigenfuction In this section we discuss a method which determines equations which serve to fix eigenfuctions µ(k) satisfying Eq. (1.6) with certain assumed analytic structure. Once these equations have been determined, we show in Sect. 4 how they can be used to solve for particular cases of eigenfuctions and related reflectionless multipole potentials which are real, nonsingular and decaying at infinity. A linear relationship is obtained. Result V. Simple poles. We begin by analyzing simple pole solutions of Eq. (1.6); i.e. special cases of Eq. (2.6). Assume that the solution to Eq. (1.6) has the following structure around k = k1 : µ(k) = ν(k) +
8 , (k − k1 )
(3.1)
where ν(k) is regular at k1 and ν(k) → 1. Define recursively the forcings hn (f, y), k→∞
n = 0, . . . n0 by hn (f, y), n = 0, . . . n0 by
h0 = 1, ∂f hn = ihn−1 , (i∂y + ∂ff )hn = 0,
(3.2a)
where f ≡ f (k1 ) = x − 2k1 y (see (1.13)). Note that the first few of them are h0 = 1, h1 = i(f + γs ), h2 = 21 (h21 − δs − 2iy), h3 = 3!1 (h31 − 3h1 (δs + 2yi) + s ) and γs , δs , s are arbitrary constants (here the subscript s stands for simple pole). We also write Hn (x, y) ≡ hn (x, y, γ = 0, δs = 0, s = 0 . . . ). Thus 1 1 H0 = 1, H1 = if, H2 = − (2iy + f 2 ), H3 = (6yf − if 3 ). 2 3!
(3.2b)
Let ρ be the solutionR of the adjoint equation at k1 and n0 be the greatest integer ¯ y)dxdy, n = 0, . . . n0 exist. Suppose also that for which the integrals Hn−1 ρ(x, DimKerGk1 = 1. In what follows we define the index Q = Q(k1 ) = Q1 . Then
Discrete Spectrum of Nonstationary Schrödinger and Kadomtsev–Petviashvili Equation
11
i) Q = n0 ; ii) 8 is related with ν as follows: Q = 1 ⇒ {ν + h1 8}k=k1 = 0, dν = 0, + h1 ν + h2 8 Q=2⇒ dk k=k1 2 dν 1d ν + h2 ν + h3 8 + h1 = 0, Q=3⇒ 2 dk 2 dk k=k1 .. . d n−1 ν dν 1 + .. + h ν + h 8 = 0; + h Q=n⇒ n−2 n−1 n (n − 1)! dk n−1 dk k=k1
(3.3.1) (3.3.2) (3.3.3)
(3.3.n)
iii) if 8 is related with µ by Eq. (3.3.n) then as r 2 ≡ x 2 + y 2 → ∞, 8≈i
∂ in ∂ log Hn = i log f n = , and ∂f ∂f f Q = winding number Hn .
Hence Q is a topological invariant for the Schrödinger operator that counts the number of zeros that Hn (f, y) has a function of f . Remarks. 1. We assume that it is possible to differentiate Eq. (1.6) (n0 − 1) times under the integral sign. 2. The proof below shows also that at k = k1 there exists solutions µ˜ n , n = 0 . . . n0 to the set of equations (3.4) Gk1 µ˜ n = An Hn−1 , n = 0 . . . n0 , where An ≡ n1 (Q − 1)...(Q − n), H−1 ≡ 0. This holds if the following Fredholm alternative conditions are satisfied: Z Z ¯ y)dxdy = 0, n = 0 . . . n0 . (3.5) Fn ≡ An Hn−1 ρ(x, These conditions typically require Q = n0 . 3. n0 is assumed to be finite. Otherwise ρ(x, y) would need to vanish faster than any power. This would contradict our assumption of rationally decreasing solutions. 4. The relationship between the Laurent coefficients and ν(k) depends on the value that the charge Q takes, which is assumed to be a positive integer. In Sect. (4) it is shown, given the pole structure of function µ around both k1 and k¯1 , and Dim Ker Gk1 = 1, that the Fredholm conditions (3.5) fix the value of Q to be n0 . 5. The case Q = 0 is not ruled out by the discussion below. We assume Q 6 = 0. 6. The results of [2] correspond to the case n0 = 1. (This is the case of the well known lumps associated with KPI, discussed briefly in Sect. 1 (ref. [2])). Proof. Proof of i and ii. For simple poles, Eqs. (2.8) are Gk1 8 = 0, {Gν +
∂G 8}k1 = 1. ∂k
12
J. Villarroel, M. J. Ablowitz
Use of (1.9.b) yields that µ1 ≡ {ν + H1 8}k=k1 solves Gk1 8 = 0, Gk1 µ1 = 1 − Q.
(3.6)
To proceed further, we need the value of the charges Q which as mentioned above is not fixed unless we specify the pole structure of function µ around both k1 and k¯1 . This matter is considered in the next section. Suppose we assume that Q = 1. With the assumption Dim Ker Gk1 = 1 this implies that µ1 = −iγ 8 which is (3.3.1). When Q 6 = 1 Eqs. (3.6) do not imply a relationship between ν, 8. The key observation that allows us to obtain equations between ν and 8 is that µ satisfies the equations obtained from formal differentiation of (1.6) under the integral sign n0 times when Q = n0 . We thus obtain, along with (2.8), the following further equations: ∂G 1 ∂ 2G dν + ν+ 8 = 0, G dk ∂k 2 ∂k 2 k=k1 2 ∂G dν 1 ∂ 3G ∂ 2G d ν 8 = 0, + 2ν+ G 2 +2 dk ∂k dk ∂k 3 ∂k 3 k=k1 .. . It is convenient to introduce the functions
µ0 ≡ 8, µ1 ≡ {ν + H1 8}k=k1 , µ2 ≡ 1 µ3 ≡ 2
1 d 2ν dν + H2 ν + H3 8 + H1 2 dk 2 dk
dν + H1 ν + H2 8 dk
k=k1
k=k1
,
... .
Elaboration of these equations gives that Gk1 µ0 = 0, Gk1 µ2 = H1 (1 − Q/2) + A, where we introduce the constants
Gk1 µ1 = 1 − Q, Gk1 µ3 = 2H2 (1 − Q/3) + H1 B + C, .. .
A ≡ −(Q0 + η/2), B ≡ (Q0 + η/3) with 1 sign (Im k1 ) Q0 ≡ 2π i
Z Z
1 sign (Im k1 ) uνdxdy, η ≡ 2π i
Z Z uH1 8dxdy.
(3.7) (The explicit value of the constant C is not important for this discussion.) Hence the linear combinations µ˜ 0 , µ˜ 1 , . . . µ˜ n0 , µ˜ 0 ≡ 8, µ˜ 1 ≡ −µ1 , µ˜ 2 ≡ (1 − Q)µ2 + Aµ˜ 1 , Q Q µ˜ 3 ≡ (Q − 1)(1 − )µ3 + B µ˜ 2 − (1 − )C µ˜ 1 , . . . , 2 2 Gk1 µ˜ n = An Hn−1 , n = 0 . . . n0 satisfy (3.4).
Discrete Spectrum of Nonstationary Schrödinger and Kadomtsev–Petviashvili Equation
13
Existence of solutions to the above equations requires that the following Fredholm conditions are satisfied: Z Z ¯ y)dxdy = 0, n = 0 . . . n0 . Hn−1 ρ(x, Fn ≡ An This is (3.5). Suppose Q = 2. Then we have from the above equation, Gk1 µ˜ 0 = Gk1 µ˜ 2 = 0. Since we are assuming that Dim Ker Gk1 = 1, this implies that µ˜ 2 must be proportional to µ˜ 0 . Using this fact and the definitions above yields (3.3.2). Besides Gk1 µ˜ 1 = 1 requires that F1 = 0. (In Sect. 4 we show that using the solution of the homogeneous adjoint equation this Fredholm condition implies that the secondary charge Q2 vanishes (see Eq. (2.10) for the definition of Q2 at k¯1 ).) More generally Eq. (3.4) results in the equation Gk1 µ˜ n = An Hn−1 , n = 1 . . . n0 − 1 which implies the Fredholm conditions: Fn = 0, n = 1 . . . n0 − 1. In Sect. 4 we will show that these equations along with the corresponding conditions at k¯1 typically determine the value of Q = n0 . In that case we also have Gk1 µ˜ 0 = Gk1 µ˜ n0 = 0, and hence that µ˜ n0 is proportional to µ˜ 0 yielding (3.3.n). iii) Integrating Eq. (1.20 ), and noting that for large r, 8 = O(1/r) Green’s theorem yields that if k1 = a + ib then Z Z 1 (i∂y + ∂xx + 2ik1 ∂x )8 sign (Im k1 )Q = − 2πi Z 1 8df, =− 2π 0∞ where f ≡ x − 2ay − 2iby, and 0∞ is a large circle (or rectangle) in the f plane. We also note that as x 2 + y 2 → ∞, ν → 1 and (3.3.n) with (3.2a) yields 8 ≈ ∂ i ∂f log Hn , hence 1 Q= 2πi
Z 0∞
∂f Hn df = winding number of Hn . Hn
t u
Result VI. Double poles. Assume that the solution to (1.6) has the following structure around k = k1 : 8 9 µ(k) = ν(k) + + , (3.8) k − k1 (k − k1 )2 where ν(k) is regular at k1 and tends to 1 as k → ∞. Note that as compared to the general formula (2.6) we simply write 9 ≡ 92,1 . Suppose also that 9 6 = 0 and that DimKerGk1 = 1. Then Z Z sign(Imk1 ) u9dxdy. i) If u = O(1/r 2 ) then Q2 = 0, where Q2 = 2π i Let Q ≥ 2.
14
J. Villarroel, M. J. Ablowitz
ii) In either of the cases below one has the following relationships between µ and its Laurent coefficients: {8 + h1 9}k=k1 = 0, (3.9.1) and also if: 1) Q = 2 ⇒
{ν + h2,d 9}k=k1 = 0,
(3.9.2)
where
1 h1 ≡ h1,d ≡ i(f + γd ), h2,d ≡ − (h21 + 2iy + δd ) 2 γ , δd are arbitrary constants (here d stands for double pole). We find it convenient to use a similar notation to that of the simple pole case. 2) Q = 3 ⇒ 1 dν = 0, (3.9.3) + (h1 + 2σ )ν + h3,d 9 dk 3 k=k1 where h3,d ≡ −(h1 + σ )3 + 3σ 2 (h1 + σ ) − 6iσy + iβ, β, δd , σ are arbitrary constants. iii) If 8, 9 are related with µ by the above equations then as r 2 ≡ x 2 + y 2 → ∞, 1 ∂ log hn,d + O( 2 ), and ∂f1 r Z ∂f hn,d 1 df = winding number of hn,d . Q= 2πi 0∞ hn,d 8=i
(3.10)
Proof. i) For double poles equations (2.8) are Gk1 9 = 0, ∂G 9}k=k1 = 0, {G8 + ∂k 1 ∂ 2G ∂G 9}k=k1 = 1. 8+ {Gν + ∂k 2 ∂k 2 The two first equations in differential form correspond to (these equations are obtained by substituting (3.8) into (1.2)) (i∂y + ∂xx + 2ik1 ∂x + u)9 = 0, (i∂y + ∂xx + 2ik1 ∂x + u)8 + 2i∂x 9 = 0.
(3.11.1) (3.11.2)
Assume that for large values of x, y, 8 ≈ f1α , 9 ≈ f1β . Then (3.11.2) shows that ∂ P 8j and we are assuming that u = O(1/r 2 ), it must be β = α + 1. Since u = −2i ∂x that α = 1. Next we integrate (3.11.1) and have, in view of the rate of decay of 9, Z Z Z Z 1 1 u9dxdy = − (i∂y + ∂xx + 2ik1 ∂x )9 = 0. sign (Im k1 )Q2 ≡ 2πi 2π i The fact that Q ≥ 2 follows by inspection of different cases and is also discussed further in the next section.
Discrete Spectrum of Nonstationary Schrödinger and Kadomtsev–Petviashvili Equation
15
ii) Using the Green’s function relations from Eq. (1.9) the above equation involving derivatives of the Green’s functions are explicitly found to be (see the appendix for relations involving the higher derivatives of the Green’s functions), Gk1 µ0 = Gk1 µ1 = 0, Gk1 µ2 = 1 − Q −
η , 2
(3.12.1)
where in a similar spirit to the simple pole case we define the functions µ0 ≡ 9, µ1 ≡ {8 + H1 9}k=k1 , µ2 ≡ {ν + H1 8 + H2 9}k=k1 and the constant η≡
1 sign (Im k1 ) 2πi
Z Z uH1 9dxdy.
Note first that Dim Ker Gk1 = 1 implies that µ1 = −iγ µ0 with γ an arbitrary constant, and also that η = −Q. The first relationship yields (3.9.1). To proceed further we need the value of the charges which is unknown unless we specify the pole structure of the function µ around both k1 and k¯1 (see Sect. 4). The following general statements can be made. If Q = 2 then µ2 solves also the homogeneous equation and hence it must be also proportional to µ0 . This yields (3.9.2). When Q 6 = 2 Eqs. (3.12) do not imply a relationship between ν, µ0 . To get around this difficulty we assume that differentiation of (1.6) under the integral sign is permitted (this may be validated a posteriori). We then obtain {G
1 ∂ 3G ∂G 1 ∂ 2G dν 8+ 9}k=k1 = 0. + ν+ 2 dk ∂k 2 ∂k 6 ∂k 3
Using the Green’s function properties (see the appendix) we get in addition to (3.12.1), Gk1 µ3 = C + H1 (1 − Q/3),
(3.12.2)
where
dν + H1 ν + H2 8 + H3 9}k=k1 . . . dk and C is a constant, the explicit value of which is not important here. Assume next that Q1 = 3. We have µ3 ≡ {
1 Gk1 µ0 = Gk1 µ1 = 0, Gk1 µ2 = − , Gk1 µ3 = C. 2 We have with µ˜ 0 ≡ µ0 , µ˜ 1 ≡ µ1 , µ˜ 2 ≡ −2µ2 , µ˜ 3 ≡ µ3 + 2Cµ2 that Gk1 µ˜ 0 = Gk1 µ˜ 1 = Gk1 µ˜ 3 = 0, Gk1 µ˜ 2 = 1 which imply µ3 = µ0 = 9 and this confirms Eq. (3.9.3). We also require the Fredholm condition F1 = 0 to be satisfied. (As mentioned earlier in Sect. 4 we show how such conditions can be used to fix the charges.) Higher order cases are treated similarly. iii) As x 2 + y 2 → ∞, we assume that ν → 1 and (3.9.n) yields in both the cases n = 2, 3 8 ≈ i ∂f∂ log h2,n ≈ i ∂f∂ log f n = in/f . The continuation of the proof is similar to that of the single pole. u t
16
J. Villarroel, M. J. Ablowitz
Result VII. Triple poles. Assume that around k = k1 the solution to (1.6) µ(k) has the structure 92 93 8 + + , µ = ν(k) + 2 k − k1 (k − k1 ) (k − k1 )3 where ν(k) is regular at k1 and tends toR 1 as k → ∞ and 93 6= 0. Let n0 − 3 be the ¯ y)dxdy, n = 0 . . . n0 exist. Suppose greatest integer for which the integrals Hn ρ(x, also that DimKerGk1 = 1. Then 2 ) then Q = Q = 0, where Q , j = 2, 3 are defined as: Q = i) If u =ZO(1/r 2 3 j j Z sign (Imk1 ) u9j dxdy and Q = n0 ≥ 3. 2π i ii) In the case Q = 3 one has {92 + h1 93 }k=k1 = 0, {8 + h2,t 93 }k=k1 = 0, {ν + h3,t 93 }k=k1 = 0.
(3.13) (3.1)
where 1 h1 ≡ h1,t ≡ i(f + γt ), h2,t ≡ − (h21 + 2iy + δt ), 2 1 3 h3,t ≡ (h1 + 3h1 (δt + 2iy) + t ). 6 γt , δt , t are arbitrary constants (here t stands for triple pole). As r → ∞, ∂ ∂ log hn,t ≈ i log f 3 = 3i/f, and ∂f1 ∂f1 Z ∂f hn,t 1 df = winding number of hn,t . Q(k1 ) = 2πi 0∞ hn,t 8≈i
The proof is similar to the other cases (see the appendix). Higher order poles can be considered in a similar way. We will not pursue this topic any more here.
4. Schrödinger Potentials In this section we construct several families of decaying potentials for the nonstationary Schrödinger operator. We shall see that the assumed pole structure of µ around both k1 and k¯1 determines the value of all the relevant charges. The results of the last section show that under the proviso Dim Ker Gk1 = 1 the assumed pole structure of µ of the form (2.6’) yields a system of equations which serve to fix the coefficients of the poles and thus µ(k). We begin considering the simplest possibility.
Discrete Spectrum of Nonstationary Schrödinger and Kadomtsev–Petviashvili Equation
17
1+1 pole. Assume that µ is meromorphic with simple poles at points k1 , k¯1 : µ(k) = 1 + µ(k) ˆ +
81¯ 81 + , (k − k1 ) (k − k¯1 )
(4.1)
where µ(k) ˆ is regular around both k1 and k¯1 and tends to 0 as k → ∞. Then the charges defined by Eq. (2.9) are given by ¯ ¯ = Q1 = 1. i) Q 1 We use the notation Q1¯ = Q(k¯1 ), Q1 = Q(k1 ). ii)
¯ ν(kα ) = −i(fα + γα )8α , α = 1, 1.
(4.2)
Note. The possibility Q1 = 0 is not ruled out by the proof below. We assume that Q1¯ , Q1 do not vanish. iii) Generalizing Eqs. (4.1–4.2) to allow µ(k) to have only simple poles at kα then real, decaying, nonsingular “reflectionless” potentials are obtained when the constants are chosen to satisfy γj +N = γ¯j , j = 1, . . . N. Proof. From Eq. (2.11.3) the charges at k¯1 Equations (3.6) now are: Gk1 81 = 0 , G µ = 1 − Q k1 1 1
are the complex conjugates of those at k1 . Gk¯1 81¯ = 0 G µ = 1 − Q 1¯ k¯1 1¯
.
The homogeneous solutions at k1 , k¯1 are 81 and 81¯ respectively and from the proof of Result I in Sect. 2 (see below Eq. (2.3.1)) the homogeneous adjoint solutions are ρ1 = u(x, y)81¯ and ρ1¯ = u(x, y)81 . Using (2.11.3) the Fredholm alternative conditions (see Eq. (3.5)) yield Z ¯ ¯ = (1 − Q1 )Q1 ⇒ Q1 = 1. 0 = F1 ≡ (1 − Q1 ) dxdy ρ¯1 = (1 − Q1 )Q 1 (Recall that we assume Q1 6 = 0.) Note. ii) follows immediately from Eq. (3.3.1). iii) From the discussion in Sect. 1 we already know that real, nonsingular potentials ¯ ¯ = Qj . However these necessary conditions are not imply that n = 2N, kj +N = k¯j Q j sufficient. We also need the constants γα to satisfy γj +N = γ¯j , j = 1, . . . N. Indeed as we pointed out in Sect. 1, the solution to (4.2) yields the potential (1.161.17). This formula shows that the potential is decaying, nonsingular and real if det(B) > 0. D E where, Following ref. [8] note first that in block form B = −E¯ D † Dij = (fi + γi )δi,j − i t u
(1 − δi,j ) −i ; Eij = ; ki − kj ki − k j
E † = E, i, j = 1, 2, . . . N.
18
J. Villarroel, M. J. Ablowitz
Note that the position of the poles occurring in pairs (kj , k¯j ) can be arranged such that I m(kj ) < 0; j = 1, 2 . . . N. We see that the determinant of B is real from the following. The matrix B has the following property: B¯ βα = Bα+N,β+N , mod 2N, hence Det Bαβ = (−1)N Det Bα+N,β = (−1)2N Det Bα+N,β+N ⇒ Det Bαβ = Det Bβα = Det Bαβ . Thus Det Bαβ is real. Then we note that E is a positive definite matrix. To show this, use the Cauchy formula for determinants: det(E) =
Y
(ki − kj )(k¯i − k¯j )
1≤i<j ≤N
=
Y 1≤i<j ≤N
N |ki − kj |2 Y
|ki − k j |2
i=1
N Y
k i,j =1 i
1 − kj
1 . (2|I m(ki )|)
Therefore det(E) > 0. Replacing N by p, 1 ≤ p ≤ N in the above formula, one can show that all the principal minors of the Hermitian matrix E are also positive. Hence it is well known that E is positive definite. It follows immediately that E¯ is also positive definite. Next we prove that det B(x, y) > 0 for all x, y. Assume that at some point (x1 , y1 ) det B(x1 , y1 ) = 0. Then at this point there exists a non null vector e ≡ (e1 , e2 ) ∈ C 2N satisfying Be = 0. This implies ¯ 1 + D † e2 = 0. De1 + Ee2 = 0; − Ee These equations by manipulation yield ¯ 1 = 0. e2 † Ee2 + e1† Ee Since E, E¯ are positive definite this yields (e1 , e2 ) = (0, 0) and hence that det B(x, y) 6 = 0 for all x, y. But det B(x, y) → ∞ if y 2 +x 2 → ∞ which in turn implies the statement. The simplest Schrödinger potential corresponds to taking µ(k) ˆ = 0 in Eq. (4.1) with two conjugate simple poles k2 = k¯1 and γ2 = γ¯1 , k1 = a + ib, z = x 0 − 2ay 0 , x 0 = x − x0 , y 0 = y − y0 , where x0 , y0 are constants related with γ1 (see Eq. (1.19). Then we have detB = F = z2 + 4b2 y 02 +
1 . 4b2
(4.3)
∂ The well known one lump solution is obtained from u(x, y, t) = 2 ∂x 2 log F . We say that the solution has index one. Solutions constructed from simple poles were well known [2,3]. The first nontrivial potentials beyond that class are discussed next. 2
Discrete Spectrum of Nonstationary Schrödinger and Kadomtsev–Petviashvili Equation
19
Simple + double pole. Assume that µ has the following structure: µ = 1 + µ(k) ˆ +
81¯ 92 81 + + , 2 k − k1 (k − k1 ) k − k¯1
(4.4)
where µ(k) ˆ is regular at both k1 , k¯1 and tends to 0 as k → ∞. Suppose also that Dim Ker Gk1 = Dim Ker Gk¯1 = 1. Then Q1 = 2. Note that from Eq. (2.11.3,4) the value of the charges at k = k¯1 follows from the charges at k = k1 . ii) The Laurent coefficients satisfy the following system of linear algebraic equations a) At k = k1 (see Eq. (3.9.1) and (3.9.2))
i)
{81 + h1 92 }k=k1 = 0, {ν + h2,d 92 }k=k1 = 0.
(4.5.1)
b) At k = k¯1 (see eq. 3.3.2) {
dν + h1 ν + h2 81¯ }k=k¯1 = 0. dk
(4.5.2)
Note that here we simply write 92,1 ≡ 92 . iii) The function (4.4) with µ(k) ˆ = 0 solves (1.2,1.6) with the Schrödinger potential u(x, y) = 2∂xx log F , γI − δI )2 F = (z2 − 4b2 y 02 + δR )2 + (2y 0 (1 + 2bz) + b 1 1 1 + 2 (z − )2 + 4b2 y 02 + 2 , b 2b 4b
(4.6)
where we define k1 ≡ a + ib, z = x 0 − 2ay 0 , x 0 = x − x0 , y 0 = y − y0 , and
γI a − γR b γI , y0 = ; b 2b δR , δI are arbitrary real constants. x0 =
Proof. Equation (3.6) for a simple pole at k¯1 and Eq. (3.12.1) for a double pole at k1 yield Q1 Gk1 92 = 0, Gk1 µ2 = 1 − 2 and Gk¯1 81¯ = 0, Gk¯1 µ1¯ = 1 − Q1¯ . As discussed in Result I, Sect. 2 (see below Eq. (2.3.1)) with the homogeneous solutions to (1.6) at k1 , k¯1 being 92 , 81¯ respectively, then the homogeneous adjoint solutions are
20
J. Villarroel, M. J. Ablowitz
14 12 10 8 6 4 2 0 -2 -6
-4 -4 -2
-2 0x
y0 2
2 4 4
6
Fig. 1a. Snapshot of solution corresponding to potential (1.20) before interaction (a = 0.5, b = 1, γ = δ = 0)
¯ 1 = Q1 , the Fredholm given by ρ1 = u(x, y)81¯ and ρ1¯ = u(x, y)92 . Also since Q alternative conditions yield a) from ρ1¯ = u(x, y)92 : Z 0 = (1 − Q1 )
dxdy ρ¯1¯ ≡ (1 − Q1 )Q2 .
This is an identity since Q2 = 0 ; b) from ρ1 = u(x, y)81¯ : Q1 ) 0 = (1 − 2
Z dxdy ρ¯1 ≡ (1 −
Q1 )Q1 ⇒ Q1 = 2 2
(recall Q1 6 = 0). Therefore Q1 = Q1¯ = 2 and by (3.3.2) and (3.9.2) i) is obtained. The simplest case corresponds to µ(k) ˆ = 0 and in this case ii) yields with δ1¯ = −δ¯1 ≡ δR + iδI : h2,d 81 − α81¯ = 1, − h1 3 1 h¯ 2,d 1 2α¯ 2 + α¯ − 8 ¯ = 1, − α¯ 81 + 2 ¯ |h1 | h1 h1 h¯ 1 1 1 . Solving this linear system we obtain the Laurent coefficients where k1 = a+ib, α ≡ 2ib 81 , 81¯ . Then from Eq. (4.5.1) we obtain 92 . Using (1.3) (which really amounts to Eq. (1.16)) the Schrödinger potential given via (4.6) follows. It is clear that this potential is real, and nonsingular. It has 6 free parameters. Note that u = 0( r12 ) as r 2 ≡ x 2 +y 2 → ∞. Hence the potential is in L2 but not in L1 . This “slow" decay is the underlying reason why the index is two. In Fig. 1 a typical plot of this function is displayed. u t
Discrete Spectrum of Nonstationary Schrödinger and Kadomtsev–Petviashvili Equation
21
25 20 15 10 5 0 -3
-6 -2
-4 -1
-2 0x
y0 2
1 4
2 3
6
Fig. 1b. Snapshot of solution corresponding to potential (1.20) during interaction – same parameters as Fig. 1
18 16 14 12 10 8 6 4 2 0 -2 -3
-6 -2
-4 -1
-2 0x
y0 2
1 4
2 3
6
Fig. 1c. Snapshot of solution corresponding to potential (1.20) after interaction – same parameters as Fig. 1
Two double poles. Assume that µ has the following structure: µ = 1 + µ(k) ˆ +
81¯ 92¯ 92 81 + + + , 2 ¯ k − k1 (k − k1 ) k − k1 (k − k¯1 )2
(4.7)
where µ(k) ˆ is regular at both k1 , k¯1 and tends to 0 as k → ∞ and Dim Ker Gk1 = 1. Then since we have assumed that Q1 cannot vanish, i)
Q1 = 3.
22
J. Villarroel, M. J. Ablowitz
ii) At k = kα , α = 1, 1¯ the Laurent coefficients satisfy the following system of linear algebraic equations: 8α = −h1,α 92.α , 1 dνα + (h1,α + 2σα )να + h3.α,d 92.α ] = 0, [ dk 3
(4.8.1) (4.8.2)
¯ 3.1 ≡ 3, 3.1¯ ≡ 3. ¯ where we define the notation as: 2.1 ≡ 2, 2.1¯ ≡ 2, ˆ =0 iii) If k1¯ = k¯1 , β1¯ = β¯1 , γ1¯ = γ¯1 + 2i σ¯ 1 , σ1¯ = σ¯ 1 the function (4.7) with µ(k) solves (1.6) with the Schrödinger potential u(x, y) = 2∂xx log F , 9 1 1 1 1 | − i(h1 + 2σ ) + |2 + 2 , F = |h3,d |2 + 2 |ih1 + |2 + 2 4b 2b 4b 2b 4b or in physical coordinates F = (z3 − 12b2 y 02 z + 3(σR2 − σI2 )z + 12bσI σR y 0 − 6σR y 0 + βˆR )2 + (6bz2 y 0 − 8b3 y 03 − 6σI σR z + 6b(σR2 − σI2 )y 0 + 6σI y 0 − βˆI )2 9 1 2 1 0 2 + 2 (z − σI − ) + (2by − σR ) + 2 4b 2b 4b 1 1 (4.9) · (z + σI + )2 + (2by 0 + σR )2 + 2 , 2b 4b where the coordinates z, x 0 , y 0 are the same as those in Eq. (4.6) except x0 =
a γI − σR (γI − σR ) − (γR + σI ), y0 = b 2b
and βˆ is a minor redefinition of the original free parameter β. This solution depends on 8 parameters. Proof. By Result VI we have that 8α = −h1,α 92α ; η1 = η1¯ = −Q1 . Hence what remains from Eq. (3.12.1) is Q1 , 2 Q¯ Gk¯1 µ2¯ = 1 − 1 . 2
Gk1 92 = 0,
Gk1 µ2 = 1 −
Gk¯1 92¯ = 0,
¯ The Fredholm It follows that the adjoint solutions are ρα = u(x, y)9α¯ , α = 2, 2. alternative conditions yield Q1 ¯ 2 1 − Q1¯ , )=Q 0 = Q2 (1 − 2 2 ¯ 2 = 0. which are identically satisfied since Q2 = Q The next equation is (3.12.2): Gk1 µ3 = C + H1 (1 − Q1 /3), G ¯ µ¯ ¯ = C ¯ + H¯ 1 (1 − Q1 /3), k1
3
1
Discrete Spectrum of Nonstationary Schrödinger and Kadomtsev–Petviashvili Equation
23
14 12 10 8 6 4 2 0 -2 -6
-6 -4
-4 -2
-2 0x
y0 2
2 4
4 6
6
Fig. 2. A three humped structure – see cf. (6.110 ), (4.9) with constants (a = 0, b = 1, γ = β = σ = 0)
where C, C1¯ are constants. Using the above adjoint solutions, the Fredholm conditions are 0 = (1 − Q1 /3)η¯ 1¯ = −Q1 (1 − Q1 /3) ⇒ Q1 = 3 since Q1 6 = 0. Thus as claimed we obtain Q1 = Q1¯ = 3. Therefore by Result VI Eq. (4.8) follow. iii) The proof requires us to solve (4.8) for 8α , α = 1, 1¯ and then obtain the potential – see (4.9) – via the second logarithmic derivative. It is a matter of straightforward but tedious linear algebra. From the potential we see that u = 0( r12 ) as r 2 ≡ x 2 + y 2 → ∞. Hence the potential is real, nonsingular and rationally decaying. Also, as are the others in this class, it is in L2 but not in L1 . This “slow" decay allows the index to be three. See Fig. 2 for a typical plot of u(x, y). u t Simple + triple pole . Assume that µ has the following structure: µ = 1 + µ(k) ˆ +
81¯ 92 93 81 + + + , 2 3 k − k1 (k − k1 ) (k − k1 ) k − k¯1
(4.10)
where µ(k) ˆ is regular at both k1 , k¯1 and tends to 0 as k → ∞, Dim Ker Gk1 = 1 and 93 6 = 0. Then i) Q1 = 3. ii) The following system of linear algebraic equations is obtained:
24
J. Villarroel, M. J. Ablowitz
a) at k = k1 from Eq. (3.13), {92 + h1 93 }k=k1 = 0, {8 + h2,t 93 }k=k1 = 0 {ν + h3,t 93 }k=k1 = 0,
(4.11.1) (4.11.2) (4.11.3)
b) at k = k¯1 , from Eq. (3.3.3), 2 1d ν dν + h2 ν + h3 81¯ + h1 = 0. 2 dk 2 dk k=k¯1 iii) If µ(k) ˆ = 0, k1 ≡ a + ib, δs = −δ¯t ≡ δR + iδI , s = −¯t ≡ R + iI , the corresponding reflectionless potential is u(x, y) = 2∂xx log F , where i2 h F = z3 − 12b2 y 02 z + 3zδR + 6by 0 (δI − 2y 0 ) + I h i2 + 8b3 y 03 − 6bz02 y 0 + 3z(δI − 2y 0 ) − 6by 0 δR + R 9 h z · 2 (z2 − 4b2 y 02 + δR − 4b b 1 2 1 1 2 1 0 2 02 + 2 ) + (4y bz − δI ) + 2 (z − ) + 4y + 4 (4.12) 2b b b 4b Proof. The proof goes along the same lines of the examples already considered where one needs to solve for 8α , α = 1, 1¯ and then obtain the potential via the second logarithmic derivative of the function F given in Eq. (4.12). Again this is a matter of substantial but straightforward linear algebra. u t See Fig. 3 for a typical plot of u(x, y). Other reflectionless potentials. More complicated potentials/solutions can be constructed by considering pairs of poles kl , k¯l of higher orders and deriving the corresponding linear algebraic equations for the residues. This procedure allows one, in principle, to find infinitely many real, nonsingular and rationally decaying potentials of the nonstationary Schrödinger operator associated with its discrete spectrum. Below we consider some of the simplest cases. Reflectionless potentials corresponding to superposition of simple and double poles with Q=2. Let C, D, E, G be N × N matrices defined by 1 1 − (4.13.1) (1 − δlj ), Clj = h2l δlj + h1l (kl − kj ) h1l (kl − kj )2 ! ! 1 1 1 1 2 − − Dlj = h1l − , (4.13.2) (kl − k¯j ) (kl − k¯j )2 h1l h¯ 1j h¯ 1j h1l (kl − k¯j )3 Elj =
−h¯ 1l , ¯ (kl − kj )
Glj = h¯ 2l δlj − h¯ 1l
(4.13.3) 1 1 + ¯ ¯ ¯ ¯ (kl − kj ) h1,j (kl − k¯j )2
! (1 − δlj ),
(4.13.4)
Discrete Spectrum of Nonstationary Schrödinger and Kadomtsev–Petviashvili Equation
25
14 12 10 8 6 4 2 0 -2 -10
-4 -3
-5 -2 y0
-1 x 0
5 1 10
Fig. 3. A multipole order 3 solution corresponding to a simple plus triple pole configuration with parameters (a = 0, b = 1, γ = δ = = 0) – c.f. (6.14), (4.12)
where the functions h1l , h2l , l = 1, . . . N (with relevant constants γs,l , δs,l ) are defined in Eq. (3.2). Let F stand for the determinant of the 2N × 2N matrix(l = 1, . . . N, j = 1, . . . N), CD B= . EG Then
∂2 log F ∂x 2 is a real nonsingular reflectionless potential for the nonstationary Schrödinger operator. The wave function for this potential is given by u(x, y) = 2
µ=1+
N X j =1
! 92,j¯ 8j¯ 8j + , + (k − k¯j )2 (k − k¯j ) (k − kj )
where the Laurent coefficients satisfy a linear algebraic system given below. Proof. Assume that the eigenfunction has the above given structure and we require that Ql = 2, l = 1 . . . N. Then according to the results (3.9.1–2), at any of the poles k¯l , l = 1 . . . N we have {8l¯ + h1l¯92,l¯}k=k¯l = 0, {ν + h2l,d ¯ 92,l¯}k=k¯l = 0. Using the first of the above equations to solve for 92,l¯ in terms of 8l¯ and substituting this equation into the second (recalling that ν stands for the nonsingular part of µ), we
26
J. Villarroel, M. J. Ablowitz
have
−h2l,d ¯ 8l¯ + h1l¯
N X
1 1 − ¯ ¯ ¯ (kl − kj ) h1j¯ (kl − k¯j )2
j =1,j 6=l
! 8j¯ + h1l¯
N X j =1
8j = −h1l¯; ¯ (kl − kj )
furthermore from (3.3.2) at any of the poles kl , l = 1 . . . N the following equations apply dν + h1l ν + h2l 8l = 0, dk k=kl and as before this yields,
h2l 8l + h1l
N X j =1
1 1 2 1 1 ( + )+ − h1j¯ (kl − k¯j ) (kl − k¯j )2 h1l h1j¯ h1l (kl − k¯j )3
N X
+ h1l
(
j =1,j 6=l
! 8j¯
1 1 − )8j = −h1l . (kl − kj ) h1l (kl − kj )2
In the above formulae for l = 1 . . . N the functions h1l¯, h2l,d ¯ (with relevant constants γd,l¯, δd,l¯) and h1l , h2l (with relevant constants δs,l , γs,l ) are defined in (3.2),(3.9). Taking ¯ 2l , h ¯ = −h¯ 1l , the parameters to satisfy γd,l¯ = γ¯s,l ; δd,l¯ = −δ¯s,l implies that h2l,d ¯ = −h 1l then the above equations are:
h2l 8l + h1l
N X ( j =1
8j
(kl − k¯j )
N X
−
j =1,j 6=1
8j ) h1l (kl − kj )2
! 1 1 2 1 1 ( − )− − 8j¯ = −h1l , + h1l (kl − k¯j ) (kl − k¯j )2 h1l h¯ 1j h¯ 1j h1l (kl − k¯j )3 j =1 ! N N X X 8j 1 1 ¯ − h + = h¯ 1l . 8 h¯ 2l 8l¯ − h¯ 1l ¯ 1l j (k¯l − k¯j ) h¯ 1,j (k¯l − k¯j )2 (k¯ − kj ) j =1,j 6=l j =1 l N X
For l = 1, . . . N, this system is equivalent to
CD EG
−h1l 8l = ¯ , 8l¯ h1l
where C, D, E, G are defined in (4.13). Using u(x, y) = −2i result follows. u t
PN
∂ j =1 ∂x (8j
+ 8j¯ ) the
Discrete Spectrum of Nonstationary Schrödinger and Kadomtsev–Petviashvili Equation
27
Reflectionless potentials corresponding to superposition of simple and triple poles with Q = 3. Let C, D, E, G be N × N matrices defined by ! h1l 1 + 2 (1 − δlj ), h2l − klj klj h¯ 1,j h1l 1 2h 3 h2l − 1l + + 2 − kl j¯ Dlj = h2l − kl j¯ kl j¯ kl j¯ kl2j¯ h¯ 2j kl j¯ 3h1l 6 1 + 2 , h2l − − kl j¯ kl j¯ h¯ 2j kl2j¯ 1 Clj = h3l δlj + klj
Elj = −
h¯ 2l , klj ¯
Glj = h¯ 3l δlj − h¯ 2l
(4.14.1)
(4.14.2)
(4.14.3) h¯ 1,j 1 1 + + 2 k¯lj h¯ 2j k¯lj h¯ 2l k¯lj3
! (1 − δlj ).
(4.14.4)
In the above formulae for l = 1 . . . N, the functions h1l , h2l , h3l (with relevant constants γs,l , δs,l , s,l ) are defined in (3.2) and we set klj ≡ kl − kj , kl j¯ ≡ kl − kj¯ , k¯lj ≡ k¯l − k¯j . Let F stand for the determinant of the 2N × 2N matrix CD B= . EG Then u(x, y) = 2
∂2 log F ∂x 2
is a real nonsingular reflectionless potential for the nonstationary Schrödinger operator. The wave function for this potential is given by µ=1+
N X j =1
! 92,j¯ 8j¯ 93,j¯ 8j + + + , (k − k¯j )3 (k − k¯j )2 (k − k¯j ) (k − kj )
where the Laurent coefficients satisfy a linear algebraic system given below. Proof. Assume that the eigenfunction has the above given structure where we require that Ql = 3, l = 1 . . . N. Then according with the result (3.3.3) at any of the poles kl , l = 1 . . . N, the following equations apply: {
dν 1 d 2ν + h1l + h2l ν + h3l 8l }k=kl = 0, 2 2 dk dk
while at k¯l , l = 1 . . . N (3.13) yields 92,l¯ + h1l¯93,l¯ = 0, 8l¯ + h2l,t ¯ 93,l¯ = 0, {ν + h3l,t ¯ 93,l¯}k=k¯l = 0.
28
J. Villarroel, M. J. Ablowitz
We take the constants to satisfy ¯s,l = −t,l¯ ; δ¯s,l = −δt,l¯ ; after tedious algebra of reducing to equations with 2N unknowns 8j , 8j¯ we obtain ! h1l 1 + 2 8j h2l − −h2l = h3l 8l + klj klj j =1,j 6=l ¯ h1,j h1l 1 h2l − 2h1l + 3 + h2l − + 2 − kl j¯ kl j¯ kl j¯ kl2j¯ h¯ 2j kl j¯ 8 3h1l 6 1 j¯ + 2 h2l − − − kl j¯ kl j¯ kl j¯ h¯ 2j kl2j¯ N X
and
N X h¯ ¯ h¯ 2l¯ = − 2l 8j − h¯ 2l¯ klj ¯
j =1,j 6=l
1 klj
h¯ 1j 1 1 + + 2 k¯lj h¯ 2j k¯lj h¯ 2l k¯lj3
! 8j¯ + h¯ 3l¯8l¯.
For l = 1, . . . N this system is equivalent to −h2l CD 8l = ¯ , 8l¯ EG h2l where C, D, E, G is defined in (4.14). Using u(x, y) = −2i follows. u t
P
x (8j
+ 8j¯ ) the result
Reflectionless potentials corresponding to superposition of double poles with Q = 3 . Let C, D, E, G be N × N matrices defined by ! 3h1l gl 1 2 − + (4.15.1) gl − (1 − δlj ), Clj = −h3l,d δlj + klj h1j klj klj h1j klj2 3h1l gl 1 2 − + Dlj = gl − , (4.15.2) kl j¯ h1j¯ kl j¯ kl j¯ h1j¯ kl2j¯ 3h1l¯ g ¯ 1 2 l g¯ l − , + − Elj = 2 klj h1j klj klj h1j klj ¯ ¯ ¯ ¯ ! 3h1l¯ g ¯ 1 2 l + − (4.15.4) g¯ l − (1 − δlj ). Glj = −h¯ 3l,d δlj + h ¯ k¯lj h ¯ k¯ 2 k¯lj k¯lj 1j
1j lj
In the above formulae for l = 1 . . . N the functions h1l,d , h3l,d (with relevant constants γd,l , βl , and σl ) are defined in (3.9.2-3) and klj ≡ kl − kj , k¯lj ≡ k¯l − k¯j , kl j¯ ≡ kl − kj¯ . Besides gl ≡ h1l,d + 2σl . Let F stand for the determinant of the 2N × 2N matrix, CD B= . EG
Discrete Spectrum of Nonstationary Schrödinger and Kadomtsev–Petviashvili Equation
29
Then
∂2 log F ∂x 2 is a real nonsingular reflectionless potential for the nonstationary Schrödinger operator. The wave function for this potential is given by ! N X 8j¯ 92,j¯ 8j 92,j + + , + µ=1+ (k − kj )2 (k − kj ) (k − k¯j )2 (k − k¯j ) u(x, y) = 2
j =1
where the Laurent coefficients satisfy a linear algebraic system given below. Proof. Assume that the eigenfunction has the above structure and require that Ql = 3, l = 1, . . . N. Then Eqs. (3.9.1–3) imply
8α = −h1α 92.α ; 1 dνα + (h1α + 2σα )να + h3α,d 92.α = 0 dk 3
at both k = kl and k = k¯l . These equations yield ! X N N X 8j¯ 92,j¯ 8j 92,j 1 + + h3l,d 92l + gl + 3 (kl − kj )2 (kl − kj ) (kl − k¯j )2 (kl − k¯j ) j =1,j 6=l j =1 ! X N N X 292,j¯ 8j¯ 8j 292,j = −gl , + + + − (kl − kj )3 (kl − kj )2 (kl − k¯j )3 (kl − k¯j )2 j =1,j 6=l
N X j =1,j 6=l¯
j =1
N X
8j¯ 92,j¯ + + 2 ¯ (kl¯ − kj ) (kl¯ − k¯j ) j =1 j =1,j 6=l¯ ! 292,j¯ 8j¯ 8j 292,j + + + = −gl¯ (kl¯ − kj )3 (kl¯ − kj )2 (kl¯ − k¯j )3 (kl¯ − k¯j )2
1 h ¯ 9 ¯ + gl¯ 3 3l,d 2l −
N X
8j 92,j + 2 (kl¯ − kj ) (kl¯ − kj )
!
for l, l¯ = 1, . . . N. Note that gl¯ = −g¯ l if the constants are chosen to satisfy βl¯ = β¯l , γl¯ = γ¯l + 2i σ¯ l , σl¯ = σ¯ l .After elimination of variables we obtain the system ! N X gl 1 2 3h1l − + gl − 8j − h3l,d 8l + k h1j klj klj h1j klj2 j =1,j 6=l lj N X gl 1 2 3h1l + − + gl − 8j¯ = −3gl h1l , l = 1, . . . N 2 kl j¯ h1j¯ kl j¯ kl j¯ h k ¯ 1 j ¯ lj j¯=1,j¯6=l N X 3h1l¯ g ¯ 1 2 l g¯ l − 8j + − − h¯ 3l,d 8l¯ + 2 klj h1j klj klj h k ¯ ¯ ¯ 1j ¯ lj j =1,j 6=l¯ ! ! 3h ¯ g¯ l 1 2 + 1l g¯ l − + − 8j¯ = −3g¯ l h1l¯. h ¯ k¯lj h ¯ k¯ 2 k¯lj k¯lj 1j
1j lj
This implies the result in the standard way. u t
30
J. Villarroel, M. J. Ablowitz
More general potentials. The cases detailed above are some of the new families that follow from our results. While we shall not elaborate any further on this, we remark that one could also consider a mixture of poles of different types as well as adding in the contribution of the continuous spectrum by taking µ(k) ˆ 6 = 0 and following the procedure discussed in refs. [1,2]. 5. Coalescence of Poles In this section we briefly describe how some of our results can be obtained by appropriately coalescing pairs of conjugate and simple poles of the eigenfunctions of the Schrödinger operator. We show this for N = 2 in formulae (1.15–16). Let us assume that the discrete spectrum consists of two pairs of simple poles, i.e. µ(k) is written: µ(k) = 1 + µ(k) ˆ +
4 X α=1
φα , (k − kα )
where kj +2 = k¯j , j = 1, 2 and all the indices Qα =1. In the limiting process we take k1 ≡ a + ib = k2 + ε, k3 = k¯1 , k4 = k¯2 , |ε| << 1, assume the following expansion for both the constants and residues of (1.10, 1.12),
γα,0
(−1)
φα + φα(0) + φα(1) ε + . . . , ε (−1) (0) (1) = γα,0 /ε + γα,0 + γα,0 ε + . . . , α = 1, . . . 4,
φα =
and substitute these expansions into Eq. (1.15). At order 1/ε2 there are various possibil(−1) (−1) (−1) (−1) ities. Reality requires γ1,0 = γ4,0 = −γ2,0 = −γ3,0 = −i. In order for µ to have (−1)
(−1)
(−1)
(−1)
= φ3 +φ4 = 0. a finite limit as ε tends to zero, we are forced to take: φ1 +φ2 Proceeding to subsequent orders is straightforward. It is convenient to define new vari(0) (0) (0) (0) (−1) (−1) (−1) ≡ 92 . We find that φ3 = φ4 = 0, ables: φ1 +φ2 ≡ 81 , φ3 +φ4 = 81¯ , φ1 (−1) (0) (0) (0) (0) (0) = i81 /(f + γ ), γ ≡ γ1,0 , f ≡ f (k1 ), φ3 = φ4 . Letting ε → 0 the φ1 above equations imply that the eigenfunction µ has the following spectral structure: µ = 1 + µ(k) ˆ +
81¯ 81 92 . + + (k − k1 )2 (k − k1 ) (k − k¯1 )
Besides this one also obtains that the new functions 92 , 81 , 81¯ satisfy the system of Eqs. (3.9.1–2) and (3.3.2). Note next that the equality Z 1 u φα dxdy 1 = Qα ≡ 2πi with the above results implies that Z Z 1 1 (0) (0) u81 dxdy = u(φ1 + φ2 )dxdy = 2 Q(81 ) = 2πi 2πi Z Z 1 1 (−1) u92 dxdy = uφ1 dxdy = 0. 2πi 2π i Thus for this case we recover our results (3.8–9) and and (3.3.2). and
Q(92 ) =
Discrete Spectrum of Nonstationary Schrödinger and Kadomtsev–Petviashvili Equation
31
6. KPI Solutions The above class of potentials is related to localized rationally decaying solutions of the KPI equation when the appropriate temporal dependence is inserted. Requiring µ(x, y, t, k) to solve also (1.4) implies a particular time evolution for the constants γ , β, ... entering in the potential. The procedure, though tedious, is standard and we will not dwell on the derivation. Here we present the solutions corresponding to potentials derived earlier and then we study their dynamics. Solutions to KPI are obtained when the constants of (3.2),(3.9),(3.13) “evolve” as k˙α (t) = 0 and δ˙α (t) = −24ikα , β˙α (t) = 12(1 + 6kα σ ),
γ˙α (t) = 12kα2 , σ˙ (t) = 0,
˙α (t) = 24i
(6.1)
α = 1 . . . 2N.
Simple poles. The simplest case corresponds to (4.1) with µˆ = 0. Using (6.1) the KPI solution u(x, y, t) is obtained by setting x → x − 12(a 2 + b2 )t, y → y − 12at
(6.2)
in (4.3). Corresponding to n = 2N simple poles the resulting solution KPI solution, called the N lump solution, is obtained by setting γα → γα (t) = γα (0) + 12kα2 t in (1.16). It describes a configuration that asymptotically is made up of N-humps all of them moving uniformly with distinct velocities vj x = 12(aj2 + bj2 ), vjy = 12aj (here kj ≡ aj + ibj ). Upon interaction there is neither a phase shift nor a deflection, hence the asymptotic dynamics of these objects is trivial. Simple + double poles. Again, the easiest case corresponds to (4.4) with µˆ = 0. The KPI solution follows from (4.6) and (6.1) by taking x → x − 12(a 2 + b2 )t, y → y − 12at, δR → δR − 24bt.
(6.3)
This gives (1.20). Study of this KP solution shows that for long times it has two maxima or humps (see Fig. 1), each of which moves with distinct asymptotic velocities and diverge from one another, proportional to |t|1/2 ( this “anomalous scattering” was first discussed in [6] with a particular choice of some of the relevant parameters). In Appendix IV we show that as t → −∞ the two maxima (+ denotes fast, - denotes slower hump) are located at q x ∼ 12(a 2 + b2 )t ± −24a 2 t + x − 1 , ± 0 b 2b q (6.4.1) y ∼ 12at ± − 6t + y ±
b
0
and at t → ∞ the two maxima are located at ( √ 1 ) x± = 12(a 2 + b2 )t ± 24bt + 0( t 1/2 y± = 12at + y0 .
(6.4.2)
32
J. Villarroel, M. J. Ablowitz
10
5
-150
-100
-50
0
50
100
-5
-10
Fig. 4. Trajectories of the maxima – see (1.20), (6.4.1)
Note that as t → −∞ the trajectories follow a hyperbola given by (Fig. 4) 2 (a 2 − b2 ) (x − 2ay) = 0. (6.5) (a 2 + b2 )y − ax + 2b The asymptotic motion of the humps can be viewed as a two particle dynamical system. To describe the dynamics of this motion it is convenient to consider a frame with the origin at (x0 , y0 ) which moves with velocities vx = 12(a 2 + b2 ), vy = 12a, with respect to the rest frame. With respect to this Galilean frame as t → −∞ one of the humps is located in the first quadrant while the other is the mirror image in the third quadrant. As t → ∞ both of them travel on the x axis, the first one moving left, the second right. The asymptotic amplitude is u(x± , y± ) = 16b2 for both maxima as |t| → ∞. We note that these “particles” are not distinguishable after scattering. Here we take the natural election that corresponds to the situation described above: i.e. say that the hump which drifts to the left as t → ∞ was the one located at the far right of the upper half quadrant at t → −∞. With this choice the angle the humps get deflected is given by: 1 . (6.6) = arc tan 2|a| By properly choosing the free parameter a we can obtain a solution that scatters with any desired angle. This is in contrast to the situation for standard lumps where this angle is zero (i.e. there is no phase shift). In this dynamical analogue the asymptotic motion of the humps can be thought of as being the result of an interaction between the humps. To determine the nature of this interaction define xi,j ≡ xi − xj , yi,j ≡ yi − yj , i, j = +, − r i ≡ (xi , yi ), r i,j ≡ (xi,j , yi,j ), ri,j ≡ |r i,j |
(6.7)
Discrete Spectrum of Nonstationary Schrödinger and Kadomtsev–Petviashvili Equation
33
2 (note r+,− = (x+ − x− )2 + (y+ − y− )2 ). The above trajectories as t → −∞ of the humps are solutions of the following system: X r i,j d 2r i = f (ri,j ), (6.8) 2 dt ri,j j,j 6=i
where the attractive force between particles is f (r) ≡ −(
24a 2 2 (1 + 4a 2 )2 . ) b 4r 3
(6.9)
These humps attract each other but they do not form a bound state since the attractive interaction is not strong enough to bind them together. They eventually split apart horizontally as they travel parallel to the y axis. Similar considerations apply as t → ∞. We note the equivalent force law only has a different constant. We remark finally that some solutions with nontrivial interaction properties of other 2+1 integrable equations have been derived using direct methods (cf. [12-14]). Two double poles. We consider (4.7) with µˆ = 0. Solutions to KPI are obtained by taking x → x − 12(a 2 + b2 )t, y → y − 12at, βˆ → βˆ + 12(1 + 6ibσ )t,
(6.10)
σ = σR + iσI , constant, in (4.9). Hence we obtain u(x, y, t) = 2∂xx log F , 9 1 2 1 1 2 1 2 F = |h3,d | + 2 |ih1 + | + 2 |−i(h1 + 2σ ) + | + 2 4b 2b 4b 2b 4b 2 3 2 02 2 2 0 0 = z −12b y z + 3(σR −σI )z + 12bσI σR y −6σR y + 12(1−6bσI )t + βˆR 2 + 6bz2 y 0 −8b3 y 03 −6σI σR z + 6b(σR2 −σI2 )y 0 + 6σI y 0 −72tbσR − βˆI 9 1 2 1 0 2 + 2 (z−σI − ) + (2by −σR ) + 2 4b 2b 4b ) ( 2 1 1 2 0 (6.11) + 2by + σR + 2 . · z + σI + 2b 4b This solution depends on 8 real parameters. Generically it describes a three humped structure with the position of the humps given asymptotically as |t| → ∞ by a ϕ + 2πj ϕ + 2πj )− sin ( ) + o(1), xj ∼ 12(a 2 + b2 )t + x0 + (ζ t)1/3 cos ( 3 b 3 yj ∼ 12at + y0 −
ϕ + 2πj (ζ t)1/3 sin ( ) + o(1), j = 0, 1, 2, 2b 3
where ζ = 12{(1 − 6bσI )2 + 36b2 σR2 }1/2 , tan ϕ = Thus the humps are located on the conic (x 0 − 2ay 0 )2 + 4b2 y 02 = (ζ t)2/3
6bσR . (1 − 6bσI )
(6.12)
34
J. Villarroel, M. J. Ablowitz
4
2
-200
-100
0
100
200
-2
-4
Fig. 5. Trajectories of the humps following (6.110 )–(6.12) in the rest frame with appropriate choices of constants
moving with uniform speed vx = 12(a 2 + b2 ), vy = 12a and whose semiaxes are first decreasing and then increasing at a rate |t|1/3 . Moreover in the moving frame the humps translate along three straight lines. After collision the formulae (6.12) shows that the particles follow their original trajectories without any scattering or phase shift. Thus there is no scattering although there is interaction that is manifest in the nontrivial trajectories (6.12) evolving as 0(t 1/3 ) as t → ∞. The expression for the solution simplifies if σ = 0 in which case F = (z3 − 12b2 y 02 z + 12t + βˆR )2 + (8b3 y 03 − 6bz2 y 0 + βˆI )2 9 1 1 1 2 1 2 2 02 2 02 + (z − ) + 4b y + 2 (z + ) + 4b y + 2 . 4b2 2b 4b 2b 4b (6.110 ) This corresponds to a three humped structure (see Fig. 2) with the position of the humps given asymptotically by (6.12), where ζ = 12, ϕ = 0 (a = 0, b = 1). The trajectories of the above humps (see for example Fig. 5) are solutions of the following system: X r i,j d 2r i = f (ri,j ) , i = 1, 2, 3. 2 dt ri,j j,j 6=i
A direct calculation similar to (6.8) shows that the “force between particles” is (for convenience, we take here a = 0) 1 1 9 f (r) ≡ − (3 + 2 )3 5 , 2 4b r which again corresponds to an attractive interaction.
Discrete Spectrum of Nonstationary Schrödinger and Kadomtsev–Petviashvili Equation
Another subcase of interest is that given by σ =
i 6b ,
35
which corresponds to
2 2 z y0 ˆR + 8b3 y 03 − 6bz2 y 0 − 5 + βˆI + β F = z3 − 12b2 y 02 z − 12b2 6b ( ) ( )! 2 2 2 2 9 1 1 2 02 2 02 z+ z− . + 2 + 4b y + 2 + 4b y + 2 4b 3b 4b 3b 4b (6.13) We note that the temporal dependence drops out and the analysis above for the hump trajectories ceases to be valid. Equation (6.13) corresponds to a three humped KP solution that is stationary (cf. also ref. [7]) which moves with uniform velocity vx = 12(a 2 + b2 ) > 0, vy = 12a and hence can be regarded as a solitary wave. Thus this solution, unlike the others considered which depend on all the variables z, y 0 , t, also solves the Boussinesq equation. Solutions that behave asymptotically as a nonlinear superposition of these multipeaked solitary waves are obtained in a straightforward way from the analysis in Sect. 4. Namely, consider the function µ=1+
N X j =1
! 92,j¯ 8j¯ 92,j 8j , + + + (k − kj )2 (k − kj ) (k − k¯j )2 (k − k¯j )
with Qj ≡ Q(kj ) = 3 and the constants kj , βj , γj , σj satisfying kj¯ = k¯j , βj¯ = β¯j , γj¯ = γ¯j + 2i σ¯ j , σj¯ = σ¯ j =
i . 6bj
All the residues satisfy
1 dνα ¯ . . . N¯ . + (h1,α + 2σα )να + h3.α,d 92.α = 0, α = 1, . . . N, 1, dk 3
As |t| → ∞ the ensuing solution behaves as a superposition of these multipeaked solitary waves: u∼2
N X
∂ 2 (log Fj )/∂x 2 ,
j =1
where Fj is given by (6.13) corresponding to parameters kj , βj , γj , σj = individual humps suffer no phase shift due to interaction.
i 6bj
. The
Solutions coming from simple + triple poles . Consider (4.10) with µˆ = 0. Using (6.1) results in the solution to KPI given by formula (4.12) with x → x − 12(a 2 + b2 )t, y → y − 12at, δR → δR − 24bt, I → I − 24t.
36
J. Villarroel, M. J. Ablowitz
Hence
i2 h F = z3 − 12b2 y 02 z + 3z(δR − 24bt) + 6by 0 (δI − 2y 0 ) + I − 24t h i2 + 8b3 y 03 − 6bz02 y 0 + 3z(δI − 2y 0 ) − 6by 0 (δR − 24bt) + R 1 9 z + 2 (z2 − 4b2 y 02 + δR − 24bt − + 2 )2 + (4y 0 bz − δI )2 4b b 2b 1 1 2 1 02 + 2 (z − ) +4y + 4 . (6.14) b b 4b
Detailed study of this multipole order-3 configuration for the KPI equation shows, in a similar way to earlier cases such as (6.4–6.5), that it is composed of three mutually interacting humps (see e.g. Fig. 3), which move with distinct asymptotic velocities. 0 , y 0 ), (x 0 , y 0 ), As t → ±∞ the three maxima are located at (assuming b > 0) (x+ + 0 0 0 , y 0 ), where x 0 = 12(a 2 + b2 )t + x , y 0 = 12at + y , (x− 0 0 − 0 0 q x 0 ∼ x 0 − 4/3b ± −72a 2 t, ± 0 b q t → −∞ y 0 ∼ y 0 ± −18t , ±
0
(
and
0 ∼ x0 ± x± 0 0 ∼ y0 , y± 0
b
√ 72bt +
1 6b ,
t → +∞.
(6.15)
From a dynamical point of view it is convenient to consider a frame with the origin at (x0 , y0 ) which moves with velocities vx = 12(a 2 + b2 ), vy = 12a, with respect to a rest frame. In this comoving frame one of the humps remains at rest at the origin (0, 0). 2 The asymptotic amplitude of the hump at rest is given by u = 720 169 b . The two remaining humps follow a nonuniform motion and undergo scattering. As t → −∞ one of the humps is located in the first quadrant, while the other is the mirror image. As t → ∞ both of them travel on the x axis, the first one moving left, the second right. The asymptotic amplitude u(x± , y± ) = 16b2 for both maxima as |t| → ∞. As before these two humps are not distinguishable after scattering. The angle by which the humps get deflected due to interaction is given again by 1 . Thus properly choosing the free parameter a we can obtain a solution = arc tan 2|a| that scatters off in any desired angle. Finally, the above trajectories of the humps are solutions of the following system: X r i,j d 2r i = f (ri,j ) dt 2 ri,j
i = +, 0, −,
j,j 6=i
where, in this case, f (r) ≡
−2 A , with A = (18/b)2 , t → −∞, and A = (72b)2 , t → +∞. 9 r3
(6.16)
Therefore the total interaction can be considered as sum of two particle interactions with force between particles given by (6.16). Note that the hump at rest is subject to forces from the other two humps. The absence of motion of this hump can be traced to the fact that the above two forces are equal in magnitude but opposite in direction.
Discrete Spectrum of Nonstationary Schrödinger and Kadomtsev–Petviashvili Equation
37
KPI solutions corresponding to superposition of simple and double poles with Q = 2. Let C, D, E, G be the N × N matrices defined in (4.13), and let F stand for the determinant of the 2N × 2N matrix, CD B= . EG Then if the constants γl (t), δl (t) evolve as required by (6.1), u(x, y, t) = 2
∂2 log F (x, y, γl (t), δl (t)) ∂x 2
is a real nonsingular KPI solution. KPI solutions corresponding to superposition of simple and triple poles with Q = 3. Let C, D, E, G be N ×N the matrices defined in (4.14) where the constants γl (t), δl (t), l (t) evolve as required by (6.1) and let F stand for the determinant of the 2N × 2N matrix, CD B= . EG Then
∂2 log F (x, y, γl (t), δl (t), l (t)) ∂x 2 is a real nonsingular KPI solution. u(x, y, t) = 2
KPI solutions corresponding to superposition of double poles with Q = 2 . Let C, D, E, G be N × N matrices defined in (4.15) where the constants γl (t), βl (t), σl (t) evolve as required by (6.1) and let F stand for the determinant of the 2N × 2N matrix, CD B= . EG Then
∂2 log F (x, y, γl (t), βl (t), σl (t)) ∂x 2 is a real nonsingular KPI solution. Assume furthermore that the constants are chosen satisfying u(x, y, t) = 2
i . 6bl Then as |t| → ∞ the ensuing solution behaves as a superposition of triply peaked multilump waves: N X ∂ 2 (log Fj )/∂x 2 , u∼2 βl¯ = β¯l , γl¯ = γ¯l + 2i σ¯ l , σl¯ = σ¯ l =
j =1
where Fj is given by (6.13) corresponding to parameters kj , βj , γl , σj = 6bi j . In each of these subgroups j = 1, . . . N the triply peaked lump is a stationary solution of KPI. The individual solitary waves suffer no phase shift due to interaction. Acknowledgements. This work was partially sponsored by the Air Force Office of Scientific Research, Air Force Materials Command, USAF, under grant number F49620-97-1-0017. This work was also partially supported by NSF grant DMS-970350 and by CICYT 0028/95 and Junta de Castila-Leon JADZ in Spain.
38
J. Villarroel, M. J. Ablowitz
Appendix I For convenience in this section we define 91,α ≡ 8α , 90,α ≡ να . Let the system of functions {9r,α }, α = 1..2N, r = 0..M, M ≥ 2 satisfy ∂G 9M,α }k=kα = · · · ∂k ∂ M−1 G ∂G 1 92,α + .. + 9M,α }k=kα = 0, = {G91,α + ∂k (M − 1)! ∂k M−1 ∂G 1 ∂ 2G 1 ∂M G 92,α + · · · + 9M,α }k=kα = 1. 91,α + {G90,α + 2 ∂k 2 ∂k M! ∂k M {G9M,α }k=kα = {G9M−1,α +
(2.8)
Then the orthogonality relationships below hold < 9s,β , 9r,α >= 0, ∀r, s = 0 . . . M, ∀β 6 = α¯ < 9s,α¯ , 9r,α >= 0, ∀r, s = 0 . . . M, satisfying r + s > M, < 9s+1,α¯ , 9r,α >=< 9s,α¯ , 9r+1,α >, ∀r, s = 0 . . . M
(A1.1) (A1.2) (A1.3)
and we define 9M+1,β ≡ 0. In particular this implies < 9r,α¯ , 91,α >=< 91,α¯ , 9r,α >, ∀r = 0 . . . M
(A1.4)
The above means the following: Consider a net of points (s, r), r, s = 0 . . . M. Then < 9s,β , 9r,α >= 0, ∀β,α if the point (s, r) is above the diagonal r + s = M, < 9s,β , 9r,α >= 0 if β 6 = α¯ and the point (s, r) is on or below the diagonal. Finally all values < 9s,α¯ , 9r,α > are equal for points (s, r) that are on a common line r + s = Constant. Then all scalar products get determined in terms of the basic ones < 91,α¯ , 9r,α >, r = 1 . . . M − 1 -which are given by the charges Qr,α . Proof. In differential form the above equations yield first (i∂y + ∂xx + u + 2ikα ∂x )9M,α = 0, (i∂y + ∂xx + u + 2ikα ∂x )9r,α + 2i∂x 9r+1,α = 0, r = 0 . . . M − 1 or simply if we define 9M+1,β ≡ 0, (i∂y + ∂xx + u + 2ikα ∂x )9r,α + 2i∂x 9r+1,α = 0, r = 0 . . . M. ¯ s,β yields through integration by parts Multiplication of these equations by 9 < 9s+1,β , 9r,α > +kα,β < 9s,β , 9r,α >=< 9s,β , 9r+1,α >, s, r = 0 . . . M (kα,β ≡ kα − k¯β ). Let first β = α¯ and find (kα,α¯ = 0), < 9s+1,α¯ , 9r,α >=< 9s,α¯ , 9r+1,α >, which is (A1.3.) Next take r = M to get < 9s+1,β , 9M,α > +kα,β < 9s,β , 9M,α >= 0, s = 1 . . . M.
Discrete Spectrum of Nonstationary Schrödinger and Kadomtsev–Petviashvili Equation
39
By induction on s we then find by letting s run from M first to 1, ¯ < 9M,β , 9M,α >= 0, β 6 = α, < 9M,β , 9M,α > +kα,β < 9M−1,β , 9M,α >= 0 ⇒ ¯ and < 9M,α¯ , 9M,α >= 0 < 9M−1,β , 9M,α >= 0, β 6 = α, .. . < 9r,β , 9M,α >= 0, ∀α, β, 1 ≤ r ≤ M, < 90,β , 9M,α >= 0, ∀β 6= α¯ which are (A1.1), (A.1.2) for r = M. The rest of the proof follows through a similar inductive argument. u t
Appendix II Result IV. Let µ(k) belong to the discrete spectrum. Then the indexes Qα satisfy the following properties: PN PN (2.11.1) i) j =1 Qj = j =1 Qj¯ , j = 1 . . . N, Qα = sign(Imkα ) < 8α¯ , 8α >,
ii)
(2.11.2)
Qr,α = sign(Imkα ) < 8α¯ , 9r,α >, ¯ α = Qα¯ ,(2.11.3a) Q ¯ r,j = Q ¯ , j = 1 . . . M Q r,j
iii)
iv) QM,α = 0, where M ≥ 2 is the multiplicity of the pole.
(2.113b) (2.11.4)
Proof. i) Recall that the position of the poles occurring in pairs (kj , k¯j ) can be arranged such that I m(kj ) > 0; j = 1, 2 . . . N. For convenience we take j = 1, 2 . . . N. P ∂ Using u(x, y) = −2i N j =1 ∂x (8j + 8j¯ ) we have that Z Z X N N X 1 u (Qj − Qj¯ ) = (8j + 8j¯ )dxdy 2πi j =1
= −1 2π This is Eq. (2.11.1).
−1 2π Z
Z Z
j =1
N ∂ X ( 8j + 8j¯ )2 dxdy ∂x j =1
N X
dy(
j =1
8j + 8j¯ )2 (x = ±∞, y) = 0.
40
J. Villarroel, M. J. Ablowitz
ii) Using reality of the potential and the orthogonality condition A.1.1 for s = 1 we get πQr,j
1 = 2i =π
Z
Z dxdyu9r,j = −
dxdy
2N X
Z ∂x 8l 9r,j =
dxdy
l=1
2N X l=1
2N X
¯ l 9r,j ∂x 8
l=1
< 8l , 9r,j >= π < 8j¯ , 9r,j > .
With r = 1 we obtain that
Qj =< 8j¯ , 8j > .
The above are Eq. (2.11.2). iii) Using the orthogonality condition (A.1.1), Z πQj¯ =
dxdy
2N X l=1
Z ∂x 8l 8j¯ = −
dxdy
Z ⇒ πQj¯ = and
2N X l=1
Z ∂x 8l 8j¯ =
dxdy8j ∂x 8j¯
dxdy8j ∂x 8j¯ ≡ πQj
¯ ¯ π Qr,j =< 8j¯ , 9r,j >=< 9r,j¯ , 8j >= −< 8j , 9r,j¯ > = π Q r,j
by (A1.4) first and then using < f, g >= −< g, f > (this is Eq. (2.11.3)). iv) We have QM,α = sign (Im kα ) < 91,α¯ , 9M,α >= 0 by (A.1.2) with r = M, s = 1. This is (2.11.4). u t Appendix III We first show that the Greens function satisfies sign kI ∂G(x, y, k) = if G − , kI 6 = 0. ∂k 2π i
(A.III.1)
This formula can be obtained by writing the Greens function of (1.8) as Z −1 2 2 ei ((l−k)x−(l −k )y) (θ(y)θ (∓(l − k)) − θ (−y)θ (±(l − k)))dl. G± (x, y, k) = 2πi ∂G ∂k
has two contributions, one stemming from the derivative of the exponential term yielding the first term in (A.III.1), the other arising from the derivative of the Heaviside functions. Next we note that (A.III.1) implies, assuming that the derivatives can be taken under the integral sign, that the operator G defined in Eq. (1.7) satisfies for kI 6= 0, Z sign kI ∂G = −if G + G(if ) + udxdy, ∂k 2π i
Discrete Spectrum of Nonstationary Schrödinger and Kadomtsev–Petviashvili Equation
i.e.
∂G ∂k
acts on any function µ(x, y) by {
sign kI ∂G }µ = −if Gµ + G(if µ) + ∂k 2π i
Similarly we obtain with λ ≡
sign kI 2πi
41
Z uµ(x, y)dxdy.
(A.III.2)
that
∂ 2 G(x, y, k) = (2iy − f 2 )G + ilf, kI 6= 0, ∂k 2 which implies {
∂ 2G }µ = (2iy − f 2 )Gµ − G((2iy + f 2 )µ) + 2f G(f µ) ∂k 2 Z Z − ilf
uµdxdy + il
f uµdxdy.
(A.III.3)
Higher order derivatives are handled similarly. Appendix IV Consider the solution (1.20). We are interested in finding the asymptotic locations of the humps. Generically for x 0 , y 0 = 0(1) then x, y = 0(t), F = 0(t 4 ) and the amplitude behaves as F Fxx − Fx2 1 = O( 2 ). u(x, y, t) = 2∂xx log F = 2 2 F t It is reasonable to expect that for long time the maxima of u are found at those points x = x(t), y = y(t) that are minima of F and thus at which the denominator has smallest order. One can obtain equations for the minima of F : Fx = Fy = 0 and solve those equations asymptotically for |t| → ∞. An alternative and more direct way is to compute the smallest order of F directly from (1.20). Inspection of Eq. (1.20) shows that the smallest terms occur when the highest polynomials vanish at leading order. Namely let us call the asymptotically dominant terms z ∼ z0 (t), y 0 ∼ y0 (t). Then we wish to satisfy (z02 (t) − 4b2 y02 (t) − 24bt)2 ∼ 0, 2y0 (t)(1 + 2bz0 (t)) ∼ 0. There are two cases as |t| → ∞, y0 = 0, r
(a)
t y0 = ± 6| |, b
(b)
(A.4.1) (A.4.2)
p z0 = ± 24|bt|,
bt > 0,
(A.4.3)
z0 = 0,
bt < 0.
(A.4.4)
Further asymptotic can be obtained by assuming fractional power expansions X corrections 1−n Cn t 2 . In both cases (a), (b) one can readily establish of the form n≥0
u∼2
1 Fzz = 16b2 + 0 . F t
(A.4.5)
The asymptotics of other multipole lumps may be calculated directly using similar ideas.
42
J. Villarroel, M. J. Ablowitz
References 1. Ablowitz, M.J. and Clarkson, P.A.: Solitons, Nonlinear Evolution Equations and Inverse Scattering. Cambridge: Cambridge University Press, 1991, p. 516 2. Fokas, A.S. and Ablowitz, M.J.: Stud. Appl. Math. 69, 211 (1983) 3. Manakov, S.V., Zakharov, V.E., Bordag, L.A., Its, A.R. and Matveev, V.B.: Phys. Lett. A 63, 205 (1977) 4. Ablowitz, M.J. and Villarroel, J.: Phys. Rev. Lett. 78, 570 (1997) 5. Johnson, R.S. and Thompson, S.: Phys. Lett. A 66, 279 (1978) 6. Gorshov, K.A., Pelinovskii, D.E. and Stepahyants, Yu.A.:JETP 77, 237 (1993) 7. Galkin, V.M., Pelinovsky, D.E. and Stepanyants, Yu.A.: Physica D 80, 246 (1995) 8. Villarroel, J., Chakravarty, S. and Ablowitz, M.J.: Nonlinearity 9, 1113 (1996) 9. Ablowitz, M.J. and Villarroel, J.: London Mathematics Society Lecture Note Series 255, Cambridge: Cambridge University Press, 1998, p. 151 10. Pelinovsky, D.E.: J. Math. Phys. 39, 5377 (1998) 11. Ablowitz, M.J., Chakravarty, S., Trubatch, A.D., Villarroel, J.: Novel potentials of the nonstationary Schrödiinger equation and solutions of the Kadomtsev–Petviashvili I equation. Department of Applied Mathematics Report #389, University of Colorado, Boulder, CO, USA 12. Mañas, M., Santini, P.: Phys. Lett. A 227, 325 (1997) 13. Ward, R.S.: Phys. Lett. A 208, 203 (1995) 14. Ioannidou, T.: J. Math. Phys. 37, 3422 (1996) Communicated by T. Miwa
Commun. Math. Phys. 207, 43 – 66 (1999)
Communications in
Mathematical Physics
© Springer-Verlag 1999
Deterministic Percolation Ilan Vardi IHES, 35 route de Chartres, 91440 Bures-sur-Yvette, France. E-mail:
[email protected] Received: 1 November 1998 / Accepted: 1 April 1999
Abstract: This paper examines percolation questions in a deterministic setting. In particular, I consider R, the set of elements of Z2 with greatest common divisor equal to 1, where two sites are connected if they are at distance 1. The main result of the paper proves that the infinite component has an asymptotic density. An “almost everywhere” sieve of J. Friedlander is used to obtain the result. 1. Introduction One can often gain insight about a deterministic problem by comparing it to a probabilistic model. For example, Hardy and Littlewood [30] made precise conjectures about the existence of twin primes based on the assumption that the prime numbers are a generic set, i.e., have “pseudo random” properties. Similarly, in [47] the theory of percolation was used to make some precise conjectures about the existence of unbounded walks of bounded step size along Gaussian primes, a question posed by Basil Gordon, see [24] for a survey. Percolation theory deals with the same question of unbounded walks of bounded step size on a lattice, but in a probabilistic context, see [14,23,26]. For example, consider the plane integer lattice {(m, n) : m, n ∈ Z}, and a fixed 0 ≤ p ≤ 1. Say that a lattice point (or site) is open with probability p and closed with probability 1−p. If these events occur independently for each lattice point, what is the probability that there is an unbounded walk using step size k or less? The main result of percolation theory discovered by Broadbent and Hammersley in 1957 [10] is that this exhibits phase transition, in other words, there is a 0 < pc < 1 for which the probability of an unbounded walk is zero if p < pc and one if p > pc (this is called the critical point). For this reason, percolation theory has been of great interest in physics, as it is one of the simplest models to exhibit phase transition. In this paper, I will examine how questions of percolation theory can be posed in a deterministic setting. Thus deterministic percolation is the study of unbounded walks on
44
I. Vardi
a single subset of a graph, e.g., defined by number theoretic conditions. This might be of interest in physics and probability theory as it studies percolation in a deterministic setting and in number theory where it can be interpreted as studying the disorder inherent in the natural numbers. Instead of just providing a conjectural framework for percolation properties of these sets as was done in [47], I would like to show what unconditional results can be obtained so I will focus on R = {(m, n) ∈ Z2 : gcd (m, n) = 1}, where two sites are connected when they are at Euclidean distance 1. This example is more tractable than Gaussian primes yet retains some similar features. Studying the connectivity properties of R was posed as a problem in [15, p. 109]. The analysis of R will use sieve methods. This seems quite natural, as sieve methods can be interpreted as an application of probabilistic methods to number theory, for example, the large sieve [6] shows that arithmetic progressions behave like independent random variables. Moreover, as Doron Zeilberger has shown [50], sieve methods like the ones used in this paper can be thought of as special cases of the Lace Expansion which has been used successfully to study percolation problems [11,29]. The sieve method used here is due to Friedlander [20] and it proves very general “almost everywhere” results. Thus, if U = U (y) is a function increasing to infinity, then the Rosser sieve [36] shows that for every sufficiently large y, the interval [y, y + U ] always contains a number whose smallest prime factor is ≥ U 1/2−ε (Corollary 5.1 below). However, Friedlander’s sieve shows that for almost every y, the interval [y, y+U ] 1/5−ε (Proposition 5.1 below). contains a number all of whose prime factors are ≥ eU One can look at other examples of deterministic percolation, for example, the generalization of R to Rn = {(a1 , . . . , an ) ∈ Zn : gcd (a1 , . . . , an ) = 1}, where two points are connected when they are at Euclidean distance 1. When n ≥ 3, this is substantially simpler since all elements (a1 , . . . , an ) for which (a1 , . . . , aj −1 , aj +1 , . . . , an ) ∈ Rn−1 lie on the infinite component. Thus the case R2 is harder since the reduction to R1 essentially corresponds to primes. Hopefully, the techniques developed here will shed some light on the original question about Gaussian primes which can also be thought of as an analogue of R1 . Sieve techniques for Gaussian primes have been developed by Coleman [12], Fouvry and Iwaniec [19], and Friedlander and Iwaniec [22]. Finally, an interesting problem might be to study deterministic percolation for Fuchsian groups, e.g., by using the recent work of Lalley [38].
2. Problems of Deterministic Percolation Consider a graph G and a subset V of vertices. A vertex, or site will be open if it belongs to V and closed, otherwise. Percolation studies the properties shared by “almost all” subsets of G which have a given density p. One should therefore compare the properties of V with those of a “generic” set of the same density (such questions are investigated for random graphs in [39]). In order to do this one must first answer
Deterministic Percolation
45
Problem 1. Does V have a probability? Since one is trying to give analogues of percolation, one needs an analogue of a probability for a deterministic set. In this paper I will use asymptotic density δ(V) as defined in the next section. As will be seen, calculating the asymptotic density of a finite event can be nontrivial, even in the simplest cases. Problem 2. Does V have an infinite connected component? One might expect there to be an unbounded component if δ(V) > pc (G), where pc (G) is the critical percolation probability of G, and not otherwise. Problem 3. How many unbounded components does V have? For site and bond percolation in Zn , it has been shown that, with probability one, there is a unique unbounded component for p > pc [1,26]. For other models [38], with probability one, there can be zero, one, or infinitely many infinite components, depending on p. Problem 4. Do the unbounded components of V have densities? Percolation theory also predicts that unbounded components have, with probability one, a density, denoted by θ(p). Problem 5. In general, let f (p) be a function defined, with probability one, in the random model. Can the corresponding quantity be defined for V? If so, how is it related to f (δ(V))? Percolation theory considers other functions related to the cluster distribution. For example, χ (p), the average size of a connected component (χ f (p), the average size of finite components when p > pc ), and κ(p), the cluster size per vertex. In the Ising model, θ, χ (or χ f ), and κ represent magnetization, susceptibility, and free energy, respectively. 3. Percolation on R The subject of this section is deterministic percolation for R. Note that R can be thought of as the points of Z2 which are visible from the origin. One can thus restate the percolation properties of R as “Lecture Hall Percolation” (compare with [9]): Consider a classroom with a regular array of tables so students sit only where they can see the teacher. The teacher starts passing out exam booklets to the closest student and each student passes on the booklets to his closest neighbors. What percentage of the class will receive an exam? To get a feeling for the problem, consider Figs. 1 through 4 which depict sites of Z2 , where a disk or radius 1/2 surrounds each open site, so connected sets of disks correspond to connected components. In Fig. 1, 3, and 4, sites are considered open if they belong to R. Figure 1 gives all sites (m, n) with −50 ≤ m, n ≤ 50. Figure 2 represent a 100 × 100 square where sites are randomly open with probability 6/π 2 . In Fig. 3, the sites were (a, b) + (m, n), where (a, b) = (8660313549, 3102586521) and 0 ≤ m, n, ≤ 100. A similar computation was done in Fig. 4, where a, b were randomly chosen 100 digit numbers 818479887666178685800484385810052015576642831879680779054188168941665 4357533055575739845797590171065, 5258627195794198788220463515375759096 523698556698348004291249422348890633805215110415985087652598121.
46
I. Vardi
Fig. 1. Relatively prime pairs near the origin
The reader is invited to explain the obvious differences between the random and deterministic models, e.g., show that independence is false for R. An interesting feature of the experimental data for R is that local properties do not seem to change when increasing the scale. Thus, a 100×100 snapshot appears roughly the same whether at distance from the origin 1010 or 10100 . One can explain this phenomenon as follows: Long unbroken lines correspond to numbers with no small prime factors. Very roughly, a number with no prime factors < k will produce a line with breaks at average spacing k, which is the expected spacing between numbers with no prime factors < ek . For example, numbers with no prime factors < log1+ε X will most likely produce unbroken lines of length ≥ log1+ε X, but the Prime Number Theorem suggests that, on average, consecutive primes are distance log X apart, so these intervals will most likely contain a prime number. Since lines corresponding to primes lie on the infinite component, these lines will most likely belong to the infinite component. Moreover, the number of numbers with no prime factors > log1+ε X is about X/ log log X, so iterating this process seems to indicate that there is a scale invariance X 7 → eX . The line segments produced by this process seem to form a regular grid and, consistent with the philosophy
Deterministic Percolation
47
Fig. 2. Random sites open with probability 6/π 2
of de Gennes [14,23], one observes that the infinite component is a large interconnected mesh with holes. These observations are used in the proof of Theorem 3.4 below and will be the key to the proof that the infinite component has an asymptotic density. Having gained some feeling for the empirical evidence, one now tries to address the problems of deterministic percolation. Problem 1. Asymptotic density. In the case of R, the answer to Problem 1 is known: The asymptotic density of R is 1/ζ (2), which is the well known result that the “probability” that two random integers are relatively prime is 6/π 2 . In order to prove this, one needs to define density Definition. The asymptotic density of an event P (z) occurring in R is δ(P ) = lim δ(P , R), where δ(P , R) = R→∞
|{z ∈ R ∩ B(R) : P (z) holds}| , |B(R)|
48
I. Vardi
Fig. 3. Relatively prime pairs near (1010 , 1010 )
whenever this limit exists. Here B(R) = B2 (R) = {z ∈ Z2 : kzk < R}, and k(m, n)k = max(|m|, |n|) gives a square summation. Remark. One can also use circular summation, i.e., the usual Euclidean norm |z| and B◦ (R) is a disk of radius R. The subtle differences between such choices is well known, e.g., in the theory of multiple trigonometric series [2,7] and will be discussed below. The density of R is easily computed by the classical estimate [44], X |m|,|n|≤R gcd(m,n)=1
1=8
X d≤R
ϕ(d) =
24 2 R + E(R), E(R) = O(R log R). π2
The error √ term cannot be improved much further as Montgomery has shown that E(R) = ± (R log log R). However, as noted above, the situation is much different if one uses
Deterministic Percolation
49
Fig. 4. Relatively prime pairs near (10100 , 10100 )
circular summation: One is then estimating X 1 = π R 2 + 1(R), S(R) = m2 +n2 ≤R 2 gcd (m,n)=1
and it is easily shown that 1(R) = O(R) using a nontrivial bound on the error term in the Gauss circle problem (the prime number theorem also implies that the error is o(R)). The error term 1(R) has been studied by Moroz [41] and by Huxley and Nowak [35]. The latter showed, using results of Baker [3] that, assuming the Riemann Hypothesis, 1(R) = O(R 3/4+ ). It seems that the actual error is O(R 1/2+ε ) and in analogy with the Gauss circle problem [33,5], one can ask whether 1(R)/R 1/2 has a limiting (logarithmic) distribution (this would require very strong assumptions as in [42]) Continuing with the analysis of R, it is also true that any finite configuration in R has a density. In other words, let be a finite subset of Z2 which contains the origin and define δ() to be the asymptotic density of in the sense that (z) is true if and only if z + ⊂ R.
50
I. Vardi
Theorem 3.1. Let be a finite subset of Z2 , then δ() exists. This result is more subtle than expected and a direct approach to compute δ() fails (I would like to thank G. Tenenbaum for explaining this point to me). For example, let = {(0, 0), (1, 1), (2, 2)}, then one is counting X 1. S() = k(m,n)k≤R gcd (m+j,n+j )=1, j =0,1,2
The direct approach uses the identity X
µ(d) =
d|n
(
1, if n = 1, 0 if n > 1,
(1)
where µ(d) is the Möbius function and this gives rise to terms of the form 2 X R µ(d1 )µ(d2 )µ(d3 ) + O(1) . d1 d2 d3 d1 ,d2 ,d3 ≤R
The term corresponding to O(1)3 will be of order R 3 and no known estimates will be able to reduce this to o(R 2 ). However, one can still hope that such a bound exists and this leads to Conjecture 1. If is a finite set containing the origin, then Y pn − | mod p| . δ() = pn p prime
This is an analogue of the Hardy–Littlewood conjectures for prime k-tuplets [30]. The conjecture states that the global density should be a product of the local densities for each prime. A similar result was proved by Hafner, Sarnak, and McCurley [27] in which the density of relatively prime values of polynomials is computed. Problems 2. Existence of an infinite component. Existence is trivial since the line {(m, 1) : m = 1, 2, 3, . . . } lies in R. Comparing R to the random model, one sees that an unbounded component should exist since 6/π 2 > .5927, the conjectured approximation to pc (Z2 ) [51] (the best rigorous bounds are are .556 < pc (2) < .679492 of [4] and [49], respectively). Problem 3. Number of infinite components.. This has an elementary answer consistent with the probabilistic model: Proposition 3.1. R has a unique infinite component. Proof. Call C1 the infinite component containing the line {(m, 1) : 1 ≤ m < ∞}. Then for each prime p, the vertical line {(p, n) : 1 ≤ n ≤ p − 1} is in C1 . Thus, a large connected component in the region {(m, n) : m > n} will eventually have to cross one of these line, and therefore belong to C1 . By symmetry, the same holds true for the other 7 lines {±1, ±k}, {±k, ±1}. Finally, these lines are all joined together, either at (±1, ±1) or by passing through (±1, 0) and (0, ±1), where, by definition, gcd(1, 0) = 1. u t
Deterministic Percolation
51
Problem 4. The asymptotic density of the infinite component. This is the main question of this paper and I prove Theorem 3.2. The infinite component of R has an asymptotic density. Let θ = θ (R) be the asymptotic density of the infinite component, i.e., θ = δ(C∞ ), where C∞ (z) is true if and only if z lies on the infinite component, which, by abuse of notation, will also be denoted by C∞ . The value of θ can examined experimentally and preliminary computations seem to indicate that θ/p(R2 ) ≈ .96 ± .01, i.e., about 96% of open sites lie on the unbounded component. In fact, I will prove Theorem 3.3. The asymptotic density of the infinite component of R is not zero. Both these theorems are quite subtle and follow from Theorem 3.4. Let f (R) be any function increasing to infinity then, except for a set of zero asymptotic density, every (m, n) ∈ B(R) is surrounded by a rectangle of perimeter < f (R) all of whose edges are contained in C∞ . As noted above, this result is consistent with de Gennes’ philosophy that the infinite component consists of a mesh with small holes [14,23]. Theorem 3.4 suggests that a stronger result is true, namely if one defines rect (m, n) to be the perimeter of the smallest rectangle containing (m, n) all of whose edges are in C∞ , then rect (m, n) should have a limiting distribution. An examination of the proof of Theorem 3.2 (Sect. 8) reveals that it works with f (R) in Theorem 3.4 replaced by (log R)1/2−ε , so Theorem 3.2 follows from the weaker result of Lemma 7.3 below. It should be noted that one can similarly show that the infinite component of Rn has an asymptotic density for n ≥ 3, and it is trivially nonzero since it is ≥ 6/π 2 in these cases. One can easily give an upper bound θ < 6/π 2 . Thus, consider γ (z) to be true if z ≡ (4, 15) (mod 30), then z is isolated if γ (z) and z ∈ R are both true. Since X X X = µ(d) 1 + O(R), k(m,n)k≤R γ (z), (m,n)=1
one gets
gcd (d,30)=1 d≤R
1 6 Y 1 1 Y , 1− 2 = 2 δ(γ ) = 2 2 30 p π p −1 p>5
so
m≡4/d (mod 30) n≡15/d (mod 30) k(m,n)k≤R/d
p≤5
1 6 6 6 = .99305¯ 2 . θ ≤ 2 − 4δ(γ ) = 1 − 2 π 144 π π
More generally, define an animal α to be a connected set containing the origin and α(z) is true if z + ω ∈ R for every ω ∈ α, but z + ω 6∈ R if ω ∈ ∂α (a site is in ∂α if it is connected to α but not in α). It should be noted that this definition of animal is not translation invariant, see [8] for general results on animals. The inclusion-exclusion principle implies that X (−1)n δ(α ∪ s), δ(α) = s⊂∂α
52
I. Vardi
where each δ(α ∪ s) on the right represents an ordinary density of a configuration. It follows that the density of animals exists and moreover, Lemma 8.4 below shows that θ=
X 6 − δ(α). 2 π α
One can therefore try to estimate θ by computing δ(α) in small cases. As noted above, this is does not seem possible using current methods. However, if one accepts Conjecture 1, then one has α1 = ζ1 (2) − 4ζ2 (2) + 8ζ3 (2) − ζ4× (2) − ζ5 (2) = .0110235396 . . . , , 62 16 α2 = 4ζ2 (2) − 32ζ3 (2) + ζ4× (2) + 36ζ5 (2) − ζ6 (2) = .0013019993 . . . , 5 3 where α1 and α2 are the densities of elements in R belonging to animals of size 1 and 2 respectively, and Y Y m m × 1 − s , ζm (s) = 1− s . ζm (s) = p p p prime
p prime ps 6 =m
Conjecture 1 therefore implies θ < 24ζ3 (2) −
57 × 16 6 ζ4 (2) − 35ζ5 (2) + ζ6 (2) ≈ .9797253010901 2 . 5 3 π
Note that the functions ζm (s) do not have an analytic continuation if m > 1 since a result of [16] states that for a polynomial f (x), the Euler product ζ (s, f (x)) = Q Estermann −s p f (p ) has an analytic continuation to the complex plane only if f (x) is a product of cyclotomic polynomials. However, one can easily find these numerically using the formula P " # 1` d|` ad µ(`/d) ∞ Y Y Y 1 f (1/ps ) , ζ (`s) 1 − `s ζ (s, f (x)) = p p≤m p≤m `=1
where log f (x) =
P∞
k=1 ak x
k,
see [18,46].
Problem 5. Functions of percolation. The evaluation of θ indicates that other functions might also have deterministic analogues. One therefore defines Number of connected components in B(R) , |B(R)| X 1 |C(z)|, χ f (R) = lim R→∞ |B(R)| κ(R) = lim
R→∞
kzk≤R z6∈C∞
if these limits exist, where C(z) represents the connected component containing z. Since the sum for κ(R) has better convergence properties than θ (R), the methods used to prove Theorem 3.2 also show that this exists. However, the existence of χ f (R) is left as an open problem.
Deterministic Percolation
53
A problem related to percolation is computing the asymptotics of the largest Rfree square. This question is a generalization the Cramer conjecture [13] which states that the largest prime gap ≤ R should tend to (log R)2 (though this is now believed to inaccurately model the primes [25]). One can compare results for the random and deterministic models. Proposition 3.2. Let Fp (R) be the area of the largest closed square inside B(R), where sites are open with probability p, then, with probability 1, Fp (R) −2 = . R→∞ log R log(1 − p) lim
Proposition 3.3. Let F (R) be the area of the largest R-free square in B(R), then F (R) 1 log R . (log log R)2 log R log log R 4. Finite Configurations have a Density This section is devoted to the proof of Theorem 3.1. I will prove this by showing that the asymptotic density of a configuration is the limit of its density modulo Q a product of initial sets of primes. The idea is that Z2 is approximated by Z2 mod p≤X p and, using an idea from ordinary percolation [26], one can think of some events as being monotonic. Definition. Two integers m, n are relatively prime modulo h if gcd (m, n, h) = 1. Furthermore, let |{(m, n) ∈ B(R) : h (m, n)}| , δh (, R) = B(R) where h (m, n) means that (m, n) holds as well as (m, n, h) = 1. Lemma 4.1. δh () = limR→∞ δh (, R) exists (and will be called the density of modulo h). Proof. This follows from the fact that this relative primality is periodic modulo h and that B(h) tiles the plane. u t Q The next step is to consider a product of an initial set of primes P (X) = p≤X p and let h = P (X). Thus the pairs relatively prime modulo P (X) will be an approximation to R. The main result thus follows from Lemma 4.2. If Y > X and X → ∞, then δP (Y ) (, P (Y )) = δP (X) (, P (X)) + O(||/X). Lemma 4.3. If both X → ∞ and R/P (X) → ∞, then δ(, R) = δP (X) (, P (X)) + O(||/X + RP (X)/|B(R)|).
54
I. Vardi
P If these results are true, then one can take an increasing sequence of Xj ’s such that 1/Xj < ∞, so that Lemma 4.2 shows that δP (Xj ) (, P (Xj )) approaches a limit. Lemma 4.3 then implies that this limit equals the limit of δ(, R) as R → ∞. This proves Theorem 3.1. Proof of Lemma 4.2. Let Y > X, then P (X)|P (Y ) and B(P (X)) exactly tiles B(P (Y )). Now consider z ∈ B(P (Y )) for which P (Y ) (z) is true. Then P (X) (z mod P (X)) is true except if the distance of z mod B(P (X)) from the boundary of B(P (X)) is less than ||. Conversely, if P (Y ) (z) is false, then P (X) (z mod B(P (X))) could be true if d|z+α, where d is divisible only by primes > X. It follows that X z∈B(P (Y )) P (Y ) (z)
1=
|B(P (Y ))| |B(P (X))|
X z∈B(P (X)) P (X) (z)
p 1 + O || |B(P (X))|
X 1 + O |B(P (Y ))| || d2
! .
d>X
Dividing this by |B(P (Y ))| gives the result. u t Proof of Lemma 4.3. Let R > P (X), where R/P (X) is large. Let z ∈ B(R), then if (z) is true, then P (X) (z mod P (X)) is true unless z mod P (X) is close to the boundary of B(P (X)). Conversely, if (z) is false, then P (X) (z mod B(P (X))) could be true if d|z + α, where d is divisible only by primes > X. One gets a similar estimate as in the above, except that the tiling of B(R) by B(P (X)) is no longer exact and there is an extra factor of size O(R |B(P (X)))|. Thus X X p |B(R)| 1= 1O + || |B(P (X))| |B(P (X))| z∈B(R) (z)
z∈B(P (X)) P (X) (z)
X 1 + O |B(R)| || d2
! + O(R P (X)).
d>X
The result follows upon dividing by |B(R)|. u t
5. Sieve Methods The sieve methods used here are due to Friedlander [20] who used them to show that for any fixed E > 5, almost every interval [y, y + logE y] contains a number with at most 4 prime factors. The term “almost every” means that the measure of y ≤ X for which this fails to hold is o(X). Such results were first obtained by A. Selberg [43] who showed that, assuming the Riemann Hypothesis, for any function f (n) → ∞, almost every interval [y, y +f (y) log2 y] contains a prime. Heath-Brown [32] later showed that the further assumption of Montgomery’s pair correlation conjectured implied the similar
Deterministic Percolation
55
result with log y replacing log2 y, and this is in some sense optimal. Unconditional results were proved by various authors and the result used here is due to N. Watt [48] and shows that almost every interval of length y 1/14+ε contains a prime. An unconditional result of Heath-Brown and Iwaniec [34] states that every interval of length y 11/20 contains a prime (this has subsequently been improved, see [48]). In [21], Friedlander improved his methods and showed that almost every interval [y, y + f (y) log y] contains a number with at most 21 prime factors. This is a much more accurate result, as the optimal interval length is f (y) log y/(log log y)21 . There are other techniques to prove such results and Harman [31] has shown that almost every interval [y, y + log7 y] contained a number with exactly two prime factors. Unlike Friedlander’s papers, this does not seem to yield techniques that are directly applicable to the problems of this paper. I will closely follow Friedlander’s paper which starts with a modified form of Rosser’s sieve [36] (see also [28, Chapter 8]). As usual [28] consider an interval A and let S(A, Z) be the set of an element of A which are not divisible by any prime p < Z. In this case A will be the interval [y, y + U ]. Theorem 5.1 (Friedlander). Let 2 ≤ Z ≤ D. There exists a function f (u) which is positive for u > 2, and a sequence λd of real numbers, this sequence depending on D and Z, and satisfying |λd | ≤ 1, such that log D U −1/3 f + O(log D) − |Ry (D)|, (2) S(A, Z) > log Z log Z where Ry (D) =
X d
λd ry (d), ry (d) = ψ
y d
−ψ
y+U d
1 , ψ(t) = t − btc − . 2
The following is the best “everywhere” result available using the Rosser sieve. Corollary 5.1. For any ε > 0, there exists a constant E such that for all sufficiently large k and X, every interval [y, y + U ] contains at least E U/ log U numbers all of whose prime divisors are greater than U 1/2−ε , This result follows by letting U = D 1+δ = Z 2+3δ in Theorem 5.1, where δ > 0 and the trivial bound |Ry (D)| ≤ 2D is used. Theorem 5.1 formulates the error term Ry (D) in analytic form which allows one to average the error as y varies over an interval [X/2, X]. Theorem 5.2 (Friedlander). Let 4 ≤ D 2 ≤ V /2 ≤ X/(2 log D), where V = y/U , then Z 1 X |Ry (D)|dy U 1/2 (log D)3/2 + (log D)2 . (3) X X/2 Remark. One can improve this result slightly by using Friedlander’s second paper [21]. This would result in replacing the U 1/2 (log D)3/2 term in (3) with U (log Z)−1/2 , under the conditions D 2 (log D)7 ≤ V ≤ X and Z 11 ≤ D. The exponent in Proposition 5.1 below would be improved from 1/5 to 1/3 (the true exponent seems to be 1). The proof of Theorem 5.2 follows exactly as in the paper of Friedlander [20] except that the log D term is everywhere substituted for log X, written as L in [20]. An examination of Friedlander’s proof reveals that this substitution is valid at each step of his argument. Friedlander used this result to prove
56
I. Vardi
Theorem 5.3 (Friedlander). For any A > 5, there is a constant c such that almost every interval [y, y + (log y)A ] contains at least c (log y)A−1 numbers x such that P− (x) > x 1/4−ε , where P− (x) is the smallest prime factor of x (so x has at most 4 prime factors). This result will be generalized to arbitrary interval lengths of small size. Proposition 5.1. Let g(R) be an unbounded increasing function such that g(R) < (log R)5 . Given any fixed ε > 0, then for all y < R, except for a set of measure O(R [g(R)]−ε/2 ), the interval [y, y + g(R)], contains an x with P− (x) > exp([g(R)]1/5−ε ). Proof. One takes a large T and has to show that almost all y ≤ T have an x ∈ [y, y + g(R)] with P− (x) > exp([g(R)]1/5−ε ). Using the above notation, one writes U = g(R) 1/5−ε . Also, in and in order to get the result of Proposition 5.1, one must choose Z = eU 2+ε order to have f ((log D)/(log Z)) > 0, one takes D = Z . Since U → ∞, assume that y > T /U and divide the interval [T /U, T ] into disjoint subintervals of the form (X, 2X]. Taking V = X/U and applying Theorem 5.1 gives a lower bound o n S(A, Z) > U 4/5+ε f (2 + ε0 ) + O((log D)−1/3 ) − |Ry (D)|, so the main term has order U 4/5+ε . Now consider G ⊂ (X, 2X] given as the set of y’s for which |Ry (D)| > U 4/5−ε . It follows that Z 2X |Ry (D)|dy > |G| U 4/5−ε . X
Since U < (log X)5 , one has Z < e(log X)
1−ε
D 2 ≤ e3(log X)
1−ε
≤
and so
V X X = , ≤ 5 2U 2 2(log X)
for all sufficiently large X. Also, since D = exp((2 + ε)U 1/5−ε ), one has U = 0 C1 (log X)5+ε D for some constants C1 , ε0 , so X X X V = ≤ C1 , ≤ 5 2 2U 2 log D (log D) for all sufficiently large X. The conditions of Theorem 5.2 are therefore satisfied, and applying it gives Z 2X |Ry (D)|dy X(U 1/2 (log D)3/2 + (log D)2 ) X U 4/5−3/2ε . X
It follows that
|G| X U −ε/2 = o(R),
i.e., the proportion of exceptional y’s goes to zero. Thus almost all y ∈ (X/2, X] have |Ry (D)| < U 4/5−ε and this is dominated by the main term U 4/5+ε of (3), so S(A, Z) is nonzero for almost all such y. Taking the union of the intervals gives the result. u t
Deterministic Percolation
57
6. R-Free Squares Before embarking on the rather daunting task of proving Theorem 3.4, I will do a “warmup” consisting of proving Propositions 3.2 and 3.3. Proof of Proposition 3.2. (a) lim sup Fp (R)/ log R ≤ −2/ log(1 − p): Let c > −2/ log(1 − p), and at each site (m, n) in B(R), consider a square of area c log k(m, n)k and with lower left hand corner at (m, n). The probability that all sites in the square are closed is (1 − p)c log k(m,n)k < k(m, n)k−2−ε , where ε > 0. Summing over all sites in Z2 gives X 1 , k(m, n)k2+ε 2 (m,n)∈Z k(m,n)k>2
which converges. The Borel-Cantelli Lemma [17] shows that the probability of infinitely many such squares being closed is zero. (b) lim inf Fp (R)/ log R ≥ −2/ log(1 − p): Let c < −2/ log(1 − p) and for each site (m, n) with k(m, n)k > 2, consider a square of area c log k(m, n)k with bottom left hand corner at b(m, n) (log k(m, n)k)2 c. As in part (a), it follows that each square has a probability k(m, n)k−2−ε of being closed. Moreover, all these events are independent so the result follows from the Borel Cantelli Lemma and the divergence of X (m,n)∈Z2
(log k(m, n)k)2 . k(m, n)k2+ε
t u
k(m,n)k>2
Proof of Proposition 3.3. (a) Lower bound: Consider an integer k, and the first k 2 primes q1 , . . . , qk 2 . By the Chinese Remainder Theorem, the following congruences have soQ2 lutions m, n ≤ ki=1 qi , k k Y Y qik+j , n ≡ −j mod qj k+i , m ≡ −i mod j =1
j =1 0
so gcd (m+i, n+j ) > 0, for 1 ≤ i, j ≤ k. The prime number theorem gives k 2(1+ε )k > k(m, n)k > k 2(1−ε)k , which yields the estimate k 2 > (1 − ε00 ) 21 log R/ log log R (one can improve the 1/2 term to π 2 /12). (b) Upper bound:Assume that there is an R-free square {(m+i, n+j ) : 1 ≤ i, j ≤ k}. By Corollary 5.1, there are x1 , . . . , x` ∈ [m+1, m+k] for which P− (xi ) > k 1/2−ε , and such Q e that ` k/ log k. Write xi = rji=1 qijij . One notes that if q is a prime, then the number of integers in a set S which are relatively prime to q is ≥ min{|S|(1 − 1/q), |S| − 1}. The asymptotic relation X 1 ∼ − log(1/2 − ε) < 1 q 1/2−ε k
2
thus implies that the number of integers in [m + 1, m + k] relatively prime to xi is (1 + log(1/2 − ε)) k − |{j ≤ ri : qij ≥ k}|, and so each ri > (1 + log(1/2 − 2ε)) k.
58
I. Vardi
One further notes that gcd (xi , xj ) ≤ k for i 6 = j , so that xi and xj can share at most 2 distinct prime factors. This implies that the number of distinct qij ’s is k 2 / log k. The prime number theorem gives ` Y i=1
xi
Y
q = eD k
2 −k 1/2−ε +o(1)
,
y
for some constant D, and so log xi k log k for all i. This therefore gives log R k log k and the result follows since log k is of order log log R. u t
7. The Structure of the Infinite Component In this section, I prove Theorem 3.4. One begins identifying some subsets of the infinite component. Lemma 7.1. The following sets are in C∞ : (1) {(m, 1) : m > 0}.
(2) {(p, n) : p > n}, p prime.
(3) {(m, q) : q < m < (q/2)20/11 , q 6 |m}, qprime. Proof. The first two results were proved above. The third follows from the unconditional result of Heath-Brown and Iwaniec [34] which states that every interval of length y 11/20 contains a prime (this has subsequently been improved, see [48]). Thus, if q 6 |m, then since q > 2 m11/20 , one of the intervals [m + m11/20 ] or [m − m11/20 , m] has none of its elements divisible by q and the corresponding line segment lies completely in R. By the result of Heath-Brown and Iwaniec, this line segment will cross a line {(p, n)}, where t p is prime, and by (2), this is in C∞ , so (m, q) is also in C∞ . u One proceeds by proving an initial version of Theorem 3.4. Lemma 7.2. All but O(R 2 / log R) pairs of B(R) are surrounded by a rectangle of perimeter O((log R)7 ) all of whose edges are contained in C∞ . Proof. Consider the set S0 = {(m, n) ∈ B(R) : ∃ n0 , P− (n0 ) > R 1/5 , (log R)7 + n > n0 > n, minimal, (d, n0 ) = 1 if |d − m| < R 1/13 , ∃ p ∈ [m − R 1/13 , m + R 1/13 ]}. By Watt’s result that almost every interval of length y 1/14+ε contains a prime, E0,1 = {(m, n) ∈ B(R) : [m − R 1/13 , m + R 1/13 ] has no primes} has density zero. In fact, his result shows that |E0,1 | R 2 /(log R)2 , since substituting E = 3 in Theorem 1 of [48] already gives this estimate for intervals of length R 1/14 (log R)22 . Next, if E0,2 = {(m, n) ∈ B(R) : P− (n0 ) ≤ R 1/5 , for all (log R)7 + n > n0 > n},
Deterministic Percolation
59
then |E1,2 | R 2 / log R. This bound follows by an argument similar to the proof of Proposition 5.1: Partition [R 1/2 , R] into intervals of the form [X, 2X] and in each of these intervals, let U = (log X)7 , D = V 1/2 /2, and for a fixed 1/10 > ε > 0, let Z = D 1/2−ε . The main term in Eq. (2) of Theorem 5.1 is then (log X)6 while Theorem 5.2 gives Z 1 2X |Ry (D)|dy (log X)5 . X X Therefore G defined as the set of y ∈ (X, 2X] for which |Ry (D)| (log X)6 satisfies |G| X/ log X. Finally, one considers E0,3 = {(m, n) ∈ B(R) : P− (n0 ) > R 1/5 , (log R)7 + n > n0 > n, minimal, ∃d |d − m| < R 1/13 , (d, n0 ) > 1}. But if P− (n0 ) > R 1/5 , then |E0,3 | R(log R)7
X
1 R 1+1/13 (log R)7 R
P− (n0 )>R 1/5 |d−m|
1
R 2+1/13−1/5 (log R)7
XR p|n0
p
R2 . log R
Since B(R) − S0 ⊂ E0,1 ∪ E0,2 ∪ E0,3 , it follows that |B(R) − S0 | R 2 / log R, and so S0 represents almost all points of B(R). An examination of the definition of S0 shows that for every (m, n) ∈ S0 there is an unbroken horizontal line segment L0,1 (m, n) ⊂ C∞ of length R 1/13 , centered at (m, n) and which passes above (m, n) within distance O((log R)7 ), and which crosses a vertical line {(p, n)}, where p is prime. By Lemma 7.1 (2), this vertical line is in C∞ , so it follows that L0,1 (m, n) is also in C∞ . Replacing the condition n0 > n with n0 < n yields the similar result with a line segment L0,2 (m, n), passing below (m, n). To construct the corresponding vertical line segments L0,3 (m, n), L0,4 (m, n), one must include the condition that m, n R 11/20 in order to apply part 3 of Lemma 7.1. The result then follows. u t The next iteration already contains most of the ideas of the general procedure and is included in order to give a self-contained proof of Theorem 3.3. Lemma 7.3. All but O(R 2 /(log log R)3 ) pairs of B(R) are surrounded by a rectangle of perimeter O((log log R)36 ) all of whose edges are contained in C∞ . Proof. Consider the set S1 = {(m, n) ∈ B(R) : ∃m0 P− (m0 ) > (log R)42 , (log log R)36 + m > m0 > m, minimal, (m0 , d) = 1, |d − n| < (log R)40 .}. Now let E1,1 = {(m, n) ∈ B(R) : P− (m0 ) < (log R)42 , for all (log log R)36 + m > m0 > m}.
60
I. Vardi
By Proposition 5.1, for almost all y, [y, y + (log log R)36 ] contains an integer x with P− (x) > exp((log log R)36(1/5−ε) ), and this is > (log R)42 for all sufficiently large R. The error term of Proposition 5.1 gives |E1,1 |
R2 R2 = , (log log R)36ε/2 (log log R)3
by letting ε = 1/6. Next, one considers E1,2 = {(m, n) ∈ B(R) : P− (m0 ) > (log R)42 , (log log R)36 + m > m0 > m, minimal, ∃d |d − n| < (log R)40 , (m0 , d) > 1}. As before, one gets X
|E1,2 | R(log log R)36
1
P− (m0 )>(log R)42 |d−n|<(log R)40 , (m0 ,d)>1
R (log log R)36 (log R)40
XR R 2 (log log R)37 p (log R)2 0
p|m
R2 log R
,
for all sufficiently large R. This last estimate used the fact that if P− (m0 ) > (log R)42 and m0 < R, then m0 can have at most O(log R/ log log R) prime factors. It follows that S1 consists of almost all elements of B(R). The definition of S1 then shows that for almost all (m, n) ∈ B(R), there is a line segment L1,1 (m, n) ⊂ R of length at least (log R)40 , centered at (m, n) and which passes by (m, n) at distance O((log log R)36 ). One now observes that S0 ∩ S1 also comprises almost all elements of B(R). The definition of S0 implies that for almost all (m, n), the line segment L1,1 (m, n) intersects a rectangle of perimeter O((log R)7 ) whose edges lie in C∞ (note that these extend in C∞ to length R 1/13 ), so L1,1 (m, n) will also be contained in C∞ . Clearly, one can similarly show that for almost all (m, n), there are line segments L1,j ⊂ R, j = 2, 3, 4, of length (log R)40 and which pass within (log log R)36 of (m, n) on all four sides, and t by the same argument L1,j ⊂ C∞ , j = 2, 3, 4. u An examination of the P proof of Lemma 7.3 reveals that the obstacle in continuing this process is the term p|m0 1/p, which must be o(1). The trivial estimate X1 ω(m) p P− (m)
(4)
p|m
was used and this will only allow one more iteration, if one first removes all m for which ω(n) > 2 log log R (as is well known, this set has zero density [44]). In fact, in this context, it is easy to improve substantially on (4), as was noted to me by G. Tenenbaum.
Deterministic Percolation
61
Lemma 7.4. p Let f (R) be any function increasing to infinity, then except for a set of m of size O(R/ f (R) log f (R)), one has the bound X p|m p≥f (R)
1 1 p . p f (R) log f (R)
Proof. One uses the simple estimate X
X
m≤R
p|m p≥f (R)
X R R 1 = . = 2 p p f (R) log f (R) p≥f (R)
P This result reflects the fact that p|m 1/p has a limiting density, as follows easily from the Erd˝os–Wintner Theorem [44]. u t This result allows one to iterate the above argument. Recall that rect (m, n) is the perimeter of the smallest rectangle surrounding (m, n) which has all its edges in C∞ . Lemma 7.5. Let log1 z = log z and logk+1 z = log(logk z), and let expk z be the inverse of logk z, then for R > expk+2 (1015 ), 1 |(m, n) ∈ B(R) : rect (m, n) ≤ (logk+1 R)36 }| = |B(R)| 1 − O , (logk+1 R)3 (5) where the O( ) term is independent of k. Proof. One proves this by induction on k. The initial step k = 1 is exactly Lemma 7.3. Now assume that (5) holds for k, then I will show that it holds for k + 1. Thus assume that one has constructed for almost all (m, n), line segments Lk,j (m, n), j = 1, 2, 3, 4, as above, but of length (logk R)40 , which pass within (logk+1 R)36 , are centered at (m, n) and lie completely in C∞ . In particular, one assumes that the set Sk = {(m, n) ∈ B(R) : ∃n0 P− (n0 ) > (logk R)100 ,
(logk+1 R)36 + n > n0 > n, minimal,
(d, n0 ) = 1, |d − m| < (logk R)40 , (m, n0 ) ∈ C∞ },
is such that |B(R) − Sk | R 2 /(logk+1 R)3 . One then considers Sk+1 = {(m, n) ∈ B(R) : ∃m0 P− (m0 ) > (logk+1 R)100 ,
(logk+2 R)36 + m > m0 > m, minimal, (m0 , d) = 1, |d − n| < (logk+1 R)40 }.
Now let Ek+1,1 = {(m, n) ∈ B(R) : P− (m0 ) < (logk+1 R)100 , for all (logk+2 R)36 + m > m0 > m}.
62
I. Vardi
By Proposition 5.1, for any ε > 0, except for O(R/(logk+2 R)36ε/2 ) values of y, the interval [y, y + (logk+2 R)36 ] contains an integer x with P− (x) > exp((logk+2 R)36(1/5−ε) ). Letting ε = 1/6, this says that except for O(R/(logk+2 R)3 ) values of y, the interval [y, y + (logk+2 R)36 ] contains an integer x with P− (x) > exp((logk+2 R)1+2/15 ). This last quantity is eventually > (logk+1 R)100 , in particular, when logk+2 R > 10015/2 , i.e., when R > expk+2 (1015 ). Thus, one concludes that |Ek+1,1 |
R2 (logk+2 R)3
, for R > expk+2 (1015 ).
Next, one considers Ek+1,2 = {(m, n) ∈ B(R) : P− (m0 ) > (logk+1 R)100 ,
(logk+2 R)36 + m > m0 > m, minimal,
∃d, |d − n| < (logk+1 R)40 , (m0 , d) > 1}.
As before, one gets X
|Ek+1,2 | R(logk+2 R)36
1
P− (m0 )>(logk+1 R)100 |d−n|<(logk+1 R)40 , (m0 ,d)>1
R (logk+2 R)36 (logk+1 R)40
X
m≤R
P− (m)>(logk+1 R)100
X1 . p p|m
Now let X
Ek+1,3 = {(m, n) ∈ B(R) :
p|m0 p>(logk+1 R)100
1 1 p ≥ p (logk+1 R)50 logk+2 R
for all (logk+2 R)36 + m > m0 > m}. Applying Lemma 7.4 gives |Ek+1,3 | so |Ek+1,2 − Ek+1,3 |
R 2 (logk+2 R)36 p , (logk+1 R)50 logk+2 R
R 2 (logk+2 R)36 (logk+1 R)40 R2 p , logk+1 R (logk+1 R)50 logk+1 R
whenever (logk+1 R)9 > (logk+2 R)36 , for example, when R > expk+2 (16). It follows that Sk+1 consists of almost all elements of B(R), where the exceptional set is R 2 /(logk+2 R)3 . The definition of Sk+1 also shows that for almost all (m, n) ∈ B(R), there is a line segment Lk+1,1 (m, n) ⊂ R of length at least (logk+1 R)40 , centered at (m, n) and which passes by (m, n) at distance O((logk+2 R)36 ).
Deterministic Percolation
63
One now observes that Sk ∩ Sk+1 also comprises almost all elements of B(R) except for an exceptional set of size
R2 (logk+1 R)3
R2
+
(logk+2 R)3
= (1 + Ak )
R2 (logk+1 R)3
,
for sufficiently large R, where k Y
(1 + Aj )
j =1
k Y
" 1+O
j =1
1
!#
(logj R)3
= O(1).
The definition of Sk+1 implies that for almost all (m, n), the line segment Lk+1,1 (m, n) of length (logk+1 R)40 will intersect the perimeter of the rectangle of perimeter (logk+1 R)36 constructed in the previous iteration (note that its sides extend to line segments in C∞ of length (logk R)40 ). Since the sides of this rectangle lie in C∞ , it follows that Lk+1,1 (m, n) will also be contained in C∞ . Thus, the line segments defined in Sk+1 also lie in C∞ . Clearly, one can similarly show that for almost all (m, n), there are line segments Lk+1,j ⊂ R, j = 2, 3, 4, of length (logk+1 R)40 and which pass within (logk+2 R)36 of t (m, n) on all four sides, and Lk+1,j ⊂ C∞ , j = 2, 3, 4. u Proof of Theorem 3.4. Assume that there is a function f (R) increasing to infinity, and a fixed λ > 0 such that for at least λ Ri2 pairs (m, n) one has rect (m, n) > f (Ri ), where Ri is a sequence increasing to infinity. One defines log∗ z to be the minimum number of iterations of log required to be ≤ 2, i.e., loglog∗ z z ≤ 2. Let k = log∗ R − log∗ f (R) − 2, then logk R log f (R) for all sufficiently large R. One can now apply Lemma 7.5 since the bound expk+2 (1015 ) = explog∗ R−logf (R)+4 (1015 ) log R = o(R) holds. Furthermore, one has 1 (logk+1
R)3
1 → 0, log4 f (R)
since logk R log3 f (R) and f (R) → ∞. Substituting this in (5) shows that for almost all (m, n) ∈ B(R) one has rect (m, n) ≤ f (R), which contradicts the above assumption. The result follows. u t Proof of Theorem 3.3. Assuming Theorem 3.2, then this follows directly from Theorem 3.4. For if the result were not true, then there would be a function f (R) increasing to infinity such that θ(R) < 1/f (R) for all sufficiently large R. However, Theorem 3.4 implies that for sufficiently large R, almost all points of B(R) are surrounded by a rectangle √ √ of perimeter f (R) all of whose edges lie in C∞ . This implies that θ (R) 1/ f (R) which is a contradiction. u t
64
I. Vardi
8. The Infinite Ccomponent has a Density As in Sect. 4, the idea is to compare the density of the infinite component in a big square θ(R) =
|B(R) ∩ C∞ | |B(R)|
with this density modulo a product of primes. Thus, let |{(m, n) ∈ B(R) : (m, n) ∈ infinite component modulo h}| . R→∞ |B(R)|
θh = lim
The following gives a local characterization of the infinite component modulo h. Lemma 8.1. z is in the infinite component modulo h if and only if z mod h is connected modulo h to all sides of B(h). Proof. In fact, this holds for the fundamental domain B ∗ (h) = {0 ≤ m, n < h} of Z2 mod h (so B(h) consists of 4 copies of B ∗ (h)). To see this, note that B ∗ (h) has reflection symmetries generated by (m, n) 7 → (n, m) and (m, n) 7 → (m, −n) mod h which preserve relatively prime pairs mod h. Thus B ∗ (h) consists of 8 triangles each of which is a reflection of its adjacent neighbor, see Fig. 5. It is clear that z is on the infinite component if and only if it is h-connected to all three sides of the triangle on which it t lies and this is clearly equivalent to being h-connected to all four sides of B ∗ (h). u Lemma 8.2. If X < Y then θP (X) ≥ θP (Y ) . Proof. If z ∈ B(P (Y )) is in the infinite component modulo P (Y ) then z mod P (Y ) is P (Y ) connected to three boundaries of B(P (Y )). Since reducing modulo P (X) does not remove any connections, it follows that z mod P (X) is P (X) connected to the boundary of B(P (X)). u t
Fig. 5. Reflection symmetries modulo h
Deterministic Percolation
65
One concludes that θP (X) is decreasing and therefore the limit θ∗ = limX→∞ θP (X) exists. Lemma 8.3. If both X → ∞ and R/P (X) → ∞, then θ (R) ≤ θP (X) + o(1). Proof. This follows exactly as in the above also following the proof of Lemma 4.3. Note that the boundary error in tiling B(R) with B(P (X)) is O(P (X)/R) = o(1) by assumption. u t One therefore has θ(R) ≤ θ∗ + o(1) and the main result follows from Lemma 8.4. limR→∞ θ(R) = θ∗ . Proof. If θ∗ = 0, then Lemma 8.3 shows that θ = 0 as well and there is nothing to prove. On the other hand, if θ∗ > 0 consider a large R and by Lemma 8.2, find a large X such that θP (X) is very close to θ∗ , with P (X) < R and R/P (X) large. Since the prime number theorem says that P (X) = eX+o(1) , one can choose X such that R 1/2 /4 < P (X) < R 1/2 . It follows that X is of order log R, i.e., there are two constants A, B, such that A log R < X < B log R. P Now going from R modulo P (X) to R in B(R) removes at most |B(R)| d>X 1/d 2 |B(R)|/X elements. So, by the above estimate, one is removing O(R 2 / log R) sites. By Lemma 7.3, given 0 < ε < 1/2, one can surround almost all sites with a C∞ rectangle of perimeter < (log R)1/2−ε . Thus apart from a set of zero density, each individual removal can disconnect at most (log R)1−2ε sites from the connected component (note that none of the elements on the rectangles specified by Lemma 7.3 are removed, since these belong to C∞ of R). It follows that at most O(R 2 (log R)1−2ε / log R) sites are disconnected, a vanishingly small percentage of the infinite connected component modulo t P (X). This implies that θ(R) is asymptotically close to θP (X) . u Acknowledgement. I would like to thank Cécile Dartyge for checking the method of Sect. 7 and Gérald Tenenbaum for helpful comments and a crucial observation (Lemma 7.4).
References 1. Aizenman, M., Kesten, H., Newman, C.M.: Uniqueness of the infinite cluster and continuity of connectivity functions for short and long range percolation. Commun. Math. Phys. 111, 505–531 (1987) 2. Ash, J.M., Wang, G.: A survey of uniqueness questions in multiple trigonometric series. In: Harper, L.H. (editor), Harmonic Analysis and nonlinear differential equations American Math. Soc. Contemp. Math. 208, 1997, pp. 35–71 3. Baker, R.C.: The square–free divisor problem. Quart. J. Math. Oxford 45, 269–277 (1994) 4. van den Berg, J., Ermakov, A.: A new lower bound for the critical probability of site percolation on the square lattice. Random Structures and Algorithms 8, (1996), 199–212 5. Bleher, P.M., Cheng, Z., Dyson, F.J., Lebowitz, J.L.: Distribution of the error term for the number of lattice points inside a shifted circle. Commun. Math. Phys. 54, 433–469 (1993) 6. Bombieri, E.: Le grand crible dans la théorie analytique des nombres. 2nd edition, Astérisque 18, 1987 7. Bourgain, J.: Spherical summation and uniqueness of multiple trigonometric series. Int. Math. Research Notices 3, 93–107 (1996) 8. Bousquet-Mélou, M.: Percolation models and animals. European J. Combin. 17, 343–369 (1996) 9. Bousquet-Mélou, M., Eriksson, K.: Lecture hall partitions. Ramanujan J. 1, 101–111 (1997) 10. Broadbent, S.R., Hammersley, J.M.: Percolation processes. Proc. Cambridge Phil. Soc. 53, 629–641 (1957) 11. Brydges, D.C., Spencer, T.: Self–avoiding walk in 5 or more dimensions. Commun. Math. Phys. 97, 125–148 (1985) 12. Coleman, M.D.: The Rosser–Iwaniec sieve in number fields, with an application. Acta Arithmetica 66, 53–83 (1993)
66
I. Vardi
13. Cramér, H.: On the order of magnitude of the difference between consecutive prime numbers. Acta Arithmetica 2, 23–46 (1937) 14. Efros, A.L.: Physics and Geometry of Disorder, Percolation Theory. Moscow: Mir, 1986 15. Erd˝os, P., Gruber, P.M., Hammer, J.: Lattice Points. New York: Longman, Wiley, 1989 16. Estermann, T.: On certain functions represented by Dirichlet series. Proc. London Math. Soc. 27, 435–448 (1928) 17. Feller, W.: An Introduction to Probability Theory and its Applications. Vol. I, New York: Wiley, 1964 18. Flajolet, P., Vardi, I.: Zeta expansions of classical constants. Preprint (1996) 19. Fouvry, E., Iwaniec, H.: Gaussian primes. Acta Arithmetica 79, 249–287 (1997) 20. Friedlander, J.B.: Sifting short intervals. Math. Proc. Camb. Phil. Soc. 91, 9–15 (1982) 21. Friedlander, J.B.: Sifting short intervals II. Math. Proc. Camb. Phil. Soc. 92, 381–384 (1982) 22. Friedlander, J.B., Iwaniec, H.: Using a parity-sensitive sieve to count prime values of a polynomial. Proc. Nat. Acad. Sci. U.S.A. 94, 1054–1058 (1997) 23. de Gennes, P.G.: La percolation: Un concept unificateur. La Recherche 7, 919–927 (1976) 24. Gethner, E., Wagon, S., Wick, B.: A stroll through the Gaussian primes. American Math. Monthly 105, 327–337 (1998) 25. Granville, A.: Unexpected irregularities in the distribution of prime numbers. Proceedings of the International Congress of Mathematicians, (Zurich, 1994). Basel: Birkhäuser, 1995, pp. 388–399 26. Grimmett, G.: Percolation. New York: Springer-Verlag, 1989 27. Hafner, J.L., Sarnak, P., McCurley, K.: Relatively prime values of polynomials. In: Knopp, M., Sheingorn, M. (eds), A tribute to Emil Grosswald: Number theory and related analysis. Contemp. Math. 143, Amer. Math. Soc., Providence, RI: 1993, pp. 437–443 28. Halberstam, H.,Richert, H.–E.: Sieve Methods. New York: Academic Press, 1974 29. Hara, T., Slade, G.: Mean-field critical behavior for percolation in high dimensions. Commun. Math. Phys. 128, 339–391 (1990) 30. Hardy, G.H., Littlewood, J.E.: Some problems of ‘Partitio Numerorum’ III: On the expression of a number as a sum of primes. Acta Math. 44, 1–70, also In: G.H. Hardy: Coll. Papers, vol. 1, 1922, pp. 561–630 31. Harman, G.: Almost primes in short intervals. Math. Ann. 258, 107–112 (1981) 32. Heath-Brown, D.R.: Gaps between primes and the pair correlation of zeros of the zeta–function. Acta Arithmetica 41, 85–99 (1982) 33. Heath-Brown, D.R.: The distribution and moments of the error term in the Dirichlet divisor problems. Acta Arithmetica 60, 389–415 (1992) 34. Heath-Brown, D.R., Iwaniec, H.: On the difference between consecutive primes. Invent. Math. 55, 49–69 (1979) 35. Huxley, M.N., Nowak, W.G.: Primitive lattice points in convex planar domains. Acta Arithmetica 76, 271–283 (1996) 36. Iwaniec, H.: Rosser’s sieve. Acta Arithmetica 36, 171–202 (1980) 37. Iwaniec, I., Mozzochi, C.J.: On the divisor and circle problems. J. Number Theory 29, 60–93 (1988) 38. Lalley, S.: Percolation on Fuchsian groups. Ann. Inst. H. Poincaré Probab. Statist. 34, 151–177 (1998) 39. Lubotzky, A., Phillips, R., Sarnak, P.: Ramanujan graphs. Combinatorica 8, 261–277 (1988) 40. Montgomery, H.L.: Fluctuations in the mean of Euler’s phi function. Proc. Indian Acad. Sci. Math. Sci. 97, 239–245 (1987) 41. Moroz, B.Z.: On the number of primitive lattice points in plane domains. Monatschrift Math. 99, 37–43 (1985) 42. Sarnak, P., Rubinstein, M.: Chebyshev’s bias. Experimental Math. 3, 173–197 (1994) 43. Selberg, A.: On the normal density of primes in short intervals, and the difference between consecutive primes. Arch. Math. Naturvid. 47, 87–105 (1943) 44. Tenenbaum, G.: Introduction to analytic and probabilistic number theory. Cambridge: Cambridge University Press, 1995 45. Turán, P.: On a theorem of Hardy and Ramanujan. J. London Math. Soc. 9, 274–276 (1934) 46. Vardi, I.: Computational Recreations in Mathematica. Reading, MA: Addison–Wesley, 1991 47. Vardi, I.: Prime Percolation. Experimental Mathematics 7, 275–289 (1998) 48. Watt, N.: Short intervals almost all containing primes. Acta Arithmetica 72, 131–167 (1995) 49. Wierman, J.C.: Substitution method critical probability bounds for the square lattice site percolation model. Combin. Probab. Comput. 4, 181–188 (1995) 50. Zeilberger, D.: The Abstract Lace Expansion. Adv. Appl. Math. 19, 355–359 (1997) 51. Ziff, R.M., Sapoval, B.: The efficient determination of the percolation threshold by a frontier–generating walk in a gradient. J. Phys A: Math. Gen. 19, L1169–L1172 (1986) Communicated by P. Sarnak
Commun. Math. Phys. 207, 67 – 80 (1999)
Communications in
Mathematical Physics
© Springer-Verlag 1999
Billiards in Tubular Neighborhoods of Manifolds of Codimension 1 Roberto Peirone Università di Roma “Tor Vergata”, Dipartimento di Matematica, via della Ricerca Scientifica, 00133, Roma, Italy. E-mail: peirone@ mat.uniroma2.it Received: 19 June 1998 / Accepted: 7 April 1999
Abstract: I prove that in (sufficiently small) tubular ρ-neighborhoods of a given C 2 manifold of codimension 1, any two points can be connected by a billiard trajectory, and 1 that in addition there exists such a trajectory having at most Hρ − 2 collision points, for 3 a suitable H > 0, provided the manifold is of class C . 0. Introduction In this paper I deal with billiard trajectories, i.e., the motion, in a region having smooth boundary, of a point which, when it reaches the boundary, moves according to the rule: incidence angle equal to reflection angle. There is extensive literature on billiards (see for example [2] and references therein). Probably, the most investigated problem on billiards is that of ergodicity of the related dynamical system (a good reference for ergodicity is [7]). Another typical problem is that of the existence of periodic trajectories (see for example [1] and [3]). Other references on billiards are [6] and [11]. In this paper I consider a problem which is not, to my knowledge, much investigated, namely that of the existence of a trajectory connecting two given points. Note that this question has a trivial positive answer if the region is convex, but an example due to L. Penrose, R. Penrose ([9]), based on an ellipse shows that there exist a region and two points of which are connected by no billiard trajectories in (see Fig. 1). If we permit the points to be in ∂ (not necessarily in the interior), it is sufficient to consider a region between two circles C1 and C2 such that C2 lies in the interior of C1 and contains the centre A of C1 , for, clearly A cannot be connected to many points of the region. In [10] a strengthened version of the result of [9] for regions having non-smooth boundary is discussed. It appears to be very difficult to give a characterization of the regions for which any two points can be connected by a billiard trajectory, thus here I consider the simpler problem of giving a not too special and yet nontrivial class of regions having this property. Namely, let M be a compact and connected C 2 manifold of codimension 1 in Rn . Then for sufficiently small positive ρ (see Prop. 1.2), the region Mρ , which is
68
R. Peirone
Fig. 1. The arc between P and Q is an arc of the ellipse whose foci are F1 and F2 . By the properties of an ellipse, no billiard trajectories connect points in A to points in B or in C
Fig. 2. Trajectory connecting two given points
the set of the points having distance from M smaller than ρ, i.e., a tubular neighborhood of M, has the above property (see for example [5], Ch. 4 for the notion of tubular neighborhood). Note that in some sense this result generalizes the trivial corresponding result for annuli. This is the main result of this paper and is proved in Sect. 2. I prove this, using a “variational approach”, i.e., I consider the polygonal having the two given points as ends, and touching the boundary at N points, each point lying on the opposite part of the boundary with respect to the previous one, and of minimum length. The nontrivial point is to prove that this polygonal, for N sufficiently large and of suitable parity, remains in the region. This polygonal is characterized by the fact that its direction is "almost orthogonal" to the boundary (see Fig. 2). By using similar considerations it is possible to prove the existence of nontrivial periodic trajectories in the two-dimensional case (see Remark 2.5). The variational approach, in a simpler form, is usual in convex regions. In Sect. 3 by supposing that is of class C3 , I get a bound for the number of collision points (which are the points at which the trajectory touches the boundary), namely I show that the number of collision points can be chosen to be at most of the 1 order of ρ − 2 for ρ → 0 (Theorem 3.5). Note that this result is trivial in the case of an annulus. The results of Sect. 2 were stated without proof in [8], for the two-dimensional case and regions of class C 4 . It could be interesting to try to extend the results of this paper to the case in which M may have codimension greater than 1.
Billiards in Tubular Neighborhoods of Manifolds of Codimension 1
69
1. Notation If a, b ∈Rn we denote by a · b the scalar product between a and b. If a ∈Rn , we denote by aˆ that vector of Rn−1 given by (a1 , . . . , an−1 ). We denote by η the euclidean distance in Rn and by || || the euclidean norm in Rn . As usual, we put η(x, A) = inf η(x, y), y∈A
when x ∈Rn , ∅ 6 = A ⊆Rn . Also, when ∅ 6 = A ⊆Rn , let Aρ = {x ∈ Rn : η(x, A) < ρ} for every ρ > 0. Given a, b ∈Rn , [a, b] (resp. ]a, b[) will denote the closed (resp. open) segment line with ends a and b, i.e., the set {ta+(1−t)b : t ∈ [0, 1]} (resp. {ta+(1−t)b : m−1 S [ai , ai+1 ]. If f is t ∈]0, 1[}). If a0 , . . . , am ∈Rn (m ≥ 1), we put [a0 , . . . , am ] = i=0
a differentiable map from an open subset U of Rn into Rm and x ∈ U , we denote by (df )x (u) the derivative of f at x calculated at u. In the following, manifold will always mean submanifold of Rn without boundary, with n ≥ 2. If M is a manifold of class C 1 , −→ 0}, x ∈ M, we denote by TM (x) the tangent space to M at x, i.e., {z ∈Rn : η(x+tz,M) t t→0
and put NM (x) = TM (x)⊥ , i.e., NM (x) is the normal space to M at x. The following lemma is well known. Lemma 1.1. If M is a C 1 -manifold, z ∈Rn , x ∈ M, and η(z, x) = η(z, M), then z − x ∈ NM (x). For the following M will denote a compact and connected (n − 1)-manifold of class C s with s ≥ 1. We consider the geodesic distance dM on M, defined by dM (x, x 0 ) = R1 inf ||γ 0 (t)|| dt, where inf is taken over all Lipshitzian curves γ : [0, 1] → M such 0
that γ (0) = x, γ (1) = x 0 , and let L = sup dM (z, z0 ). It is known that Rn \ M z,z0 ∈M
has exactly two connected components, one bounded which we denote by i(M) and the other unbounded, which we denote by e(M), and also, there exists a continuous function νM : M →Rn such that νM (x) is the outer unit normal vector to M at x, and we write simply ν instead of νM when the manifold is clear from the context. I guess that the following proposition is known. However we sketch a proof of it ; i) and part of iii) can be found in [4], Appendix. It regards, in substance, the existence of a tubular neighborhood of M. For our convenience, we here consider a standard simple local representation of M, namely, given x ∈ M, there exist an isometry 3x of Rn , and Ux open in Rn and Vx open in Rn−1 containing 0, and ϕx : Vx →R of class C s , such that x ∈ Ux , 3x (x) = 0, and ˆ ϕx (0) = 0, 3x (Ux ∩ M) = {x ∈ Rn : xˆ ∈ Vx , xn = ϕx (x)}, ∂ϕx ∂ϕx − ∂x1 (x), ˆ . . . , − ∂xn−1 (x), ˆ 1 ∂ϕx (0) = 0, ν3x (M) (x) = ∀ x ∈ 3x (Ux ∩ M), x x ∂xi || − ∂ϕ ˆ . . . , − ∂x∂ϕn−1 (x), ˆ 1 || ∂x1 (x), in particular, ν3x (M) (0) = en (= (0, . . . , 0, 1)). Proposition 1.2. If M is of class C s , s ≥ 2, then there exists ρ > 0 such that (i) For every z ∈ Mρ there exists a unique p1 (z) ∈ M such that z−p1 (z) ∈ NM (p1 (z)) and ||z − p1 (z)|| < ρ; also p1 (z) is the unique point of M of minimum distance from z. (ii) p1 has class C s−1 , and we have (dp1 )z (v) = v − (v · νM (z))νM (z) for all z ∈ M.
70
R. Peirone
(iii) If φ :Rn →R is defined by φ(z) =
(
η(z, M) if z ∈ e(M), −η(z, M) if z ∈ / e(M),
then φ is continuous on Rn and of class C s on Mρ , and we have (dφ)z (v) = νM (p1 (z)) · v for every z ∈ Mρ . (iv) If |t| < ρ, then φ −1 (t) is a (nonempty) compact and connected C s manifold of dimension n − 1. (v) Given t, t 0 with |t| < ρ, |t 0 | < ρ and z ∈ φ −1 (t), then z0 = p1 (z) + t 0 νM (p1 (z)) is the unique point of φ −1 (t 0 ) of minimum distance from z and ||z − z0 || = |t − t 0 |. Also, z − p1 (z) ∈ Nφ −1 (t) (z). (vi) ||z − z0 || ≥ |φ(z) − φ(z0 )| ∀ z, z0 ∈ Mρ . (vii) If x1 , x2 ∈ M, y ∈ NM (x1 ), then |(x2 − x1 ) · y| ≤
||x2 −x1 ||2 ||y|| . 2ρ
Proof. Given x ∈ M, consider a local representation of M at x as above, and for ρ > 0 consider ψx : Vx ×] − ρ, ρ[→Rn defined by ˆ ϕx (x)) ˆ + xn ν3x (M) (x, ˆ ϕx (x)). ˆ ψx (x) = (x, A simple calculation yields (1.1) (dψx )0 = I d, thus, by possibly restricting Vx , we can and do assume that there exists ρx > 0 such that ψx : Vx ×] − ρx , ρx [ is a C s−1 diffeomorphism, in particular one-to-one, onto its image. By a compactness argument, there exists ρ > 0 such that the map ψ : M×] − ρ, ρ[→Rn defined by ψ(x, t) = x + tνM (x) is a one-to-one map whose image is an open neighborhood of M, which is, in view of Lemma 1.1, exactly Mρ . Thus, i) easily follows, in view of Lemma 1.1 again, and by (1.1) we also get ii). Since η(z) = ||z − p1 (z)||, iii) is a simple consequence of ii). To prove iv), observe that φ −1 (t) = {x + tν(x) : x ∈ M} = Ct (M), where Ct (x) = x + tν(x), so φ −1 (t) is compact and connected, and is also a C s manifold, because, by ii), (dφ) 6 = 0 there. v) is a simple consequence of Lemma 1.1, and vi) follows from v). For vii), suppose y 6= 0. y y − x2 || ≥ ρ, thus ||x1 − x2 ||2 + ρ 2 ± 2ρ(x2 − x1 ) · ||y|| ≥ ρ2, Then, we have ||x1 ± ρ ||y|| and vii) follows at once. u t We remark that we can express any z ∈ Mρ as z = p1 (z) + p2 (z), where p2 (z) = z − p1 (z) ∈ NM (p1 (z)), and in fact we have p2 (z) = φ(z)νM (p1 (z)), in other words z = p1 (z) + φ(z)νM (p1 (z)), for all z ∈ Mρ . Note also that when 0 < ρ < ρ we have Mρ = {z ∈Rn : η(z, M) ≤ ρ}. We now study the billiard trajectories. If U is a connected open set in Rn whose boundary is a C 1 -manifold of dimension n − 1, and given a, b, c ∈Rn , with a 6= b 6= c, we say that [a, b, c] is a billiard trajectory in U , if b ∈ ∂U , ]a, b[∪]b, c[⊆ U , and, the half-line with origin at b and passing through a and the half-line with origin at b and passing through c are symmetric with respect to the normal to ∂U at b, in formula c−b a−b ·z= · z ∀z ∈ N∂U (b), ||a − b|| ||c − b|| c−b a−b ·z=− · z ∀z ∈ T∂U (b), ||a − b|| ||c − b||
Billiards in Tubular Neighborhoods of Manifolds of Codimension 1
71
||c−b|| or, equivalently, c − b = ||a−b|| (2n · (a − b)n − (a − b)), where n is any unit normal vector to ∂U at b. We say that [a0 , . . . , am ], m ≥ 2, is a billiard trajectory in U if [ai , ai+1 , ai+2 ] is a billiard trajectory in U for i = 0, . . . , m − 2. In such a situation we say that a1 , . . . , am−1 are the collision points (or c.p.) of the trajectory, and that the trajectory connects a0 and am . The following proposition gives a useful characterization of billiard trajectories, and is well known.
Proposition 1.3. Let a, b, c ∈Rn with a 6 = b 6 = c, let U be a nonempty open set in Rn whose boundary is a C 1 manifold of dimension n − 1. Suppose ]a, b[∪]b, c[⊆ U , b ∈ ∂U . Then (i) [a, b, c] is a billiard trajectory in U if and only if the function S : ∂U →R defined by S(x) = ||a − x|| + ||x − c||, is stationary at b, in the sense that (for example) S(x)−S(b) ||x−b|| −→ 0. x→b
(ii) If there exists V open in ∂U containing b such that S(x) ≥ S(b) for all x ∈ V , then [a, b, c] is a billiard trajectory. From now on, we assume that M is of class C 2 (at least). 2. Trajectories Connecting Two Given Points Fix ρ with 0 < ρ < ρ. In this section we prove the main result of this paper, i.e., that any two points in Mρ can be connected by a billiard trajectory in Mρ having N collision points, provided N is sufficiently large and of appropriate parity, the parity depending on the two points (Theorem 2.4). We use the following notation. Let ζ (1) , ζ (2) ∈ Mρ . Then put if φ(ζ (1) )φ(ζ (2) ) < 0 2N P(ζ (1) , ζ (2) ) = 2N + 1 if φ(ζ (1) )φ(ζ (2) ) > 0 N if φ(ζ (1) )φ(ζ (2) ) = 0 , PN (ζ (1) , ζ (2) ) = {z ∈ (Rn )N+2 : z0 = ζ (1) , zN +1 = ζ (2) , φ(zi )φ(zi+1 ) ≤ 0 ∀ i = 0, . . . , N, |φ(zi )| = ρ ∀ i = 1, . . . , N}. Here, and in the following, we indicate a point z of (Rn )N +2 as (z0 , . . . , zN +1 ). When convenient, we will identify z with the polygonal [z0 , . . . , zN +1 ]. Note that PN (ζ (1) , ζ (2) ) 6 = ∅ provided N ∈ P(ζ (1) , ζ (2) ) (and this is the motivation to introduce P(ζ (1) , ζ (2) )). The idea is now to minimize the length of a polygonal in PN (ζ (1) , ζ (2) ) when N is a fixed element of P(ζ (1) , ζ (2) ), sufficiently large. Thus we are lead to N P ||zi+1 − zi ||; also put 30N (ζ (1) , ζ (2) ) = define the length 3N of z as 3N (z) = i=0
|φ(ζ (1) )|+|φ(ζ (2) )|+2Nρ. Observe that we have 30N (ζ (1) , ζ (2) )
=
N P
|φ(zi+1 )−φ(zi )| i=0 30N (ζ (1) , ζ (2) ) for all
for every z ∈ PN (ζ (1) , ζ (2) ), and by Prop. 1.2, vi, 3N (z) ≥ z ∈ PN (ζ (1) , ζ (2) ). On the other hand, the geometry of a tubular domain suggests that, by choosing z ∈ PN (ζ (1) , ζ (2) ) in such a way that ||p1 (zi+1 ) − p1 (zi )|| is small, for every ε > 0 the length of the polygonal can be smaller than 30N (ζ (1) , ζ (2) ) + ε (for sufficiently large N ∈ P(ζ (1) , ζ (2) )). Thus, for this N, the minimizing polygonal remains in Mρ (in the contrary case its length is bigger than 30N (ζ (1) , ζ (2) ) + σ for some
72
R. Peirone
σ > 0, independent of N). To realize this idea we need three (simple) lemmas. For our convenience, we put H = {(ζ (1) , ζ (2) ) : ζ (1) , ζ (2) ∈ Mρ , |φ(ζ (1) )| = ρ, φ(ζ (1) )φ(ζ (2) ) ≤ 0} and observe that if N ≥ 1 and z ∈ PN (ζ (1) , ζ (2) ), then for every i = 0, . . . , N, either (zi , zi+1 ) or (zi+1 , zi ) is in H. Lemma 2.1. There exists σ > 0 such that if (ζ (1) , ζ (2) ) ∈ H and |p1 (ζ (1) )−p1 (ζ (2) )| < σ , then ]ζ (1) , ζ (2) [⊆ Mρ . Proof. In this proof we always assume (ζ (1) , ζ (2) ) ∈ H. By Prop. 1.2, iii, by putting α(t)(= αζ (1) ,ζ (2) (t)) = |φ(tζ (1) + (1 − t)ζ (2) )|, we see that, if p1 (ζ (1) ) = p1 (ζ (2) ), and φ(ζ (2) ) 6 = 0, then α 0 (0) < 0, α 0 (1) > 0. So, by compactness, there exist β ∈]0, 21 [ and σ 0 > 0 such that if ||p1 (ζ (1) ) − p1 (ζ (2) )|| < σ 0 , then α(t) < ρ for every t ∈ ]0, β[∪]1 − β, 1[ (for t ∈]0, β[ we see this directly if |φ(ζ (2) )| ≥ θρ with 0 < θ < 1, for example θ = 21 , but this is also obvious when |φ(ζ (2) )| < ρ2 ). On the other hand, if β ≤ t ≤ 1 − β and p1 (ζ (1) ) = p1 (ζ (2) ), then we have |α(t)| ≤ (1 − β)ρ, thus there exists σ 00 > 0 such that if β ≤ t ≤ 1 − β and ||p1 (ζ (1) ) − p1 (ζ (2) )|| < σ 00 , then t α(t) < ρ. We conclude by taking σ = min{σ 0 , σ 00 }. u Lemma 2.2. For every ε > 0 there exists λ(ε) > 0, such that, if (ζ (1) , ζ (2) ) ∈ H, and |p1 (ζ (1) ) − p1 (ζ (2) )| ≥ ε, then ||ζ (1) − ζ (2) || ≥ |φ(ζ (1) ) − φ(ζ (2) )| + λ(ε). Proof. This easily follows from Prop. 1.2, vi, and a compactness argument. u t Lemma 2.3. There exists α : [0, +∞[→ [0, +∞[ with α(0) = 0, α continuous and increasing such that if (ζ (1) , ζ (2) ) ∈ H, then ||ζ (1) − ζ (2) || ≤ |φ(ζ (1) ) − φ(ζ (2) )| + ||p1 (ζ (1) ) − p1 (ζ (2) )||α(||p1 (ζ (1) ) − p1 (ζ (2) )||), more precisely ||ζ
(1)
−ζ
(2)
|| ≤ |φ(ζ
(1)
) − φ(ζ
(2)
r ρ )| + ||p1 (ζ (1) ) − p1 (ζ (2) )||2 (1 + ) + ρ 2 − ρ. ρ
Proof. Let (ζ (1) , ζ (2) ) ∈ H. Then ||ζ (1) − ζ (2) || − |φ(ζ (1) ) − φ(ζ (2) )| ≤ ||ζ (1) − p1 (ζ (2) )|| + ||ζ (2) − p1 (ζ (2) )|| − |φ(ζ (1) )| − |φ(ζ (2) )| = ||ζ (1) − p1 (ζ (2) )|| − ρ, and, by Prop. 1.2, vii), we get ||ζ (1) − p1 (ζ (2) )||2 = ||ζ (1) − p1 (ζ (1) )||2 + ||p1 (ζ (1) ) − p1 (ζ (2) )||2 + 2(ζ (1) − p1 (ζ (1) )) · (p1 (ζ (1) ) − p1 (ζ (2) )) ρ ≤ ||p1 (ζ (1) ) − p1 (ζ (2) )||2 (1 + ) + ρ 2 . ρ Now, the thesis easily follows. u t
Billiards in Tubular Neighborhoods of Manifolds of Codimension 1
73
Theorem 2.4. Suppose M is of class C 2 . Then there exists N ≥ 1 such that, if ζ (1) , ζ (2) ∈ Mρ , N ≥ N , N ∈ P(ζ (1) , ζ (2) ), then there exists a billiard trajectory in Mρ with N c.p. connecting ζ (1) and ζ (2) . Proof. The function 3N , being continuous, does take a minimum at some z ∈ PN (ζ (1) , ζ (2) ). By Prop. 1.3 it suffices to prove that there exists N ≥ 1 such that if ζ (1) , ζ (2) ∈ Mρ , N ≥ N, N ∈ P(ζ (1) , ζ (2) ), and i = 0, . . . , N, then ]zi , zi+1 [⊆ Mρ . By Lemma 2.1 it suffices to prove that ||p1 (zi ) − p1 (zi+1 )|| < σ . Suppose this is not true. Then by Lemma 2.2, we have 3N (z) ≥ 30N (ζ (1) , ζ (2) ) + λ(σ ). Now, given N ∈ P(ζ (1) , ζ (2) ) sufficiently large, we exhibit z˜ ∈ PN (ζ (1) , ζ (2) ) with 3N (˜z) < 30N (ζ (1) , ζ (2) ) + λ(σ ), in contrast to the definition of z. Choose xi ∈ M such that x0 = p1 (ζ (1) ), xN+1 = p1 (ζ (2) ), and ||xi+1 − xi || ≤ NL+1 for i = 0, . . . , N, and for i = 1, . . . , N let yi = (−1)i+c ρν(xi ) with c ∈ {0, 1} such that, putting y0 = p2 (ζ (1) ), yN+1 = p2 (ζ (2) ) and z˜ i = xi + yi , we have z˜ ∈ PN (ζ (1) , ζ (2) ). Note that this choice of yi is possible because N ∈ P(ζ (1) , ζ (2) ). By Lemma 2.3 ||˜zi+1 − z˜ i || ≤ |φ(˜zi+1 ) − φ(˜zi )| +
L L α( ). N +1 N +1
L ) < λ(σ ), to have that Now, it suffices to take N so large that Lα( N+1
3N (˜z) < 30N (ζ (1) , ζ (2) ) + λ(σ ).
t u
Remark 2.5. We can obtain similar results for different problems. For example, we can prove that given ζ (1) , ζ (2) ∈ Mρ and 1 > 0, there exists a billiard trajectory z ∈ PN (ζ (1) , ζ (2) ) in Mρ connecting ζ (1) and ζ (2) such that ||p1 (zi+1 ) − p1 (zi )|| ≤ min{1, σ } for i = 0, . . . , N. In order to get this, it suffices to consider the minimum of 3N over the set of z ∈ PN (ζ (1) , ζ (2) ) for which ||p1 (zi+1 ) − p1 (zi )|| ≤ min{1, σ } for i = 0, . . . , N,
(2.1)
and to prove, by proceeding in a similar way as in Theorem 2.4, that for this minimum z the strict inequality holds in (2.1), thus by Prop. 1.3 we have a billiard trajectory. By choosing 1 smaller and smaller, we see that the billiard trajectory connecting ζ (1) and ζ (2) can be taken to be arbitrarily close to the normal to ∂Mρ at all its collision points. Also, we can consider the problem min 3N (z), z∈A
where A can be a given closed subset of (Rn )N +2 , and proceed similarly. For example if A = {z : p1 [[z0 , . . . , zN+1 ]] is homotopic to γ , z0 = ζ (1) , zN +1 = ζ (2) }, where γ is a given curve on M connecting p1 (ζ (1) ) to p1 (ζ (2) ), and [[z0 , . . . , zN +1 ]] denotes the curve which runs along the polygonal [z0 , . . . , zN +1 ] with unitary velocity. We get in this way the existence of a billiard trajectory connecting ζ (1) and ζ (2) in a given homotopy class. If, instead, γ is a closed curve in M and consider A = {z : p1 [[z0 , . . . , zN+1 ]] is homotopic to γ , z0 = zN +1 }, we get the existence of a periodic billiard trajectory in a given homotopy class. This type of result is interesting especially in case M is a Jordan curve in R2 , for in this case there exist many mutually non-homotopic curves.
74
R. Peirone
3. The Case ρ Small In Sect. 2 we have proved the existence of a billiard trajectory connecting two given points, and it is possible in fact to estimate explicitly the number of collisions. For example, by making more explicit the arguments of Sect. 2, we get that if N ≥ 1 is such that s s L 2 ρ ρ 2ρ ) (1 + ) + ρ 2 − ρ < 4ρ 2 + (1 − )ρ 2 − 2ρ, (N + 1) ( N +1 ρ ρ ρ+ρ then for every ζ (1) , ζ (2) ∈ Mρ , if N ∈ P(ζ (1) , ζ (2) ), there exists a billiard trajectory in Mρ connecting ζ (1) and ζ (2) having N collision points. The proof of this estimate is not very complicated, but a bit technical, and will be omitted. However, this estimate is far from the best, in particular for ρ small. In fact, the case where M is a circle in 1 R2 suggests that the number of collision points has the order of ρ − 2 . In this section we prove that this type of estimate holds, provided M is of class C 3 (Theorem 3.5). Since we work in a compact subset of Mρ we a priori consider only ρ not bigger than some specific ρˆ < ρ, for example ρˆ = ρ2 . We now sketch the idea of the proof. First of all, we consider a variant of the problem considered in Sect. 2. Namely, given ζ (1) , ζ (2) ∈ Mρ , let PN,T (ζ (1) , ζ (2) ) = {z ∈ PN (ζ (1) , ζ (2) ) : ||p1 (zi+1 ) − p1 (zi )|| √ ≤ T ρ ∀ i = 0, . . . , N}, mN,T (ζ (1) , ζ (2) ) = {z ∈ PN,T (ζ (1) , ζ (2) ) : 3N (z) ≤ 3N (z) ∀ z ∈ PN,T (ζ (1) , ζ (2) )}. Once ζ (1) and ζ (2) are fixed, we omit the dependence on ζ (1) and ζ (2) , and write simply PN,T , mN,T . In the following we will use without explicit mention the simple observation that, if z ∈ mN,T (ζ (1) , ζ (2) ), and i = 0, . . . , N − 1, then (zi , zi+1 , zi+2 ) ∈ m1,T (zi , zi+2 ). Now, it is easy to see that mN,T (ζ (1) , ζ (2) ) is nonempty provided N is sufficiently large and N ∈ P(ζ (1) , ζ (2) ). Moreover, in view of Prop. 1.3, every z ∈ mN,T (ζ (1) , ζ (2) ) is a billiard trajectory in Mρ , provided it lies in fact in Mρ and √ (3.1) ||p1 (zi ) − p1 (zi+1 )|| < T ρ for each i = 0, . . . , N. √ In Lemma 3.1 we see that if ρ and T are sufficiently small (for example, T = ρ), then the trajectory lies in Mρ . On the other hand, if (3.1) does not hold then, by Lemmas 3.2 and 3.3, (for sufficiently small ρ and T ) ||p1 (zi ) − p1 (zi+1 )|| is, roughly speaking, "rather far" from 0 for all i, provided N is not too large. Therefore, using Lemma 3.4 and a not completely obvious argument, we can find a point in PN,T whose length is strictly smaller than that of z, and this contradicts the hypothesis z ∈ mN,T . Lemma 3.1. There exists ρ1 ∈]0, ρ2 [ such that if ρ ∈]0, ρ1 [ and (ζ (1) , ζ (2) ) ∈ H and √ ||p1 (ζ (1) ) − p1 (ζ (2) )|| ≤ ρρ, then ]ζ (1) , ζ (2) [⊆ Mρ . Proof. For t ∈ [0, 1] let zt = ζ (2) +t (ζ (1) −ζ (2) ), (so in particular z1 = ζ (1) , z0 = ζ (2) ), let β(t) = φ(zt ), and let α be defined as in Lemma 2.1, so that α(t) = |β(t)|. By the hypothesis we have that α : [0, 1] → R is continuous, α(0) ≤ ρ, α(1) ≤ ρ, β(0)β(1) ≤ 0. We are going to prove that if for some t ∈ [0, 1] and for some i = 0, 1 we have α(t) = ρ, β(t)β(i) ≤ 0, (3.2)
Billiards in Tubular Neighborhoods of Manifolds of Codimension 1
( >0 α 0 (t) <0
then
75
if i = 0 if i = 1.
(3.3)
Thus, a standard argument implies α(t) < ρ for all t ∈]0, 1[, thus the thesis. Fix t ∈ [0, 1] satisfying (3.2) and let ν = ν(p1 (zt )). Observe that ν · (zt − zi ) = ν · (zt − p1 (zt )) + ν · (p1 (zt ) − p1 (zi )) + ν · (p1 (zi ) − zi ) = φ(zt ) − φ(zi )ν · ν(p1 (zi )) + ν · (p1 (zt ) − p1 (zi )). √ On the other hand, if ρ1 is sufficiently ||zt −zi || ≤ ||ζ (1) −ζ (2) || ≤ ρρ+2ρ √ small, from√ we deduce ||p1 (zt ) − p1 (zi )|| ≤ ρρ + 4ρ < 2ρρ, and also ||ν(p1 (zi )) − ν|| ≤ 1. Thus, ν · ν(p1 (zi )) ≥ 0, and by (3.2) and Prop. 1.2, vii, ν · (zt − zi ) has the same sign as β(t). Now, (3.3) easily follows from Prop. 1.2, iii. u t In the following lemma we prove that, if (z1 , z2 , z3 ) ∈ P1,T (z1 , z3 ), then, roughly 1 (z1 )−p1 (z2 )|| speaking, the ratio ||p ||p1 (z3 )−p1 (z2 )|| cannot be too large. The geometrical interpretation of this statement is that if the previous ratio is very large, then the angle between [z1 , z2 ] and ν(p1 (z2 )) is bigger than the angle between [z3 , z2 ] and ν(p1 (z2 )), thus, moving z2 "towards" z1 (remaining on φ −1 (φ(z2 ))), we can make the length of the polygonal / m1,T (z1 , z3 ). strictly smaller, thus (z1 , z2 , z3 ) ∈ Lemma 3.2. Suppose M is of class C 3 . Then for every T > 0 there exist K(T ) ∈]0, 1[ and ρ2 (T ) ∈]0, ρ1 ] such that, if ρ ∈]0, ρ2 (T )[, and (z1 , z2 , z3 ) ∈ m1,T (z1 , z3 ) and √ ||p1 (z1 ) − p1 (z2 )|| = T ρ, (3.4) then
||p1 (z3 ) − p1 (z2 )|| ≥ K(T )T
√ ρ.
Proof. √ Given (z1 , z2 , z3 ) ∈ P1,T (z1 , z3 ) satisfying (3.4) with ||p1 (z3 ) − p1 (z2 )|| = AT ρ, let ν = νM (p1 (z2 )), and put v = z2 − z1 − (z2 − z1 ) · ν ν, α(t) = ||z2 + tv − z1 || + ||z2 + tv − z3 ||,
u = z2 − z3 − (z2 − z3 ) · ν ν, (3.5) β(t) = ||p1 (z2 + tv) − p1 (z1 )||.
It suffices to prove that if ρ and A are sufficiently small then α 0 (0) > 0,
(3.6)
β 0 (0) > 0.
(3.7)
In fact, if (3.6) and (3.7) hold, it is easy to see that, by moving the point z2 with direction −v, remaining on φ −1 (φ(z2 )) we get a polygonal in P1,T (z1 , z3 ) whose length is strictly smaller than that of (z1 , z2 , z3 ). Therefore, if (z1 , z2 , z3 ) ∈ m1,T (z1 , z3 ), then for sufficiently small ρ, A must be greater than (or equal to) some K(T ) > 0. We prove ||v||2 (3.6). We have α 0 (0) = ||z3u·v −z2 || + ||z1 −z2 || . Thus, in order to have (3.6) it is sufficient to prove
||v||2 ||z1 −z2 ||2
>
||u||2 , ||z3 −z2 ||2
or equivalently, in view of (3.5),
|(z3 − z2 ) · ν| |(z1 − z2 ) · ν| < . ||z1 − z2 || ||z3 − z2 ||
(3.8)
76
R. Peirone
Using Prop. 1.2, vii and (3.4) we have |(z1 − z2 ) · ν| ≤ 2ρ +
T 2ρ 2ρ ;
also √ ||z1 − z2 || ≥ ||p1 (z1 ) − p1 (z2 )|| − ||p2 (z2 )|| − ||p2 (z1 )|| ≥ T ρ − 2ρ. √ Thus, if ρ is so small that T ρ − 2ρ > 0, we have ρ(2 + T2ρ ) |(z1 − z2 ) · ν| ≤ √ . ||z1 − z2 || T ρ − 2ρ
(3.9)
2
(3.10)
Similarly, (z3 − z2 ) · ν = (z3 − p1 (z3 )) · ν + (p1 (z3 ) − p1 (z2 )) · ν + (p1 (z2 ) − z2 ) · ν. √ Now, since ν(p1 (z2 )) · ν = ν · ν = 1, and ||p1 (z3 ) − p1 (z2 )|| ≤ T ρ, if we take ρ2 (T ) so small that p x1 , x2 ∈ M, ||x1 − x2 || ≤ T ρ2 (T ) ⇒ ||ν(x1 ) − ν(x2 )|| ≤ 1 , we have, in view of Prop. 1.2, vii, |(z3 − z2 ) · ν| ≥ |(p1 (z2 ) − z2 ) · ν| − |(p1 (z3 ) − p1 (z2 )) · ν| ≥ ρ − also ||z3 − z2 || ≤ AT
√ ρ + 2ρ, hence we get
A2 T 2 ρ ; 2ρ
ρ(1 − A2ρT ) |(z3 − z2 ) · ν| ≥ . √ ||z3 − z2 || AT ρ + 2ρ 2
2
In conclusion, in view also of (3.10), if ρ and A are sufficiently small, we get (3.8), hence (3.6). We are now going to prove that (3.7) holds for sufficiently small ρ and A. We have (p1 (z2 ) − p1 (z1 )) · γ10 (0) , β 0 (0) = ||p1 (z2 ) − p1 (z1 )|| where γ1 (t) = p1 (z2 + tv). Put γ2 (t) = p1 (p1 (z2 ) + tv). We have ||γ10 (0) − γ20 (0)|| ≤ Hρ, where H is an appropriate constant > 0 related to the second derivatives of p1 in M ρ . Also, by Prop. 1.2, ii, γ20 (0) = v. Thus, using (3.4) and (3.5), we have
2
||p1 (z2 ) − p1 (z1 )||β 0 (0) ≥ (p1 (z2 ) − p1 (z1 )) · v − Hρ||p1 (z2 ) − p1 (z1 )|| = √ (z2 − z1 ) · v + (p1 (z2 ) − z2 ) · v + (z1 − p1 (z1 )) · v − H T ρ ρ ≥ 3 √ ||v||2 − 2ρ||v|| − H T ρ ρ = ||v||(||v|| − 2ρ)H T ρ 2 . √ By (3.9) and (3.10) ||v|| ≥ r ρ for suitable r > 0 and sufficiently small ρ > 0, and this concludes the proof. u t In the next lemma we estimate the difference between the length of successive segment lines in a billiard trajectory provided all the ends lie on ∂Mρ . To interpret geometrically this result in dimension two, observe that when M is a circle then the segment lines have the same length, and in the general case, M is approximated by some circle (or straight line) up to the third order.
Billiards in Tubular Neighborhoods of Manifolds of Codimension 1
77
Lemma 3.3. Suppose M is of class C 3 . Then there exist T > 0, B > 0, ρ3 ∈]0, ρ1 [ such that, if ρ ∈]0, ρ3 [ and [z1 , z2 , z3 ] is a billiard trajectory in Mρ with (z1 , z2 , z3 ) ∈ P1,T and |φ(z1 )| = |φ(z3 )| = ρ, then ||z1 − z2 ||3 . ||z3 − z2 || − ||z1 − z2 || ≤ B √ ρ
(3.11)
Proof. Suppose given T , ρ3 , ρ, z1 , z2 , z3 satisfying the hypotheses and prove that if ρ3 , T are appropriate, there exists B > 0 satisfying (3.11), with ρ3 , T , B independent of z1 , z2 , z3 . Put ν = ν(p1 (z2 )), z1 = z2 + aν + v with v · ν = 0. Since [z1 , z2 , z3 ] is a billiard trajectory we have ∃ t > 0 : z3 = z2 + t(aν − v). Put z4 = z2 + (aν − v), λ(t) = φ(z2 + aν + tv) − φ(z2 + aν − tv). It follows ||z3 − z2 || − ||z1 − z2 || = ||z3 − z2 || − ||z4 − z2 || = ||z3 − z4 ||.
(3.12)
We now choose ρ3 so small that all the segment lines [z2 + aν, z2 + aν ± v], [z2 , z4 ], [z3 , z4 ], are contained in M ρ , and it is sufficient that 2
T
ρ √ ρ3 + 3ρ3 ≤ . 3
(3.13)
Indeed, in this case, if P is one of the points z2 , z2 + aν, z2 + aν ± v, z3 , z4 , then we have ||P − p1 (z2 )|| ≤ ρ2 , for, ||P − z2 || ≤ ||z1 − z2 || unless P = z3 , and, clearly, √ √ ||z1 − z2 || ≤ T ρ3 + 2ρ3 , ||z3 − z2 || ≤ T ρ3 + 2ρ3 . Since by Prop. 1.2, iii) we easily get λ0 (0) = λ00 (0) = 0, for a suitable B1 > 0 related to the third derivatives of φ, we have |φ(z4 ) − φ(z3 )| = |φ(z4 ) − φ(z1 )| = |λ(1)| ≤ B1 ||z1 − z2 ||3 .
(3.14)
Thus, 2ρ − B1 ||z1 − z2 ||3 ≤ |φ(z4 ) − φ(z2 )| = gradφ(ξ1 ) · u||z4 − z2 || , B1 ||z1 − z2 ||3 ≥ |φ(z3 ) − φ(z4 )| = gradφ(ξ2 ) · u||z3 − z4 || , where u =
z4 −z2 ||z4 −z2 || , ξ1
(3.5) (3.16)
is a suitable point in ]z2 , z4 [, and ξ2 is a suitable point in ]z3 , z4 [.
Note that max{||ξ1 − p1 (z2 )||, ||ξ2 − p1 (z2 )||} ≤
ρ 2,
thus [ξ1 , ξ2 ] ⊆ M ρ . Moreover, 2
78
R. Peirone
√ ||ξ1 − ξ2 || ≤ max{||z3 − z2 ||, ||z4 − z2 ||} ≤ T ρ + 2ρ. Hence if (for example) n P 2φ max | ∂x∂i ∂x |, where the max is taken over M ρ , using (3.15) we have α= j i,j =1
2
|gradφ(ξ2 ) · u| ≥ |gradφ(ξ1 ) · u| − α||ξ1 − ξ2 || 2ρ − B1 ||z1 − z2 ||3 √ − α(T ρ + 2ρ) ||z1 − z2 || √ 2ρ − B1 (T ρ + 2ρ)3 √ − α(T ρ + 2ρ). ≥ √ T ρ + 2ρ
≥
Hence, if ρ3 and T satisfy for some β > 0, √ 2ρ − B1 (T ρ + 2ρ)3 √ √ − α(T ρ + 2ρ) ≥ β ρ, √ T ρ + 2ρ
(3.17)
for every ρ ∈]0, ρ3 [, then by (3.12) and (3.16) we get B ||z − z ||3 1 1 2 . ||z3 − z2 || − ||z1 − z2 || ≤ √ β ρ In conclusion, it suffices to take T , ρ3 satisfying (3.13) and (3.17), and to put B = t u
B1 β .
We will prove Theorem 3.5 by choosing N in such a way that, on one hand, it is sufficiently large (thus PN,T 6 = ∅), on the other is not too large (see (3.28)), thus if (3.1) does not hold for z ∈ PN,T , then, √ using Lemmas 3.2 and 3.3, every segment line of z is not much shorter than K(T )T ρ, hence z is too long to be in mN,T . This argument works for sufficiently small ρ and T . However, we need a final lemma to prove that in fact Lemmas 3.2 and 3.3 imply this consideration about the length of z. Lemma 3.4. There exists r > 1 such that every (N − 1)-tuple (l1 , . . . , lN −1 ) of real numbers (with N > 1) satisfying, for some α > 0, λ > 0, formulas (3.18), (3.19), (3.20) given by h r 1 i , (3.18) l1 , . . . , lN−1 ∈ 0, 3α |li+1 − li | ≤ α(min{li , li+1 })3 if min{li , li+1 } < λ, i = 1, . . . , N − 2, (3.19) ∃ i = 1, . . . , N − 1 such that li ≥ λ,
(3.20)
also satisfies N−1 X
li ≥ (N − 1)λr −(N−1)αλ . 2
i=1
Proof. Let r > 1 be such that (1 − x1 )x ≥ 1r for every x ≥ 3. Suppose i, i ± 1 ∈ {1, . . . , N − 1}. By (3.19) if li < λ we have li±1 ≥ li − αli3 ≥ li (1 − αλ2 ); if instead
Billiards in Tubular Neighborhoods of Manifolds of Codimension 1
79
li ≥ λ > li±1 we have li±1 ≥ li − αli3 ≥ λ(1 − αλ2 ) because the function l 7→ l − αl 3 h q i 1 is increasing in 0, 3α . In any case li±1 ≥ min{li , λ} (1 − αλ2 ). By (3.18) and (3.20) we deduce αλ2 ≤ 13 , and also, given j = 1, . . . , N − 1, 1
lj ≥ λ(1 − αλ2 )N−1 = λ (1 − αλ2 ) αλ2
(N−1)αλ2
≥ λr −(N−1)αλ , 2
and we immediately get the thesis. u t Theorem 3.5. Suppose M is of class C 3 . Then there exist ρ˜ ∈]0, ρ[ and H > 0 such that for every ρ ∈]0, ρ[, ˜ any two points in Mρ can be connected by a billiard trajectory in Mρ having no more than √Hρ collision points. Proof. Let T > 0 and ρ˜ > 0 be such that
p T < min{T , ρ}, ρ˜ ≤ min{ρ2 (T ), ρ3 }.
(3.21) (3.22)
Given ρ ∈]0, ρ[ ˜ and ζ (1) , ζ (2) ∈ Mρ , suppose N satisfies N≥
L √ , T ρ
N ∈ P(ζ (1) , ζ (2) ).
(3.23) (3.24)
Then we can construct z˜ ∈ PN,T as in Theorem 2.4, thus PN,T 6= ∅. Consider z ∈ mN,T . Note that z is in Mρ by (3.21), (3.22), and Lemma 3.1. We are going to prove that z satisfies (3.1). We will prove in fact that if (3.1) does not hold, then the length of z is strictly greater than that of z˜ , a contradiction. Note that if we assume K(T )T 2 ) , (3.25) 4 √ √ √ we have ||zi+1 −zi || ≤ T ρ +2ρ ≤ 23 T ρ, and also ||zi+1 −zi || ≥ K(T )T ρ −2ρ ≥ √ √ 1 2 K(T )T ρ, if in addition, ||p1 (zi+1 ) − p1 (zi )|| ≥ K(T )T ρ. Thus, if for some i = 0, . . . , N, (3.1) does not hold, by Lemmas 3.2 and 3.3, by putting li = ||zi+1 − zi ||, √ α = √Bρ , λ = 21 K(T )T ρ, we can apply Lemma 3.4 provided ρ˜ < (
27 p 2 B ρT ˜ ≤ 1, 4
(3.26)
and we get 3N (z) ≥
N−1 X i=1
||zi+1 − zi || ≥
B N −1 2 2√ √ K(T )T ρr −(N−1) 4 K(T ) T ρ . 2
We conclude that we can make 3N (z) arbitrarily large; for example we have 3N (z) > 2L,
(3.27)
80
R. Peirone
√ √ provided K(T )T (N − 1) ρ > 8L and BK(T )2 T 2 (N − 1) ρ < inequalities are satisfied if
ln2 lnr ,
and these
ln2 8L √ < (N − 1) ρ < . K(T )T BK(T )T 2 lnr
(3.28)
Now, if T satisfies (3.21) and the inequality between the left and the right side of (3.28), and for this T , ρ˜ satisfies (3.22), (3.25), (3.26), and also p
ρ˜ <
ln2 1 8L , − 2 2 BlnrK(T )T T K(T )
then there exists N satisfying (3.23), (3.24), (3.28), and for this N z satisfies (3.27). On the other hand if we suppose these conditions on T , ρ, ˜ N are satisfied, and use Lemma 2.3 and (3.28) we get s L 2 ρ ) (1 + ) + ρ 2 + (N + 1)ρ ≤ 2L, 3N (˜z) ≤ (N + 1) ( N +1 ρ provided ρ˜ is sufficiently small. In such a case, by using also (3.27) we have 3N (˜z) < 2ln2 , see (3.28). u t 3N (z). The theorem is proved if we take for example H = BlnrK(T )T 2 Acknowledgement. I thank E. De Giorgi1 for useful discussions and suggestions on this subject.
References [1]
Benci, V., Giannoni, F.: Periodic Bounce Trajectories with a low number of Bounce Points. Ann. Inst. Henri Poincaré, Anal. non linéaire, 6 n. 1, 73–93 (1989) [2] Cornfeld, I.P., Fomin, S.V., Sinai, Ya.G.: Ergodic Theory. New York: Springer-Verlag, 1982 [3] Degiovanni, M.: Multiplicity of Solutions for the Bounce Problem. J. Diff. Equations, 54, 414–428 (1984) [4] Gilbarg, D., Trudinger, N.: Elliptic Partial Differential Equations of Second Order. Berlin: SpringerVerlag, 1983 [5] Hirsch, M.: Differential Topology. New York: Springer, 1976 [6] Kozlov, V., Treshchev, D.: Billiards. A Genetic Introduction to the Dynamics of Systems with Impacts. Translations of Math. Monographs No 89, Providence, RI: Amer. Math. Soc., 1991 [7] Markarian, R.: Introduction to the Ergodic Theory of Plane Billiards. In: Dynamical Systems, Pitman Research Notes in Math. Series n. 285, pp. 327–439 [8] Peirone, R.: Bounce Trajectories in Plane Tubular Domains, Atti Accad. Naz. Linc. Rend. Cl. Sc. Fis. Mat. Natur., Ser. 8, Vol. LXXXIII, 39–42 (1989) [9] Penrose, L., Penrose, R.: Puzzles for Christmas. The New Scientist, 25 Dec. 1958, pp. 1580–81 and 1597 [10] Rauch, J.: Illumination of Bounded Domains. Am. Math. Month. 85, 359–361 (1978) [11] Tabachnikov, S.: Billiards. SMF Panoramas et Synthese, No 1, 1995 Communicated by Ya. G. Sinai
1 who died in 1996
Commun. Math. Phys. 207, 81 – 106 (1999)
Communications in
Mathematical Physics
© Springer-Verlag 1999
Remarks on the Time Dependent Periodic Navier–Stokes Flows on a Two-Dimensional Torus Zhi-Min Chen1,? , W. Geraint Price2 1 Department of Mathematics, Tianjin University, Tianjin 300072, P.R. China. E-mail: [email protected] 2 Department of Ship Science, Southampton University, Southampton SO17 1BJ, UK.
E-mail: [email protected] Received: 4 August 1998 / Accepted: 9 April 1999
Abstract: This note is a continuation of the earlier paper [Chen, Z. M., Price, W. G.: Time dependent periodic Navier–Stokes flows on a two-dimensional torus. Commun. Math. Phys. 179, 577–597 (1996)], where the evidences of the occurrence for the time dependent periodic Navier–Stokes flows were found based on a combination of analysis and computation. This investigation is now confirmed via rigorous analysis. The existence of the time dependent periodic Navier–Stokes flows on a two-dimensional torus is proved.
1. Introduction From a general point of view, the occurrence of time dependent periodic flows is the inherent phenomenon to exist in the viscous incompressible fluid motions governed by the Navier–Stokes equations. Especially, Hopf bifurcation or the transition from a steady-state flow to time dependent periodic flows may be considered to be the primary stage in understanding the nature of turbulence (cf. [7,10]). The existence theorem of Hopf bifurcation for general Navier–Stokes equations was established by Joseph and Sattinger [5] under the assumption of the eigenvalue simplicity and eigenvalue transversal crossing conditions for the linearized Navier–Stokes equations. Nevertheless, these conditions are difficult to verify. One may refer to [6] for the comment given in 1975. The possibility of a time periodic motion is well established if certain conditions on the linearization are satisfied. However, they have not yet been realized for a single model. Rigorous mathematical analysis on the existence of time dependent periodic solutions to the Navier–Stokes equations is still in its early stage. For example, a critical Reynolds number can be found for parallel Navier–Stokes motions by studying the OrrSommerfeld equations or the linearization in the space of functions spatially periodic in ? Research partially supported by the National Natural Science Foundation of China.
82
Z.M. Chen, W. G. Price
the horizontal direction (cf. [8]). There still exists a large gap to understand the eigenvalue conditions with respect to the Orr-Sommerfeld equations in solving bifurcation problems. In 1959, Kolmogorov (cf. [1]) gave a model of parallel fluid motions, which are spatially periodic and excited by the external force k 2 (sin ky, 0) for an integer k. This force gives rise to the steady-state flow (sin ky, 0). In connection with this model, the gap is narrowed via an approach due to Meshalkin and Sinai [9] by transforming the OrrSommerfeld equations to algebraic equations. The spectral behavior becomes easier to be examined from the investigation on the algebraic equations. An important extended study on this problem is due to Iudovich [4], where the existence of a sequence of steady-state bifurcation points is obtained. Consider Kolmogorov’s model with the forcing term 4k 2 (sin 2ky, 0). This is formulated by the following Navier–Stokes equations on the two-dimensional torus T 2 = R 2 /(2πZ): ∂t u − 1u + λ(u·∇)u + λ∇p = 4k 2 (sin 2ky, 0), ∇ ·u = 0, Z u dx dy = 0,
(1)
T2
where u = (u1 , u2 ), p and λ represent respectively the velocity, the pressure and the Reynolds number defining the fluid motions. The work made here is an continuation of the earlier investigation (cf. [3]) √ on the occurrence of time dependent periodic flows, which were found to exist in [ 3k] flow invariant subspaces by truncating Eq. (1) reduced to each of these subspaces into an ordinary differential system and by numerical experiments. Here [a] denotes the integer part of a. In this paper we obtain the existence of time dependent periodic solutions of Eq. (1) observed in [3]. To state the results more precisely, we rewrite Eq. (1) in the vorticity form: ∂t ψ = 1ψ − λ1−1 (∂y ψ∂x 1ψ − ∂x ψ∂y 1ψ) − 2k cos 2ky,
Z T2
ψdxdy = 0,
(2)
(−1ψ = ∂y u1 − ∂x u2 ),
and we introduce the following flow invariant subspaces of Eq. (2) given in [3]: H 2 = the Hilbert space {ψ| ψ, 1ψ ∈ L2 (T 2 ; R)}/R, which is a Banach algebra with respect to the function multiplication. ( 2 Hl,k
= ψ ∈ H2 |ψ = +
∞ X n=1
∞ ∞ X X
ηm,n cos(2mlx m=1 n=−∞ ∞ ∞ X X
ξn cos 2nky +
− lx + 2nky + ky) )
ξm,n cos(2mlx + 2nky) ,
m=1 n=−∞
which is a Banach subalgebra of H 2 generated by the three modes cos 2ky, cos(lx − ky) and cos(lx + ky) for an integer vector (l, k).
Periodic Navier–Stokes Flows
83
2 can be formulated as follows: Likewise, a symmetric subspace with respect to Hl,k ( ∞ ∞ X X 2 ˜ ηm,n sin(2mlx − lx + 2nky + ky) Hl,k = ψ ∈ H 2 | ψ =
+
∞ X n=1
m=1 n=−∞ ∞ ∞ X X
ξn cos 2nky +
) ξm,n cos(2mlx + 2nky) ,
m=1 n=−∞
which is a Banach subalgebra of H 2 generated by the three modes cos 2ky, sin(lx − ky) and sin(lx + ky). These subspaces are flow invariant of Eq. (2) in the following sense: 2 ∪ H˜ 2 with l ≥ 0 and k ≥ 1, Eq. (2) Lemma 1.1 ([3, Lemma 2.1]). For every ψˆ ∈ Hl,k l,k admits a unique solution with ψ(0) = ψˆ such that 2 2 ), when ψˆ ∈ Hl,k ψ ∈ C([0, ∞); Hl,k
and
2 2 ), when ψˆ ∈ H˜ l,k . ψ ∈ C([0, ∞); H˜ l,k
We are now in the position to state the existence result on time dependent periodic flows. √ Theorem 1.1. Let the integers k ≥ 1 and l = 1, . . . , [ 3k]. There exist a critical Reynolds number λl,k and a continuous function λl,k, with respect to small > 0 such that Eq. (2) with λ = λl,k, admit two different time dependent periodic solutions ψl,k, and ψ˜ l,k, satisfying 2 2 ψl,k, ∈ C([0, ∞); Hl,k ), ψ˜ l,k, ∈ C([0, ∞); H˜ l,k ),
ψl,k, , ψ˜ l,k, → ψ0 , λl,k, → λl,k as → 0, where ψ0 = −(1/2k) cos 2ky is the basic steady-state solution of Eq. (2). In approaching this result, we use the Hopf bifurcation theorem due to Joseph and Sattinger [5]. For the reader’s convenience, we simplify this theorem concerning Eq. (2) as follows: Theorem 1.2 (Joseph and Sattinger[5]). Assume that the spectral problem, the linerization of Eq. (2) around the steady-state solution ψ0 in H 2 , ρψ = 1ψ − λ1−1 (∂y ψ0 ∂x 1ψ − ∂x ψ0 ∂y 1ψ + ∂y ψ∂x 1ψ0 − ∂x ψ∂y 1ψ0 ) admits an eigenvalue ρ = ρ(λ) subject to the eigenvalue simplicity condition 1 = dim{ψ(λ) ∈ H 2 | (ρ(λ), ψ(λ)) satisfies Eq. (3)} and the eigenvalue transversal crossing condition Imρ(λ0 ) 6 = 0, Reρ(λ0 ) = 0 and Re
dρ(λ0 ) >0 dλ
(3)
84
Z.M. Chen, W. G. Price
for some critical Reynolds number λ0 > 0. Then there exist continuous functions λ and ν for small > 0 such that Eq. (2) with λ = λ admits a solution ψ = ψ ∈ C([0, ∞); C 2 (T 2 ; R)) satisfying ψ (x, y, t) = ψ (x, y, t +
2π ) for small > 0, λ ν
ψ → ψ0 , λ → λ0 and ν → Imρ(λ0 ) as → 0. Here 2π/(λ ν ) is the time period of ψ . In fact, the eigenvalues of Eq. (3) always have multiplicity 2m for some integer m, and the whole space H 2 is not suitable for applying the Hopf bifurcation theorem. However, the simplicity condition can be verified in the above-mentioned flow invariant subspaces, and Theorem 1.2 remains true by reducing Eq. (2) to these subspaces. Consequently, taking Theorem 1.2 and Lemma 1.1 into account, we are led to the proof of the following result on the linear spectral problem presented in Eq. (3). √ Theorem 1.3. Let the integers k ≥ 1 and l = 1, . . . , [ 3k] and let Aλ ψ be the righthand side of Eq. (3). Then the following assertions hold true. (i) The spectral problems ∞ X
Aλ ψ = ρψ, ψ =
n=−∞
2 ξn cos(lx + ky + 2nky) ∈ Hl,k
(4)
2 ξn sin(lx + ky + 2nky) ∈ H˜ l,k
(5)
and ∞ X
Aλ ψ = ρψ, ψ =
n=−∞
admit an eigenvalue ρ = ρl,k (λ) such that lim ρl,k (λ) = −l 2 − k 2 and Imρl,k (λ) < 0, λ > 0.
λ→0
(6)
(ii) ρl,k satisfies the eigenvalue positivity condition lim Reρl,k (λ) > 0, λ > 0.
λ→∞
(iii) The eigenvalue ρl,k is smooth and unique with respect to the spectral problems described in Eqs. (4–5) and satisfies the monotonicity condition Re
dρl,k (λ) > 0, λ > 0. dλ
(7)
Periodic Navier–Stokes Flows
85
2 (resp. H˜ 2 ) contains a flow invariant subspace X (resp. X) ˜ of Eq. (2) such that, (iv) Hl,k l,k for λ > 0,
˜ Aλ ψ = ρl,k (λ)ψ} 1 = dim{ψ ∈ X| Aλ ψ = ρl,k (λ)ψ} = dim{ψ ∈ X| and ˜ Aλ ψ = ρl,k (λ)ψ}. 2 = dim{ψ ∈ X ∪ X| For convenience, we may suppose that H 2 is a complex space when it is concerned with the spectral problem. This result implies the existence of critical Reynolds number λl,k so that Reρl,k (λl,k ) = 0. Thus it follows from Theorem 1.2 that a pair of time dependent periodic solutions branch off the bifurcation point (λl,k , ψ0 ). These two different solutions are contained ˜ This gives Theorem 1.1. respectively in the spaces C([0, ∞); X) and C([0, ∞); X). To show Theorem 1.3, it is necessary to formulate the spectral problem to algebraic equations, which can be deduced essentially by using the method of [9]. Now we state this result as follows: √ Theorem 1.4. Let λ > 0, k ≥ 1, l = 1, . . . , [ 3k] and Z denote the integer set. Then Eqs. (4–5) admit an eigenvalue ρ = ρl,k (λ) satisfying Eq. (6), and both Eq. (4) and Eq. (5) are equivalent to the following coupled set of algebraic equations 2βn (βn + ρ)ξn + λαn−1 ξn−1 − λαn+1 ξn+1 = 0, n ∈ Z,
(8)
where βn = βn (l, k) = l 2 + (2nk + k)2 , αn = αn (l, k) = l(βn − 4k 2 ), and {ξn }n∈Z is defined by the following conditions: (−1)n ξn = cγ1 · · · γn ξ0 = c, ξ−1 = ic, (−1)n−1 ξ−n = iξn−1 , γn =
α0 , n ≥ 1, αn (9) n ≥ 2,
1 , 1 2βn (βn + ρ) + 1 2βn+1 (βn+1 + ρ) λαn + λαn+1 .. .
and, in particular, 2β0 (β0 + ρ) 1 1 = i. = + 2 2 1 2β (β + ρ) γ0 λl(l − 3k ) 1 1 + 1 2β2 (β2 + ρ) λα1 + λα2 .. . √ Here c is an arbitrary complex constant and i = −1.
(10)
86
Z.M. Chen, W. G. Price
In connection with Eq. (4), this result follows from the proof of [3, Theorem 3.1]. It is readily seen that Eqs. (4) and (5) are equivalent. Thus this result with respect to another spectral problem Eq. (5) can be shown in the same way. The purpose of this paper is to prove Theorem 1.3. We see Assertion (i) is implied in Theorem 1.4. Thus it remains to show Assertions (ii)–(iv). The proof for these assertions is a bit lengthy, and is to be completed by combining the examinations of the remaining sections. Section 2 contains preliminary lemmas devoted to the estimation of the algebraic equation, Eq. (10). Assertion (ii) is to be verified in Sect. 3 based on the study in Sect. 2. Assertion (iii) is to be shown in Sect. 4 via estimation on the coupled set of algebraic equations, Eq. (8). With the help of Lemma 1.1 and Assertions (i)–(iii), Assertion (iv) becomes an easy problem, which will be shown in Sect. 5. It should be noted that Assertion (ii) was stated in [3, Theorem 3.1] without detailed proof, since the authors considered it as an easy problem then. In fact, the critical Reynolds numbers λl,k vary sharply and are unbounded when k and −3k 2 + l 2 increase. This phenomenon can be observed easily via numerical experiments. Hence it is difficult to show the existence of all Hopf bifurcation values λl,k , and, especially, to show Assertion (ii) as k increases. The difficulty arises from the fact that ρl,k (λ) is complex and disjoint to the real axis for all λ > 0. To overcome the above-mentioned difficulty, we present some technical analysis in estimating the algebriac equations, Eqs. (8–10). Let us mention that, among other things, Theorem 1.1 has been extended to a threedimensional Navier–Stokes model (cf. [2]) by developing the approach of this paper, and extra instabilities are found therein. 2. Prelimary Lemmas The proof of Assertion (ii) is based on the following technical lemmas in estimating Eq. (10) when λ is close to infinity. In what follows, let ρ(λ) = ρl,k (λ) and (l, k) be given in Theorem 1.4, and let µ(λ) > −β0 and ν(λ) < 0 be the real functions so that µ(λ) + iν(λ) = ρ(λ). Let us begin with an elementary lemma. Lemma 2.1. The following holds −
ν(λ) 3 < < 0 for λ > 0. 2 λl
Proof. Recalling from Theorem 1.4 that γn (λ) =
1 , n≥0 1 2βn (βn + ρ(λ)) + 1 2βn+1 (βn+1 + ρ(λ)) λαn + λαn+1 .. .
and ν(λ) < 0, we see that |γn (λ)| ≤
λαn λαn ≤ → 0 as n → ∞, 2βn (βn + µ(λ)) 2βn (βn − β0 ) Imγn+1 (λ) =
2βn |ν(λ)| 1 + Im , n ≥ 0, αn λ γn (λ)
(11)
Periodic Navier–Stokes Flows
87
≥
1 2|ν(λ)| + Im , n ≥ 1, lλ γn (λ)
and, by Eq. (10), Imγ1 (λ) = 1 −
2(l 2 + k 2 )|ν(λ)| 2|ν(λ)| 2β0 |ν(λ)| = 1 − <1− . 2 2 2 2 l(3k − l )λ lλ(3k − l ) 3lλ
If ν(λ)/(λl) ≤ −3/2 for some λ > 0, we see that Imγ1 (λ) < 0, and thus Imγn (λ) > 3 for n ≥ 2. This contradicts Eq. (11), and completes the proof. Without loss of generality, the limits of ν(λ)/(λl) and µ(λ) may be supposed to exist as λ tends to infinity. Otherwise, we may reduce our study in this section to a sequence {λn > 0| n ≥ 1} instead of the half line {λ| λ > 0}. Now we adopt the notations: 2ν(λ) ≤ 0, λl µ0 = lim µ(λ) ≥ −β0 , ν0 = lim
λ→∞ λ→∞
κn + iθn κn (λ) + iθn (λ) = λ λ =
1 , n ≥ 0, 1 2βn (βn + ρ(λ)) + 2βn+1 (βn+1 + ρ(λ)) 1 αn λ + αn+1 λ .. .
τn + iσn = τn (λ) + iσn (λ) =
(12)
λ2 κn (λ) + iθn (λ)
2βn (βn + ρ(λ)) + κn+1 (λ) + iθn+1 (λ), n ≥ 0, (13) αn where κn , θn , τn and σn are real functions. Obviously, κn (λ) > 0, τn (λ) > 0 for n ≥ 1 since αn > 0 and βn + Reρ > 0 for t n ≥ 1, and Lemma 2.1 implies that ν0 is finite. u =
Lemma 2.2. Assume that µ0 < ∞, and there exists an integer n ≥ 1 so that limλ→∞ κn (λ) < ∞. Then (i)
σn+1 (λ) βn ν0 6 = 1, lim βn − 4k 2 λ→∞ λ lim κn (λ)
λ→∞
σn+1 (λ) 2 2βn+1 (βn+1 + µ0 ) 2βn (βn + µ0 ) lim + + lim κn+2 (λ) λ→∞ λ→∞ αn λ αn+1 = 2 βn ν0 σn+1 (λ) −1 lim 2 βn − 4k λ→∞ λ
holds whenever lim
λ→∞
and
|τn+1 (λ) + iσn+1 (λ)| < ∞, λ
88
Z.M. Chen, W. G. Price
(ii) 2βn (βn + µ0 ) + lim κn+1 (λ) λ→∞ αn lim κn (λ) = 2 λ→∞ βn ν0 βn − 4k 2 holds whenever lim
λ→∞
|τn+1 (λ) + iσn+1 (λ)| = ∞. λ
Proof. (i) Note that κn (λ) = Re
=
λ
1 2βn (βn + ρ(λ)) + τn+1 (λ) + iσn+1 (λ) αn λ λ τn+1 λ2 2βn (βn + µ(λ)) + 2 2 αn τn+1 + σn+1 !2 !2 2βn ν(λ) τn+1 λ σn+1 λ 2βn (βn + µ(λ)) + + 2 − 2 2 2 αn λ αn λ τn+1 + σn+1 τn+1 + σn+1 1
=
2βn ν(λ) σn+1 λ − 2 2 αn λ τn+1 + σn+1
!2 .
τn+1 2βn (βn + µ(λ)) + 2 + 2 2 αn λ τn+1 λ2 2βn (βn + µ(λ)) τn+1 + σn+1 + 2 2 αn τn+1 + σn+1 Hence τn+1 2βn (βn + µ(λ)) 2 αn + σn+1
2 τn+1
2βn (βn + µ(λ)) 1 − − κn (λ) αn λ2 τn+1 λ2 2βn (βn + µ(λ)) + 2 2 αn τn+1 + σn+1 !2 2 λ2 τn+1 σn+1 λ 2βn ν(λ) − 2 + 2 2 )2 2 αn λ (τn+1 + σn+1 τn+1 + σn+1 = τn+1 λ2 2βn (βn + µ(λ)) + 2 2 αn τn+1 + σn+1 2 2 2 τn+1 2βn ν(λ) 2βn ν(λ) 1 + σn+1 − λ 2 2 2 2 αn λ αn λ τn+1 + σn+1 τn+1 + σn+1 = τn+1 λ2 2βn (βn + µ(λ)) + 2 2 αn τn+1 + σn+1
Periodic Navier–Stokes Flows
2 2βn ν(λ) σn+1 −1 + λ2 αn λ λ = , 2 2 + σn+1 2βn (βn + µ(λ)) τn+1 + τn+1 αn λ2 which is bounded by 2 1 2βn ν(λ) σn+1 2βn ν(λ) 2 τn+1 −1 . + αn λ λ2 τn+1 αn λ λ
2βn ν(λ) αn λ
2
89 2 τn+1
(14)
If limλ→∞ τn+1 (λ) = ∞, this term vanishes as λ → ∞. This implies lim κn (λ) = ∞,
λ→∞
which leads to a contradiction. Thus τn+1 (λ) is bounded. Passing to the limit as λ → ∞ in Eq. (14), therefore, yields 1 λ→∞ κn (λ) lim
2 σn+1 (λ) βn ν0 lim −1 βn − 4k 2 λ→∞ λ , = 2 (λ) σn+1 2βn+1 (βn+1 + µ0 ) 2βn (βn + µ0 ) lim + + lim κn+2 (λ) λ→∞ λ→∞ αn λ2 αn+1
and gives (i). (ii) Noticing that lim
λ→∞
|κn+1 (λ) + iθn+1 (λ)| = λ
1 = 0, |τn+1 (λ) + iσn+1 (λ)| lim λ→∞ λ
and λ 2βn (βn + ρ(λ)) κn+1 (λ) + iθn+1 (λ) + αn λ λ 1 = , 2βn ν(λ) θn+1 (λ) 2 + 2βn (βn + µ(λ)) κn+1 (λ) αn λ λ + + 2βn (βn + µ(λ)) αn λ2 λ2 + κn+1 (λ) αn we find that limλ→∞ κn (λ) = ∞ whenever limλ→∞ κn+1 (λ) = ∞, and κn (λ) = Re
2βn (βn + µ0 ) + lim κn+1 (λ) λ→∞ αn lim κn (λ) = β ν λ→∞ n 0 ( )2 βn − 4k 2 whenever limλ→∞ κn+1 (λ) < ∞. This gives (ii) and thus completes the proof. With the use of Lemmas 2.1 and 2.2, we can now estimate Eq. (10) as λ tends to infinity. u t
90
Z.M. Chen, W. G. Price
Lemma 2.3. Assume µ0 < ∞. Then ∞
β0 (β0 + µ0 ) X 2 βn (βn + µ0 ) ≥ an , 3k 2 − l 2 βn − 4k 2
(15)
n=1
where, and in what follows, a0 = 1, a1 = 1 +
an = an (l, k) = (−1)n
(l 2 + k 2 )ν0 , 3k 2 − l 2
βn−1 ν0 an−1 + an−2 for n ≥ 2. βn−1 − 4k 2
(16)
Proof. For an integer n > 1, we recall from Eqs. (10,12,13) that 2β0 (β0 + ρ(λ)) κ1 (λ) + iθ1 (λ) =i+ λ λl(3k 2 − l 2 ) =
1 2β1 (β1 + ρ(λ)) + λα1 .. .+
1
1 1 2βn (βn + ρ(λ)) + τ + iσn+1 λαn n+1 λ 1 . = 1 2β1 (β1 + ρ(λ)) + λα1 1 .. .+ 2βn (βn + ρ(λ)) κn+1 + iθn+1 + λαn λ Since µ0 = limλ→∞ µ(λ) < ∞, passing to the limit as λ → ∞ yields lim
λ→∞
τn+1 (λ) + iσn+1 (λ) = (−1)n+1 i if ν0 = 0, λ
lim κ1 (λ) =
λ→∞
1+
β0 ν0 = 3k 2 − l 2
2β0 (β0 + µ0 ) , l(3k 2 − l 2 )
1 −β1 ν0 + β1 − 4k 2 .. .+
(17)
(18)
(19)
1 1 (−1)n−1 βn−1 ν0 βn−1 − 4k 2
Periodic Navier–Stokes Flows
91
if limλ→∞ |τn+1 (λ) + iσn+1 (λ)|/λ = 0 and ν0 < 0, 1+
β 0 ν0 = 3k 2 − l 2
1 −β1 ν0 + β1 − 4k 2 .. .+
(20)
1 1 (−1)n βn ν0 βn − 4k 2
if limλ→∞ |τn+1 (λ) + iσn+1 (λ)|/λ = ∞ and ν0 < 0, τn+1 (λ) = 0, λ→∞ λ lim
1+
β 0 ν0 = 3k 2 − l 2
1 −β1 ν0 + β1 − 4k 2 .. .+
(21)
1 1 (−1)n βn ν0 βn − 4k 2
+
(−1)n+1 σn+1 (λ) lim λ→∞ λ
if 0 < limλ→∞ |τn+1 (λ) + iσn+1 (λ)|/λ < ∞ and ν0 < 0. We now start to check the desired inequality based on these equations. If ν0 = 0, Eq. (17) together with Lemma 2.2 implies lim κn (λ) =
λ→∞
2βn (βn + µ0 ) 2βn+1 (βn+1 + µ0 ) + + lim κn+2 (λ). λ→∞ αn αn+1
Combining this and Eq. (18) yields ∞
X 2βn (βn + µ0 ) 2β0 (β0 + µ0 ) = lim κ (λ) ≥ . 1 λ→∞ l(3k 2 − l 2 ) αn n=1
If ν0 < 0, it follows from Eq. (18) and Lemma 2.2 that σ2 (λ) β1 ν0 − 1 6 = 0. lim β1 − 4k 2 λ→∞ λ Addtionally, Eq. (18) and Lemma 2.2 yield lim κ3 (λ) < ∞, when lim
λ→∞
λ→∞
|τ2 (λ) + σ2 (λ)| < ∞, λ
and
|τ2 (λ) + σ2 (λ)| = ∞. λ→∞ λ→∞ λ Thus for limλ→∞ |τ2 (λ) + σ2 (λ)|/λ < ∞, from Eqs. (19) and (21) we find that lim κ2 (λ) < ∞, when lim
β1 ν0 a2 = a2 = 1 + a1 = a0 β1 − 4k 2
1 6= 0 β1 ν0 σ2 (λ) − lim + 1 β1 − 4k 2 λ→∞ λ
92
Z.M. Chen, W. G. Price
and
a1 β0 ν0 σ2 (λ) , = a1 = 1 + 2 = a2 lim λ→∞ a0 3k − l 2 λ and hence, by Eq. (18) and Lemma 2.2, we have 2β0 (β0 + µ0 ) = lim κ1 (λ) λ→∞ l(3k 2 − l 2 )
σ2 (λ) 2 2β2 (β2 + µ0 ) 2β1 (β1 + µ0 ) lim + + lim κ3 (λ) λ→∞ λ→∞ α1 λ α2 = 2 β1 ν0 σ2 (λ) −1 lim β1 − 4k 2 λ→∞ λ 2β1 (β1 + µ0 ) 2β2 (β2 + µ0 ) = a12 + a22 + a22 lim κ3 (λ). λ→∞ α1 α2
For another case, limλ→∞ |τ2 (λ) + iσ2 (λ)|/λ = ∞, from Eq. (20) and the fact ν0 < 0 it follows that β0 ν0 1 a1 = a1 = 1 + 2 = 6= 0. β1 ν0 a0 3k − l 2 − β1 − 4k 2 Using Eq. (18) and Lemma 2.2 again, we have 2β0 (β0 + µ0 ) = lim κ1 (λ) λ→∞ l(3k 2 − l 2 ) 2β1 (β1 + µ0 ) + lim κ2 (λ) λ→∞ α1 = 2 β1 ν0 β1 − 4k 2 2β1 (β1 + µ0 ) = a12 + a12 lim κ2 (λ). λ→∞ β1 − 4k 2 To estimate κn for all n, we may suppose that an−1 6 = 0 and lim κn (λ) < ∞ λ→∞
for an integer n ≥ 2. Applying an elementary calculation to Eqs. (19-21), we have σn+1 lim an λ→∞ λ = , β ν σn+1 an−1 n 0 n+1 (−1)n lim + (−1) βn − 4k 2 λ→∞ λ 2 an+1
1 6= 0, β σn+1 ν n 0 n+1 (−1)n lim + (−1) βn − 4k 2 λ→∞ λ whenever limλ→∞ |τn+1 + iσn+1 |/λ < ∞, and 2 an−1
=
an = an−1
1 (−1)n
βn ν0 βn − 4k 2
6 = 0,
Periodic Navier–Stokes Flows
93
2 whenever limλ→∞ |τn+1 + iσn+1 |/λ = ∞. Thus an2 + an+1 > 0. By Lemma 2.2, we have 2 lim κn (λ) = an2 an−1 λ→∞
2βn (βn + µ0 ) 2 2βn+1 (βn+1 + µ0 ) 2 + an+1 + an+1 lim κn+2 (λ) λ→∞ αn αn+1
when an+1 6 = 0, and 2 lim κn (λ) = an2 an−1 λ→∞
2βn (βn + µ0 ) + an2 lim κn+1 (λ) when an+1 = 0. λ→∞ αn
Hence the desired assertion follows by induction. The proof is complete. u t Assertion (ii) of Theorem 1.3 will be deduced in the next section via careful analysis on Eq. (15). Of course, this essentially depends of the estimation on an defined by Eq. (16). For convenience, we first set r=
βn k2 , bn = , n ≥ 1, 2 −l βn − 4k 2
(22)
3k 2
which gives bn =
(2n + 1)2 r + 3r − 1 > 0, a1 = 1 + (4r − 1)ν0 , (2n + 1)2 r − r − 1
(23)
(24) an = (−1)n bn−1 v0 an−1 + an−2 , n ≥ 2. √ Furthermore, the choice of the integer l with l = 1, . . . , [ 3k] is now implied in the condition 1 < r ≤ k2 . 3 With these observations in mind, we can now estimate an when 2 < r ≤ k 2 . Lemma 2.4. Let 2 < r ≤ k 2 , a1 < 0 and n ≥ 1. Then 2 1 −|a1 | ≤ a2n−1 ≤ −|a1 | + (n − )|ν0 | + |a1 |(n2 + n)ν02 , 3 2 provided that a2n0 ≥ 0 for n0 = 0, . . . , n − 1. Proof. Let us begin with the proof of the following estimates: −|a1 | ≤ a2n−1 ≤ −|a1 | + |ν0 |
n−1 X
b2m + |a1 |ν02
b2m0 −1 b2m ,
(25)
m=1 m0 =1
m=1
a2n ≤ 1 + |a1 ||ν0 |
m n−1 X X
n X m=1
b2m−1 .
(26)
94
Z.M. Chen, W. G. Price
We argue by induction. When n = 1, these estimates are obviously valid. Assume that Eqs. (25–26) hold true for any n0 < n. By (24), a2n−1 = a2n−3 + a2n−2 b2n−2 |ν0 | ≤ − |a1 | + |ν0 |
n−2 X
b2m + |a1 |ν02
m n−2 X X
b2m0 −1 b2m
m=1 m0 =1
m=1
+ |ν0 |(1 + |a1 ||ν0 |
n−1 X
b2m−1 )b2n−2
m=1
= − |a1 | + |ν0 |
n−1 X
b2m + |a1 |ν02
m n−1 X X
b2m0 −1 b2m ,
m=1 m0 =1
m=1
a2n−1 = a2n−3 + a2n−2 b2n−2 |ν0 | > a2n−3 ≥ −|a1 |, since a2n−2 ≥ 0, a2n = a2n−2 − a2n−1 b2n−1 |ν0 | n−1 X
≤ 1 + |a1 ||ν0 | = 1 + |a1 ||ν0 |
m=1 n X
b2m−1 + |a1 ||ν0 |b2n−1 b2m−1 .
m=1
We thus obtain Eqs. (25–26). Furthermore, we note that n−1 X
b2m
m=1
n−1 X (4m + 1)2 r + 3r − 1 = (4m + 1)2 r − r − 1 m=1
n−1 X
4 (4m + 1)2 − 2 m=1 ∞ X 1 2 1 − =n− , ≤ n−1+ 4m − 1 4(m + 1) − 1 3
≤ n−1+
m=1
n−1 X m=1
b2m
m X
b
2m0 −1
=
m0 =1
n−1 X
b2m
m=1
≤
n−1 X
b2m
=
m=1
≤
n−1 X m=1
m =1 m X m0 =1
m=1 n−1 X
m X (4m0 − 1)2 r + 3r − 1 (4m0 − 1)2 r − r − 1 0
b2m
(4m0 − 1)2 + 2 (4m0 − 1)2 − 2
m X 4 4 m+ + 0 7 (4m − 1)2 − 2 0
m+
4 b2m 5
m =2
!
Periodic Navier–Stokes Flows
95
≤
X n−1 2 n(n − 1) 4 4m + n− + 2 5 3 (4m + 1)2 − 2 m=1
≤
n(n − 1) 4 + n+ 2 5
n−1 X m=1
1 1 ≤ (n2 + n), 4m + 1 2
and thus the desired assertion follows from Eq.(25). The proof is complete. u t Lemma 2.5. Let 2 < r ≤ k 2 . Then a2n > 1 −
n 1 , when 0 > a1 ≥ −1 and 1 ≤ n ≤ (4r − 1), 4r − 1 2
a2n > 1, when − 1 > a1 > −
1√ 3√ 4r − 1 and 1 ≤ n ≤ 4r − 1. 2 4
(27)
(28)
Proof. First, we prove Eq. (27). In order to use Lemma 2.4, we note a0 = 1 > 0 and a2 = 1 + b1 |a1 ν0 | > 1. Thus we may suppose that n ≥ 2 is an integer satisfying 2n ≤ 4r − 1 and a2n0 ≥ 0 for n0 = 1, . . . , n − 1. Recalling that a1 = 1 + (4r − 1)ν0 , we have for m = 1, . . . , n, m 1 |a1 |(|a1 | + 1) 1 m|a1 ||ν0 | = m ≤ ≤ , 2 2(4r − 1) 4r − 1 2 and so, by Lemma 2.4, 1 a2m−1 ≤ −|a1 | + m|ν0 | + |a1 |m2 ν02 . 2
(29)
On the other hand, we see that 2|a1 | + 1 z0 = √ 1 + |a1 |(2|a1 | + 1) + 1 is the positive solution to the algebraic equation 1 1 −|a1 | + z + |a1 |z2 = , 2 2 and the condition 2m ≤ 4r − 1 becomes m|ν0 | = m
1 + |a1 | 2|a1 | + 1 1 + |a1 | ≤ ≤√ = z0 . 4r − 1 2 1 + |a1 |(2|a1 | + 1) + 1
Combining this with Eq. (29) gives a2m−1 ≤ 1/2. This together with Eq. (24) yields, by induction, a2m = a2m−2 − a2m−1 |ν0 |b2m−1 1 ≥ a2m−2 − |ν0 |b2m−1 2 m X 1 ≥ a2 − |ν0 | b2m0 −1 2 0 m =2
96
Z.M. Chen, W. G. Price m X 1 = 1 + |a1 ν0 |b1 − |ν0 | b2m0 −1 2 0 m =2
m X 1 (4m0 − 1)2 + 2 ≥ 1 − |ν0 | 2 (4m0 − 1)2 − 2 0 m =2
1 ≥ 1 − m|ν0 | 2 m m 1 + |a1 | >1− >0 = 1− 2 4r − 1 4r − 1 for m = 1, . . . , n. Thus applying this argument repeatedly, we obtain Eq. (27). √ Next, we prove Eq. (28) when −1 ≥ a1 ≥ −3/2 and n ≤ (3/4) 4r − 1. Similar√to the proof of Eq. (27), we may suppose that n ≥ 2 is an integer such that n ≤ (3/4) 4r − 1 and a2n0 ≥ 0 for n0 = 1, . . . , n − 1. Since r > 2, we see that for m = 1, . . . , n, m|a1 |(|a1 | + 1) 1 m|a1 ||ν0 | = 2 2(4r − 1) 3m 2 3m(|a1 | + 1) ≤ √ ≤ . ≤ √ 3 10 4r − 1 4 4r − 1 This together with Lemma 2.4 implies 1 a2m−1 ≤ −|a1 | + m|ν0 | + |a1 |m2 ν02 . 2
(30)
Additionally, observing 3 3√ 4r − 1, −1 ≥ a1 ≥ − , r > 2, and m ≤ 2 4 we have m|ν0 | = m
|a1 | + 1 3(|a1 | + 1) ≤ √ ≤ 4r − 1 4 4r − 1
q 1 + 2a12 − 1 |a1 |
,
for which we denote by z0 the positive solution of the algebraic equation 1 −|a1 | + z + |a1 |z2 = 0. 2 Consequently, from Eq. (30) it follows that a2m−1 ≤ 0, and so, by Eq. (24), a2m = a2m−2 + |a2m−1 ||ν0 |b2m−1 ≥ a2m−2 . By induction, this yields a2m ≥ a2 = 1 + |a1 ||ν0 |b1 > 1 for m = 1, . . . , n. Thus we reach Eq. (28) with −1 ≥ a1 ≥ −3/2 by following the above argument repeatedly.
Periodic Navier–Stokes Flows
97
√ Finally, we prove Eq. (28) with −3/2 > a1 > − 4r − 1/2. We see |a1 | + 1 1 |a1 |(|a1 | + 1) 1 |a1 ||ν0 | = ≤ √ ≤ . 2 2(4r − 1) 4 4 4r − 1 Combining this with Lemma 2.4 gives 5 1 a2m−1 ≤ −|a1 | + m|ν0 | + |a1 |m2 ν02 , m = 1, . . . , n, 4 2 where the integer n is supposed to satisfy the condition a2n0 ≥ 0 for n0 = 1, . . . , n − 1 and n <
3√ 4r − 1. 4
Hence a2m−1 ≤ 0 is valid if the inequality 5 1 −|a1 | + m|ν0 | + |a1 |m2 ν02 ≤ 0 4 2 holds or m|ν0 | ≤
q (5/4)2 + 2a12 − 5/4 |a1 |
.
Indeed, we see m|ν0 | =
3(|a1 | + 1) 3(|a1 | + 1) m(|a1 | + 1) ≤ √ ≤ ≤ 4r − 1 8|a1 | 4 4r − 1
q (5/4)2 + 2a12 − 5/4 |a1 |
,
where use is made of the condition m≤
3 1√ 3√ 4r − 1 and − ≥ a1 ≥ − 4r − 1. 4 2 2
Hence, by Eq. (24) and the fact a2m−1 ≤ 0, a2m = a2m−2 + |a2m−1 ||ν0 |b2m−1 > a2m−2 > a2 > 1, m = 1, . . . , n. By induction, this implies the desired assertion. The proof is complete. u t 3. Proof of Theorem 1.3 (ii) With the aid of the lemmas in the preceding section, we can now carry out the proof of Assertion (ii), that is, µ0 = limλ→∞ Reρ(λ) > 0. We argue by contradiction. If µ0 ≤ 0 holds, it follows from Lemma 2.3 that ∞
∞
n=1
n=1
X a2β 2 X a 2 (l 2 + (2nk + k)2 )2 β02 (l 2 + k 2 )2 n n n = ≥ = . 3k 2 − l 2 3k 2 − l 2 βn − 4k 2 l 2 + (2nk + k)2 − 4k 2
98
Z.M. Chen, W. G. Price
This becomes, by Eqs. (22–23), ∞ X ((2n + 1)2 r + 3r − 1)2 2 a (4r − 1) ≥ (2n + 1)2 r − r − 1 n 2
(31)
n=1
(12r − 1)2 2 (28r − 1)2 2 (52r − 1)2 2 a + a + a + ··· , 8r − 1 1 24r − 1 2 48r − 1 3 √ and the condition l = 1, . . . , [ 3k] is implied in the condition 1/3 < r ≤ k 2 . Consequently, it suffices to give the contradiction that the right-hand side of Eq. (31) is strictly greater than (4r − 1)2 for 1/3 < r ≤ k 2 . Indeed, this will be achieved in the following by considering r in the interval (1/3, k 2 ] through four steps. =
Step 1. The case 1/3 < r ≤ 3/8. First, when a2 ≤ 0, we see that a1 > 0 and thus a2 = 1 + (1 − (4r − 1)|ν0 |)
12r − 1 9 ν0 ≥ 1 + (1 + (4r − 1)ν0 )ν0 . 8r − 1 5
This gives ν0 >
−3 −
√ √ 1 9 − 20(4r − 1) 21 and a1 = 1 + (4r − 1)ν0 > − , 6(4r − 1) 2 18
which together with Eq. (31) leads to the contradiction √ !2 1 3 3 9 21 2 − (12r − 1) > (4r − 1)2 . > (4r − 1) ≥ (12r − 1)a1 > (12r − 1) 2 2 2 18 100 2
Next, when a2 ≥ 0 and a1 ≤ 0, we see that a2 ≥ 1. Hence, by Eq. (31), (4r − 1)2 ≥
7 7 (28r − 1)a22 > (28r − 1) > (4r − 1)2 . 6 6
Finally, when a2 ≥ 0 and a1 > 0, we see that a32 = (|ν0 |b2 a2 + a1 )2 > a12 , and, by Eq. (31), (4r − 1)2 > This yields
3 13 (12r − 1)a12 + (52r − 1)a32 > (74r − 3)a12 . 2 12 4r − 1 , a1 = 1 + (4r − 1)ν0 ≤ √ 74r − 3
that is, ν0 ≤ ν00 with ν00 = −
1 1 +√ . 4r − 1 74r − 3
(32)
Periodic Navier–Stokes Flows
99
Consequently, we readily see that 12r − 1 ν0 8r − 1 12r − 1 0 ≥ 1 + (1 + (4r − 1)ν00 ) ν 8r − 1 0 12r − 1 1 4r − 1 ≥ 1− − √ 8r − 1 74r − 3 74r − 3 9 , ≥ 1− √ 5 74r − 3
a2 = 1 + (1 + (4r − 1)ν0 )
(33)
and hence, by Eq. (31), 7 (28r − 1)a22 6 2 9 > 1− √ (28r − 1) 5 74r − 3 2 > (28r − 1) > (4r − 1)2 . 5
(4r − 1)2 >
Thus µ0 > 0 holds true in the case 1/3 < r ≤ 3/8. Step 2. The case 3/8 < r ≤ 2. If a1 ≤ 0, we see a2 > 1. Thus Eq. (32) follows from Eq. (31). If a1 > 0, we also have Eq. (33), or 1 4r − 1 12r − 1 , when r > 3/8. − a2 > 1 − √ 8r − 1 74r − 3 74r − 3 This gives 7 3 3 , when < r < , a2 > 1 − √ 8 2 4 74r − 3 17 5 1 17 1 3 4r − 1 >1− , when ≤ r ≤ 2. − a2 > 1 − √ √ − 11 66 18 2 74r − 3 74r − 3 3 Taking this observation and Eq. (31) into account, we have (4r − 1)2 >
7 49 3 (28r − 1)a22 > (28r − 1) > (4r − 1)2 , when ≤ r < 1, 6 100 8
(4r − 1)2 >
7 7 3 (28r − 1)a22 > (28r − 1) > (4r − 1)2 , when 1 ≤ r < , 6 10 2
(4r − 1)2 >
7 99 3 (28r − 1)a22 > (28r − 1) > (4r − 1)2 , when ≤ r ≤ 2. 6 100 2
Hence µ0 > 0 holds true for 3/8 < r ≤ 2.
100
Z.M. Chen, W. G. Price
Step 3. The case a1 ≥ 0 and 2 < r ≤ k 2 . To begin with, we see that if there exists an integer n satisfying a2n ≥ 1 −
n 4r − 1 4r − 1 and ≥n> − 1. 4r − 1 2 2
(34)
Since r ≥ 2, Eq. (31) gives 2 ((4n + 1)2 r + 3r − 1) ≥ (4r − 1)2 ≥ a2n
1 (2(4r − 1) − 3)2 2 > (4r − 1)2 . 4
This leads to a contradiction and yields the assertion µ0 > 0. Thus it remains to prove Eq. (34). Let us note that 0 ≤ a1 = 1 − (4r − 1)|ν0 | ≤ 1 and 1 ≥ a2 = 1 − (1 + (4r − 1)ν0 )
12r − 1 |ν0 | 8r − 1
1 12r − 1 8r − 1 4(4r − 1) 2 > 0. ≥ 1− 5(4r − 1)
≥ 1−
By induction, for an integer n ≤ 4r − 1, if the estimates holds true 0 ≤ a2n−3 ≤ 1 − (4r − 1)|ν0 | +
n−2 X (4m + 1)2 r + 3r − 1 |ν0 | ≤ 1, (4m + 1)2 r − r − 1
(35)
m=1
1 ≥ a2n−2
1 ≥1− 4r − 1
n−2
2 X (4m + 3)2 r + 3r − 1 + 5 (4m + 3)2 r − r − 1
! ≥ 0,
(36)
m=1
we have (4n − 3)2 r + 3r − 1 |ν0 |a2n−2 ≥ a2n−3 ≥ 0, (4n − 3)2 r − r − 1 (4n − 3)2 r + 3r − 1 ≤ a2n−3 + |ν0 | (4n − 3)2 r − r − 1 n−1 X (4m + 1)2 r + 3r − 1 |ν0 | ≤ 1 − (4r − 1)|ν0 | + (4m + 1)2 r − r − 1
a2n−1 = a2n−3 + a2n−1
m=1
n−1 X (4m + 1)2 + 2 (4m + 1)2 − 2 m=1 ∞ X 1 1 − |ν0 | ≤ 1 − (4r − 1)|ν0 | + (n − 1)|ν0 | + 4m − 1 4(m + 1) − 1
≤ 1 − (4r − 1)|ν0 | + |ν0 |
≤ 1 − (4r − 1)|ν0 | + n|ν0 | ≤ 1,
m=1
Periodic Navier–Stokes Flows
101
and (4n − 1)2 r + 3r − 1 |ν0 |a2n−1 ≤ a2n−2 ≤ 1, (4n − 1)2 r − r − 1 (4n − 1)2 r + 3r − 1 |ν0 | a2n−2 − (4n − 1)2 r − r − 1 ! n−2 2 X (4m + 3)2 r + 3r − 1 (4n − 1)2 r + 3r − 1 1 + − |ν | 1− 0 4r − 1 5 (4m + 3)2 r − r − 1 (4n − 1)2 r − r − 1 m=1 ! n−1 2 X (4m + 3)2 r + 3r − 1 1 + 1− 4r − 1 5 (4m + 3)2 r − r − 1 m=1 ! n−1 3 X 1 4 n− + 1− 4r − 1 5 (4m + 3)2 − 2 m=1 n ≥ 0. 1− 4r − 1
a2n = a2n−2 − a2n ≥ ≥ ≥ ≥ ≥
This shows Eqs. (35–36) with n − 1 replaced by n. In particular, we obtain the bound a2n > 1 −
n , whenever n ≤ 4r − 1, 4r − 1
which gives Eq. (34) by choosing the integer n0 such that (4r − 1)/2 − 1 < n0 ≤ (4r − 1)/2. Step 4. The case a1 < 0 and 2 < r ≤ k 2 . When 0 ≥ a1 ≥ −1, we may choose an integer n0 so that 2n0 < 4r − 1 ≤ 2n0 + 2. By Lemma 2.5 and Eq. (31), we have a2n0 > 1/2 and 2 (4n0 + 1)2 r (4r − 1)2 ≥ a2n 0 r ≥ (2(4r − 1) − 3)2 > (4r − 1)2 . 4 √ When −1 ≥ a1 > −(1/2) 4r − 1, we may choose an integer n0 so that
n0 ≤
3√ 4r − 1 < n0 + 1. 4
By Lemma 2.5 and Eq. (31) , we have a2n0 > 1 and 2 ((4n0 + 1)2 r + 3r − 1) (4r − 1)2 ≥ a2n 0 √ > 9r( 4r − 1 − 1)2 + 3r − 1 > (4r − 1)2 . √ Finally, when a1 ≤ −(1/2) 4r − 1, it follows from Eq. (31) that
(4r − 1)2 >
3 2 3 a1 (12r − 1) > (4r − 1)(12r − 1) > (4r − 1)2 . 2 8
Thus Eq. (31) is√ not valid for µ0 ≤ 0 and r ∈ (1/3, k 2 ]. Hence we conclude that µ0 > 0 for l = 1, . . . , [ 3k]. The proof is complete.
102
Z.M. Chen, W. G. Price
4. Proof of Theorem 1.3 (iii) Let us begin with the proof the smoothness and uniqueness. By Theorem 1.4, we see that coefficients {ξn }n∈Z of the eigenfunctions from the spectral problems Eqs. (4-5) satisfy the coupled set of the algebraic equations 2βn (βn + ρ(λ))ξn + λαn−1 ξn−1 − λαn+1 ξn+1 = 0, n ∈ Z.
(37)
Multiplying this nth equation by (−1)n αn ξn /2 and and summing the resultant equations, we have ∞ X
n
(−1)
n=−∞
βn (βn + ρ)αn ξn2
∞ X
+λ
(−1)n+1 αn αn+1 ξn ξn+1 = 0.
(38)
n=−∞
Denote by F (ρ(λ), λ) the left-hand side of this equation. Recalling from Eq. (6) that lim ρ(λ) = −l 2 − k 2 = −β0 ,
λ→0
we see that by the implicit function theorem ρ(λ) is a smooth and unique solution of Eq. (37), if ∂F /∂ρ 6 = 0 holds true. Indeed, by Eqs. (9) and (37), ∞ ∞ X X ∂ξn ∂F = (−1)n βn αn ξn2 + (−1)n 2βn (βn + ρ)αn ξn ∂ρ ∂ρ n=−∞ n=−∞ ∞ X ∂ξn+1 ∂ξn (−1)n+1 αn αn+1 ξn+1 + ξn +λ ∂ρ ∂ρ n=−∞ ∞ X
=
(−1)n βn αn ξn2
n=−∞ ∞ X
+
(−1)n αn
n=−∞ ∞ X
=
∂ξn (2βn (βn + ρ)ξn + λαn−1 ξn−1 − λαn+1 ξn+1 ) ∂ρ
(−1)n βn αn ξn2
n=−∞ ∞ X
=2
n=0
(−1)n βn αn ξn2 .
(39)
On the other hand, multiplying the nth equation of Eq. (37) by αn ξ¯n , and summing the resultant equations, we have ∞ X
2βn (βn + ρ)αn |ξn |2 +
n=−∞
∞ X
λαn αn+1 (ξn ξ¯n+1 − ξn+1 ξ¯n ) = 0,
n=−∞
which yields, by Eq. (9), ∞ X n=−∞
βn (βn + Reρ)αn |ξn |2 =
∞ X n=0
2βn (βn + Reρ)αn |ξn |2 = 0.
(40)
Periodic Navier–Stokes Flows
103
Using this equation, we have ∞
|
X ∂F (−1)n βn αn ξn2 | | ≥ 2β0 l(3k 2 − l 2 )|ξ0 |2 − 2| ∂ρ n=1
2 = β0 + Reρ ≥ = ≥
2 β0 + Reρ 2 β0 + Reρ 1 β0 + Reρ
∞ X n=1 ∞ X n=1 ∞ X n=1 ∞ X
∞ X βn (βn + Reρ)αn |ξn | − 2| (−1)n βn αn ξn2 | 2
n=1
βn (βn + Reρ)αn |ξn |2 − 2
∞ X
βn αn |ξn |2
n=1
βn (βn − β0 )αn |ξn |2 , for − β0 < Reρ,
(41)
βn (βn + Reρ)αn |ξn |2 , for − β0 < Reρ < 0,
n=1
= β0 l(3k 2 − l 2 )|ξ0 |2 . Since ξ0 6 = 0 is a constant, we thus have, for λ > 0 and Reρ ≥ −β0 , ∂F (ρ, λ) ∂F (ρ, λ) > 0. 6 = 0 and lim inf ∂ρ ∂ρ Reρ&−β0 Now we proceed to the proof of Eq. (7). By Eqs. (39) and (41), we have ∞ X (−1)n βn αn ξn2 6= 0. n=0
By Eq. (9), we may choose the constant ξ0 so that ∞ X
n
(−1)
n=−∞
βn αn ξn2
∞ X =2 (−1)n βn αn ξn2 = 2.
(42)
n=0
Moreover, by Eq. (40), we see ∞ X n=0
∞
X 1 βn αn |ξn | = − βn (βn − β0 )αn |ξn |2 < 0. β0 + Reρ(λ) 2
n=1
Differentiating Eq. (38) with respect to λ yields 0=
∞ X
(−1)n 2βn (βn + ρ(λ))αn ξn
n=−∞ ∞ X
dξn dλ
∞ X dρ + (−1)n+1 αn αn+1 ξn ξn+1 dλ n=−∞ n=−∞ ∞ X dξn dξn+1 n+1 +λ (−1) αn αn+1 ξn+1 + ξn dλ dλ n=−∞
+
(−1)n βn αn ξn2
(43)
104
Z.M. Chen, W. G. Price
=2 +
∞ X dρ(λ) (−1)n+1 αn αn+1 ξn ξn+1 , by Eq. (42), + dλ n=−∞ ∞ X
dξn (2βn (βn + ρ(λ))ξn + λαn−1 ξn−1 − λαn+1 ξn+1 ) dλ
(−1)n αn
n=−∞
=2
∞ dρ(λ) 1 X − (−1)n βn (βn + ρ(λ))αn ξn2 , by Eqs. (37–38), dλ λ n=−∞
=2
dρ(λ) 2 X (−1)n βn (βn + ρ(λ))αn ξn2 , by Eq. (9), − dλ λ
∞
n=0
∞
=2
2X dρ(λ) 2 (−1)n βn (βn − β0 )αn ξn2 , by Eq. (42). − (β0 + ρ(λ)) − dλ λ λ n=1
Therefore, by Eq. (43), ∞
λRe
X dρ (−1)n βn (βn − β0 )αn Re(ξn2 ) = β0 + Reρ(λ) + dλ n=1
> β0 + Reρ(λ) −
∞ X
βn (βn − β0 )αn |ξn |2
n=1
= (β0 + Reρ(λ))(1 +
∞ X
βn αn |ξn |2 ).
n=0
Now we show 1+
∞ X
βn αn |ξn |2 > 0.
n=0
Indeed, by Eq. (43), we denote by h the positive constant such that h=−
∞ X
βn αn |ξn |2 = β0 l(3k 2 − l 2 )|ξ0 |2 −
n=0
∞ X
βn αn |ξn |2 .
n=1
Using Eq. (42), we have β0 l(3k 2 − l 2 )(|ξ0 |2 − Im(ξ02 )) = h +
∞ X n=1
and β0 l(3k − l )(|ξ0 | 2
2
2
+ Im(ξ02 ))
=h+
∞ X n=1
Thus it follows from the Schwartz inequality that β02 l 2 (3k 2 − l 2 )2 |Re(ξ02 )|2
βn αn (|ξn |2 − (−1)n Im(ξn2 )),
βn αn (|ξn |2 + (−1)n Im(ξn2 )).
Periodic Navier–Stokes Flows
= h+
∞ X
105
! n
βn αn (|ξn | + (−1) 2
n=1
≥
h+
∞ X n=1
Im(ξn2 ))
h+
! n
βn αn (|ξn | − (−1) 2
n=1
!2 βn αn |Re(ξn2 )|
∞ X
Im(ξn2 ))
.
This implies, by Eq. (42), h ≤ β0 l 2 (3k 2 − l 2 )|Re(ξ02 )| −
∞ X n=1
βn αn |Re(ξn2 )|
∞ ∞ X X (−1)n βn αn Re(ξn2 )| − βn αn |Re(ξn2 )| < 1, ≤ 1+| n=1
n=1
where we have used the fact ξn2 6 = 0 for all n which is due to Eq. (9) and the choice ξ0 6 = 0. The proof is complete. 5. Proof of Theorem 1.3 (iv) 2 is an orthogonal sum of the following three subspaces Let us note that Hl,k
( ψ ∈ H |ψ = 2
∞ ∞ X X
) ξn cos(2mlx + 2nky) ,
m=0 n=−∞
ψ ∈ H 2| ψ =
X
∞ X
√ (2m−1)l> 3k n=−∞
ψ ∈ H 2| ψ =
√ [X 3k]
∞ X
(2m−1)l=l n=−∞
ξn cos((2m − 1)lx + ky + 2nky) ,
ξn cos((2m − 1)lx + ky + 2nky) ,
and each of these subspaces is invariant with respect to the operator Aλ . If the spectral problem Aλ ψ = ρψ is reduced to the first subspace, the absence of of the complex eigenvalue ρ with Imρ 6 = 0 and Reρ = 0 follows from [4]. For the second subspace, we readily see that Reγ0 > 0, which contradicts Eq. (10). This gives the absence of such an eigenvalue as well. Here we have used the equivalence given in Theorem 1.4 with the condition l 2 < 3k 2 replaced by the condition l 2 > 3k 2 . This equivalence is shown in [3]. However, as far as the last subspace is concerned, it has been obtained in Sects. 3 and 4 that the spectral problem Aλ ψ = ρψ in the space ( ψ| ψ =
∞ X n=−∞
) ξn cos((2m − 1)lx + ky + 2nky) ∈ H
2
106
Z.M. Chen, W. G. Price
√ for each m with l ≤ (2m − 1)l < [ 3k] has a unique and simple eigenvalue ρ = ρ(2m−1)l,k (λ) such that Re and Let
dρ(2m−1)l,k (λ) > 0, Imρ(2m−1)l,k (λ) < 0 for λ > 0 dλ
Reρ(2m−1)l,k (λ(2m−1)l,k ) = 0 for some λ(2m−1)l,k > 0. m0
be the positive integer such that m0 = max{m| λ(2m−1)l,k = λl,k , (2m − 1)l <
√ 3k}.
2 It is readily seen that ρl,k is simple in the flow invariant subspace H(2m 0 −1)l,k , which is 2 contained in Hl,k . Likewise, we obtain that ρl,k is simple in the flow invariant subspace 2 ˜2 H˜ (2m 0 −1)l,k ⊂ Hl,k .
The corresponding eigenfunction ψ is in the following form: ψ=
∞ X
ξn sin((2m0 − 1)lx + ky + 2nky) ∈ H 2 .
n=−∞
By Eq. (9), we thus obtain 2 1 = dim{ψ|Aλ ψ = ρl,k ψ, ψ ∈ H(2m 0 −1)l,k }, 2 1 = dim{ψ| Aλ ψ = ρl,k (λ)ψ, ψ ∈ H˜ (2m 0 −1)l,k }, 2 ˜2 2 = dim{ψ| Aλ ψ = ρl,k (λ)ψ, ψ ∈ H(2m 0 −1)l,k ∪ H(2m0 −1)l,k }.
This shows Assertion (iv) and completes the proof of Theorem 1.3. Acknowledgement. This research was partially made when the first author was visiting the Department of Mathematics at Indiana University.
References 1. Arnold, V.I., Meshalkin, L.D.: A. N. Kolmogorov’s seminar on selected problems of analysis (1958-1959). Russ. Math. Surv. 15 (1), 247–250 (1960) 2. Chen, Z.M.: Hopf bifurcation of the three-dimensional Navier–Stokes equations. J. Math. Anal. Appl., to appear 3. Chen, Z.M., Price, W.G.: Time dependent periodic Navier–Stokes flows on a two-dimensional torus. Commun. Math. Phys. 179, 577–597 (1996) 4. Iudovich, V.I.: Example of the generation of a secondary stationary or periodic flow when there is loss of stability of the laminar flow of a viscous incompressible fluid. J. Math. Mech. 29(3), 453–467 (1965) 5. Joseph, D.D., Sattinger, D.H.: Bifurcating time periodic solutions and their stability. Arch. Rational Mech. Anal. 45, 75–109 (1972) 6. Kirchgässner. K.: Bifurcation in nonlinear hydrodynamic stability. SIAM Rev. 17, 652–683 (1975) 7. Landau, L.: On the problem of turbulence. Comptes Rend. Acad. Sci. USSR 44, 311–316 (1944) 8. Lin, C.C.: The Theory of Hydrodynamic Stability. Cambridge: Cambridge University Press, 1955 9. Meshalkin, L.D., Sinai,Ya.G.: Investigation of the stability of a stationary solution of a system of equations for the plane movement of an incompressible viscous fluid. J. Math. Mech. 19 (9), 1700–1705 (1961) 10. Ruelle, D., Takens, F.: On the nature of turbulence. Commun. Math. Phys. 20, 167–192 (1971) Communicated by Ya. G. Sinai
Commun. Math. Phys. 207, 107 – 129 (1999)
Communications in
Mathematical Physics
© Springer-Verlag 1999
Fluid Dynamical Profiles and Constants of Motion from d-Branes R. Jackiw1,? , A. P. Polychronakos2,3 1 Center for Theoretical Physics, Massachusetts Institute of Technology, Cambridge, MA 02139–4307, USA 2 Theoretical Physics Department, Uppsala University, 75108 Uppsala, Sweden 3 Physics Department, University of Ioannina, 45110 Ioannina, Greece
Received: 3 February 1999 / Accepted: 9 April 1999
Abstract: Various fluid mechanical systems enjoy a hidden, higher-dimensional dynamical Poincaré symmetry, which arises owing to their descent from a Nambu–Goto action. Also, for the same reason, there are equivalence transformations between different models. These interconnections, summarized by the diagram below, are discussed in our paper.
Parameterization-invariant Nambu–Goto d-brane in d + 1, 1 spacetime
Light-cone parameterization
Cartesian parameterization
Nonrelativistic limit Poincaré-invariant
Galileo-invariant Chaplygin gas in (d, 1) spacetime
Exact transformation
Born–Infeld model in (d, 1) spacetime
? This work is supported in part by funds provided by the U.S. Department of Energy (D.O.E.) under contract #DE-FC02-94ER40818.
108
R. Jackiw, A. P. Polychronakos
1. Introduction In this paper we shall be concerned with nonlinear dynamical systems that are described by a density of “matter” ρ, flowing in time {t} with local velocity v on a d-dimensional surface coordinated by {r}. The vectorial nature of v is not unrestricted: v is a function of ∇θ , where θ is a velocity “potential”, and we shall examine several such functions. (When v is linear in ∇θ, the flow is irrotational, ∇ × v = 0.) The density is linked to the velocity by a continuity equation involving the matter current j = vρ, ρ(t, ˙ r) + ∇ · v(t, r)ρ(t, r) = 0, (1.1) while the velocity satisfies an “Euler” equation v˙ (t, r) + v(t, r) · ∇v(t, r) = f(ρ, v).
(1.2)
[The over-dot always indicates differentiation with respect to the temporal argument, while the gradient ∇ (unless further specified) differentiates the spatial argument.] We shall examine theories with various expressions for the force (per unit mass) f, which lead to Galileo, Poincaré and additional unexpected kinematical symmetries of the equations, and which sometimes produce completely integrable systems, with an infinite number of local conserved quantities. The above equations can to be formulated with an action principle. Consequently the symmetries that we find are in fact Noether symmetries, which leave the action invariant. Additionally, we shall present limiting and equivalence transformations between different models, which allow mapping solutions of one model onto solutions of another. Various topics that we discuss have already appeared in the literature. A common feature unites the diverse models that we study: they have a common antecedent in that they can be gotten from a parameterization-invariant Nambu–Goto action for a d-brane on a d + 1-dimensional space, moving in (d + 1, 1)-dimensional space-time. [A “dbrane” is a d-dimensional extended object: d = 1 is a string, d = 2 is a membrane, etc. A d-brane inhabiting d + 1-dimensional space divides that space in two.] When a lightcone parameterization is selected for the Nambu–Goto problem, one derives the Euler and continuity equations for a d-dimensional “Chaplygin gas” [in Eq. (1.2), f ∝ ρ13 ∇ρ]. Alternatively, a Cartesian parameterization produces the d-dimensional “Born–Infeld” model (see below). The relation between membranes and planar fluid mechanics (the d = 2 case) was known to Goldstone [1], and was developed by his student Hoppe (with collaboration) [2]. Similar connections, yielding equations in one spatial dimension, were discussed by Nutku (and collaborators) [3]. Here we consider the general d-dimensional case, and use the common ancestry of the various fluid-mechanical models to posit unexpected transformations between them, and to identify hidden, dynamical symmetries in each model, which derive from the high degree of symmetry of the Nambu–Goto parent theory. For d = 1, the models that we study become especially simple for two reasons. First, their common antecedent is a string (1-brane) moving on a plane (2-space), and for this system the Nambu–Goto equations are integrable [4]. Second, the requirement that velocity be expressible in terms of a potential poses no restriction in one dimension, where any function can be related to the derivative of another function. In this way one may understand that the d = 1 models are completely integrable, as has been noted previously [3–5].
Fluid Dynamical Profiles and Constants of Motion from d-Branes
109
In Sect. 2 we consider noninteracting systems, with Galileo- and Poincaré-invariant kinetic terms. Specific interactions that preserve Galileo and Poincaré symmetry, as well as higher, dynamical symmetries are discussed in Sect. 3. The relation of these to the Nambu–Goto theory is explained in Sect. 4, where we also exhibit mappings between the Galileo-invariant and the Poincaré-invariant models. The last Sect. 5 is devoted to models in one spatial dimension. 2. Force-Free Motion The force-free problem, f = 0, describes the free flow of dust. Equations (1.1) and (1.2) are readily solved in terms of initial data, specified (without loss of generality) at t = 0, = ρ0 (r), (2.1) ρ(t, r) v(t, r)
t=0
t=0
= v0 (r).
(2.2)
Upon defining the retarded position q(t, r) by the equation q + tv0 (q) = r,
(2.3)
one verifies that (1.1) and (1.2), with vanishing right side, are solved by v(t, r) = v0 (q)
(2.4)
∂q i ρ(t, r) = ρ0 (q) det j . ∂r
(2.5)
The free equations, with v restricted to a function of ∇θ , possess a variational action formulation, which was first given by Eckart for Galileo-invariant nonrelativistic motion [6]. We reproduce and generalize his argument. The Lagrangian for N point-particles of mass m in free nonrelativistic motion is the P 2 Galileo-invariant kinetic energy 21 N n=1 mvn (t). In a continuum description, the particle counting index n becomes the continuous variable particles are distributed R d r, and the PN 2 2 with density ρ, so that n=1 vn (t) becomes d r ρ(t, r)v (t, r). But we also wish to link the density with the current by the continuity equation (1.1), which can be enforced with the help of a Lagrange multiplier θ. We thus arrive at the free continuum Lagrangian Z h i Galileo = dd r ρ 21 mv 2 + θ ρ˙ + ∇ · (vρ) . (2.6) L Since L is first-order in time, and the canonical 1-form v may be varied, evaluated and eliminated [7]. We find
R
˙ does not contain v, dd r θ ρdt
ρmv − ρ∇θ = 0, showing that ∇θ is the local momentum p = mv, and the velocity is irrotational: v=
1 ∇θ, m
(2.7)
110
R. Jackiw, A. P. Polychronakos
∇ × v = 0.
(2.8)
The Lagrangian (2.6) becomes Z LGalileo 0
=
(∇θ )2 , d r θ ρ˙ − ρ 2m d
(2.9)
where the subscript 0 denotes absence of interaction. Varying θ in (2.9) regains the continuity equation (1.1), while varying ρ produces the free “Bernoulli” equation for the velocity potential θ θ˙ +
1 (∇θ)2 = 0. 2m
(2.10)
This is also recognized as the free Hamilton-Jacobi equation. The gradient of (2.10) gives rise, in view of (2.8), to the free Euler equation (1.2) (with f = 0). Remarkably the same equations emerge for a kinetic energy T that is an arbitrary function of v. If we generalize (2.6) to Z h i (2.11) L0 = dd r ρT (v) + θ (ρ˙ + ∇ · (vρ) we get, instead of (2.7), ∂T (v) ≡ p = ∇θ, ∂v
(2.12)
and (2.9) becomes Z L0 =
∂T (v) dd r θ ρ˙ − ρ v · − T (v) , ∂v
(2.13)
where it is understood that the Legrendre transform of T is expressed in terms of ∇θ by inverting (2.12). Varying θ in (2.13) again gives the continuity equation, Z δ ∂T (v) δL0 d = ρ˙ − d r ρv · 0= δθ δθ ∂v Z δ = ρ˙ − dd r ρv · ∇θ δθ = ρ˙ + ∇ · (vρ), (2.14) while varying ρ leaves a generalization of the free Bernoulli equation: 0=
∂T δL0 = −θ˙ − v · + T (v). δρ ∂v
(2.15)
With the help of (2.12), this implies 2 ∂ 2T ∂ j ∂ θ ˙ = −v j = −v θ ∂r i ∂r i ∂v j ∂r i ∂r j 2 ∂ T ∂v k ∂ 2 T = −v j j i = −v j j . ∂r ∂v ∂r ∂v k ∂v i
(2.16a)
Fluid Dynamical Profiles and Constants of Motion from d-Branes
111
On the other hand, from (2.12) it follows that ∂ 2T ∂ θ˙ = i k v˙ k . i ∂r ∂v ∂v
(2.16b)
Equations (2.16a) and (2.16b) are consistent only if the free Euler equation (1.2) holds 2T (provided the matrix ∂v∂i ∂v j has an inverse). With a general form for T (v), the local momentum (2.12) p ≡ ∂T ∂v remains irrotational while the velocity, as determined by inverting (2.12), becomes a nonlinear function of ∇θ . Evidently the solution (2.1)–(2.5) works with arbitrary kinetic energy, whose specific form enters only in the fixing the relation between v and ∇θ . However, the initial data for the velocity must be consistent with the expression of the velocity in terms of ∇θ. One may present a family of constants of motion: Z (2.17) C = dd r ρ(t, r)C v(t, r), r − tv(t, r) . The time independence of C is established either by differentiating with respect to t and using the free equations of motion to prove that dC dt = 0, or more easily, by inserting the solution (2.3)–(2.5) into (2.17) and changing integration variables from r to q. (Carrying out these manipulations requires assuming that ρ0 and v0 obey appropriate regularity conditions and drop off sufficiently at large distances.) Various constants of motion arise from invariance against time and space translation (energy E and total momentum P, respectively) as well as space rotation (angular momentum Lij ), provided T (v) carries no explicit time and coordinate dependence, and does not depend on any external vectors, i.e., T (v) = T (v). These constants are Z ∂T − T (v) , (2.18) E = H = dd r H H = ρ v · ∂v Z ∂T (v) = ρp, (2.19) P = dd r P P = ρ∇θ = ρ ∂v Z (2.20) Lij = dd r (r i P j − r j P i ). Also shifting θ by a constant is a symmetry, leading to conservation of Z N = dd r ρ.
(2.21)
To recognize that these constants of motion involve particular forms for C(v, r − tv) in (2.17), we recall that according to (2.12) ∇θ is a function of v, and the two are colinear when T (v) = T (v). The densities H and P are components of an energy-momentum tensor, T 00 and T 0i respectively, which satisfy continuity equations with energy flux T i0 and momentum flux T ij , T 00 T i0 T 0i T ij
= = = =
H, v i H, Pi, vi P j .
(2.22a) (2.22b) (2.22c) (2.22d)
112
R. Jackiw, A. P. Polychronakos
[T ij is symmetric when T (v) = T (v).] The continuity equations ∂ T˙ 00 + i T i0 = 0, ∂r ∂ T˙ 0j + i T ij = 0 ∂r
(2.23a) (2.23b)
are entirely equivalent to the free dynamical equations (1.1) and (1.2), with vanishing force. The symplectic structure, which is determined by the canonical 1-form in (2.13), indicates that the only nonvanishing bracket is [7] or equivalently
{θ(t, r), ρ(t, r0 )} = δ(r − r0 )
(2.24)
{p(t, r), ρ(t, r0 )} = ∇δ(r − r0 ).
(2.25)
With these, one verifies that the constants of motion (2.18)–(2.21) generate the appropriate infinitesimal transformation on the variables θ and ρ. Specific forms for T (v) support additional, kinematical symmetries and lead to further constants of motion. In the nonrelativistic case presented in Eqs. (2.6)–(2.10), we have Galileo invariance against boosts by velocity u. The transformation law for the fields reads ρ(t, r) → ρu (t, r) = ρ(T , R), θ(t, r) → θu (t, r) = θ(T , R) + m(u · r − u2 t/2),
(2.26)
where the transformed coordinates are boosted: t → T = t, r → R(t, r) = r − tu.
(2.27)
The inhomogenous terms in θu are recognized as the well-known 1-cocycle of field theoretic realizations of the Galileo group. Also they ensure that the transformation for the velocity v = ∇θ/m, v(t, r) → vu (t, r) = v(t, r − tu) + u, is appropriate for the co-moving velocity. The conserved quantity arising from the Galileo symmetry is Z B = tP − m dd r rρ Z = −m dd r (r − tv)ρ,
(2.28)
(2.29)
where the last equality casts B in the form (2.17). With the help of the bracket (2.24), B generates the infinitesimal transformation on the fields, and its bracket with P closes on N, thereby exposing the familiar Galileo 2-cocycle, which provides an extension of the algebra: {B i , P j } = δ ij mN.
(2.30)
Fluid Dynamical Profiles and Constants of Motion from d-Branes
113
The free Galileo-invariant theory possesses further symmetries, which survive even in the presence of a particular interaction. Hence we postpone discussing them until later, when interactions are included. Subsequently, in addition to the Galileo-invariant case, we shall also be concerned with a relativistic, p Poincaré-invariant model for which the point-particle kinetic energy P is −mc2 nn=1 1 − vn2 (t)/c2 . Upon passing to a continuum description, as in the nonrelativistic case, we find q (2.31) T (v) = −mc2 1 − v 2 /c2 , mv ∂T (v) ≡p= p = ∇θ, ∂v 1 − v 2 /c2 c∇θ , v= p 2 m c2 + (∇θ )2
(2.32) (2.33)
leading to q p mc2 , H = ρc m2 c2 + (∇θ)2 = ρc m2 c2 + p2 = ρ p 1 − v 2 /c2 and Lagrangian
Z = LLorentz 0
p dd r θ ρ˙ − ρc m2 c2 + (∇θ )2 .
(2.34)
(2.35)
In the nonrelativistic limit this becomes → −mc2 N + LGalileo . LLorentz 0 0 Under Lorentz boosts, with velocity u, the fields transform as p ˙ , R) + c m2 c2 + (∇θ (T , R))2 θ(T p ρ(t, r) → ρu (t, r) = ρ(T , R) ∂t θ(T , R) + c m2 c2 + (∇r θ (T , R))2 θ (t, r) → θu (t, r) = θ(T , R)
(2.36)
(2.37)
with Lorentz-boosted coordinates 1 β · r sinh β, t → T (t, r) = t cosh β + b c r → R(t, r) = r + b β ct sinh β + b β · r(cosh β − 1) ,
(2.38)
where β = u/c. Invariance is most easily verified by writing the action corresponding to (2.35) as Z n o p Lorentz = − dt dd r ρ θ˙ + c m2 c2 + (∇θ )2 . (2.39) I0 The infinitesimal version of the field transformation (2.37) r ∂ v + ct∇ ρ − β · ρ, δρ = β · c ∂t c r ∂ + ct∇ θ δθ = β · c ∂t
(2.40)
114
R. Jackiw, A. P. Polychronakos
is generated by the Lorentz constant of motion Z L = tP − dd r rH/c2 Z r p = dd r tρ∇θ − ρ m2 c2 + (∇θ )2 c Z 1 = −m dd r (r − tv)ρ p 1 − v 2 /c2
(2.41)
with the last equality exhibiting the form (2.17). The transformation laws for ρ and θ ensure that j α = (1, v/c)ρ and U α = (1, v/c) √ 1 2 2 transform as Lorentz vectors, so 1−v /c p that ρ 1 − v 2 /c2 and θ are scalars [8]. The equation of motion (2.15) for θ˙ together with (2.31)–(2.33) implies that the Lorentz vector ∂α θ satisfies a Lorentz-invariant equation (∂α θ)2 = m2 c2 .
(2.42)
Note that the Lorentz transformation law (2.37) for θ does not involve a 1-cocycle, which is a nonrelativistic effect. It is interesting to see how (2.26) arises in the nonrelativistic limit. By comparing the relativistic action (2.39) to the nonrelativistic one, we see that the relationship between θR and θNR – the relativistic and nonrelativistic variables – is θR = θNR − mc2 t.
(2.43)
Applying the Lorentz transformation law (2.37) to θR implies that θNR (t, r) − mc2 t → θNR (T , R) − mc2 T or θNR (t, r) → θNR (T , R) + mc2 (t − T ). We evaluate the nonrelativistic limit of the last quantity from (2.38) and find mc2 (t − T ) → m(u · r − u2 t/2), which matches the 1-cocycle in (2.26). Similar to the Galileo-invariant theory, this Poincaré-invariant model possesses further symmetries, which we shall discuss below, when we include an interaction that preserves them. 3. Motion in the Presence of Interactions 3.1. Nonrelativistic motion. Interactions that preserve the Galileo symmetry of the free, nonrelativistic motion can be included by adding a θ -independent potential V (ρ) to the Lagrangian (2.9): Z = LGalileo V
(∇θ )2 dd r θ ρ˙ − ρ − V (ρ) . 2m
(3.1)
With nonvanishing V , (2.3)–(2.5) are no longer solutions, and the quantity (2.17) with arbitrary C(v, r − tv) is no longer constant. Of course the Galileo generators (2.18)– (2.21) and (2.29), with H=ρ
(∇θ)2 + V (ρ) 2m
(3.2)
Fluid Dynamical Profiles and Constants of Motion from d-Branes
115
remain time independent. The energy-momentum tensor retains the form (2.22), with H 0i ij given by (3.2); in T i0 , H of (3.2) is diminished by V − ρ ∂V ∂ρ ; T is unchanged and T acquires the addition δ ij V − ρ ∂V ∂ρ . The dynamics implied by (3.1) arise in diverse physical contexts. The most directly physical application is to isentropic, irrotational motion in fluid mechanics with the 2 1/2 V (ρ) is the speed “force” f (ρ) = − ∂V∂ρ(ρ) corresponding to the enthalpy and ρ ∂ ∂ρ 2 of sound [9]. Alternatively one finds (3.1) (with V depending also on ∇ρ) in the hydrodynamical formulation of quantum mechanics, which emerges when the wave function is presented as [10], ψ = ρ 1/2 eiθ/h¯ .
(3.3)
[In this context, the inhomogenous Galileo transformation of θ (2.26) corresponds to the familiar change of phase in a wave function under Galileo boosts, while shifting θ by a constant is just the phase-invariance of quantum mechanics, which leads to probability conservation, i.e., constant N in (2.21).] But we shall be especially concerned with the case V (ρ) = λ/ρ
(3.4)
which arises in the study of “d-branes” – d-dimensional extended objects – moving on a (d + 1)-dimensional space, in (d + 1, 1)-dimensional space-time, and descending from a Nambu–Goto action (see Sect. 4) [1,2]. Furthermore, (3.4) arises in the nonrelativistic limit of a Poincaré-invariant model with interactions, which we shall also describe below. The equations of this theory, which follow from Z (∇θ )2 λ d = d r θ ρ ˙ − ρ − (3.5) LGalileo λ 2m ρ read in their Bernoulli form
ρ˙ + ∇ ·
θ˙ +
∇θ ρ m
= 0,
λ (∇θ)2 = 2, 2m ρ
(3.6)
(3.7)
while their Euler form is gotten by recalling that v = ∇θ/m, ρ˙ + ∇ · (vρ) = 0, v˙ + v · ∇v =
−2λ ∇ρ. mρ 3
(3.8) (3.9)
These are the equations for a “Chaplygin gas”. For thisR model there exist further symmetry transformations [11]. The action is invariant against a rescaling of time, with parameter ω, unIλGalileo = dtLGalileo λ der which the fields change according to ρ(t, r) → ρω (t, r) = e−ω ρ(eω t, r), θ(t, r) → θω (t, r) = eω θ (eω t, r),
(3.10)
116
R. Jackiw, A. P. Polychronakos
and the time-independent generator of the infinitesimal transformation is Z D = tH − dd r ρθ.
(3.11)
Furthermore, the action is also invariant against a field-dependent diffeomorphism, implicitly defined by t → T (t, r) = t + ω · r + 21 ω2 θ (T , R)/m, r → R(t, r) = r + ωθ(T , R)/m,
(3.12)
where the transformed fields are 1 , |J | θ(t, r) → θω (t, r) = θ (T , R),
ρ(t, r) → ρω (t, r) = ρ(T , R)
(3.13)
and |J | is the Jacobian of the transformation.
∂T ∂t J = det ∂R i ∂t
∂T −1 2 ∂r j = 1 − ω · ∇θ(T , R) − ω θ˙ (T , R) . m 2m ∂R j ∂r j
(3.14)
Here ω is a (vectorial) parameter of the transformation, with dimension of (velocity)−1 , and the time-independent quantity Z (3.15) G = dd r (rH − θ P/m) generates the infinitesimal transformation. Note that the generators D and G remain time-independent even in the absence of the interaction (3.4), hence these symmetries are also present for the free theory. The generators are not of the form (2.17): they involve θ , and cannot be written in terms of v = ∇θ/m. Finally we note that bracketing of the additional generators with the Galileo generators of the nonrelativistic theory on a (d, 1)-dimensional space-time produces an algebra which is isomorphic to the Poincaré group in (d + 1, 1) dimensions, [2,12] under which (t, θ, r) transforms as a (d +2)-Lorentz vector Xµ in light-cone components (t = X + , θ = X− ) [4]. Using (3.7), we may eliminate ρ, and describe the model solely in terms of θ , whose dynamics is governed by the Lagrangian √ Z Lλ = −2 λ dd r
s θ˙ +
(∇θ )2 . 2m
(3.16)
It is seen that the “interaction strength” λ in fact disappears from the equations of motion for θ ; λ serves only to normalize the Lagrangian. In the free theory it is not possible to achieve this compact formulation. Furthermore, the interacting dynamical equations
Fluid Dynamical Profiles and Constants of Motion from d-Branes
117
can be summarized by an equation for θ, which follows from (3.6), once (3.7) is used to eliminate ρ, or alternatively the equation is derived from (3.16), ∇θ/m 1 ∂ = 0. +∇· r r ∂t 2 2 (∇θ) (∇θ ) θ˙ + 2m θ˙ + 2m
(3.17)
In spite of their awkward appearance, (3.16) and (3.17) are Galileo invariant in (d, 1) space-time, and possess a hidden, nonlinearly realized Poincaré symmetry in (d + 1, 1) space-time (which is a descendant of the symmetries of the Nambu–Goto action; see Sect. 4). Apart from the intrinsic interest in this nonlinear realization of a kinematical/dynamical Poincaré symmetry, which is provided by (3.10)–(3.15) supplementing the linearly realized Galileo symmetry, the new symmetries have the useful consequence of generating new solutions to the equations of motion (3.6)–(3.9) from old ones. For example, the time-rescaling invariant, particular solution θ(t, r) = − r ρ(t, r) =
mr 2 , 2(d − 1)t |t| 2λ (d − 1) md r
(3.18)
can be transformed by (3.12)–(3.14) into new solutions θω and ρω , which are very different in character from (3.18) [11]. Note that in (3.18) we must have d > 1 and λ > 0. At d = 1, we can obtain general, time-rescaling invariant solutions. With the timerescaling–invariant Ansatz θ ∝ 1/t, (3.17) leads to a second order differential equation for the x-dependence of θ. Therefore solutions involve two arbitrary constants, one of which fixes the origin of x, and can be ignored. The other, which we call k, appears in two distinct families of solutions (which are related by an imaginary shift of x): m cosh2 kx, 2k 2 t m θ(t, x) = 2 sinh2 kx. 2k t
θ(t, x) = −
(3.19a) (3.19b)
For real θ, k must be real or imaginary. When a real ρ is computed from (3.7), we find that k must be real for λ > 0, imaginary for λ < 0, and a nonsingular density exists only with (3.19a) for λ > 0, r 2λ k|t| . (3.20) ρ(t, x) = m cosh2 kx 0
The current j = θm ρ exhibits a kink profile (derivation with respect to the single spatial variable is indicated by a dash) r 2λ tanh kx, (3.21) j (t, x) = ∓ m
118
R. Jackiw, A. P. Polychronakos
where the sign is fixed by the sign of t. In the last section we shall further review the d = 1 case. Another interesting solution, which is essentially one-dimensional, even though it exists in arbitrary spatial dimension, is given by n · u)2 . (3.22) θ(t, r) = 2(b n · r) + mu · r − 21 mt u2 − (b Here b n is a spatial unit vector, and u is an arbitrary vector with dimension of velocity, while 2 is an arbitrary function with static argument, which can be boosted by the Galileo transformation (2.26). The corresponding charge density is time-independent: √ 2λ/m (3.23) ρ(t, r) = b n · u + 20 (b n · r)/m and the static current becomes j(t, r) =
r
2λ u −b n(b n · u) b n+ . m b n · u + 20 (b n · r)/m
(3.24)
, which pre3.2. Relativistic motion. We seek an interacting generalization of LLorentz 0 serves Poincaré invariance. To find this, proceed as follows. Let Z (3.25) LLorentz = dd r {θ ρ˙ − H(ρ, p)} with p given by (2.32) and H to be determined. The symplectic structure is as in the free theory, hence the Poisson brackets retain the form (2.24), (2.25). We calculate the Poisson bracket between two Hamiltonian densities; in one the fields are evaluated at r, in the other at r0 : Z 0 δH(r) 00 000 00 000 δH(r ) 0 − r ) . (3.26a) · ∇δ(r − r ↔ r {H(r), H(r0 )} = dr dr 00 000 δp(r ) δρ(r ) We assume that H is a local function of p and ρ, so that the functional derivatives lead to ordinary derivatives, and (3.26a) becomes ∂H(r) ∂H(r) ∂H(r0 ) ∂H(r0 ) + · ∇δ(r − r0 ). (3.26b) {H(r), H(r0 )} = ∂p ∂ρ ∂p ∂ρ On the other hand, the Dirac-Schwinger condition for Lorentz invariance states that the bracket (3.26) should give rise to c2 times the momentum density P, which in this problem is given in (2.19) as P = ρp.
(3.27)
Rotational invariance requires that the p dependence of H is only on the magnitude p. Thus we conclude that 4
∂ ∂ H 2 H = c2 . 2 ∂p ∂ρ
(3.28)
Fluid Dynamical Profiles and Constants of Motion from d-Branes
119
While many forms for H can solve (3.28), we take a solution that is relevant to the present context, i.e., it generalizes in a simple manner the free Hamiltonian density (2.34), leads to a theory that descends from the Nambu–Goto action, and coincides in the nonrelativistic limit with the Galileo-invariant Chaplygin gas model: H=
q p p mc2 ρ 2 c2 + a 2 m2 c2 + (∇θ)2 = ρ 2 + a 2 /c2 p . 1 − v 2 /c2
(3.29)
Evidently the parameter a measures strength of “interaction”. Thus an interacting, Poincaré-invariant theory is described by the Lagrangian Z h i p p Lorentz = dd r θ ρ˙ − ρ 2 c2 + a 2 m2 c2 + (∇θ )2 . (3.30) La The corresponding conserved Lorentz generator takes the same form as in the first equality of (2.41), with H given by (3.29), and it generates the infinitesimal transformation q r ∂ v + ct∇ ρ − β · ρ 2 + a 2 /c2 , δρ = β · c ∂t c r ∂ + ct∇ θ. (3.31) δθ = β · c ∂t The finite transformation law is gotten by iterating (3.31), and the generalization of (2.37) becomes q ρ(t, r) → ρu (t, r) = ρ(T , R) 21 (+ + − ) + ρ 2 (T , R) + a 2 /c2 21 (+ − − ), θ (t, r) → θu (t, r) = θ(T , R),
(3.32)
where the Lorentz transformed coordinates (T , R) are given in (2.38) and p ˙ , R) ± c m2 c2 + (∇θ (T , R))2 θ(T p . ± ≡ ∂t θ(T , R) ± c m2 c2 + (∇r θ (T , R))2 p , vc √ It follows that j α = ρ, vc ρ 2 + a 2 /c2 and U α = √ ρ ρ 2 +a 2 /c2
(3.33) 1 1−v 2 /c2
are
Lorentz vectors, while θ is a scalar. The equations of motion are q q c∇θ ρ 2 + a 2 /c2 = ρ˙ + ∇ · (v ρ 2 + a 2 /c2 ) = 0, (3.34a) ρ˙ + ∇ · p m2 c2 + (∇θ)2 p ρ m2 c2 + (∇θ)2 mc2 p = θ˙ + p = 0. (3.34b) θ˙ + ρc p ρ 2 + a 2 /c2 ρ 2 + a 2 /c2 1 − v 2 /c2 Using (3.34b) to express ρ in terms of θ , ρ=−
∂0 θ a p , 2 c m2 c2 − (∂µ θ )2
(3.35)
120
R. Jackiw, A. P. Polychronakos
and substituting this in (3.34a) yields a second order, Lorentz covariant equation for θ . That equation may also be gotten by eliminating ρ from (3.30) and deriving a Lagrangian for θ , Z p (3.36) Lθ = −a dd r m2 c2 − (∂α θ )2 . This is the “Born–Infeld” Lagrangian, leading to the equation of motion 1 α ∂α θ = 0. ∂ p m2 c2 − (∂µ θ)2
(3.37)
As in the nonrelativistic theory [see (3.16)] the possibility of expressing ρ in terms of θ requires presence of the “interaction”, a 6 = 0, whose nonvanishing strength disappears from the nonlinear, interacting equations for θ . The energy-momentum tensor for the theory (3.30) is Lorentz covariant, of second rank, and symmetric. After eliminating ρ with (3.35), the resulting expression, depending solely on θ, bears the usual relation to its Lagrangian (3.36). Since the interacting Lorentz-invariant model is a descendant of the Nambu–Goto Lagrangian (see Sect. 4), it comes as no surprise that it too possesses additional kinematic symmetries, whose generators supplement the generators of the linearly realized Poincaré group in (d, 1) dimensions to give a nonlinear realization of dynamical Poincaré algebra in (d + 1, 1) dimensions [13]. The additional symmetry transformations, which leave (3.30) or (3.36) invariant, involve a field-dependent reparameterization of time, defined implicitly by t → T (t, r) =
θ (T , r) t + tanh mc2 ω, cosh mc2 ω mc2
(3.38)
under which the field transforms according to θ(t, r) → θω (t, r) =
θ(T , r) − mc2 t tanh mc2 ω. cosh mc2 ω
(3.39)
[We record only the action of the transformations on θ ; their effect on ρ can be read off from (3.35).] The infinitesimal generator, which is time independent by virtue of the equation of motion (3.37), is Z Z p p d 2 4 2 2 2 2 2 2 D = d r m c tρ + θ ρ c + a m c + (∇θ ) = dd r (m2 c4 tρ + θ H). (3.40) A second class of invariances involves a reparameterization of the spatial variable, implicitly defined by 1 − cos mcω tan mcω +b ωb ω·r , (3.41) r → R(t, r) = r − b ωθ(t, R) mc cos mcω θ(t, r) → θω (t, r) =
θ(t, R) − mcb ω · r sin mcω . cos mcω
(3.42)
Fluid Dynamical Profiles and Constants of Motion from d-Branes
121
√ Here ω is a vectorial parameter, b ω = ω/ω, ω = ω2 . The time-independent generator of the infinitesimal transformation reads Z Z (3.43) G = dd r (m2 c2 rρ + θρ∇θ) = dd r (m2 c2 rρ + θ P). With the addition of D and G to the previous generators, the Poincaré algebra in (d + 1, 1) dimension is reconstructed, and the transformation laws (3.38), (3.39), (3.41), (3.42) ensure that (t, r, θ) transforms as a (d + 2)-dimensional Lorentz vector (in Cartesian components) [2]. Note that this symmetry also holds in the free, a = 0, theory. Because of the extended symmetry, one can generate new solutions from old ones since both θω and θω in (3.39) and (3.42) solve the equation of motion if θ does. A remarkable fact is that the nonrelativistic limit of the above relativistic and interacting model precisely corresponds to the nonrelativistic interacting model discussed previously. This is easily seen from (3.30), which in the limit gives (3.1), with the help of (2.43), Z Z d (∇θN R )2 a2 d 2 d → − r mc tρ + d r θ ρ ˙ − ρ d − LLorentz NR a dt 2m 2mρ . = −mc2 N + LGalileo a 2 m/2 Equivalently, when ρ is eliminated, we have from (3.36) and (2.43), Z q Lθ → −a dd r 2mθ˙NR − (∇θN R )2 = Lλ=a 2 m/2 .
(3.44)
(3.45)
Correspondingly, the equation of motion (3.37) goes over into (3.17). It is easy to exhibit solutions of the relativistic theory, which reduce to solutions of the nonrelativistic equations that were given previously. The following profiles solve (3.37): q (3.46) θ(t, r) = −mc c2 t 2 + r 2 /(d − 1). With (2.43), this reduces to (3.18). In one dimension we have q θ(t, x) = −mc c2 t 2 + cosh2 kx/k 2
(3.47)
reducing to (3.19a). The relativistic analog of the lineal solution (3.22) is p n · u)2 . θ(t, r) = 2(b n · r) + mu · r − mct c2 + u2 − (b
(3.48)
Note that the above profiles continue to solve (3.37), even when the sign of the square root is reversed; but then they no longer possess a nonrelativistic limit. Additionally, there exists an essentially relativistic, chiral solution describing massless propagation in one direction: θ can satisfy the wave equation 2θ = 0 when
(∂µ θ)2 = constant
(3.49a) (3.49b)
122
R. Jackiw, A. P. Polychronakos
as, for example, with plane waves θ(t, r) = f (b n · r ± ct),
(3.50)
where (∂µ θ )2 vanishes. Then ρ reads from (3.35) ρ=∓
a 0 f . mc2
(3.51)
4. Relation to Nambu–Goto Action The Nambu–Goto action for a d-brane in (d + 1) spatial dimensions, moving in time on (d + 1, 1) Minkowski space is Z √ (4.1) IN-G = − dφ 0 dφ 1 · · · dφ d G, where G is (−1)d times the determinant of the induced metric Gαβ =
∂Xµ ∂Xµ . ∂φ α ∂φ β
(4.2)
Here Greek letters, from the beginning of the alphabet, label the quantities φ α = (φ 0 , φ), with which the d-brane coordinate X µ is parameterized; φ 0 is the evolution parameter, and φ = {φ a , a = 1, . . . , d} are the fixed-time, spatial parameters. These d-brane coordinates carry a Greek-letter index from the middle of the alphabet, with value 0 for the temporal coordinate X0 and m for the d-brane’s d + 1 spatial coordinates, X = {X m , m = 1, . . . , d, d + 1}. The action is invariant against reparameterizations of the φ α , and we make the parameterization choice that the d coordinates X m , m = 1, . . . , d, are given by φ m , which we rename r m . For the remaining parameters we use one of two options, “light-cone” and “Cartesian”. In the light-cone option, we define 1 (4.3) X± = √ (X 0 ± Xd+1 ) 2 √ and for X + choose the parameterization X+ = √ 2λm φ 0 ; also we rename X+ as t. The remaining coordinate X − , a function of φ 0 = t/ 2λm and φ = r, is renamed θ (t, r)/m. Upon evaluating the determinant G, we see that the Nambu–Goto action coincides with the action for (3.16) [14]. This identity also explains the higher symmetry noted in Eqs. (3.10) and (3.13). Our choice of parameterization does not interfere with invariance against the (d + 1, 1) Poincaré group, which acts on Xµ . In the chosen parameterization, the Poincaré transformation acts nonlinearly, mixing coordinates (t, r) with the field θ. (However, the higher symmetry is also enjoyed by the noninteracting theory, λ = 0, which is not equivalent to the Nambu–Goto model.) For the second, Cartesian option the chosen parameterization permits writing X0 , which is renamed ct, as amcφ 0 , while the last coordinate, Xd+1 , which is a function of φ 0 = t/am and φ = r, is called θ(t, r)/mc. With these choices the Nambu–Goto action coincides with that for the Born–Infeld Lagrangian, Eq. (3.36). Again the higher dynamical symmetry, described by Eqs. (3.38)–(3.43), is now understood as the covariance of
Fluid Dynamical Profiles and Constants of Motion from d-Branes
123
the Nambu–Goto variables Xµ against (d +1, 1)-dimensional Poincaré transformations. (But once again, the similar invariance of the free theory, a = 0, cannot be related to properties of a Nambu–Goto action.) The Nambu–Goto action is normaized by the d-brane tension, which has been scaled to unity in (4.1). Thus the non-relativistic and relativistic free models (λ = 0 = a) may be viewed as descriptions of ”tension-less” d-branes. Since both the (d, 1)-dimensional Galileo-invariant Chaplygin gas equations and the (d, 1)-dimensional Poincaré-invariant Born–Infeld equations correspond to different parameterizations of the (d + 1, 1) Nambu–Goto action, there must be a transformation – recognized as a reparameterization – that takes solutions of one model into the other. This transformation may be formulated as follows. Given a solution θNR (t, r) to (3.17), we solve for T (t, r) from the equation 1 1 θNR T (t, r), r = t. (4.4) √ T (t, r) + mc2 2 Then a solution θR (t, r) to (3.37), is given by 1 1 1 θ θR (t, r) T (t, r), r = √ T (t, r) − NR 2 mc mc2 2 or θR (t, r) = mc2
√ 2T (t, r) − t .
(4.5a)
(4.5b)
Indeed this mapping produces solutions (3.46), (3.47) and (3.48) from (3.18), (3.19a) and (3.22) respectively. [In fact both signs of the square root are obtained; also (3.48) emerges from (3.22) only after a redefinition of 2 and u.] Oppositely, given as solution θR (t, r) to the relativistic Born–Infeld equation (3.37), we reparameterize by solving for T (t, r) from 1 1 θ T (t, r), r =t (4.6) √ T (t, r) + R mc2 2 and then find θNR (t, r) from 1 1 1 θ T (t, r), r θ (t, r) = √ T (t, r) − 2 R 2 NR mc mc 2 or θNR (t, r) = mc2
√ 2T (t, r) − t .
The two transformations are collected in the statement √ θNR (t, r) = mc2 ( 2T − t), √ θR (T , r) = mc2 ( 2t − T )
(4.7a)
(4.7b)
(4.8)
with the instruction that obtaining an expression for θN R in terms of θR , or vice-versa, requires that one of T or t be eliminated in favor of the other.
124
R. Jackiw, A. P. Polychronakos
The interrelationships may be summarized by the following diagram: Parameterization-invariant Nambu–Goto d-branes in d + 1, 1 spacetime
Light-cone parameterization
Cartesian parameterization
Nonrelativistic limit Poincaré-invariant
Galileo-invariant Chaplygin gas in (d, 1) spacetime
Exact transformation
Born–Infeld model in (d, 1) spacetime
It is striking that there exists a two-fold relationship between the Chaplygin gas and the Born–Infeld model. First, there is the exact mapping, given in (4.8), of one into the other. Second, the nonrelativistic limit (3.44), (3.45) of the latter produces the former. 5. One-Dimensional Motion In one spatial dimension, the motion of these systems simplifies and they become integrable. We shall give a self-contained demonstration of integrability and derive the integrals of motion in a compact form, stressing the connection between the relativistic and nonrelativistic case. In one dimension the requirement that the local momentum is irrotational poses no restriction. We shall use as phase space variables the local momentum p(t, x) and the particle density ρ(t, x). The equal-time Poisson bracket (2.25) becomes {p(x), ρ(x 0 )} = δ 0 (x − x 0 ).
(5.1)
Note that p and ρ in this R relation are on an equal footing and are governed by a nonlocal ˙ − y)p(y)dt, where is the antisymmetric step canonical 1-form 21 dx dy ρ(x)(x function. We shall consider local integrals, that is, quantities of the form Z F = dxF(p(x), ρ(x)) (5.2) with F a local function of p and ρ. The Poisson bracket of two such integrals F and G is calculated through (5.1) as Z i h (5.3) {F, G} = − dx (Fρρ Gp + Fρp Gρ )∂x ρ + (Fρp Gp + Fpp Gρ )∂x p , where we suppressed the dependence on x and subscripts indicate partial derivative. If the above integrand is a total x-derivative then (with appropriate boundary conditions)
Fluid Dynamical Profiles and Constants of Motion from d-Branes
125
the integral will vanish and F and G will be in involution. For this we need the curl-free condition (Fρρ Gp + Fρp Gρ )p = (Fρp Gp + Fpp Gρ )ρ
(5.4)
Gρρ Fρρ = . Fpp Gpp
(5.5)
or, finally
Choosing one of the integrals to be the Hamiltonian H = condition
R
dxH, the well-known
Hρρ Fρρ = Fpp Hpp
(5.6)
guarantees that F is a constant of the motion. If we recover a set of such integrals satisfying (5.6) then they will obviously also satisfy (5.5) among themselves. Therefore, constants of motion will automatically be in involution. For the nonrelativistic case the Hamiltonian density (3.2), (3.4) is H=ρ
λ p2 + , 2m ρ
(5.7)
and therefore the integrals of motion are generated by functions that satisfy ρ 4 Fρρ = 2λmFpp .
(5.8)
This can readily be solved by separation of variables. We prefer, however, simply to give its general solution in terms of two arbitrary functions f and g of one variable: √ √ 2λm 2λm + ρg p − . (5.9) F = ρf p + ρ ρ We essentially get two infinite towers of integrals. Choosing, e.g., f (z) = zn , g = 0 or g(z) = zn , f = 0, we get the integrals √ Z 2λm n . (5.10) In± = dx ρ p ± ρ All the integrals presented in [5,16] can be identified as linear combinations of the In± . As stated, the In± are all in involution and demonstrate the complete integrability of the system. The Hamiltonian, in particular, is included as 4H = I2+ + I2− and the total momentum as 2P = I1+ + I1− . The quantities √ 2λm (5.11) R± = p ± ρ appearing above are known as Riemann coordinates. The equations of motion for this system (continuity and Euler) are summarized in terms of R± : 1 0 . R˙ ± = − R∓ R± m
(5.12)
126
R. Jackiw, A. P. Polychronakos
Although this formulation for the equations is known, the relation to the constants of motion does not seem to appear in the literature. Another fluid system for which the equations of motion, expressed in terms of Riemann coordinates, take a form similar to (5.12) possesses a potential cubic in ρ: V (ρ) = `ρ 3 /3. This also arises in a collective, semiclassical description of nonrelativistic free fermions, where the cubic potential reproduces fermion repulsion [17]. In this case, the Riemann corrdinates read √ (5.13) R± = p ± 2`m ρ and, in contrast to (5.12), they decouple in the the equations of motion: 1 0 . R˙ ± = − R± R± m
(5.14)
Indeed it is seen that R± satisfy essentially the free Euler equation [compare (1.2) and identify R± with v]. Consequently (5.14) is solved by analogs of (2.2)–(2.4). Both these examples are special limiting cases of a more general system, with potential λ ρ + 23 α + Aρ + B. (5.15) V (ρ) = (ρ + α)2 (The terms A, B correspond to a dynamically trivial part that does not alter the equations of motion.) The Chaplygin gas corresponds to α = A = B = 0, while the cubic potential is recovered in the limit λ = `α 4 , A = 13 `α 2 , B = − 23 `α 3 , α → ∞. The Riemann coordinates are R± = p ±
√ 2λm ρ+α
and the conserved integrals of this system are given by functions of R± : √ √ 2λm 2λm + (ρ + α)g p − . F = (ρ + α)f p + ρ+α ρ+α
(5.16)
(5.17)
(5.18)
The physical meaning of this general system is not clear. We conclude the discussion on the one-dimensional nonrelativistic Chaplygin gas by presenting a set of moving solutions, which are the Galileo boosts of the static solutions (3.22), (3.23) in Sect. 3. These read √ 2λm (5.19) p = p(x − ut) , ρ = |p − mu| with p(x − ut) an arbitrary function of x − ut (provided it never equals mu). Clearly this is a constant-profile solution moving with a velocity u. For the relativistic Born–Infeld system, the Hamiltonian density (3.29) is q p (5.20) H = ρ 2 c2 + a 2 m2 c2 + p2 .
Fluid Dynamical Profiles and Constants of Motion from d-Branes
127
Relation (5.6) for the conserved integrals reads m2 (ρ 2 c2 + a 2 )2 Fρρ = a 2 (m2 c2 + p2 )2 Fpp .
(5.21)
This can be solved by separation of variables. We prefer again, however, to define Riemann coordinates and give the solution in terms of arbitrary functions of one variable, just as in the nonrelativistic case. The relevant combinations here are R± = φρ ± φp , φρ = arctan
ρc p , φp = arctan , a mc
(5.22)
and the equations of motion read, in terms of these, 0 , R˙ ± = ∓c(cos R∓ )R±
(5.23)
while the general solution to (5.21) is F=
i h mca f (φρ + φp ) + g(φρ − φp ) . cos φρ cos φp
(5.24)
By choosing, as before, simple monomials for f and g we get the tower of conserved quantities Z (φρ ± φp )n . (5.25) In± = mca dx cos φρ cos φp The Hamiltonian is included as H = I0± . The momentum, on the other hand, is an infinite series in the above integrals, requiring f and g to be exponential functions. We can give an alternative tower of complex conserved integrals involving only algebraic functions of p and ρ: Z q 1−n p 1−n m2 c2 + p2 (ρc ± ia)n (p + imc)n . (5.26) I˜n± = dx ρ 2 c2 + a 2 Then the Hamiltonian is H = I˜0± while the total momentum, total particle number and integral of the local momentum are contained in the real and imaginary parts of I1+ and I1− (as well as a trivial constant). It can be checked that the integrals (5.26) go over to the nonrelativistic ones (5.10) in the limit c → ∞ upon proper rescaling. In the relativistic model ρ need not be constrained to be positive (negative ρ could be interpreted as antiparticle density). The transformation p → −p, ρ → −ρ is a symmetry and can be interpreted as charge conjuguation. Further, p and ρ appear in an equivalent way. As a result, this theory enjoys a duality transformation: ρ→±
mc2 a ρ. p, p → ± mc2 a
(5.27)
Under the above, both the canonical structure and the Hamiltonian remain invariant. Solutions are mapped in general to new solutions. Note that the nonrelativistic limit is mapped to the ultra-relativistic one under duality. Self-dual solutions ρ = ± mca 2 p satisfy ρ˙ = ∓cρ 0
(5.28)
128
R. Jackiw, A. P. Polychronakos
and are, therefore, the chiral relativistic solutions that were presented at the end of Sect. 3. In the self-dual case, when p is eliminated from the canonical 1-form and from the Hamiltonian, one arrives at an action for ρ, which coincides (apart from irrelevant constants) with the self-dual action, constructed some time ago [18]: Z Z q p 1 2 2 2 2 2 2 dx dy ρ(x)(x ˙ − y)p(y) dt − dx ρ c + a m c + p dt 2 2
=
2mc a
2 Z 1 4
dx dy ρ(x)(x ˙ − y)ρ(y) dt −
c 2
Z
p= mca ρ
a2 dx ρ 2 (x) + 2 dt . (5.29) c
A set of constant-profile solutions can be found by Lorentz-boosting the static solution (3.48). Their most general form is √ a up ± mc c2 − u2 (5.30) p = p(x − ut) , ρ = √ c p c2 − u2 ∓ mcu with p(x − ut), again, a general function. Note that the two choices of sign in the above formula are related by charge conjuguation. In the extreme relativistic case u → c this solution goes over to the chiral relativistic solution (3.51). The set of these solutions is closed under duality transformations. We shall conclude by presenting an explicit mapping between the relativistic and the nonrelativistic theories in one dimension, which demonstrates their kinematical equivalence. Note that the relativistic equations of motion (5.23) become, in terms of cos R± , ∂t cos R± = ∓c(cos R∓ )∂x cos R± .
(5.31)
These are essentially identical to the equations of motion for the nonrelativistic Riemann coordinates (5.12). In fact, putting R¯ ± = ±mc cos R± ,
(5.32)
we see that the R¯ ± obey the nonrelativistic equations (5.12). Expressing R¯ ± and R± in terms of the corresponding nonrelativistic and relativistic variables produces a mapping between the two sets. Call pNR , ρNR and pR , ρR the local momentum and density of the nonrelativistic and the relativistic theory, respectively. Then the mapping is √ 2λm ρR pR HR , pN R = mc2 , (5.33) ρNR = 2 2 am c HR q q 2 is the relativistic Hamiltonian density. As can where HR = ρR2 c2 + a 2 m2 c2 + pR be checked by direct algebra, this maps the relativistic equations of motion to the nonrelativistic ones. Since the combinations of pR and ρR that appear in (5.33) are dualityinvariant, the mapping of solutions is two-to-one. Note that the constant-profile relativistic solutions (5.30) are mapped to the corresponding nonrelativistic ones (5.19). We stress that the transformation (5.33) is not canonical, since it does not preserve the Poisson brackets. Accordingly, it does not map the relativistic Hamiltonian into the nonrelativistic one. This is a manifestation of the bi-Hamiltonian structure of these systems, since there are now two pairs of Hamiltonian and canonical structure (the standard nonrelativistic one and the one obtained through this mapping) that lead to the same nonrelativistic equations of motion.
Fluid Dynamical Profiles and Constants of Motion from d-Branes
129
We note that a similar mapping between the nonrelativistic system and the relativistic one in light-cone coordinates was presented previously by Verosky [19]. The mapping (5.33), then, can be considered as the analog of Verosky’s transformation for the Lorentz system, although it cannot be obtained from it in any straightforward way. The existence of an infinite set of constants of motion for both the nonrelativistic Chaplygin gas, (5.9) and (5.10), as well as for the relativistic Born–Infeld model, (5.24)– (5.26), signals the complete integrability of these theories. The actual integration of the equations of motion can be carried out only indirectly. First of all, since both these d = 1 models descend, via alternate parameterization choices, from the Nambu–Goto string (1-brane) on the plane (2-space), the explicit integration of the latter [4] allows presenting solutions of the former two in terms of two arbitrary functions. Alternatively, the Chaplygin gas equations can be combined, after a Legendre transformation, into a linear, second-order partial differential equation [9], whose general solution, in terms of two arbitrary functions, is known explicitly [20]. The Born–Infeld solution can then be constructed with the help of transformations described in Sects. 4 and 5. Also, the Born– Infeld equations have been solved with a hodograph transformation without reference to the Chaplygin gas [21]. References 1. Goldstone, J.: Unpublished 2. Bordemann, M. and Hoppe, J.: Phys. Lett. B317, 315 (1993); B325, 359 (1994) 3. Nutku, Y.: J. Math. Phys. 28, 2579 (1987); Olver, P. and Nutku, Y.: J. Math. Phys. 29, 1610 (1988); Arik, M., Neyzi, F., Nutku, Y., Olver, P. and Verosky, J.: J. Math. Phys. 30, 1338 (1989) 4. See e.g. Bazeia, D.: Phys. Rev. D 59, 085007 (1999) 5. Brunelli, J. and Das, A.: Phys. Lett. A235, 597 (1997) 6. Eckart, C.: Phys. Rev. 54, 920 (1938); Recent work: Schakel, A.: Mod. Phys. Lett. B10, 999 (1996); Ogawa, N.: Preprint, hep-th/9801115 7. Faddeev, L.D. and Jackiw, R.: Phys. Rev. Lett. 60, 1692 (1988) 8. In conventional treatments of relativistic hydrodynamics, as in Weinberg, S.: Gravitation and Cosmology. New York: Wiley, 1972, ρ is a scalar rather than the time-component of a Lorentz vector. This distinction makes no difference in the noninteracting case, but our choice allows introducing self-interactions in a Lorentz-invariant manner. 9. Landau, L. and Lifschitz, E.: Fluid Mechanics. 2nd ed. Oxford, UK: Pergamon, 1987 10. Madelung, E.: Z. Phys. 40, 322 (1926); Recent work: Merzbacher, E.: Quantum Mechanics. 3rd ed. New York: Wiley, 1998 11. Bazeia, D. and Jackiw, R.: Ann. Phys. (NY) 270, 246 (1998); Jackiw, R. and Polychronakos, A.: Faddeev Festschrift. Preprint, hep-th/9809123 12. Jevicki, A.: Phys. Rev. D 57, 5955 (1998) 13. Bordemann and Hoppe: Second paper in Ref. [2]. 14. This result was established with a hodographic transformation for d = 2 by Bordemann and Hoppe (first paper in Ref. [2]) and for arbitrary d by Jackiw and Polychronakos, Ref. [11]; see also Bazeia, Ref. [4]. The present, simple approach was shown to us by Hoppe; see also Hoppe, J.: Phys. Lett. B329, 10 (1994) 15. This result was established with a hodographic transformation for d = 2 by Bordemann and Hoppe (second paper in Ref. [2]); see also Hoppe, Ref. [14] 16. Nutku, Olver and Nutku, Ref. [3] 17. Jevicki, A. and Sakita, B.: Phys. Rev. D 22, 467 (1980) and Nucl. Phys. B165, 511 (1980); Polchinski, J.: Nucl. Phys. B362, 25 (1991); Avan, J. and Jevicki, A.: Phys. Lett. B266, 35 (1991) and B272, 17 (1991) 18. Floreanini, R. and Jackiw, R.: Phys. Rev. Lett. 59, 1873 (1987) 19. As cited in Olver and Nutku, Ref. [3] 20. Jackiw and Polychronakos, Ref. [11] 21. Barbishov, B. and Chernikov, N.: Zh. Eksp. Teor. Fiz. 51, 658 (1966) [Eng. trans.: Sov. Phys. JETP 24, 437 (1967)] Communicated by R. H. Dijkgraaf
Commun. Math. Phys. 207, 131 – 143 (1999)
Communications in
Mathematical Physics
© Springer-Verlag 1999
The Principal Eigenvalue of a Conformally Invariant Differential Operator, with an Application to Semilinear Elliptic PDE Matthew J. Gursky? Department of Mathematics, Indiana University, Bloomington, IN 47405-5701, USA. E-mail: [email protected] Received: 1 November 1998 / Accepted: 16 April 1999
Abstract: In this paper we give conditions under which the Paneitz operator, a conformally invariant fourth order differential operator, is positive. We also prove a sharp estimate for a related conformal invariant. These results are combined to study the existence of a semilinear PDE with critical Sobolev growth. Introduction Our aim in this paper is to study an important fourth order differential operator along with a related conformal invariant from four-dimensional Riemannian geometry. The operator and associated invariant have arisen in the study of semi-linear elliptic PDEs with critical Sobolev growth ([Li,CY1,CY2,WX]), the zeta-functional determinant ([BO,CY1, BCY]), vanishing theorems for the first Betti number of a compact four-manifold of positive scalar curvature ([Gu]), and in a gauge choice for Maxwell’s equations ([ES1, Pa]). Let us begin by defining the objects of our study. Let (M 4 , g) be a smooth compact four-dimensional Riemannian manifold. We let Rmg denote the Riemannian curvature tensor of g, and drop the subscript whenever there is no possibility of confusion. We also let W, Ric, and R denote the various components of the curvature tensor: Weyl, Ricci, and scalar. Now define Q = Qg =
1 (−1R − 3|Ric|2 + R 2 ), 12
(0.1)
where 1 denotes the Laplace–Beltrami operator with respect to g (in our case a negative operator). Our first observation is that the quantity Z (0.2) κg = Qg dµg , ? Supported in part by NSF Grant DMS-9801046 and an Alfred P. Sloan Research Fellowship
132
M. J. Gursky
where dµg is the Riemannian measure, is a conformal invariant. There are (at least) two ways of seeing this. First, by the Chern–Gauss–Bonnet formula ([Be, (6.31)]), Z Z (0.3) 4π 2 χ(M 4 ) = |W |2 dµ + Qdµ, where χ (M 4 ) is the Euler characteristic of M 4 . Since the integral of the norm squared of the Weyl tensor is conformally invariant in four dimensions, we see that (0.2) must be also. Indeed, (0.3) indicates the analogy between the quantity Q of a four-manifold and the Gauss curvature of a surface – a theme which we will repeatedly encounter in this paper. In the latter case, the Gauss–Bonnet formula is Z (0.4) 2πχ (6 2 ) = KdA so that the Euler characteristic is given by a conformally invariant integral. The corresponding formula in four dimensions (0.3) shows that the Euler characteristic of a four-manifold is also a sum of conformally invariant integrals, one of which (the Weyl term) measures whether the manifold is locally conformally flat. Since every surface is locally conformally flat this latter term is absent in (0.4). Another way to see that (0.2) is conformally invariant is by writing down the formula for the transformation of Q under a conformal change of metric. In doing this we also introduce the aforementioned differential operator, called the Paneitz operator, which will be a focus of this paper. Let d : 30 (M 4 ) = C ∞ (M 4 ) → 31 (M 4 ) denote the exterior derivative, and δ : 31 (M 4 ) → 30 (M 4 ) its L2 -formal adjoint. We define P = Pg by P = 12 + δ
2 3
Rg − 2Ric d.
(0.5)
P is clearly fourth order, and turns out to be conformally invariant ([Pa]). More precisely, if g˜ = e2w g then Pg˜ = e−4w Pg . Having defined P we can now give the transformation law for Q : if g˜ = e2w g then Pg w + 2Qg˜ e4w = 2Qg .
(0.6)
From (0.6) it is again easy to see that the integral of Q is a conformal invariant: just use the divergence theorem along with the fact that dµg˜ = e4w dµg . Returning to our analogy with the theory of surfaces, we see that P is, from the point of view of conformal geometry, the proper generalization of the Laplace–Beltrami operator to four dimensions. For, in the case of surfaces if g˜ = e2w g then 1g˜ = e−2w 1g ; i.e., the Laplace–Beltrami operator is also conformally invariant. The two-dimensional counterpart to (0.6) is the Gauss curvature equation: 1w + Kg˜ e2w = Kg .
(0.7)
There is by now a vast literature on (0.7), but the study of (0.6) is still in many respects quite preliminary. For example, given a Riemannian manifold (M 4 , g), under what conditions is there a conformal metric g˜ = e2w g with Qg˜ identically constant? The analogous question for (0.7) is the famous uniformization theorem, and there are many approaches one can take to resolve this. In four dimensions the problem was undertaken
Principal Eigenvalue of a Conformally Invariant Differential Operator
133
by Chang and Yang ([CY1]) using the direct variational approach. That is, they studied the convergence of minimizers of the functional Z Z Z Z (0.8) F [w] = wP wdµ + 4 Qwdµ − Qdµ log e4w dµ Z (here
=
Z
dµ
−1 Z
) whose critical points satisfy (0.6) with Qg˜ = κg . In the
course of their work they discovered two things. First, if the principal eigenvalue λ1 (P ) is allowed to be negative, then a minimizing sequence for F can diverge.1 For example, it is easy to see that some hyperbolic metrics have constant Q but are saddle points of F . Second, to establish the compactness of a minimizing sequence the value of κg as defined in (0.2) is crucial. More precisely, they showed that when λ1 (P ) = 0 with Ker P = {constants}, κg < 8π 2 , then a minimizing sequence always converges. The condition κg < 8π 2 is completely analogous to the compactness criteria involved in theYamabe problem ([LP]), where one requires that the conformally invariant Sobolev constant be less than its value on the sphere (and equal only in the case of the sphere). Indeed, κg¯ = 8π 2 when g¯ is the round metric. The sharp inequality appearing in the work of Chang–Yang is the Moser–Trudinger type inequality due to Adams [Ad]. In this context we can now state the two main results of this paper. The first gives conditions under which λ1 (P ) = 0. Since P is a conformally invariant operator, our conditions should be conformally invariant. Now in four dimensions there are two natural (signed) conformally invariant quantities to consider; κg and the Yamabe invariant: Z Y (g) = inf
g=e ˜ 2w g
Rg˜ dµg˜
. Z
1 dµg˜
2
.
We then have: Theorem A. Let (M 4 , g) be a smooth compact four-dimensional Riemannian manifold. If κg ≥ 0 and Y (g) ≥ 0 then λ1 (P ) = 0 and Ker P = {constants}. Remarks. 1. In [ES2] another criterion for λ1 (P ) = 0 is given. However, this requires checking in a pointwise fashion the eigenvalues of the Ricci tensor. In particular, the condition is not conformally invariant. 2. There are hyperbolic manifolds (necessarily with Y (g) < 0) for which λ1 (P ) < 0. So the assumption of non-negativity of the Yamabe invariant cannot be relaxed. On the other hand, in [ES2] examples are constructed of manifolds with Y (g) > 0, κg < 0, and λ1 (P ) = 0. Our next result involves estimating κg when Y (g) ≥ 0. Theorem B. Let (M 4 , g) be a smooth compact four-dimensional Riemannian manifold. If Y (g) ≥ 0 then κg ≤ 8π 2 . Moreover, Y (g) ≥ 0 and κg = 8π 2 if and only if (M 4 , g) is conformally equivalent to the round sphere. Remark. Hyperbolic manifolds provide examples with Y (g) < 0 but κg 8π 2 . From Theorems A and B along with the work of Chang and Yang we have 1 Since P annihilates constants, λ (P ) is always non-positive. 1
134
M. J. Gursky
Corollary 0.1. Let (M 4 , g) be a smooth compact four-dimensional Riemannian manifold with κg ≥ 0 and Y (g) ≥ 0. Then there exists a metric g˜ = e2w g with Qg˜ a constant. In [CY1] Chang and Yang also studied extremals of the zeta-functional determinant. We will describe this work in some detail in Sect. 2. For now we only point out that the above estimate for κg leads to a much simpler existence criteria for extremals than those stated in [CY1]. For example, if L denotes the conformal Laplacian and 6 ∇ 2 the spin Laplacian, then Corollary 0.2. Let (M 4 , g) be a smooth compact four-dimensional Riemannian (spin) manifold with Y (g) ≥ 0. Then an extremal for log det L (log det 6 ∇ 2 ) exists in the conformal class of g. Since our criteria for the existence of solutions to (0.6) depends on checking the sign of two conformal invariants, it seems reasonable to ask whether there are any obstructions to both invariants being simultaneously positive. Recent work in [Gu] on the Weyl functional provides necessary conditions involving the first deRham cohomology group. Before we state the result more precisely, let us make a simple observation about the case when Y (g) = 0. Lemma (See [Gu, Proposition E]). If Y (g) = 0 then κg ≤ 0. Furthermore, κg = 0 if and only if g is conformal to a Ricci-flat metric. This lemma neatly summarizes the zero case and allows us to restrict our attention to the case where Y (g) > 0 (see Sect. 3 for details). Let H 1 (M 4 ) denote the first deRham cohomology group. Assuming M 4 is oriented we let σ (M 4 ) denote the signature of M 4 . Theorem C. Let (M 4 , g) be a smooth compact oriented four-dimensional Riemannian manifold with κg ≥ 0. (i) Then χ (M 4 ) ≥ 23 |σ (M 4 )| with equality if and only if (M 4 , g) is either self-dual or anti-self-dual. (ii) If Y (g) > 0 then H 1 (M 4 ) = 0 unless (M 4 , g) is conformally equivalent to a quotient of S 3 × R endowed with the product metric, in which case κg = 0. Remark. Recall that χ (M 4 ) ≥ 23 |σ (M 4 )| is a necessary condition for M 4 to admit an Einstein metric ([Hi,Th]). Moreover, all Einstein metrics have κg ≥ 0. As a concrete application of Theorem C we have: Corollary 0.3. (i) For m ≥ 9 the manifold CP 2 ]m(−CP 2 ) does not admit a metric with κg ≥ 0. (ii) Let N 4 be a smooth compact oriented four–manifold. Then for any m ≥ 1 the manifold M 4 = N 4 ]m(S 3 × S 1 ) does not admit a metric with Y (g) > 0 and κg > 0. Remark. If 3 ≤ m ≤ 8 the manifold CP 2 ]m(−CP 2 ) admits a Kähler–Einstein metric ([Ti]). These metrics obviously have constant Q.
Principal Eigenvalue of a Conformally Invariant Differential Operator
135
We conclude the introduction with a few words about the proofs of Theorems A–C. For Theorem A our argument relies on some of the existence work for the functional determinant in [CY1], along with subsequent regularity work in [CGY] and a maximum principle result in [Gu]. Thus the argument is rather involved, and relies on a fair amount of existing technical work. The proof of Theorem B is conceptually easier. Indeed, if we are willing to cite Schoen’s work on the Yamabe problem, it is practically automatic. However, we will use a sharp result of Cheng [Ch] from comparison geometry to bypass Schoen’s work on the positive mass theorem. Our ability to do this reveals a curious feature of this subject which belies all of the impressive analogies with the Yamabe problem: Namely, the case of positive Yamabe invariant is much easier here, while the negative case is poorly understood. Theorem C and its corollary rely on the results in [Gu]. This is explained in Sect. 4. A final note on the organization of the paper: for technical reasons it will simplify matters if we first prove Theorem B. This is done in Sect. 1. In Sect. 2 we give some technical background to the proof of Theorem A, and give the proof proper in Sect. 3. Finally, in Sect. 4 we give the proof of Theorem C and Corollary 3. 1. The Proof of Theorem B Let us begin by recalling some basic facts about the Yamabe problem; details may be found in the survey article of Lee and Parker [LP] or in the monograph of Schoen [Sc1]. Since we are only interested in four dimensions we will specialize our account to this case. For p ∈ [2, 4] define the functional R (6|∇u|2 + Ru2 )dµ Yp [u] = R 2/p . |u|p dµ It follows from the Rellich compactness theorem that for each p ∈ [2, 4) there exists a smooth minimizer up > 0 with Yp [up ] = µp ≡ inf Yp [u]. u∈W 1,2
If we normalize so that kup kp = 1 then up satisfies the Euler equation p−1
61up + µp up
= Rup .
Since µp is continuous from the left and non-decreasing [Au], we may choose a sequence pk % 4 such that µk = µpk % Y (g). Let uk = upk and gk = u2k g. The scalar curvature Rk of gk is then given by h i − 61uk + Ruk Rk = u−3 k p −4
= µ k uk k
.
(1.1)
If we let Ek denote the trace-free Ricci tensor of gk , then as κg is a conformal invariant we have Z 1 1 2 2 − |Ek | + Rk dµgk . (1.2) κg = 4 48
136
M. J. Gursky
Using (1.1) along with the fact that dµgk = u4k dµ we have Z Z 1 1 2 |Ek | dµgk + Rk2 dµgk κg = − 4 48 Z 1 2 2p −4 µk uk k dµ ≤ 48 2(pk −2) Z Z 4−pk pk pk 1 2 p . µk uk k dµ dµ ≤ 48
(1.3)
Taking the limit as κ → ∞ in (1.3) we get κg ≤
1 (Y (g))2 . 48
(1.4)
√ By the energy estimate of Aubin [Au], Y (g) ≤ Y (S 4 ) = 8 6π . Thus if Y (g) ≥ 0 from (1.4) we conclude κg ≤ 8π 2 . Now suppose κg = 8π 2 . If we do not omit the term involving the trace–free Ricci tensor in (1.3) then we get 1 8π + 4 2
Z |Ek | dµgk 2
1 2 µ ≤ 48 k
Z
2(pk −2) Z
p uk k dµ
pk
4−pk dµ
pk
.
(1.5)
Taking the limit as κ → ∞ in (1.5) and using Aubin’s estimate once again we conclude Z (1.6) |Ek |2 dµgk = 0. lim k→∞
There are now two possibilities to consider, depending on whether the sequence {uk } is precompact or not. First let us assume that Z u2k dµ = 0. lim k→∞
In this case, we can find a subsequence (still denoted {uk }) and points {P1 , . . . , Pm } ⊂ M 4 such that uk → 0 in C ∞ on compact sets K ⊂ M 4 \{P1 , . . . , Pm } (see [Sc1] for details). Choose another point x0 6 ∈ {P1 , . . . , Pm } and let vk = uk /uk (x0 ). By the Harnack inequality vk → v > 0 in C ∞ on compact K ⊂ M 4 \{P1 , . . . , Pm }, and v satisfies −61v + Rv = 0. (1.7) Since we may assume that R > 0 it follows that v is singular at some of the Pi ’s; say P1 , . . . , Pm0 i.e., v is a sum of Green’s functions for the conformal Laplacian L = −61+R. If we consider the sequence of metrics hk = vk2 g, then by the scale–invariance of the L2 -norm of the Ricci tensor we have Z Z |Ehk |2 dµhk = |Ek |2 dµgk . Thus by (1.6) we see that Ehk → 0 on compact K ⊂ M 4 − {P1 , . . . , Pm0 }, and conclude that h = v 2 g is Einstein. By (1.7) h is also scalar-flat (because Rh = v −3 (−61v+Rv) =
Principal Eigenvalue of a Conformally Invariant Differential Operator
137
0), so in fact h is Ricci-flat. It then follows from a result of Schoen ([Sc2]) that (M 4 , g) is conformally equivalent to the round sphere. On the other hand, suppose Z u2k > 0. lim k→∞
It is then standard that {uk } contains a subsequence which converges to a non-zero smooth solution of the Yamabe problem; call it v. Just as before we can argue that the metric h = v 2 g is in fact Einstein. Moreover, Z 1 2 Rh2 dµh . (1.8) 8π = 48 Let us normalize h so that Rich = 3h; then Rh = 12 and by (1.8) vol(M 4 , h) = 83 π 2 . Let D denote the diameter of (M 4 , h). By Myer’s theorem, D ≤ π. Also, by the Bishop comparison theorem, Z D 8 2 4 3 π = vol(M , h) ≤ vol(S ) sin3 r dr 3 0 2 2 1 3 cos D − cos D + = 2π 3 3 so that the diameter satisfies 1 2 ≤ cos3 D − cos D. 3 3 It is easy to verify that the function f (x) = 13 cos3 x − cos x attains its maximum value of 2/3 on the interval [0, π] at x = π; thus D = π . It then follows from a result of Cheng [Ch] that (M 4 , g) is isometric to the round sphere. 2. Technical Preliminaries to Theorem A In this section we provide some necessary background material for the proof of Theorem A, beginning with the work of [CY1] on the functional determinant. Let (M 4 , g0 ) be a smooth compact four-dimensional Riemannian manifold. We then define the functionals Z Z Z 2 2 |W0 | dµ0 log e4w dµ0 , (2.1) I [w] = 4 |W0 | wdµ0 − I I [w] =
Z
Z
Z wP0 wdµ0 + 4
I I I [w] = 12
Q0 wdµ0 −
Q0 dµ0 log
Z e4w dµ0 ,
Z 2 10 w + |∇0 w|2 dµ0 − 4 R0 |∇0 w|2 dµ0 Z − 4 10 R0 wdµ0 ,
(2.2)
Z
(2.3)
and linear combinations of the above: 8[w] = γI [w] + γ2 I I [w] + γ3 I I I [w].
(2.4)
138
M. J. Gursky
In (2.1)–(2.4) terms have the subscript 0 to indicate that they are computed with respect to g0 . Also, notice that the functional I I is just the functional F we defined in (0.8). The “sub-functionals” I–III arise in the calculation of the functional determinant of a conformally covariant operator A due to Branson and Orsted [BO]. For example, if A = L = −61 + R, the conformal Laplacian, then log
det Lg 2 = I [w] − 4I I [w] − I I I [w], det L0 3
where g = e2w g0 . If M 4 is spin and A = 6∇ 2 , then log
det 6∇g2 det 6 ∇02
= 7I [w] − 88I I [w] −
14 I I I [w]. 3
In general, any operator A which satisfies certain “conformal” and “naturality” assumptions defined in [BO] has log det Ag / det A0 = γ1 I [w]+γ2 I I [w]+γ3 I I I [w] for some choice of γi = γi (A). The main existence result of [CY1] for extremals of the functional determinant can be stated as follows: Theorem 2.1 ([CY1, Theorem 1.1]). Let 8 be defined as in (2.4). If γ2 < 0, γ3 < 0, and Z γ1 |W0 |2 dµ + 8π 2 , κg0 < − γ2 then supw∈W 2,2 8[w] is attained by some w ∈ W 2,2 . Moreover, the extremal metric g = e2w g0 satisfies the Euler equation γ1 |W |2 + γ2 Q − γ3 1R = constant.
(2.5)
Remarks. 1. It follows from [CGY] that extremal metrics are smooth. 2. In (2.5) the curvature terms and Laplace–Beltrami operator are with respect to the extremal metric. 3. Using Theorem B we can now give the proof of Corollary 2. As we observed above, when A is the conformal Laplacian then (γ1 , γ2 , γ3 ) = (1, −4, − 23 ). Thus existence will follow if we can show that Z 1 (2.6) |W0 |2 dµ0 + 8π 2 . κg0 < 4 By Theorem B, (2.6) holds unless (M 4 , g0 ) is conformally equivalent to the round sphere. However, this case was treated separately in [BCY]. If M 4 is spin and A = 6∇ 2 then (γ1 , γ2 , γ3 ) = (7, −88, − 14 3 ), so existence follows if we can show κg0 <
7 88
Z |W0 |2 dµ0 + 8π 2 .
(2.7)
Again, (2.7) is satisfied except in the case of the round sphere, and this was treated in [BCY] as well.
Principal Eigenvalue of a Conformally Invariant Differential Operator
139
For our purposes it will be helpful to rewrite the Euler equation (2.5). Let E denote the trace-free Ricci tensor; then 1R = λ − α|W |2 − β|E|2 + with α = γ1 (γ3 +
1 βR 2 12
1 γ2 )−1 , 12
1 1 γ2 (γ3 + γ2 )−1 , 4 12 Z −1 n Z λ= dµ α |W0 |2 dµ0 − 4βκg0 }. β=
(2.8)
(2.9) (2.10) (2.11)
Notice that if γ2 < 0 and γ3 < 0 as required by Theorem 1.1, then β > 0. Furthermore if γ1 ≥ 0 and κg0 ≥ 0 then α ≤ 0 and λ ≤ 0. This is important, for the following reason: Proposition 2.2 (See [Gu, Lemma 1.2]). Suppose Y (g0 ) > 0. If α ≥ 0, 0 ≤ β ≤ 2, and λ ≤ 0, then R > 0. We will see that the pointwise positivity of the scalar curvature when Y (g0 ) > 0 is crucial to our argument. The following proposition assembles all of the material thus reviewed into one concise – and for our purposes useful – statement. Proposition 2.3. Let (M 4 , g0 ) be a smooth compact four-dimensional Riemannian manifold with Y (g0 ) > 0 and κg0 ≥ 0. Then there is a conformal metric g = e2w g0 which satisfies 1 (2.12) 1R = λ − |E|2 + R 2 12 with λ ≤ 0 and R > 0. Proof. Just take γ1 = 0, γ2 = 6, and γ3 = 1. By Theorem B, κg0 < 8π 2 unless (M 4 , g0 ) is conformally equivalent to the sphere. In the latter case, the round metric satisfies (2.12). In the former case, we can quote the existence result of [CY1] as stated in Theorem 2.1, along with the maximum principle result Proposition 2.2 to get the positivity of R. u t 3. The Proof of Theorem A Let (M 4 , g0 ) satisfy κg0 ≥ 0 and Y (g0 ) ≥ 0. The case where Y (g0 ) = 0 was addressed in the introduction; recall that in this case κg0 ≤ 0 with equality if and only if g0 is conformal to a Ricci-flat metric g = e2w g0 . Thus if Y (g0 ) = 0 and κg0 = 0 then with respect to g one has P = 12g . Clearly Theorem A follows. So from now on assume Y (g0 ) > 0 and κg0 ≥ 0. By Proposition 1.3 there is a metric g = e2w g0 satisfying 1 (3.1) 1R = λ − |E|2 + R 2 12
140
M. J. Gursky
with λ ≤ 0 and R > 0. Let φ ∈ C ∞ (M 4 ) be any function; we wish to estimate D
P φ, φ
Z
E
=
L2
φPg φdµg .
Using the definition of P and integrating by parts we have (see [CY1, (1.3)]) D E P φ, φ
L2
=
Z h i 1 (1φ)2 + R|∇φ|2 − 2E(∇φ, ∇φ) 6
(3.2)
(from now on we will omit the subscript g and the volume form dµ in order to simplify notation). Theorem A will be an easy consequence of the following lemma: Lemma 3.1. For any φ ∈ C ∞ (M), Z
Z −2E(∇φ, ∇φ) ≥
−(1φ)2 −
1 R|∇φ|2 . 48
(3.3) √
Proof. Since E is trace–free, we have the inequality −E(∇φ, ∇φ) ≥ − 23 |E||∇φ|2 . Thus, Z Z √3 |E||∇φ|2 −2E(∇φ, ∇φ) ≥ −2 2 Z |E|2 3 ≥ −2 |∇φ|2 − R|∇φ|2 , R 8 √ where we have used the arithmetic–geometric mean inequality −2 23 xy ≥ −2x 2 − 3 2 8y .
Now using (3.1) we have Z
Z
1 13 |∇φ|2 (−|E|2 + R 2 ) − R|∇φ|2 R 12 24 Z 13 |∇φ|2 (1R − λ) − R|∇φ|2 = 2 R 24 Z 13 |∇φ|2 1R − R|∇φ|2 , ≥ 2 R 24
−2E(∇φ, ∇φ) ≥
2
(3.4)
the last line following from the fact that λ ≤ 0. Considering the first term on the right hand side of (3.4), we integrate by parts to get Z 2|∇φ| R 2
−1
Z 1R = Z =
E D −2|∇φ|2 ∇(R −1 ), ∇R E D − 2R −1 ∇|∇φ|2 , ∇R 2
D ∇R E |∇R|2 2 2 |∇φ| − 2 . , ∇|∇φ| R2 R
Principal Eigenvalue of a Conformally Invariant Differential Operator
D
Using the simple inequality | Z
∇R 2 R , ∇|∇φ|
E
2 | ≤ 2 |∇R| R |∇φ||∇ φ|, we have
Z
|∇R|2 |∇R| |∇φ||∇ 2 φ| |∇φ|2 − 4 2 R R Z |∇R|2 |∇R|2 2 |∇φ| − 2 |∇φ|2 − 2|∇ 2 φ|2 ≥ 2 R2 R2 Z = −2|∇ 2 φ|2 .
2|∇φ|2 R −1 1R ≥
141
2
Substituting (3.5) into (3.4) we get Z Z 13 −2E(∇φ, ∇φ) ≥ −2|∇ 2 φ|2 − R|∇φ|2 . 24
(3.5)
(3.6)
We now use the integrated Bochner formula: Z Z |∇ 2 φ|2 = (1φ)2 − Ric(∇φ, ∇φ) Z 1 = (1φ)2 − E(∇φ, ∇φ) − R|∇φ|2 . 4 Substituting this into (3.6) gives Z Z 1 −2E(∇φ, ∇φ) ≥ −2(1φ)2 + 2E(∇φ, ∇φ) − R|∇φ|2 24 ⇒ Z Z 1 t u −2E(∇φ, ∇φ) ≥ −(1φ)2 − R|∇φ|2 . 48 To complete the proof of Theorem A, combine (3.2) and (3.3) to get Z D E 7 R|∇φ|2 . P φ, φ 2 ≥ L 48 t Thus λ1 (P ) = 0 and Ker P = {constants}. u 4. The Proof of Theorem C The basis for Theorem C is the Chern–Gauss–Bonnet formula along with sharp L2 estimates for the self-dual part of the Weyl tensor in [Gu]. Recall that if (M 4 , g) is oriented, then the bundle of two–forms splits 32 = 32+ ⊕ 32− into its self-dual and anti-self-dual sub–bundles. This in turn induces a splitting of the Weyl curvature W = W + ⊕ W − , viewed as a bundle endomorphism W : 32 → 32 . To prove (i) of Theorem C, we combine the Hirzebruch signature formula Z 12π 2 σ (M 4 ) = (|W + |2 − |W − |2 )dµ with the Chern–Gauss–Bonnet formula (0.3) to get Z Z 2π 2 2χ (M 4 ) ± 3σ (M 4 ) = |W ± |2 dµ + Qdµ.
(4.1)
142
M. J. Gursky
Thus, if κg ≥ 0 we conclude 2χ(M 4 ) + 3σ (M 4 ) ≥ 0, with equality if and only if either W + ≡ 0 or W − ≡ 0. Part (i) follows. To prove part (ii), we appeal to the following result in [Gu]: Theorem ([Gu, Theorem 2]). Let M 4 be a closed, compact, oriented four-dimensional manifold with H 1 (M 4 ) 6 = 0. Then for any metric g on M 4 satisfying Y (g) > 0, Z (4.2) |W + |2 dµ ≥ 2π 2 2χ (M 4 ) + 3σ (M 4 ) . Furthermore, equality is attained in (4.2) by some metric g0 with Y (g0 ) > 0 if and only if (M 4 , g0 ) is conformal to a quotient of S 3 × R equipped with the product metric. Assuming Y (g0 ) > 0 and H 1 (M 4 ) 6 = 0, combining (4.1) and (4.2) we get Z Z 2π 2 2χ (M 4 ) + 3σ (M 4 ) = |W + |2 dµ + Qdµ ≥ 2π 2 2χ(M 4 ) + 3σ (M 4 ) + κg so that κg ≤ 0. Furthermore, if κg = 0 then equality is attained in (4.2), and (M 4 , g0 ) is conformal to a quotient of the product S 3 × R. The proof of Corollary 3 is immediate. u t Acknowledgements. It is a pleasure to thank S.Y.A. Chang and P.Yang for bringing some of problems addressed in this article to our attention, and for their encouragement during the preparation of the manuscript.
References [Ad] [Au] [BCY] [Be] [BO] [CGY] [Ch] [CY1] [CY2] [ES1] [ES2] [Gu] [Hi] [Li] [LP]
Adams, D.: A sharp inequality of J. Moser for higher order derivatives. Ann. of Math. 128, 385–398 (1988) Aubin, T.: Equation differentielles non lineaires et probleme de Yamabe concernant la courbure scalatre. J. Math. Pures Appl. 55, 269–296 (1976) Branson, T., Chang, S.Y.A. and Yang, P.: Estimates and extremals for zeta function determinants on four-manifolds. Commun. Math. Phys. 149, 241–262 (1992) Besse, A.: Einstein manifolds. Berlin: Springer-Verlag, 1987 Branson, T. and Orsted, B.: Explicit functional determinants in four dimensions, Proc. Am. Math. Soc. 113, 669–682 (1991) Chang, S.Y.A., Gursky, M.J. and Yang, P.: Regularity of a fourth order non-linear PDE with critical exponent. To appear in Am. J. Math. Cheng, S.Y.: Eigenvalue comparison theorems and its geometric application. Math. Z. 143, 289–297 (1975) Chang, S.Y.A. and Yang, P.: Extremal metrics of zeta function determinants on 4-manifolds. Ann. of Math. 142, 171–212 (1995) Chang, S.Y.A. and Yang, P.: On uniqueness of solutions of a n-th order differential equation in conformal geometry. Math. Res. Letters 4, 91–102 (1997) Eastwood, M.G. and Singer, M.A.: A conformally invariant Maxwell gauge. Phys. Lett. A 107, 73–74 (1985) Eastwood, M.G. and Singer, M.A.: The Frohlicher spectral sequence on a twistor space. J. Diff. Geom. 38, 653–669 (1993) Gursky, M.J.: The Weyl functional, deRham cohomology, and Kahler–Einstein metrics. Ann. of Math. 148, 315–337 (1998) Hitchin, N.: On compact four-dimensional Einstein manifolds. J. Diff. Geom. 9, 435–442 (1974) Lin, C.S.: A classification of solutions of a conformally invariant fourth order equation in R n . Commun. Math. Helv. 73, 206–231 (1998) Lee, J.M. and Parker, T.H.: The Yamabe problem. Bull. Am. Math. Soc. 17, 37–91 (1987)
Principal Eigenvalue of a Conformally Invariant Differential Operator
[Pa] [Sc1] [Sc2] [Th] [Ti] [WX]
143
Paneitz, S.: A quartic conformally covariant differential operator for arbitrary pseudo-Riemannian manifolds. Preprint, 1983 Schoen, R.: Variational theory for the total scalar curvature functional for Riemannian metrics and related topics. In: Topics in Calculus of Variations, Seminar 1987 (ed. M. Giaquinta), LNM 1365, New York: Springer-Verlag, 1989, pp. 120–154 Schoen, R.: Conformal deformation of a Riemannian metric to constant scalar curvature. J. Diff. Geom. 20, 479–495 (1984) Thorpe, J. Some remarks on the Gauss–Bonnet integral. J. Math. Mech. 18, 779–786 (1969) Tian, G.: On Calabi’s conjecture for complex surfaces with positive first Chern class. Invent. Math. 101, 101–172 (1990) Wei, J. and Xu, X.: On conformal deformations of metrics on S n . J. Funct. Anal. 157, 292–325 (1998)
Communicated by A. Connes
Commun. Math. Phys. 207, 145 – 171 (1999)
Communications in
Mathematical Physics
© Springer-Verlag 1999
Multifractal Analysis of Lyapunov Exponent for Continued Fraction and Manneville–Pomeau Transformations and Applications to Diophantine Approximation Mark Pollicott1 , Howard Weiss2,? 1 Department of Mathematics, The University of Manchester, Oxford Road, M13 9PL, Manchester, UK.
E-mail: [email protected]
2 Department of Mathematics, The Pennsylvania State University, University Park, PA 16802, USA.
E-mail: [email protected] Received: 13 October 1998 / Accepted: 19 April 1999
Abstract: We extend some of the theory of multifractal analysis for conformal expanding systems to two new cases: The non-uniformly hyperbolic example of the Manneville– Pomeau equation and the continued fraction transformation. A common point in the analysis is the use of thermodynamic formalism for transformations with infinitely many branches. We effect a complete multifractal analysis of the Lyapunov exponent for the continued fraction transformation and as a consequence obtain some new results on the precise exponential speed of convergence of the continued fraction algorithm. This analysis also provides new quantitative information about cuspital excursions on the modular surface. 1. Introduction In this paper we extend some aspects of the multifractal analysis which are useful in studying problems in Diophantine approximation, in studying the behavior of geodesics on the modular surface, and in studying an important non-uniformly hyperbolic dynamical system. In particular, we first study the continued fraction (Gauss) transformation T1 : [0, 1] → [0, 1] defined by 1 1 1 = , T1 x ≡ − x x x for x 6 = 0 and T1 (0) ≡ 0. Here [1/x] denotes the integer part of 1/x. This map (see Fig. 1) is uniformly hyperbolic, but being naturally coded by an infinite alphabet and having infinite topological entropy, the usual theory of multifractal analysis for conformal maps [PW1,PW2] does not directly apply. ? The work of the second author was partially supported by a National Science Foundation grant DMS9704913. The manuscript was completed during the second author’s sabbatical visit at IPST, University of Maryland, and he wishes to thank IPST for their gracious hospitality.
146
M. Pollicott, H. Weiss 1
0
1/4
1/3
1/2
1
Fig. 1. Graph of the map T1
For the continued fraction transformation, our multifractal analysis of the Lypaunov exponent will yield new detailed information about the precise rates of Diophantine approximation to irrational numbers. This has immediate implications for cuspital excursions of geodesics on the modular surface. We also study the Manneville–Pomeau transformation [MP1] defined by (
T2 : [0, 1] → [0, 1] T2 x = x + x 1+α mod 1,
where α is a non-negative constant. This important model (see Fig. 2a) is a non-uniformly hyperbolic transformation having the most benign type of non-hyperbolicity: an indifferent fixed point at 0, i.e., T2 (0) = 0 and T20 (0) = 1, and exhibits intermittent behavior [MP1]. In general one could not expect the full force of multifractal analysis to apply for general non-uniformly hyperbolic systems. However, we show that part of the theory carries over to this, and similar, transformations. A key aspect of our analysis is a reduction, via inducing, of this system to a countable state uniformly hyperbolic system (i.e., using the Schweiger jump transformation). In particular, we show that the Lyapunov exponent attains an interval of values, realized on dense sets with positive Hausdorff dimension. This quantifies the size of the set of points whose orbits spend a disproportionate amount of time near the indifferent fixed point. A major tool in the multifractal analysis is the use of symbolic dynamics and thermodynamic formalism, i.e., pressure (and its derivatives), equilibrium states, etc. In the present context of infinite state subshifts of finite type, we are fortunate in having at our disposal a theory worked out by Walters [Wa1].
Multifractal Analysis of Lyapunov Exponent for Continued Fraction
147
1
0
a
1
0
Fig. 2a. Graph of the map T2 1
0
a
3 I 3
a
a
2 I 2
a
1
1
0
I1
Fig. 2b. Graph of the map Tb2
I0
148
M. Pollicott, H. Weiss
Finally, while we state our main results for these two model transformations, our analysis works in greater generality. Our results are valid for interval maps which have induced maps (coded by an infinite alphabet) which satisfy the EMR (expanding-MarkovRényi) property (see Sect. 2) and potentials (for equilibrium states) which satisfy properties (W1) and (W2) (see Sect. 2). In particular, our analysis works for the family of functions, which Prellberg calls Cs [Pr], containing certain piecewise monotone transformations of the interval with an indifferent fixed point at the boundary. For instance, the Farey map ( x/(1 − x), 0 ≤ x < 1/2 F (x) = (1 − x)/x, 1/2 ≤ x ≤ 1 belongs to the class C1 . There are other examples to which a similar sort of analysis extends, particularly in the realm of number theoretic analysis. Our analysis applies to the family of maps Ts (x) = {1/(s(1 − x))} for 0 < s < 4 and s 6 = 4 cos2 (π/q), q = 3, 4, · · · , and there is a theory of s−backward continued fractions based on this family of maps [GH]. Our analysis of the Manneville–Pomeau map easily extends to the complex continued fraction algorithm [Sc, §23.6.], where the underlying map is conformal and has an indifferent fixed point. The underlying map for the Jacobi Perron Algorithm [Sc, §23.1] for continued fractions in several variables is hyperbolic with infinitely many branches, but is not conformal. In this case, a multifractal analysis based on local entropy rather than pointwise dimension should routinely follow. Shortly after the completion of this manuscript, the authors were given a recent preprint [N] which contains an argument corresponding to Proposition 3(4). 2. Markov Maps and Inducing Let I denote an interval of real numbers. For the maps T1 and T2 the interval I will be [0, 1]. The study of our two transformations can be reduced to studying transformations of the following general form: Definition. We say that a transformation T : I → I is an EMR Transformation if we may write I = ∪∞ n=0 In as a countable union of closed intervals (with disjoint interiors ◦
In ), which we call basic subintervals, such that ◦
(1) The map T is C 2 on ∪∞ k=1 I n . (2) Some power of T is uniformly expanding, i.e., there exists a positive integer r and ◦
α > 0 such that |(T r )0 (x)| ≥ α > 1 for all x ∈ ∪∞ n=1 I n . (3) The map T is Markov. (4) The map T satisfies Renyi’s condition, i.e., there exists a positive number K such that |T 00 (x)| ≤ K < ∞. sup sup 0 0 n x,y,z∈In |T (y)||T (z)| ◦
It is easy to verify that |(T12 )0 (x)| ≥ 4 for all x ∈ ∪∞ n=1 I n . Proposition 1 ([Sc, Wa1, p. 148]). The continued fraction transformation T1 satisfies EMR.
Multifractal Analysis of Lyapunov Exponent for Continued Fraction
149
For α = 0 the Manneville–Pomeau transformation is the usual doubling map of the circle. For 0 < α < 1, Thaler [T] constructed an finite absolutely continuous invariant measure. For α > 1, Thaler also constructed a sigma-finite but not finite absolutely continuous invariant measure. However, we shall not consider this range. The Manneville–Pomeau map has two branches for [0, a0 ] and [a0 , 1] where 1 = a0 + a0α . There is a natural topological conjugacy of this map to the standard doubling map and has topological entropy equal to log 2. Although this map is not hyperbolic since the derivative T20 (0) = 1, the induced map on [a0 , 1] is hyperbolic [Sc]. More precisely, we choose the monotone decreasing sequence an → 0 such that T (an+1 ) = an and define In = [an , an−1 ], for n ≥ 1, and I0 = [a0 , 1]. For an+1 < x < an we define Tb2 (x) = T2n (x) (see Fig. 2B). The map Tb2 is piecewise analytic on countably many intervals {In } and is uniformly hyperbolic. Each point which is not an end point of an interval In has two pre-images under T2 . Proposition 2 ([T, pp. 312–313, I, §2]). For 0 < α < 1, the transformation Tb2 satisfies EMR. We will first effect a multifractal analysis for the map Tb2 and then transfer the multifractal analysis to the map T2 . −k (∪∞ ∂I ), where For an EMR transformation T we define the set O = ∪∞ k=0 T k=0 k subset of I . ∂Ik denotes the two endpoints of the interval Ik . Clearly O is a countable −i {I } ∞ of the original T We denote by In (x) the element of the refinement ∨n−1 n i=0 n=1 partition containing x. Every x ∈ I \ O has a unique symbolic coding since for every k ∈ N there is a unique basic subinterval that contains T k (x). A fundamental property of an EMR transformation is that it satisfies what is usually called the Jacobian Estimate [CFS, p. 171], i.e., there exists positive K such that for all x ∈ I \ O one has n 0 (T ) (x) 1 ≤ sup sup n 0 ≤ K < ∞. (J) 0< K n≥0 y∈In (x) (T ) (y) This property will be exploited many times in the proof of Proposition 3 and is also an essential component in the proof of Proposition 2. For the reasons mentioned in the introduction, it is natural that for infinitely coded maps like T1 and T2 , the class of potentials which admit unique equilibrium states is more restrictive than in the usual case. We will now discuss a class of potentials, which we denote W = W(T ) (for Walters), for which the usual results in thermodynamic formalism are valid. Let T : I → I be an EMR transformation and φ : I → R a function such that exp φ is continuous and satisfies the following two properties: P (W1) There exits a constant C > 0 such that the sum T y=x exp φ(y) ≤ C, for all x ∈ I. (W2) The function Cφ (x, x 0 ) = sup sup
n−1 X φ(T i y) − φ(T i y 0 )
n≥1 y∈T −n x i=0
is bounded by a constant Cφ and Cφ (x, x 0 ) tends to zero as |x − x 0 | → 0.
150
M. Pollicott, H. Weiss
Such potentials will be said to belong to the class W . The thermodynamic Pressure P (φ) can be defined for a continuous function φ via the variational principle as Z P (φ) = sup hµ (T ) + φdµ , µ T −inv
I
where hµ (T ) denotes the measure theoretic entropy of T , and the supremum is taken over all T −invariant Borel probability measures µ [Wa2]. Example 1. It is an important feature of allowable potentials for the continued fraction transformation T1 that φ(x) → −∞ as x → 0, in contrast to the usual boundedness of potentials for EMR maps having only finitely many intervals. Observe also that for T1 , the non-zero constant function never satisfies (W1). For T1 with the piecewise analytic potential φ = −t log |T10 | we have that X
exp φ(y) ≤
∞ X 1 < ∞, n2t n=1
T1 y=x
and condition (W1) holds for t > 21 . Condition (W2) can similarly be seen to hold on the same range [Wa1]. Let x, x 0 ∈ X, y ∈ T1−n x and let y 0 be the corresponding point of T1−n x 0 . Two applications of the Mean Value Theorem yield that T100 (zi ) i i 0 |φ(T y) − φ(T y )| = t 0 |x − x 0 |, T (zi )(T n−i )0 (wi ) 1
1
where wi and zi lie between T1i y and T1i y 0 . Using properties (2) and (4) of EMR we obtain that |φ(T i y) − φ(T i y 0 )| ≤ Kt|x − x 0 |/α n−i−1 and thus n−1 X φ(T i y) − φ(T i y 0 ) ≤ Kt|x − x 0 |/(α − 1). i=0
Condition (W2) immediately follows. On the range t > 21 , the function t 7 → P (−t log |T10 |) is analytic since the function exp(P (−t log |T10 |)) is an isolated eigenvalue for the associated transfer operator, about which we shall say more in [M1, Sect. VI]. It is also strictly convex (see Fig. 3). Moreover, as in the usual theory, one can use perturbation theory to compute the second derivative of exp(P (−t log |T10 |)) and deduce this is strictly convex. It follows from the standard Rohlin equality that P (− log |T10 |) = 0. Mayer [M1] has also shown that the function t 7 → P (−t log |T10 )| has a logarithmic singularity at t = 21 . This will explain the range of values in the statement of Theorem 1. Example 2. For the induced Manneville–Pomeau transformation Tb2 with the piecewise analytic potential φ = −t log |Tb20 |, we have that X Tb2 y=x
exp φ(y) =
X Tb2 y=x
∞
X 1 = |G0n (x)|t , 0 t |Tb2 (y)| n=1
Multifractal Analysis of Lyapunov Exponent for Continued Fraction
1/2
151
1
Fig. 3. Graph of t 7 → P (−t log |T10 |)
where Gn = F1 F0n−1 , and F0 , F1 denote the two branches of the inverse of T2 . Properties (W1) and (W2) for Tb2 are established in [T, p. 312] and [I, Lemma 2.2] for t < 1. Prellburg [Pr,PS,V] showed that the function P (−t log |T20 |) = 0 for t > 1, and that on the range 1/α < t < 1, the function t 7 → P (−t log |T20 |) is analytic and strictly convex Fig. 4). It follows from the Rohlin equality that P (− log |T20 |) = 0. Using the variational principle, it is easy to see that for all values of t we have that P (−t log |T20 |) ≥ R0, since if we take µ to be the Dirac measure supported at 0, then hµ (T2 ) = 0 and I log |T20 |dµ = 0. Lopes [L] has studied the precise nature of the singularity of the map t 7 → P (−t log |T10 |) at t = 1. This explains the range of values in the statement of Theorem 2. Lyapunov exponents. Lyapunov exponents measure the exponential rate of divergence of infinitesimally close orbits of a smooth dynamical system. These exponents are intimately related with the global stochastic behavior of the system and are fundamental invariants of a smooth dynamical system. For a transformation T : I → I we define the Lyapunov exponent λ(x) of T by n−1
Y 1 1 log |(T n )0 (x)| = lim log |T 0 (T i x)|, n→∞ n n→∞ n
λ(x) ≡ lim
(3)
i=0
when the limit exists. The function λ(x) is clearly T −invariant. There is a natural decomposition of the interval I by level sets of the Lyapunov exponent Lβ = {x ∈ I : λ(x) = β}, [ Lβ ∪ {x ∈ I | λ(x) does not exist} . I= −∞<β<∞
152
M. Pollicott, H. Weiss
1
0
1
Fig. 4. Graph of t 7 → P (−t log |T20 |)
To study this complicated decomposition we introduce the Lyapunov spectrum by considering the level sets of the Lyapunov exponent and by defining g(β) = dimH (Lβ ), where dimH (Lβ ) denotes the Hausdorff dimension of Lβ . In the next section we relate the Lyapunov spectrum to a related spectrum for local dimension and prove several remarkable properties about it. 3. The Multifractal Analysis of Equilibrium States The general concept of a multifractal analysis for a dynamical system concerns a detailed study of the exceptional behavior of asymptotically defined dynamical quantities such as pointwise dimension, Lyapunov exponent, local entropy, Birkhoff average, etc. In many examples with hyperbolic structure these quantities are constant almost everywhere, with respect to an appropriate ergodic measure. We consider two notions of local or pointwise dimension with respect to an invariant measure. The pointwise dimension of a Borel probability measure µ defined on I is defined by log µ(B(x, r)) , r→0 log r
dµ (x) = lim
(1)
when the limit exists. Here B(x, r) denotes the ball of radius r centered at the point x. The Markov pointwise dimension of a T −invariant Borel probability measure µ is defined on I , defined by log µ(In (x)) , δµ (x) ≡ lim n→∞ − log `(In (x))
Multifractal Analysis of Lyapunov Exponent for Continued Fraction
153
when the limit exists. Here the intervals {In } are those in the definition of ERM transformation and `(In (x)) denotes the length of the interval In (x). In [PW1,PW2] the authors establish deep relationships between these two notions of local dimension for equilibrium states for conformal expanding maps. By definition conformal expanding maps are local homeomorphisms, which the maps we consider in this paper are not. Not surprisingly, some of relationships that are valid at every point in certain sets are no longer true, and this is a major obstacle in generalizing the usual multifractal theory to these more general transformations. There are natural decompositions of the interval I by level sets Kα = {x : dµ (x) = α}, [ Kα ∪ {x ∈ I | dν (x) does not exist} , I= −∞<α<∞
and by the level sets KαM = {x : δµ (x) = α}, [ KαM ∪ {x ∈ I | δν (x) does not exist} . I= −∞<α<∞
Since we are mostly interested in effecting a multifractal analysis of the Lyapunov exponent, we only analyze the latter decomposition. To study this decomposition we define the Markov dimension spectrum fµ (α) = dimH (KαM ), where dimH (KαM ) denotes the Hausdorff dimension of the set KαM . This is similar to the Lyapunov spectrum we defined in Sect. 2. The parts of the multifractal analysis which we establish are that under suitable hypotheses, the function fµ (α) is real analytic and strictly convex (on a suitable interval) and is given in terms of thermodynamic formalism. More precisely, let T : I → I be an EMR transformation, and let φ ∈ W. The equilibrium state µ is a T -invariant probability measure µ such that there exists a positive C such that µ(In (x)) 1 ≤ C, ≤ P j C exp(−nP (φ) + n−1 j =0 φ(T y))
(2)
for all x ∈ I and y ∈ In (x). We assume that φ is not cohomologous to log |T 0 | [PW1]. Let ψ be the positive function such that log ψ = φ − P (φ). Clearly ψ ∈ W, the pressure P (log ψ) = 0, and µ is also the equilibrium state for log ψ. For these potentials it follows from the Jacobian estimate (J) that the Markov pointwise dimension satisfies Q i log n−1 log µ(In (x)) i=0 ψ(T x) = lim , δµ (x) ≡ lim n→∞ − log `(In (x)) n→∞ − log |(T n )0 (x)| when the limits exist. The following proposition establishes relationships between the two notions of local dimension for equilibrium states and the Lyapunov exponent. Proposition 3. Let T : I → I be an EMR transformation and let µ be the equilibrium state corresponding to the potential φ ∈ W. Let φ(x) denote the Birkhoff average P k limn→∞ 1/n n−1 i=0 φ(T (x)) at x and consider x ∈ I \ O.
154
M. Pollicott, H. Weiss
(1) Suppose that δµ (x) exists. Then d µ (x) ≤ δµ (x). If φ(x) also exists, then d µ (x) ≥ δµ (x). In this case dµ (x) = δµ (x) =
−log ψ(x) P (φ) − φ(x) = , λ(x) λ(x)
where λ(x) denotes the Lyapunov exponent at x. (2) Suppose that dµ (x) exists. Then δ µ (x) ≥ dµ (x). If φ(x) also exists, then δ µ (x) ≤ dµ (x). In this case δµ (x) = dµ (x). (3) If exp φ is uniformly bounded away from zero, then the pointwise dimension dµ (x) = γ if and only if the Markov pointwise dimension δµ (x) = γ . (4) If T = T1 or T = T2 and log ψ = φ − P (φ) is bounded away from zero, then the Markov pointwise dimension δµ (x) = γ implies that the pointwise dimension dµ (x) = γ . Remark. We caution that the hypothesis in (3) does not hold for the important family of potentials φ = φs = −s log |T10 | since exp φs = x 2s . As noted in Example 1 of Sect. 2, every potential φ ∈ W for the continued fraction map T 1 must approach 0 as x → 0 and thus (3) will not hold for any such potential for this map. However, Property (4) is satisfied (at least for s close to 1). This is an important difference from the multifractal analysis of the classical conformal expanding maps where we consider Hölder continuous potentials and where this condition is always satisfied. Proof of Proposition 3. The Jacobian estimate (J) allows us to estimate the lengths (`) of the intervals In (x) using the derivative of T [PW1], i.e., there exist positive constants C1 and C2 such that for all x ∈ I \ O and all n ∈ N, we have C1 ≤
`(In (x)) ≤ C2 . |(T n )0 (x)|−1
(4)
Suppose that the Markov pointwise dimension δµ (x) exists at a point x ∈ I \O. Given r > 0 there exists a unique n = n(r) such that C1 |(T n )0 (x)|−1 < r ≤ C1 |(T n−1 )0 (x)|−1 . It immediately follows from (4) that log µ(B(x, C1 |(T n−1 )0 (x)|−1 ) log µ(In−1 (x)) log µ(B(x, r)) ≥ ≥ (5) log r log r log r log µ(In−1 (x)) log µ(In (x)) log µ(In−1 (x)) = . (6) ≥ log(C1 |(T n )0 (x)|−1 ) log µ(In (x)) log(C1 |(T n )0 (x)|−1 ) It follows from the definition of equilibrium state that P j exp(−(n − 1)P (φ) + n−2 µ(In−1 (x)) j =0 φ(T x)) , Pn−1 µ(In (x)) exp(−nP (φ) + j =0 φ(T j x)) and thus if φ(x) exists, then lim
n→∞
log µ(In−1 (x)) = 1. log µ(In (x))
We obtain that if δµ (x) exists, then d µ (x) ≥ δµ (x). This proves the first part of (1).
Multifractal Analysis of Lyapunov Exponent for Continued Fraction
155
We note that by replacing the potential φ by log ψ, we obtain the following expression: Qn−2 j 1 µ(In−1 (x)) j =0 ψ(T x) = Qn−1 , n−1 x) j µ(In (x)) ψ(T j =0 ψ(T x) and if we assume that ψ (or exp φ) is uniformly bounded away from zero, then clearly the quotient µ(In−1 (x))/µ(In (x)) will be uniformly bounded for all x and all n, and thus for all x we have log µ(In−1 (x)) = 1. lim n→∞ log µ(In (x)) Under this uniform boundedness away from zero assumption, we also obtain that d µ (x) ≥ δµ (x) contributing to the proof of (3). Next given r > 0 there exists a unique n = n(r) such that C2 |(T n )0 (x)|−1 < r ≤ C2 |(T n−1 )0 (x)|−1 . It immediately follows from (4) that log µ(B(x, C2 |(T n )0 (x)|−1 ) log µ(B(x, r)) ≤ log r log r log µ(In (x)) log µ(In (x)) ≤ . ≤ log r log(C2 |(T n )0 (x)|−1 ) We obtain that d µ (x) ≤ δµ (x) and hence dµ (x) = δµ (x). Now suppose that the pointwise dimension dµ (x) exists at a point x ∈ I \ O. Equation (4) immediately implies that log µ(B(x, C2 |(T n )0 (x)| log µ(In (x)) ≥ , log |(T n )0 (x)|−1 log(|(T n )0 (x)|−1 ) and we obtain that δ µ (x) ≥ dµ (x). Finally, choose an increasing sequence of positive integers {nk } such that lim
nk →∞
log µ(Ink (x)) = δ µ (x), log(C1 |(T nk )0 (x)|−1 )
and for each nk choose some rk > 0 such that C1 |(T nk )0 (x)|−1 < rk ≤ C1 |(T nk −1 )0 (x)|−1 . From (5) and (6) we have that log µ(Ink (x)) log µ(Ink −1 (x)) log µ(B(x, rk )) . ≥ log rk log µ(Ink (x)) log(C1 |(T nk )0 (x)|−1 ) Again, under the assumption that φ(x) exists we showed that lim log µ(Ink −1 (x))/ log µ(Ink (x)) = 1,
n→∞
and thus we obtain that dµ (x) ≥ δ µ (x). We conclude that δµ (x) = dµ (x) completing the proof of (2). Finally, as above, we obtain the same estimate if we assume that ψ is uniformly bounded away from zero on I . Part (3) now easily follows from parts (1) and (2).
156
M. Pollicott, H. Weiss
For the final part, let δµ (x) = γ . It suffices to show that for T = T1 or T = T2 then log |T 0 (T n x)| −→ 0 as n → ∞, Pn−1 i i=0 log ψ(T x) which, using δµ (x) = γ easily implies that log ψ(T n x) −→ 0 as n → ∞, Pn−1 i i=0 log ψ(T x) from which the conclusion easily follows. We shall concentrate on the case of the continued fraction transformation T1 ; the case of T2 being similar. Fix > 0. If inf x∈I | log ψ(x)| = δ > 0, then choose n0 ∈ N such that log(n0 + 1)2 /(δn0 ) ≤ . If 1/(k + 1) ≤ T1n (x) < 1/k (where k ∈ N), then k 2 ≤ |T10 (T1n x)| < (k + 1)2 . If k ≥ n0 , then provided n ≥ n0 , we can estimate log(n0 + 1)2 log |T 0 (T n x)| ≤ ≤ . Pn−1 i δn0 i=0 log ψ(T x) If k < n0 then we can still bound log(n0 + 1)2 log |T 0 (T n x)| ≤ ≤ , Pn−1 i δn i=0 log ψ(T x) provided n is sufficiently large. u t The essential feature of the proof of (4) above is the need for polynomial bounds on the derivative of the transformation. The polynomial bounds on the derivative for T2 may be found in [I,Pr]. We define the two parameter family of functions φq,t = −t log |T 0 | + q log ψ in W. Define the function t (q) by requiring that P (φq,t (q) ) = 0 and let µq be the equilibrium state for φq,t (q) . Definition. We say that a triple (T , φ, µ) satisfies a multifractal analysis1 if (1) The Markov pointwise dimension δµ (x) exists for µ-almost every x ∈ I . Moreover, R δµ (x) = δµ ≡ hµ (T )/ I log |T 0 | dµ for µ-almost every x ∈ I . (2) The function t (q) is the Legendre transform of the dimension spectrum, i.e., we have that fµ (α(q)) = t (q) + qα(q), where Z Z α(q) = −t 0 (q) = − log ψdµq / log |T 0 |dµq . I
I
In particular, t (q) is smooth and strictly convex on some interval (qmin , qmax ). 1 This multifractal analysis should properly be called a Markov multifractal analysis although this is not standard terminology. The term multifractal analysis should refer to an analysis effected using the pointwise dimension. However, in order not to introduce non-standard terminology, we will refer to this analysis as a multifractal analysis.
Multifractal Analysis of Lyapunov Exponent for Continued Fraction
- α
2
157
f(α)
T(q)
f( α) not defined
-α 0
- α 1
Vertical Slope
1
1/2 T(q) not defined
0
q q
0
α
0
α
1
α
α 2
Fig. 5. Multifractal analysis for the continued fraction transformation
An immediate consequence is that if (T , φ, µ) satisfies a multifractal analysis, then the dimension spectrum fµ (α) is smooth and strictly convex on an interval, and hence the Markov pointwise dimension δµ (x) attains the interval of values (α(qmax ), α(qmin )), where each value is attained on an uncountable dense set which supports an equilibrium state. We note that this is only a partial multifractal analysis in the sense of [PW1,PW2], in that a complete multifractal analysis also establishes analogous results for the Rényi spectrum of dimensions for µ and then establishes a Legendre transform relation between the dimension spectrum and the Rényi spectrum. Here we only extend those aspects of the theory which we use in our applications. One technical complication in extending the entire theory is that the equilibrium state µ may not be included in the family of equilibrium states µq , where in the usual theory µ1 = µ. In other words, the interval (α(qmax) , α(qmin )) may not contain dµ . This happens for the continued fraction transformation T1 with potential φ = −s log |T10 |, s > 21 (see Corollary 5). Henceforth for T1 we shall only consider potentials φ which are elements of W(T1 ), and for T2 we shall only consider potentials φ which are elements of W(Tb2 ). For these classes of potentials we establish a multifractal analysis for the pointwise dimension of the associated equilibrium state. For applications to Diophantine approximation, we only require a multifractal analysis for the very special class of potentials of the form −s log |T 0 |. Theorem 1. A multifractal analysis holds for the continued fraction transformation T1 in the range of q such that t (q) > 21 (see Fig. 5). Theorem 2. A multifractal analysis holds for the Manneville–Pomeau transformation Tb2 for 1/α < q < 1 (see Fig. 6).
158
M. Pollicott, H. Weiss f( α )
T(q)
q t(q) not defined
1/α
1
- α
α
1
α
1
Fig. 6. Multifractal analysis for the Manneville–Pomeau transformation
If the pointwise dimension (or Lyapunov exponent) exists at a point x for an EMR transformation T , then the same limit exists for the induced map Tb. In particular, the estimates for Tb provide a lower bound on the dimensions of the set of values with the same limit for T . However, the converse need not necessarily be true. It is plausible that there exist uncountably many points (comprising a set of positive Hausdorff dimension) for which the limit defining the pointwise dimension or Lyapunov exponent for T does not exist, but the subsequential limit, which corresponds to the pointwise dimension or Lyapunov exponent for Tb does exist. The following is an immediate consequence of Proposition 3 and Theorem 1, and allows us to relate the Lyapunov spectrum to the Markov dimension spectrum for a special class of equilibrium states (see [We]). Corollary 1. Let T : I → I be an EMR transformation. If φ(x) = −s log |T 0 |, then λ(x) = P (−s log |T 0 |)/(δµs (x) − s), where µs is the equilibrium state for −s log |T 0 |. In the case s = 0 we obtain that except on a countable set, λ(x) = hT OP (T )/δµMAX (x), where hT OP (T ) denotes the topological entropy of the map T and µMAX denotes the measure of maximal entropy 2 . Since countable sets have zero Hausdorff dimension, we have that P (−s log |T 0 |) . fµs (α) = g α−s The Lyapunov spectrum g(β) is smooth and strictly convex on an interval. We will see in Sect. 4 that the Lyapunov exponent for the continued fraction transformation measures the precise exponential rate of rational approximation for the continued fraction algorithm. 4. Continued Fractions, Diophantine Approximation, and Cuspital Excursions on the Modular Surface For a wealth of classical results about continued fractions we recommend the superb books [C], [HW] and [K]. The books [B,CFS] contain an excellent introduction to the 2 We remind the reader that h T OP (T1 ) = ∞ and thus care must be taken in the allowable range of s.
Multifractal Analysis of Lyapunov Exponent for Continued Fraction
159
dynamics of the continued fraction transformation and the connection with Diophantine approximation. Every irrational number 0 < x < 1 has a continued fraction expansion of the form 1
x=
=
1
a1 + a2 +
[a1 , a2 , a3 , . . . ],
1 a3 + · · ·
where a1 , a2 , · · · are positive integers. For every positive integer n define the n-th approximant pn /qn to be the rational number pn = qn
1
.
1
a1 + a2 +
1 ... +
1 an
There is a simple recursive relationship between pn , qn and an : p0 = 0, q0 = 1,
p−1 = 1, q−1 = 0,
pn = an pn−1 + pn−2 , qn = an qn−1 + qn−2 ,
k = 1, 2, . . . , k = 1, 2, · · · .
(7)
The continued fraction transformation can be considered as a simple algorithm for associating to irrational numbers 0 < x < 1 a sequence of rational numbers pn /qn . It is well known that the approximants pn /qn satisfy 1 1 1 1 1 x − pn ≤ < < 2 . (8) ≤ = 2 qn (qn + qn+1 ) qn qn (an+1 qn + qn−1 ) qn qn+1 qn 2qn+1 There is an intimate connection between the numbers a1 , a2 , · · · and the continued fractions transformation T1 . Given 0 < x < 1 we can write x= =
1 1 x
= 1 x
1 +
1 = x
1 a1 +
1 a2 +
1 T1 x
1 1 = a1 + T1 x a1 + =
1 a1 +
1
1 1 T1 x
=
1 a1 +
h
1 T1 x
i 1n o + T 1x 1
= ··· .
a2 + T12 x
h i Thus a1 = [1/x] , a2 = [1/T1 x] , · · · , ak = 1/T1k−1 x . Alternatively, if x = [a1 , a2 , · · · ], then T1n (x) = [an+1 , an+2 , an+3 , · · · ]. From this relation one immediately sees a close connection between the distribution of the values of ak and the ergodic properties of the map T1 . It also easily follows from the recursion that pn (x) = qn−1 (T1 x). To see this 1 1 pn (x) = [a1 , · · · , an ] = = qn (x) a1 + [a2 , · · · , an ] a1 + [pn−1 (T1 x)/qn−1 (T1 x)] qn−1 (T1 x) . (9) = a1 qn−1 (T1 x) + pn−1 (T1 x)
160
M. Pollicott, H. Weiss
Since all fractions are irreducible, the result immediately follows. There is an absolutely continuous T1 -invariant probability measure µG on [0, 1], usually called the Gauss measure, defined by Z 1 1 dx, µG (B) = log 2 B (1 + x) where B is a Borel subset of I . Clearly 1 1 `(B) ≤ µG (B) ≤ `(B), 2 log 2 log 2 where ` denotes Lebesgue measure. The map T1 is ergodic with respect to µG [K]. For x ∈ I \ O the set In (x) consists of all points 0 ≤ y ≤ 1 whose nth approximant is the same as for x. Dynamically this is equivalent to T k x, T k y ∈ [1/(ak (x) + 1), 1/ak (x)] for 1 ≤ k ≤ n. An easy calculation shows that `(In (x)) = 1/(qn (x)(qn (x) + qn−1 (x))). Applying (4) and (8) we obtain λ(x) = − lim
n→∞
1 1 log `(In (x)) = 2 lim log qn (x) ≡ 2q(x), n→∞ n n
(10)
when the limits exist, where q(x) denotes the exponential growth rate of the sequence {qn (x)}. Moreover, when one or the other limit exists, the other limit must also exist. An immediate consequence of (8) is that pn (x) 1 (11) λ(x) = − lim log x − n→∞ n qn (x) when one or the other limit exists. It also follows from the simple estimate on the density of the Gauss measure µG that 1 1 pn (x) λ(x) = − lim log µG (In (x)) = 2q(x) = − lim log x − n→∞ n n→∞ n qn (x) when the limit exists. Again, when one or the other limit exists the other limit must also exist. Since T1 is ergodic with respect to µG , it follows from the Birkhoff ergodic theorem applied to the function log |T10 x| and a simple calculation [B] that for µG -almost all x ∈ I, λ(x) = λ0 ≡
π2 π2 = 2.37314 · · · and q(x) = q0 ≡ = 1.18657 · · · . 6 log 2 12 log 2 (12)
Thus the Lyapunov exponent of the continued fraction transformation measures both the precise exponential speed of approximation of a number by its approximants and the exponential growth rate of the sequence {qn (x)}. For µG -almost all x ∈ I , |x − pn (x)/qn (x)| exp(−nπ 2 /6 log 2) and qn (x) exp(nπ 2 /12 log 2). Our aim now is to understand all the possible values which the Lyapunov exponent attains on the exceptional set of zero measure, as well understanding the distribution and structure of the sets of points where the exceptional values are realized. By studying periodic points of T1 one can easily find points x such that λ(x) 6 = λ0 . The fixed points of T1 correspond to numbers of the form x = [a, a, a, · · · ]. The Lyapunov
Multifractal Analysis of Lyapunov Exponent for Continued Fraction
161
exponent at a fixed point x√ is −2 log x, and since fixed points exist arbitrarily close to 0 ([2a, 2a, 2a, · · · ] = a 2 + 1 − a), it follows that λ(x) attains arbitrarily large values. On the other hand, it is easy to see that√the minimum value that λ(x) can attain is 2 log(γ ) = 0.962424 · · · , where γ = (1 + 5)/2 is the Golden Mean. This follows from the fundamental recursion (7) since qn = an qn−1 + qn−2 ≥ qn−1 + qn−2 , and thus qn (x) ≥ cγ n for all x and all n, where c is a constant which is determined by initial conditions. This value of λ(x) is attained, for example, at the fixed point x = γ − 1 = [1, 1, 1, · · · ]. This value of λ(x) is also attained at any number x whose continued fraction expansion consists of all 1’s from some point on, i.e., x = [a1 , · · · , an , 1, 1, 1, · · · ]. Such numbers are sometimes called noble numbers. More generally, this value will be realized by precisely those numbers whose continued fraction expansions have a proportion of 1s which increases to 100 percent. The set of such numbers is dense and uncountable. We know that λ(x) can realize the values γ and λ0 . In the case of γ we can find a periodic orbit x∞ such that λ(x∞ ) = γ . We claim that the Lyapunov exponent for T1 possesses an intermediate value property: any intermediate value can also be realized as a Lyapunov exponent. Lemma 1. Given any value 2 log γ < ξ < λ0 , there exists a point x ∈ I such that λ(x) = ξ . Proof. Fix any value 2 log γ < ξ < λ0 . We first note that since the T1 -invariant measures are convex, then Z 0 log |T1 |dµ : µ is T1 − invariant = [2 log γ , λ0 ]. [2 log γ , λ0 ] ∩ Moreover, since the periodic point measures are weak star dense in the T1 -invariant measures, we can choose a sequence of periodic orbits T1Nn xn = xn , n ≥ 1, such that the associated Lyapunov exponent λ(xn ) satisfies |λ(xn ) − ξ | < 1/n. We can write each periodic orbit in terms of its continued fraction expansion, i.e. xn = [a0 (xn ), a1 (xn ), · · · , aNn −1 (xn )]. We can choose an increasing sequence nk inductively P such that (1/nk ) k−1 i=1 ni Ni → 0. Finally, we define the point x ∈ I having the continued fraction expansion x = [a0 (x1 ), · · · , aN1 −1 (x1 ), a0 (x2 ), · · · , aN2 −1 (x2 ), · · · ], {z } | {z } | n1 times
n2 times
i.e., we concatenate the repeated block in the continued fraction expansion of x1 (n1 times), followed by the repeated block in the continued fraction expansion of x2 (n2 times), etc. By construction we have that λ(x) = ξ . u t To study the distribution of values of λ(x) more precisely, we define the (T1 −invariant) level sets of the Lyapunov exponent 3α = {x ∈ I : λ(x) = α}. These sets (along with the set on which λ(x) does not exist) provide a decomposition of the interval. From (10) and (11) we see that for x ∈ 3α , |x − pn (x)/qn (x)| exp(−nα) and qn (x) exp(nα/2). The following proposition on the distribution of values of λ and the precise Hausdorff dimension of the level sets are easy consequences of Lemma 1, Theorem 1, and Proposition 3 applied to the function φ = −s log |T10 | for s > 1/2:
162
M. Pollicott, H. Weiss
Proposition 4. Let T1 : I → I be the continued fraction transformation. (1) The Lyapunov exponent λ(x) attains the interval of values [2 log γ , ∞) = [2 log γ , λ0 ) ∪ [λ0 , ∞). R (2) For α ∈ [λ0 , ∞) the value α = I log |T10 |dµs is attained R by the Lyapunov exponent on a set of (positive) Hausdorff dimension hµs (T1 )/ I log |T10 |dµs = hµs (T1 )/α. This level set is uncountable and also dense in I . Proof. Lemma 1 implies that λ(x) attains the interval of values [γ , λ0 ]. Consider the family of potentials φ s = −s log |T10 |, s > 1/2 and let µs be the corresponding family of equilibrium states. To prove Proposition 4, we shall only require (1) in our definition of multifractal analysis, that δµs (x) = δs ≡ R
I
P (−s log |T10 |) hµs (T1 ) =s+ R >0 0 0 log |T1 |dµs I log |T1 |dµs
for µs -almost all x ∈ I . Let KδMs = {x ∈ I : δµs (x) = δs }. From Proposition 3 we see that d µs (x) ≤ δs for all x ∈ KδMs and d µs (x) ≥ δs for µs -almost all x ∈ KδMs . By standard arguments in dimension theory [PW1, pp. 253–254] it follows that dimH Kδs = δs [PW1, pp. 253–254]. PropositionR 1 and the Variational PrincipleR immediately imply that if δµs (x) = δs then λ(x) = I log |T10 |dµs , and thus λ(x) = I log |T10 |dµs for µs -almost all x ∈ I . Thus dimH 3R log |T 0 |dµs = δs > 0. I
1
Since µs (KdMs ) = 1 and µs (3R log |T 0 |dµs ) = 1, and equilibrium states are positive 1 I on open sets, we have that each of the sets KdMs and 3R log |T 0 |dµs are dense in I . 1 I R Recall that µ1 = µG and thus I log |T10 |dµ1 = λ0 . It immediately follows from the first derivative formula for pressure [R] that Z d log |T10 |dµs = − P (−s log |T10 |). ds I
Since lims&1/2 P (−s log |T10 |) = ∞ and P (−s log |T10 |) is smooth and strictly convex on (1/2, ∞), it follows that lims&1/2 (d/ds)P (−s log |T10 |) = −∞. Thus Z lim
s&1/2 I
log |T10 |dµs = ∞.
R The map s 7 → I log |T10 |dµs is smooth on (1/2, ∞), by analytic perturbation theory since µs corresponds to an isolated maximal eigenvalue for a transfer operator [M1], and this implies that the Lyapunov exponent λ attains all values between λ0 and ∞, each on an uncountable dense set of positive Hausdorff dimension. u t Proposition 4 immediately implies the following two number theoretic corollaries. Corollary 2. The asymptotic quantity limn→∞ (1/n) log |x − pn (x)/qn (x)| attains the interval of values [λ0 , ∞) and each value in this interval is attained on an uncountable
Multifractal Analysis of Lyapunov Exponent for Continued Fraction
163
dense set of positive Hausdorff dimension. There is an explicit formula for the Hausdorff dimension of each level set in [λ0 , ∞): Z For α = log |T10 |dµs , I 1 pn (x) = α = hµs (T1 )/α. dimH x ∈ [0, 1] : lim log x − n→∞ n qn (x) It easily follows from this formula that the Hausdorff dimension of the level sets vary smoothly. Corollary 3. The asymptotic quantity limn→∞ (1/n) log qn (x) attains the interval of values [ 21 λ0 , ∞) and each value in this interval is attained on an uncountable dense set of positive Hausdorff dimension. There is an explicit formula for the Hausdorff dimension of each level set, which in particular shows that the Hausdorff dimension of the level sets vary smoothly in α. We have seen that pn (x) = qn−1 (T1 x) and this easily implies that 1 1 log pn (x) = lim log qn−1 (T1 x) = q(T1 x). n→∞ n n→∞ n
p(x) ≡ lim
It follows from (12) that p(x) = q0 = π 2 /12 log 2 for µG −almost all x ∈ I . The next corollary, which analyzes the exceptional set, follows immediately. Corollary 4. The asymptotic quantity limn→∞ (1/n) log pn (x) attains the interval of values [ 21 λ0 , ∞) and each value in this interval is attained on an uncountable dense set of positive Hausdorff dimension. There is an explicit formula for the Hausdorff dimension of each level set, which in particular shows that the Hausdorff dimension of the level sets vary smoothly. Remark. Consider the function (x) log x − pqnn(x) pn (x) 1 = − lim log x − τ (x) = − lim n→∞ n→∞ n log qn (x) qn (x)
1 1 n
log qn (x)
,
when the limits exist. It immediately follows from (10) and (11) that for µG -almost all x the function τ (x) = τ0 ≡ λ0 /q0 = 2. More precisely, if q(x) = limn→∞ (1/n) log qn (x) exists at a point x ∈ I \ O then τ (x) = 2. We now consider the setP of points for which λ(x) does not exist. An example is ∞ −k! . It is easy to see that |x − p (x)/q (x)| = the Liouville number x = n n k=1 10 P∞ −k! and thus 10−(n+1)! < |x − p (x)/q (x)| < 2 × 10−(n+1)! . It follows that n n k=n+1 10 λ(x) = ∞. It is also easy to construct numbers for which the limit defining λ(x) does not exist and is not infinite. The construction uses the trick in Lemma 2. Consider the number x with continued fraction expansion x = [1, · · · , 1, 2, · · · , 2, 1, · · · , 1, 2, · · · , 2, · · · ], | {z } | {z } | {z } | {z } n1 times
m1 times
n2 times
m2 times
164
M. Pollicott, H. Weiss
with each ni and mi being much larger than the sum of all the proceeding choices. A routine argument shows that for suitable choices of ni and mi , the Lyapunov exponent λ(x) does not exist. A straightforward extension of an argument by Shereshevsky [Sh] gives that the set of points for which λ(x) does not exist has positive Hausdorff dimension. A natural problem is to compute the precise Hausdorff dimension. In [BaS] the authors show that for a conformal expanding map (they assume that their map is a local homeomorphism), the Hausdorff dimension of the set of points where the Lyapunov exponent does not exist is maximal, i.e., the same dimension as the limit set (repeller). The essential hypothesis for their result is a smooth map which possesses a sequence µk of ergodic invariant measures such that limk→∞ dimH (µk ) = dimH (J ), where J is the limit set and dimH (µ) ≡ inf{dimH (U ), µ(U ) = 1} [Pe, p. 42] is the Hausdorff dimension of the measure µ. As will be noted after the proof of Theorems 1 and 2, the existence of such sequences of measures for the maps T1 and T2 follow immediately from the proof of Theorems 1 and 2. Thus a straightforward extension of the proof of Barreira and Schmeling proves the following result. Theorem 3. The set of points x ∈ I for which λ(x) does not exist has Hausdorff dimension equal to 1. There are classical results on Hausdorff dimension and Diophantine approximation, due to Jarnik, to which our results can be viewed as complimentary. Recall that the continued fraction approximants pn /qn of a number x all satisfy the Diophantine condition x − p ≤ 1 . q q2 Let us now consider the set of numbers which admit a faster approximation by rational numbers. For τ > 2 let Fτ denote the set of τ −well approximable numbers, i.e., those that satisfy p 1 Fτ = x ∈ I : x − ≤ τ infinitely often . q q Here infinitely often means that there are infinitely many distinct rational p/q which satisfy the relation. Legendre showed that if p/q is any rational approximation to an irrational number x satisfying |x − p/q| ≤ 1/2q 2 , then p/q must be an approximant for x. Thus the rationals p/q in Fτ are all approximants. It is easy to show that this set has zero measure for each τ > 2. Jarnik [J] computed the Hausdorff dimension of Fτ and showed that dimH (Fτ ) = 2/τ . While Jarnik explicitly computes the Hausdorff dimension of the set of numbers x such the approximation error |x − p/q| is bounded above by q −τ for infinitely many approximants pn /qn , Corollary 2 is a statement about the Hausdorff dimension of sets of numbers x such that the error |x − pn /qn | admits precise asymptotic limiting behavior. Corollary 2 also quantifies the precise speed of convergence of the continued fraction algorithm on exceptional sets. Although our results are related, they do not seem to be obtainable from each other. Application to Geodesics on the Modular Surface. The map T1 is also closely related to the symbolic description of the geodesic flow φt : P SL(2, R)/P SL(2, Z) → P SL(2, R)/P SL(2, Z) defined on the unit tangent bundle of the modular surface M = H2 /P SL(2, Z) by φt (g)P SL(2, Z) = ggt P SL(2, Z) with gt = diag(exp(t), exp(−t)) and H2 = {x + iy : y > 0}. There is an interesting connection between Corollary 1
Multifractal Analysis of Lyapunov Exponent for Continued Fraction
165
(x+y)/2
-1
-y
0
a
0
x
a +1 0
Fig. 7. Continued fractions and geodesic excursions
and such geodesic flows. The Lyapunov exponent quantifies the proportion of time that a geodesic spends in excursions on cuspital excursion [St,Su]. Given a geodesic γ on the modular surface we can consider a lift γˆ to the Poincaré upper half-plane H2 . Such geodesics on H2 correspond to Euclidean semi-circles which meet the real line perpendicularly. In particular, we can choose our lift so that the endpoints satisfy and 1 < x ≡ γˆ (∞) < ∞ and −1 < y ≡ γˆ (−∞) < 0. Let us consider the continued fraction expansion x = [a0 (x), a1 (x), a2 (x), . . . ] then we see that the Euclidean height of the arc is equal to (x − y)/2 and lies between a0 /2 and a0 /2 + 1. In particular, the hyperbolic distance of the cuspital excursion can be estimated by log a0 (x) (see Fig. 7). Fix p ∈ M. Given a unit tangent vector v ∈ T1 M = P SL(2, R)/P SL(2, Z) at p, we let γ : R → M be the unique unit speed geodesic with γ˙ (0) = v. We denote by sn , n ≥ 1, the times s > 0 at which γs (v) maximizes d(γs (v), p) on each successive excursion into the cusp (i.e., s1 < s2 < s3 < . . . are local maxima for d(γs (v), p)), where d denotes the hyperbolic distance on M. By the above observations we see that these heights can be estimated with the function − log an (x). Unfortunately, this function does not satisfy condition (W1), so that we need to consider instead functions of the form φ(x) ≡ −β log an (x), for β > 1. Moreover, sn can be estimated by log |(T1n )0 (x)|. We can therefore interpret the quantity An (v) ≡ Pn−1 Pn−1 n 0 i=0 log ai (x)/ log |(T1 ) (x)| ∼ i=0 d(γsi (v), p)/sn as an estimate on the average height of the first n geodesic excursions, compared with the time required for the excursions. For β > 1 let us denote ( ) n−1 X α α d(γsi (v), p) = = v : lim An (v) = . 3α ≡ v : lim (1/sn ) n→∞ n→∞ β β i=0
Theorem 1 implies the following result.
166
M. Pollicott, H. Weiss
Theorem 4. There exists an interval of values (αmin , αmax ) such that for α in this interval the set 3α is an uncountable dense set of positive Hausdorff dimension, and the Hausdorff dimension dimH (3α ) varies smoothly (analytically). This can be compared with the results of Sullivan [Su], Melián-Pestana [MP2], and Stratmann [St], which are of a somewhat complementary nature. In our notation these authors’ results relate to the subsequence tm = snm of successive farthest geodesic excursions (i.e., d(γt1 (v), p) ≤ d(γt2 (v), p) ≤ . . . ). Sullivan [Su] shows that a typical geodesic extends a distance at most log t into a geodesic at time t. More precisely, Sullivan shows that for almost every unit tangent vector v at p, d(γtm (v), p) = 1. m→∞ log tm lim
Stratmann and Melián-Pestana compute the Hausdorff dimension of the sets 5α = {v : lim d(γtm (v), p)/tm = α}. m→∞
Remark. As one would imagine, a similar study can be made of the behaviour of geodesics on other surfaces with cusps, generalizing those for the modular surface. In this case, one needs to use the general analysis described in [BoS]. There is also an analogous notion of Diophantine for more general Fuchsian groups [Pa], to which our results would naturally apply. Remark. We do not yet have a multifractal analysis of Birkhoff sums for T1 . With such P machinery one could make similar statements about n−1 i=0 d(γsi (v), p)/n as we can for Pn−1 d(γ (v), p)/s . si n i=0 5. Thermodynamic Formalism for Infinite State Systems The following proposition contains useful formulas for the derivatives of pressure. Proposition 5. Let T I → I be an EMR transformation. Let f, g and h be functions on I such that for sufficiently small 1 , 2 the family of functions f +1 g +2 h satisfy (W1), (W2), and P (f + 1 g + 2 h) > −∞. Then the function (1 , 2 ) 7 → P (f + 1 g + 2 h) is analytic, convex (in each variable), strictly convex if f is not cohomologous to a constant, and satisfies the following derivative formulas: Z d P (f + g) = g dµf , 1 d1 =0 I and
∂ 2 P (f + 1 g + 2 h) ≡ Qf (g, h), ∂1 ∂2 1 =2 =0
where Qf is the bilinear form on C α (I, R) defined by Z Z ∞ Z X g · (h ◦ T k ) dµf − g dµf h dµf , Qf (g, h) = k=0
I
I
I
and µf is the equilibrium state for f . Also Qf (g, g) ≥ 0 for all g and Qf (g, g) > 0 if and only if f is not cohomologous to a constant function.
Multifractal Analysis of Lyapunov Exponent for Continued Fraction
167
Proof. The proof is very similar to the proof in the usual setting where the interval map is piecewise smooth and expanding on finitely many intervals [R]. There is an additional potential complication which needs to be addressed in that there may be an infinite sum, rather than just a finite sum, for the transfer operator. This occurs for T1 and Tb2 . This complication is handled with conditions (W1) and (W2) which ensure that the infinite sum is well defined and can be treated in the usual ways. The key to the proof is the study of the transfer operator Lφq,t : C 0 (I ) → C 0 (I ) given by X exp(φq,t (y))k(y), Lφq,t k(x) = T y=x
whose maximal eigenvalue is exp(P (φq,t )). Moreover, to obtain an isolated eigenvalue in the appropriate range, we study Lφq,t : BV (I ) → BV (I ) acting on the space BV of functions of bounded variation. Prellburg [Pr] has shown that for Tb2 , the spectrum of this operator consists of the closed unit ball, plus at most a countable number of isolated eigenvalues of modulus strictly greater than 1. Thus when the quantity P (φq,t ) > −∞, its exponential is the maximal positive isolated eigenvalue for Lφq,t . The result follows by analytic perturbation theory. The convexity is a direct consequence of the second derivative formula. u t The next proposition is an immediate consequence of Proposition 5 and the implicit function theorem [PW1]. Proposition 6. Let T I → I be an EMR transformation. Assume that for a range of values (q, t) the family of functions φq,t , satisfy (W1), (W2), and P (φq,t ) > −∞. Then the function (q, t) 7 → P (φq,t ) is analytic and convex (in each variable). Furthermore, for the one parameter family of potentials φq = φq,t (q) , where t = t (q) is defined by requiring that P (φq,t (q) ) = 0, we have that (1) The function t (q) is real analytic and convex. It is strictly convex if log ψ is not cohomologous to − log |T 0 |. (2) The derivative R log ψ dµq , t 0 (q) = − R I log |T 0 | dµq I where µq is the equilibrium state for φq . (3) The second derivative satisfies 2 2 2 ∂ P (φq,r ) 0 (q) ∂ P (φq,r ) + ∂ P (φq,r ) t 0 (q)2 − 2t ∂q ∂r ∂r 2 ∂q 2 , t 00 (q) = ∂P (φq,r ) ∂r
evaluated at (q, r) = (q, −t (q)). The following useful corollary follows easily from Proposition 2. Here we collect many results which we will require in Sect. 7. Corollary 5. In the special case φ = −s log |T 0 | we have that the function φq = −(t (q) + qs) log |T 0 | − qP (s log |T 0 |) and that (1) The function t (q) is defined implicitly by P (−(t (q) + qs) log |T 0 |) = qP (−s log |T 0 |).
168
M. Pollicott, H. Weiss
(2) The derivative t 0 (q) = − R
I
hνq (T ) log |T 0 |dνq
,
where νq is the equilibrium state for φq . (3) We have the special values t (0) = 1, t (1) = 0, and t 0 (0) = −1. (4) For the continued fraction transformation T1 the expressions in (1)-(3) are well defined (satisfy W1, W2, and P 6 = −∞) provided that t (q) > 1/2 and s > 1/2. (5) For the induced Manneville–Pomeau transformation Tb2 the expressions in (1)-(3) are well defined (satisfy W1, W2, and P 6 = −∞) provided that t (q) < 1 [Pr]. Example 1. The transfer operator for the continued fraction transformation T1 for the family φq,t , can be written explicitly as q 2t ∞ X 1 1 1 ψ k . Lφq,t k(x) = x+n x+n x+n n=1
For the potentials φq,t (for t > 1/2), conditions (W1) and (W2) apply. In the special case that φ is real analytic, the operator Lφq,t preserves the smaller space of analytic functions and is compact. In particular, we can waive the assumption that P (φq,t ) > −∞. Example 2. The transfer operator for the induced Manneville–Pomeau transformation Tb2 for the family φq,t , can be written explicitly as Lφq,t k(x) =
∞ X ψ q (x) X k(x) = ψ(Gn (x))q |G0n (x)|t k(Gn (x)), |Tb0 (x)|t
Tb2 y=x
n=1
2
where Gn = F1 F0n−1 , and F0 , F1 denote the two branches of the inverse of Tb2 . In the case of the Manneville–Pomeau transformation, the induced transformation Tb2 : I → I is a smooth map on each of the intervals [an , an+1 ] [Pr]. For the potentials φq,t (for t > 1), conditions (W1) and (W2) apply. The next result describes the construction of equilibrium states using transfer operators. Proposition 7. Let T I → I be an EMR transformation. Assume that φ satisfies (W1) and (W2). Then there exists a unique equilibrium state. Sketch of Proof. The associated transfer operator satisfies the following [Wa1, p. 128] (1) There exists λ > 0 and h ∈ C 0 (I ) with Lψq h = λh; 0 (2) There exists a T -invariant R probability measure µ such that for f ∈ C (I ) we have n −n that λ Lψq f (x) → h f dµ. R If we denote g = exp(ψ)h/(λh ◦ T ) we have that λ−n Lnlog g f (x) → f dµ [Wa1, p. 128]. Moreover, dT ∗ µ/dµ = 1/g [Wa1, p. 124], where T ∗ µ denotes the pull-back of the measure µ by T . The proof that µ is the required equilibrium state comes from (2). A useful property of g is that there exists C0 > 0 such that for all x, y ∈ I , exp(−C0 d(x, y)) ≤
n−1 Y i=0
|g(T i x)| ≤ exp(−C0 d(x, y)) |g(T i y)|
Multifractal Analysis of Lyapunov Exponent for Continued Fraction
169
[Wa1, p. 130]. In particular, this implies that there exists C > 0 such that for all x ∈ I , µ(In (x)) 1 ≤ C. ≤ Qn−1 i C i=0 g(T x) Uniqueness comes by the ergodicity of µ and the fact that any two solutions are absolutely continuous. u t
6. Proofs of Theorems 1 and 2 Consider the transformation T = T1 or Tb2 . Statement (1) in our definition of multifractal analysis easily follows from the Birkhoff ergodic theorem. Given x ∈ I \ O, it follows state µ is an from Proposition 3 that dµ (x) = −logR ψ(x)/λ(x). Since the equilibrium R ergodic measure we have that λ(x) = I log |T 0 |dµ and log ψ(x) = I log ψdµ for µalmost everyRx ∈ I . Since P (log ψ) = 0, it follows from the R Variational Principle that hµ (T ) = − I log ψdµ. It follows that dµ (x) = hµ (T )/ I log |T 0 | dµ for µ-almost every x ∈ I . Recall that t (q) is the unique solution to P (φ(q,t (q) ) = P (−t (q) log |T 0 |+q log ψ) = 0 and µq is the equilibrium state for φq = φ(q,t (q)) . For each map T and each potential φ, the function t (q) is defined for a certain range (qmin , qmax ) of q which we have discussed in Sect. 6. By Proposition 6 the function t (q) is smooth and strictly convex on this interval. Using the derivative formula in Proposition 6, we define for each q the function R log ψdµq 0 , α(q) ≡ −t (q) = RI − I log |T 0 |dµq and consider the level sets M = x : δµ (x) = α(q) . Kα(q) We remind the reader that log µ(In (x)) M Kα(q) = x ∈ I \ O : lim n→∞ log `(In (x)) R Pn−1 i log ψdµq i=0 log ψ(T x) =RI . = lim Pn−1 0 0 i n→∞ I log |T |dµq i=0 log |T (T (x))| An immediate consequence of the Birkhoff ergodic theorem is that the Markov pointwise M ) = 1. dimension satisfies µq (Kα(q) M , Proposition 3 immediately implies that for all x ∈ Kα(q) δµq (x) = −
φ q (x) λ(x)
=
t (q)log |T 0 |(x) − qlog ψ(x) = t (q) + qα(q). λ(x)
It follows from Proposition 3(2) that the upper pointwise dimension d µq (x) ≤ δµq (x) = M . t (q) + qα(q) for all x ∈ Kα(q) M (and equals Since the Birkhoff average log ψ(x) exists for µq -almost every x ∈ Kα(q) R I log ψdµq ), it follows from Proposition 3(3) that d µq (x) ≥ δµq (x) = t (q) + qα(q)
170
M. Pollicott, H. Weiss
M . By standard arguments in dimension theory [PW1, for µq -almost every x ∈ Kα(q) M ) ≥ t (q) + qα(q), and these pp. 253–254], this a˙e˙inequality implies that dimH (Kα(q) M two estimates imply that dimH (Kα(q) ) = t (q) + qα(q). The smoothness and convexity properties of t (q) follow from Proposition 6. u t
Remark. As mentioned before the statement of Theorem 3, the above proof of Theorems 1 and 2 provides a sequence of ergodic invariant measures for T1 or Tb2 such that lim dimH (µk ) = dimH (I ) = 1. k→∞
M and µ (K M ) = 1, a weaker Since δµq (x) = t (q) + qα(q) for µq almost-all x ∈ Kα(q) q α(q) conclusion than we obtained in the theorem above is that for the (ergodic) equilibrium state µq we have that dimH (µq ) = t (q) + qα(q) [Pe, p. 42]. Since t (0) = 1 we have that limq→0 dimH (µq ) = 1. Thus any allowable potential φ ∈ W gives rise to such a one parameter family (and hence such a sequence) of measures.
References [B] [BaS] [BoS] [C] [CFS] [GH] [HW] [I] [J] [K] [L] [M1] [MP1] [MP2] [N] [Pa] [Pe] [Pr] [PS] [PW1] [PW2] [PW3]
Billingsley, P.: Ergodic Theory and Information. Krieger, 1978 Bareirra, L. and Schmeling, J.: Sets of “non-typical” points have full topological entropy and full Hausdorff dimension. Preprint Bowen, R. and Series, C.: Markov maps associated with Fuchsian groups. Publ. Math.(IHES) 50, 153–170 ( 1979) Cassels, J.: An Introduction to Diophantine Approximation. CUP, 1957 Cornfeld, I., Fomin, S., Sinai, Ya.: Ergodic theory. Berlin–Heidelberg–New York: Springer-Verlag, 1982 Gröchenig, K. and Haas, A.: Backwards Continued Fractions, Hecke Groups and Invariant Measures for Transformations of the Interval. Ergodic Theory Dynam. Systems 16, 241–1274 (1996) Hardy, G. and Wright, E.: An Introduction to the Theory of Numbers. Fifth edition, Oxford: Oxford University Press, 1979 Isola, S.: Dynamical Zeta Functions and Correlation Functions for Non-uniformly Hyperbolic Systems. Preprint Jarnik, V.: Über die simultanen diophantischen Approximationen. Math. Zeit. 33, 505–543 (1931) Khinchin, A.: Continued fractions. Chicago: University of Chicago Press, 1964 Lopes, A.: The Zeta Function, Nondifferentiability of Pressure, and the Critical Exponent of Transition. Adv. Math. 101 2, 133–165 (1993) Mayer, D.: On the Thermodynamics Formalism for the Gauss Map. Commun. Math. Phys. 130, 311–333 (1990) Pomeau, Y. and Manneville, P.: Intermittent Transition to Turbulence in Dissipative Dynamical Systems. Commun. Math. Phys. 74 189–197 (1980) Melián, M. and Pestana, D.: Geodesic Excursions into Cusps in Finite-Volume Hyperbolic Manifolds. Michigan Math. J. 40, 77–93 (1993) Nakaishi, K.: Multifractal Formalsim For Some Parabolic Maps. Preprint Patterson, S.: Diophantine approximation in Fuchsian groups. Philos. Trans. Roy. Soc. London, Ser. A 282, 1976, pp. 527–563 Pesin, Y.: Dimension Theory in Dynamical Systems. CUP, 1997 Prellberg, T.: Maps of Intervals with Indifferent Fixed Points: Thermodynamic Formalism and Phase Transitions Va. Polytechnique Institute Theses, 1991 Prellberg, T. and Slawny, J.: Maps of Intervals with Indifferent Fixed Points: Thermodynamic Formalism and Phase Transitions. J. Stat. Phys. 66, 503–514 (1992) Pesin, Y. and Weiss, H.: Multifractal Analysis of Equilibrium Measures for Conformal Expanding Maps and Moran-like Geometric Construction. J. of Stat. Phys. 86, 233–275 (1997) Pesin, Y. and Weiss, H.: The Multifractal Analysis of Gibbs Measures: Motivation, Mathematical Foundation and Examples. Chaos 7, 89–106 (1997) Pesin, Y. and Weiss, H.: On the Dimension of Deterministic and Random Cantor-like sets, Symbolic Dynamics, and the Eckmann–Ruelle Conjecture. Commun. Math. Phys. 182, 105–153 (1996)
Multifractal Analysis of Lyapunov Exponent for Continued Fraction
[R] [Sc] [Sh] [St] [Su] [T] [V] [Wa1] [Wa2] [We]
171
Ruelle, D.: Thermodynamic Formalism. Reading, MA: Addison-Wesley, 1978 Schweiger, P.: Ergodic Theory of Fibred Systems and Metric Number Theory. Oxford: Oxford University Press, 1995 Shereshevsky, M.: A Complement to Young’s Theorem on Measure Dimension: The Difference Between Lower and Upper Pointwise Dimension. Nonlinearity 4, 15–25 (1991) Stratmann, B.: Fractal Dimensions for Jarnik Limit Sets of Geometrically Finite Kleinian Groups; The Semi-Classical Approach: Ark. Mat. 33, 385–403 (1995) Sullivan, D.: Disjoint Spheres, Approximation by Imaginary Quadratic Numbers and the Logarithmic Law for Geodesics. Acta. Math. 149, 215–237 (1982) Thaler, M.: Estimates on the invariant densities of endomorphisms with indifferent fixed points. Israel J. Math. 37, 303–314(1980) Verbitski, E.: Personal communication Walters, P.: Invariant Measures and Equilibrium States for Some Mappings Which Expand Distances. Transactions of the AMS 236, 121–153 (1978) Walters, P.: Introduction to Ergodic Theory. Berlin–Heidelberg–New York: Springer Verlag, 1982 Weiss, H.: The Lyapunov Spectrumof Equilibrium Measures for Conformal Expanding Maps and Axiom-A Surface Diffeomorphisms. J. Stat. Physics 95, (1999)
Communicated by P. Sarnak
Commun. Math. Phys. 207, 173 – 195 (1999)
Communications in
Mathematical Physics
© Springer-Verlag 1999
Normal Forms and Quantization Formulae Dario Bambusi1 , Sandro Graffi2 , Thierry Paul3 1 Dipartimento di Matematica, Università di Milano, 20133 Milano, Italy. E-mail: [email protected] 2 Dipartimento di Matematica, Università di Bologna, 40127 Bologna, Italy. E-mail: [email protected] 3 Ceremade, Université de Paris-IX, 75776 Paris, France. E-mail: [email protected]
Received: 9 January 1998 / Accepted: 21 April 1999
Abstract: We consider the Schrödinger operator Q = −h¯ 2 1+V in Rn , where V (x) → +∞ as |x| → +∞, is Gevrey of order ` and has a unique non-degenerate minimum. A −a quantization formula up to an error of order e−c| ln h¯ | is obtained for all eigenvalues of −b Q lying in any interval [0, | ln h¯ | ], with a > 1 and 0 < b < 1 explicitly determined 1/` and c > 0. For eigenvalues in [0, h¯ δ ], 0 < δ < 1, the error is of order e−c/h¯ . The proof is based upon uniform Nekhoroshev estimates on the quantum normal form constructed quantizing the Lie transformation.
1. Introduction and Statement of the Results Consider the Schrödinger operator Q(h¯ ) = −h¯ 2 1/2 + V (x), V ≥ 0. Assume: (A1) V ∈ C ∞ (Rn ), n ≥ 1; V → +∞, |x| → ∞; V has a unique non-degenerate minimum. (Without loss we can take V (0) = ∇V (0) = 0, HessV (0) = diag[ω12 , . . . , ωn2 ].) By (A1) Q defined on H 2 (Rn ) ∩ D(V ) is self-adjoint in L2 (Rn ) with discrete spectrum (see e.g. [RS], §XIII.14). Finding an approximate quantization formula for its eigenvalues when the symbol q(x, ξ ) := ξ 2 /2 + V (x) is a non-integrable Hamiltonian is an old problem. It is long known that the eigenvalues of Q in [0, C h¯ ], C > 0 are approximated by those of the harmonic oscillator up to order h¯ 2 ; the mathematical justification of a quantization formula valid for all such eigenvalues "near the bottom of the well" is however much more recent (see e.g.[CDS,HS,Si]). On the other hand, classically these systems admit a normal form. Write q = p0 (x, ξ ) + V3 (x), where p0 (x, ξ ; ω) :=
1 X 2 1X 2 2 ωk x k , ξk + ωk2 xk2 , V3 (x) := V (x) − 2 2 n
n
k=1
k=1
(1.1)
174
D. Bambusi, S. Graffi, T. Paul
p0 is an integrable system (the harmonic oscillator), and V3 (x, ξ ) = O(|x|3 ) as |x| → 0. If (x, ξ ) ∈ := {(x, ξ ) | |(x, ξ )| ≤ } the “size” kV k := sup V3 (x) of the |x|≤
perturbation V3 can be made small: one then constructs on a canonical transformation close to the identity, such that the image of q has an expansion in powers of whose coefficients depend only on the actions Ik = (ξk2 + ωk2 xk2 )/2ωk , k = 1, . . . , n. This procedure has been implemented in quantum mechanics via microlocal analysis by Sjöstrand [Sj], yielding a quantization formula mod h¯ ∞ for the “semi-excited levels”; i.e. he proved that there is a smooth function F (t1 , . . . , tn ; h¯ ) admitting a full asymptotic expansion in h¯ such that, ∀δ > 0, all eigenvalues of Q in [0, h¯ δ ] are given by F ((k1 + 1/2)h¯ , . . . , (kn + 1/2)h¯ ; h¯ ) + O(h¯ ∞ ), ki = 1, . . . , i = 1, . . . , n. (1.2) The 0th order term in h¯ is the normal form of q truncated at a suitable δ dependent order. The purpose of this paper is twofold: (i) To find a quantization formula valid further away from the bottom of the well, and hence closer to the chaotic regime.Actually our formula, which holds for eigenvalues in the interval [0, | ln h¯ |−b ], does not have the standard form of an expansion in powers of h¯ with coefficients depending on (k + 1/2)h¯ ; β (ii) To turn the O(h¯ ∞ ) error estimate valid in [0, h¯ δ ] into an O(e−c/h¯ ) one, 0 < β < 1. To prove these results we proceed as follows. First we perform the unitary rescaling U : L2 (Rn ) 3 ψ(x) 7 → (U ψ)(x) := n/2 ψ(x) ∈ L2 (Rn ). The unitary image S := U QU−1 of the operator Q(h¯ ) under U has the form S = 2 P (h¯ 0 , ), P (h¯ 0 , ) := P0 (h¯ 0 , ω) + V (x), h¯ 0 = h¯ / 2 , X (∂x3h xk xl V )(0)xh xk xl + . . . . V (x) := −3 V3 (x) =
(1.3) (1.4)
h+k+l=3
Hence Spec(Q(h¯ )) = 2 Spec(P (h¯ 0 , )). To study Spec(P (h¯ 0 , )) we replace V by a,R (x, ξ ) := V (x)t (|x|)t (|ξ |), where t is a Gevrey function with t (u) ≡ 1 if u ∈ [0, R], t (u) ≡ 0 if u > 2R, R > 0 fixed, and prove that this replacement modifies any finite 1/(`+1) . Finally, using the quantum Lie part of the spectrum only by terms of order e−c/h¯ transform (see e.g.[DGH,Ba1]), we construct (and estimate) the normal form for the semiclassical pseudodifferential operator of Weyl symbol p0 (x, ξ ) + a,R (x, ξ ). An exponentially small remainder is obtained by implementing in this context the Nekhoroshev[Ne] technique: the normal form is truncated at an order chosen as a function of so as to minimize the remainder. The error estimate is uniform in h¯ (as in [Ba1,BV]). We further remark that here the Fourier integral operator construction of [Sj] is not required, and the results of [BV] are obtained as a by-product (see Proposition 3.1). Our first result is: Theorem 1.1. In addition to (A1) above assume (A2) There exist ν > 0, Cj > 0, α > 0, K(U ) > 0 and 1 < ` < ∞ such that, ∀ j = (j1 , . . . , jn ), |j | := |j1 | + . . . + |jn |, and for any bounded U ⊂ Rn : |j | |j | sup (1 + |x|2 )|j |/2 ∂xj V (x) ≤ Cj (1 + |x|2 )ν , sup ∂xj V (x) ≤ K(U )α −|j | (|j |!)` . x∈Rn
x∈U
(1.5)
Normal Forms and Quantization Formulae
175
(A3) There exist τ > n − 1, γ > 0 such that hω, ki ≥ γ |k|−τ , ∀k ∈ Zn \ {0}, |k| := |k1 | + . . . + |kn |, ω := (ω1 , . . . , ωn ). (1.6) Let furthermore b := [(τ + 2)(2` + 3)]−1 . Then: 1. There exist h∗ > 0, ∗ > 0, A > 0, B > 0 (independent of h¯ and ) and a smooth function Z(t1 , . . . , tn ; h¯ , ) : Rn × [0, h∗ ] × [0, ∗ ] → R with asymptotic expansion in h¯ : Z(t1 , . . . , tn ; h¯ , ) ∼ Z0 (t1 , . . . , tn ; ) + h¯ Z1 (t1 , . . . , tn ; ) + . . .
(1.7)
such that, ∀ ≤ ∗ , h¯ ≤ h¯ ∗ 2 the eigenvalues of Q(h¯ ) in [0, 2 ] have the representation h(k + 1/2), ωih¯ + 3 Z (k + 1/2)h¯ / 2 , h¯ / 2 , (1.8) 2 1/(`+1) b + O 3 e−A/ . + 2 O e−(B /h¯ ) 2. There exists F(t1 , . . . , tn ; h¯ ) such that, for any fixed integer M < N () := (∗ /)b : h(k + 1/2), ωih¯ + 3 Z((k + 1/2)h¯ / 2 , h¯ / 2 , ) = F ((k + 1/2)h¯ , h¯ ) + OM ( M ) + O((h¯ / 2 )∞ ).
(1.9)
Here F := h(k + 1/2), ωih¯ + F((k + 1/2)h¯ , h¯ ) is as in Theorem 0.1 of [Sj]. Remarks. 1. Z0 is the classical Birkhoff normal form computed up to order N () = (∗ /)b : Ik = (ξk2 + ωk2 xk2 )/2ωk are the action variables of the harmonic oscillator, and Z0 (I1 , . . . , In , ) =
N() X
ζs (I1 , . . . , In ; ) s .
(1.10)
s=0
The coefficients ζs are given by the Birkhoff transformation. The quantum corrections Zj : j ≥ 1 are likewise explicitly computable recursively. 2. The normal form (1.8) depends on because the scale invariance of the perturbation is broken by the modifications of the potential at infinity. Thus (1.8) is different from that of [Sj], Theorem 0.1. Assertion 2 states that the difference is O (h¯ / 2 )∞ . Choosing 2 = ϕ(h¯ ), within the requirement h¯ /ϕ(h¯ )b/2−1 → 0, h¯ → 0 we immediately recover Theorem 0.1 of [Sj], namely: Corollary 1.1. Let ϕ :]0, 1[→ R+ be positive and increasing. Then 1. If lim ϕ(x)b ln x = 0 all eigenvalues of Q(h¯ ) in [0, ϕ(h¯ )] admit the representation x→0
h(k + 1/2), ωih¯ + ϕ(h¯ ) Z 3
(k1 + 1/2)h¯ (kn + 1/2)h¯ h¯ p ,... , , , ϕ(h¯ ) ϕ(h¯ ) ϕ(h¯ ) ϕ(h¯ ) b/2
+ O(e−A/ϕ(h¯ ) ). (1.11)
176
D. Bambusi, S. Graffi, T. Paul
2. If ϕ(h¯ ) = h¯ δ/2 , 0 < δ < 1, (1.11) becomes, for all eigenvalues in [0, h¯ δ/2 ]: h(k + 1/2), ωih¯ + h¯ 3δ/2 Z (k1 + 1/2)h¯ 1−δ , . . . , (kn + 1/2)h¯ 1−δ , h¯ 1−δ , h¯ δ/2 δb/2
+ O(e−c/h¯
). (1.12)
3. If an error O(h¯ ∞ ) is allowed the fractional dependence of Z on h¯ disappears and (1.12) reduces to Theorem 0.1 of [Sj], i.e. the eigenvalues admit the representation h(k + 1/2), ωih¯ + F ((k1 + 1/2)h¯ , . . . , (kn + 1/2)h¯ , h¯ ) + O(h¯ ∞ ).
(1.13)
Remarks. 1. If ϕ(x) = | ln b +η (x)|−1 , η > 0 , (1.11) yields h a quantization formula i δ more general than Theorem 0.1 of [Sj], because [0, h¯ ] ⊂ 0, | ln2/b+η (h¯ )|−1 ∀ δ > 2
0, η > 0. Moreover, the error estimate O(exp(−A/| ln h¯ |1+ηb/2 )) is sharper than O(h¯ ∞ ).
2. Formula (1.13) follows from (1.9): for so small that 1/δ 2 < (∗ /)b , and any √ M 2 M ). M > 1/δ , one has O( ) = O(h¯ δ
3. A more general result, i.e. a quantization formula up to order e−c/h¯ , δ < 1, for all eigenvalues in any interval [0, M], M > 0, has been recently stated by Popov [Po], by a quantization of KAM theory.
2. Reduction to a Perturbation of Compact Support In this section we estimate the variation undergone by the eigenvalues of P (h¯ 0 , ) defined by (1.3, 1.4) when V is approximated by an operator with compact support symbol (and therefore trace-class). Denote by IB the open interval ]0, B[, 0 < B < ∞, and I¯B any compact subinterval of IB . Let H1 (κ), H2 (κ) be norm-resolvent continuous, self-adjoint operator families in a Hilbert space H ([Ka], §VII.1) for κ ∈ IB , with compact resolvents. Let J ⊂ R be a fixed bounded interval, and set SpecJ (H1,2 ) := Spec(H1,2 ) ∩ J . Definition 2.1. Let 0 < δ, c < +∞. We say that δ
SpecJ (H1 ) = SpecJ (H2 ) (mod e−c/κ ) as κ → 0 in I if H1 (κ)φ1 (κ) = λ1 (κ)φ1 (κ), φ1 6= 0, λ1 ∈ J imply φ1 (κ) ∈ δ D(H2 ) and k[H2 (κ) − λ1 (κ)]φ1 (κ)k = O(e−c/κ ) and, conversely, H2 (κ)φ2 (κ) = λ2 (κ)φ2 (κ), φ2 6 = 0, λ2 ∈ J imply φ2 (κ) ∈ D(H1 ) and k[H1 (κ) − λ2 (κ)]φ2 (κ)k = δ O(e−c/κ ). δ
Remark. If SpecJ (H1 ) = SpecJ (H2 ) (mod e−c/κ ), then for any λ1 ∈ SpecJ (H1 ) there δ exists λ2 ∈ SpecJ (H2 ) such that λ1 (κ) = λ2 (κ) + O(e−c/κ ) and conversely. Let P (h¯ 0 , ) be the operator in L2 (Rn ) defined by P0 (h¯ 0 ) + V on D(P0 ) ∩ D(V ). P0 (h¯ 0 , ω), the harmonic oscillator, is defined on D(P0 ) := H 2 (Rn ) ∩ L22 (Rn ).
Normal Forms and Quantization Formulae
177
Let R > 0 be fixed, u 7 → t (u) ∈ GR`,α (R+ ) be a Gevrey function such that t (u) ≡ 1 for |u| ≤ R, t (u) ≡ 0 for |u| ≥ 2R, and set: p(x, ξ ) := p0 (x, ξ ) + V (x), a,R (x, ξ ) := V (x)t (|x|)t (|ξ |), p,R (x, ξ ) := p0 (x, ξ ) + a,R (x, ξ ).
(2.1)
Denote furthermore QR := {(x, ξ )} | |x| ≤ R, |ξ | ≤ R}. Remarks. 1. a,R (x, ξ ) ∈ G `,α (R2n ) uniformly in ∈ I¯B ; namely, for some K, α independent of : |j | (2.2) sup ∂zj a,R,h¯ 0 (z) ≤ K(|j |!)` /α |j | , z := (x, ξ ). z∈R2n
/ Q2R . Moreover a,R (x, ξ ) ≡ V (x), (x, ξ ) ∈ QR , a,R (x, ξ ) ≡ 0, (x, ξ ) ∈ 2. Let A := OpW (a) denote the operator in L2 (Rn ) defined by the Weyl quantization of the symbol (equivalently, classical observable) a,R (x, ξ ): Z a,R ((x + y)/2, ξ )eihx−y,ξ i/h¯ ψ(y) dydξ. (2.3) (Op W (a)ψ)(x) = (2π h¯ )−n Rn ×Rn
In this case ([Ro], Th.III.44) A = Op W (a,R ) extends from S(Rn ) to a trace-class selfadjoint operator in L2 (Rn ), realized as an h¯ - pseudodifferential operator ([Ro], II.4). 3. By standard results of perturbation theory ([Ka], §VII.1) the operator family PR (h¯ 0 , ) := Op W (p,R ) = P0 + A,R , A,R := OpW (a,R ) defined on D(P0 ) is positive selfadjoint in L2 (Rn ), and norm-resolvent continuous with compact resolvents for (h¯ 0 , ) in a neighbourhood I¯B × I¯B of zero because A,R is a fortiori continuous. We now compare SpecJ P and SpecJ P,R for |J | < +∞: Proposition 2.1. Let J ⊂ R+ be bounded and fixed. Then there exist R > 0 and ∗ > 0 independent of , h¯ 0 , such that, provided 0 ≤ ≤ ∗ one has 0 1/`
SpecJ (P ) = SpecJ (P,R ) (mod e−c/h¯
).
The proof of the proposition requires a decay property of the eigenfunctions of the Weyl operator corresponding to a Gevrey symbol, well known in the analytic case (see e.g. [HS,Ma]). We prove this property in three steps (in what follows C and c denote positive constants which may sometimes assume different values). Given a Gevrey function a ∈ GR`,α , supp a ⊂ QR , denote kak`,α the smallest constant K such that the following inequality holds: |j | (2.4) sup ∂zj a(z) ≤ K(|j |!)` /α |j | , z := (x, ξ ). z∈R2n
Lemma 2.1. Let a(x, ξ ) ∈ GR`,α be a Gevrey function, supp a ⊂ QR . Let η ∈ C ∞ (Rn ; R+ ) fulfill 0 ≤ η(x) < (σ |x|)β with β := 1/` and 0 < σ < α. Assume also that |η(x) − η(y)| ≤ (σ |x − y|)β . Define the operator β β ˜ [Aψ](x) := eη(x)/h¯ OpW ae−η(x)/h¯ ψ(x).
(2.5)
Then there is a positive CA = CA (α, `, kak`,α , R, σ ) < +∞ such that A˜ ≤ CA .
178
D. Bambusi, S. Graffi, T. Paul
Proof. Denote aˆ 2 (x, k) :=
1 (2π)n/2
Z
a(x, ξ )eihk,ξ i dξ.
By the Gevrey property there is a compactly supported smooth positive function A(x) such that β |aˆ 2 (x, k)| ≤ A(x)e−(α|k|) . Given ψ ∈ S, by definition of Weyl quantization (see (2.3)) we have Z η(x)−η(y) ˜ a((x + y)/2, ξ )eihx−y,ξ i/h¯ e h¯ β ψ(y) dydξ [Aψ](x) = (2π h¯ )−n Rn ×Rn Z η(x)−η(y) aˆ 2 ((x + y)/2, (x − y)/h¯ )e h¯ β ψ(y) dy, = (2π h¯ 2 )−n/2 Rn
and therefore Z ˜ 2 −n/2 [Aψ](x) ≤ (2π h¯ )
x+y A n 2 R
Define
1 f (ξ ) := (2π)n/2
Z
e
β − α x−y h¯
|ψ(y)|e
x−y β σ h¯
dy.
(2.6)
β
e−(α−σ )|z| e−ihξ,zi dz,
then the r.h.s. of (2.6) is equal to Z −n A((x + y)/2)f (ξ )eihx−y,ξ i/h¯ |ψ(y)| dydξ (2π h¯ ) Rn ×Rn i h = OpW (A(x)f (ξ )) |ψ| (x). Since A(x)f (ξ ) has an L1 Fourier transform OpW (A(x)f (ξ )) is bounded uniformly in h¯ ([Ro], Cor.II.19). This proves the assertion. u t Given a positive constant E define the classically allowed domain of |ξ |2 + V (x) by E := x ∈ Rn : V (x) ≤ E . E is compact. Denote cE its complement. Proposition 2.2. Let b(x, ξ ) := |ξ |2 + V (x) + a(x, ξ ), B = OpW (b), where V (x) fulfills (1.5) and a(x, ξ ) ∈ GR`,α , supp a ⊂ QR . Denote β := 1/`; let the interval J and 0 < σ < α be fixed. Then there exist constants η˜ = η(β, ˜ n, σ ) > 0, ς = ς(β, n, σ, V ) > 0 and ∗ = ∗ (β, n, σ, R, α, `, kak`,α ) > 0 such that, provided < ∗ , any eigenfunction ψ of B corresponding to an eigenvalue λ ∈ J satisfies the estimate
(2.7)
ψ exp (σ |x|/h¯ )1/` 2 c ≤ CE , L (E )
where E := max{η˜ + 2 sup λ, ς }. λ∈J
(2.8)
Normal Forms and Quantization Formulae
179
Proof. For E fixed, denote E − δ the set of the points x contained in E together with the ball Bδ (x) of radius δ and center x. Likewise, denote E + δ the set of the set ∪x∈E Bδ (x). Fix E and δ such that E − 2δ 6 = ∅. Let ηN (x) be a sequence of positive, smooth, compactly supported function such that (1) ηN (x) = 0, x ∈ E − δ; ηN (x) = (σ |x|)β , x ∈ E + N, |ηN (x) − ηN (y)| ≤ (σ |x − y|)β . (2) ∇ηN is bounded uniformly with respect to x ∈ Rn and N ∈ N. (1) and (2) can be fulfilled choosing E and δ large and small enough, respectively. Let furthermore χ(x) ∈ C0∞ (Rn ) be such that χ(x) = 1 for x ∈ E − 2δ, χ(x) = 0 for x 6 ∈ E − δ. We omit the dependence on N in the forthcoming argument which β β holds for N fixed. For λ ∈ J , denote T := −h¯ 2 4 +V , S(λ) := eη/h¯ T e−η/h¯ − λ, β β A0 := eη/h¯ OpW (a)e−η/h¯ , so that β
β
R := eη/h¯ OpW (b + a)e−η/h¯ − λ = S(λ) + A0 , with kA0 kL2 →L2 < C(R) by Lemma 2.1 (we drop the dependence on the Gevrey constants). The Weyl symbol of S(λ) is clearly s(x, ξ ; λ) = (ξ − i∇η)2 + V (x) − λ, whence D(S) = D(T ) by the boundedness of ∇η. Let u ∈ D(S) = D(T ). Then, since kRuk ≥ kS(λ)uk − C(R)kuk: kRuk2 ≥ (1 − C(R))kS(λ)uk2 − C(R)(1 + C(R))kuk2 ≥ (1 − C(R))k(1 − χ)S(λ)uk2 − C(R)(1 + C(R))kuk2 , whence kRuk2 ≥ (1 − C(R))hu, S(λ)∗ (1 − χ)2 S(λ)ui − C(R)(1 + C(R))kuk2 . (2.9) We estimate from below hu, S(λ)∗ (1−χ)2 S(λ)ui by the Garding inequality ([Ro, Th.III4] or [He, §2.4]). Remark that the principal symbol of S(λ)∗ (1 − χ)2 S(λ) is (1 − χ )2 |s(x, ξ ; λ)|2 . Define M(E, δ, λ) :=
inf
x∈E −2δ
[V (x) − λ]2 ,
and remark that, by assumption (1.5) on the potential and the Lagrange mean value theorem, there exists ς 0 which tends to zero as E → ∞ such that M(E, δ, λ) ≥ [E(1 − ς 0 δ) − λ]2 . Choose now the constant ς in such a way that ς 0 δ < 21 , so that M > (e/2 − λ)2 . Then one has (1 − χ )2 |s(x, ξ ; λ)|2 > M(E, δ, λ)(1 − χ )2 − η, η := sup |∇η(x)|2 . x∈Rn
(2.10)
¯ Then Th.III-4 of [Ro] yields Let C1 satisfy C1 > η. 1 hu, S(λ)∗ (1 − χ )2 S(λ)ui ≥ M k(1 − χ)uk2 − C1 kuk2 . 2
(2.11)
180
D. Bambusi, S. Graffi, T. Paul
Substituting in (2.9) we get kRuk2 ≥ M(1 − C(R))k(1 − χ)uk2 − C2 kuk2 , C2 (R, η, ) := C1 (η) + C(R)(1 + C(R)). Define now ∗ according to
1 C1 min 1, , ∗ := C(R) 2
so that C2 < 2C1 . Now kuk2 ≤ 2kχuk2 + 2k(1 − χ)uk2 , whence kRuk2 ≥ (M − 4C1 )k(1 − χ)2 uk − 4C1 kχuk2 .
(2.12)
Now take u := uN = eηN ψ, where ψ is an eigenfunction of B corresponding to the eigenvalue λ. Since ηN ≡ 0 on supp χ we have kχuN k ≤ kψk. On the other hand, RuN = 0 for all N. This yields β
(M − 4C1 )k(1 − χ )eηN /h¯ ψk ≤ 4C1 kψk
(2.13)
which holds provided M > 4C1 which is true provided the constant η˜ in (2.8) is chosen large enough. This estimate is independent of N , which proves the proposition. u t Corollary 2.1. Let ψ be an eigenfunction of B corresponding to the eigenvalue λ ∈ J . Fix a positive σ < α; then ∃C < ∞ such that |ψ(x)| ≤ C exp −(σ |x|/h¯ 0 )1/` , x ∈ cE (2.14) with E given by (2.8). Proof. ψ fulfills (2.7). Op W (aτ ) is bounded. Hence the eigenvalue equation −h¯ 0 4 2 ψτ + V ψτ + OpW (aτ )ψτ = λτ ψτ yields h¯ 0 k4ψτ k ≤ C kψτ k + kV ψτ k. Moreover: 2 β β Z Z − hx0 hx0 2 2 |V (x)ψτ (x)| dx + kV ψτ k = V (x)e ¯ e ¯ ψ(x) dx x∈E x ∈ / E 2
≤ sup |V (x)|2 kψk2 + sup |V (x)e x∈E
β − hx0 2 ¯
x ∈ / E
κ|x|
| kψτ e 2h¯ 0 k2 ≤ C.
Iterating the argument we get the existence of a constant C such that k1N ψτ k ≤ C N , ∀N ≥ 1, whence the thesis by (2.7) on account of the Sobolev embedding theorem. u t We come now to the heart of the proof. We have, defining p1 (x, ξ ) := p0 (x, ξ ) + V (x)t (|x|) , P1 := OpW (p1 ).
Normal Forms and Quantization Formulae
181
Proposition 2.3. There exists R∗ < ∞ and a function ∗ (R) such that, provided R > R∗ and < ∗ (R) one has 0β
Spec(P1 ) = Spec(P ) mod(e−c/h¯ ). Proof. Since p(x, ξ ) − p1 (x, ξ ) = [1 − t (|x|)]V (x), it is enough to prove that, if ψ is an eigenfunction of P or P1 , then 0β
k[1 − t (|x|)]V (x)ψk = O(e−c/h¯ ). Both p and p1 fulfill the assumptions of Corollary 2.1, and [1 − t (|x|)]V (x) is supported outside the domain B2R := |x| ≤ 2R. Hence the inequality holds provided B2R ⊃ E . This can be ensured by taking R large enough, which is possible since the constants η˜ and ς in Proposition 2.2 do not depend on R. On the contrary the constant ∗ will depend on R, and this yields the assertion. u t Proposition 2.4. Let ψ be an eigenfunction of P1 or of P,R . Then, there exist R∗ , two functions ∗ (R) and h(R) and suitable constants ς, C, c such that, provided R > R∗ , < ∗ (R), h¯ 0 < h(R) one has:
−1/`
(P − P,R )ψ ≤ C h¯ −ς e−ch¯ 0 . Proof. Instead of comparing P1 and P1 we compare theirZunitary images under the h¯ 0 de-
pendent Fourier transform Fh¯ 0 : (Fh¯ 0 )(ξ ) = (2π h¯ 0 )−n/2
0
Rn
u(x)e−ihx,ξ i/h¯ dx. Given a
symbol g(x, ξ ), it is well known (see e.g. [BS], §5.4) that the symbol of Fh¯ 0 OpW (g)Fh¯−1 0 is simply g(−ξ, x). Hence we compare the spectrum of the operators of Weyl symbols p(−ξ, x) ≡ p0 (ξ, x) + V (ξ )t (|ξ |)t (|x|) and p1 (ξ, x) ≡ p0 (ξ, x) + V (ξ )t (|ξ |). Both these operators fulfill the assumptions of Corollary 2.1; therefore their eigenfunctions fulfill (2.14). We denote W (ξ ) := V (ξ )t (|ξ |). Let ψ be an eigenfunction of either one of these operators; fix R1 such that BR1 ⊂ E . We now estimate −1 [(Fh¯ 0 P1,τ Fh¯−1 0 − Fh ¯ 0 Pτ Fh¯ 0 ψ)](x) Z i hx−y,ξ i = (2π h¯ 0 )−n W (ξ )[1 − t (|x + y|/2)]e h¯ 0 ψ(y) dydξ
Rnξ ×Rnx
−n/2 0 −n
Z
Wˆ ((x − y)/h¯ 0 )[1 − t (|x + y|/2)] ψ(y)dy Z Z −n + Wˆ ((x − y)/h¯ 0 )[1 − t (|x + y|/2)] ψ(y)dy. = (2π)−n/2 h¯ 0
= (2π)
h¯
Rn
|y|≤R1
|y|≥R1
Hence, by the decay properties of ψ and of the Fourier transform of W the second integral is estimated by β Z − σ (|x−y|β +|y|β ) −ς e h¯ 0 |1 − t (|x + y|/2)| dy. (2.15) C h¯ 0 Rn
Here ς is a suitable constant. Choose R = 3R1 , and consider first the case |x| > R/2; by the estimate |x − y|β + |y|β ≥ |x|β /2 + |y|β /2 our integral is majorized by Z 0 β 0 β −ς 0 −ς −(σ |x|/h¯ 0 )β e e−(σ |y|/h¯ ) dy < C h¯ 0 e−(σ |x|/h¯ ) . C h¯ |y|>R1
182
D. Bambusi, S. Graffi, T. Paul
For |x| < R/2 we need a stronger estimate. To obtain it remark that for |y| < 3R/2 one has |x + y| < R, and therefore, by the occurrence of 1 − t, the argument of the second integral in (2.15) vanishes unless |y| > 3R/2 which implies |x − y| > R. From this 0 β −ς it follows that the second integral of (2.15) is also estimated by C h¯ 0 e−(σ R/h¯ ) . We now estimate the first integral of (2.15). To this end we remark that it vanishes unless |(x + y)/2| ≥ R and |y| ≤ R1 . This gives |x − y| ≥ |x| − |y| ≥ |x + y| − 2|y| |x| R1 5 |x| |x + y| |x + y| + − 2|y| ≥ − + R − 2R1 ≥ + R − R1 . = 2 2 2 2 2 2 t Since R = 3R1 , proceeding as above we get the result. u 3. Perturbation Theory Estimates Consider the pseudodifferential operator PR (h¯ 0 , ) = P0 + A,R with Weyl symbol p,R (x, ξ ) = p0 (x, ξ ; ω) + a,R,h¯ 0 (x, ξ ) as in Proposition 2.1. We now construct the quantum normal form for PR (h¯ 0 , ) with uniform bounds for the remainder. The result is (dropping the dependence on R fixed): Theorem 3.1. There exist positive constants (independent of h¯ 0 , ) ∗ , C1 , . . . , C6 and, for < ∗ , a unitary transformation U (, h¯ 0 ) : L2 → L2 such that 1. The unitary image S(h¯ 0 , ) := U PR (h¯ 0 , )U −1 of PR (h¯ 0 , ) has a Weyl symbol σ (h¯ 0 , ) of the form σ (h¯ 0 , ) = p0 + Z(h¯ 0 , ) + R(h¯ 0 , ).
(3.1)
2. Z(h¯ 0 , ) admits the representation Z(h¯ 0 , ) =
N() X
ζs (x, ξ ; , h¯ 0 ) s .
(3.2)
s=1
Here N () := (∗ /)b , b := [(τ + 2)(2` + 3)]−1 ; ζs , s = 1, . . . , N() is a smooth, bounded semiclassical symbol depending on (x, ξ ) only through the actions Ik , k = 1, . . . , n. 3. The Weyl quantization of the symbols Z(h¯ 0 , ), R(h¯ 0 , ) fulfills the estimates: h i kOp W (Z(h¯ 0 , ))k ≤ C2 , kOpW (R(h¯ 0 , ))k ≤ C4 exp − (C5 /)b . (3.3) 4. For every fixed integer k < N() let Z (k) (I (x, ξ ); , h¯ 0 ) :=
k X
ζs (I (x, ξ ); , h¯ 0 ) s .
(3.4)
s=1
Then 3 Z (k) is independent of mod (h¯ 0 )∞ , i.e. there is Z˜ (k) (I (x, ξ ); h¯ ) such that 3 Z (k) (I (x, ξ ); , h¯ / 2 ) = Z˜ (k) (I (x, ξ ); h¯ ) + Ok ((h¯ 0 )∞ ).
(3.5)
Normal Forms and Quantization Formulae
183
Remark. Denoting Z = OpW (Z), R = OpW (R) we have S = P0 + Z + R, where Z commutes with P0 . Hence by (3.2,3.3) the eigenvalues of PR (h¯ 0 , ) have the expression hω, (l + 1)h¯ 0 i +
N() X
b
ζs ((l1 +1/2)h¯ 0 ω1 , . . . , (ln +1/2)h¯ 0 ωn ; , h¯ 0 ) s + O(e−(C5 /) ),
s=1
(3.6) where li ≥ 0, i = 1, . . . , n. We prove this result first for a particular class of analytic perturbations f of p0 , and then we recover the conditions of the theorem through an approximation procedure. To specify our assumptions on the perturbation f we need some preliminary notions. Define an analytic action 9 of Tn into R2n through the flow of p0 : 9 : Tn × R2n → R2n , (φ, (x, ξ )) 7 → (x 0 , ξ 0 ) = 9φ (x, ξ ), ξk sin φk + xk cos φk , ξk0 := ξk cos φk − ωk xk sin φk . xk0 := ωk
(3.7) (3.8)
If z := (x, ξ ), the flow of initial datum z0 is indeed z(t) = 9ωt (z0 ), ωt := (ω1 t, . . . , ωn t). If g ∈ L1loc (R2n ), its angular Fourier coefficient of order k is defined by Z 1 g(9φ (z))e−ihk,φi dφ, k ∈ Zn . g˜ k (z) := (2π)n Tn If g ∈ C1 one has, as is well known, X X g˜ k (z)eihk,φi H⇒ g(z) = g˜ k (z). g(9φ (z)) = k∈Zn
k∈Zn
Remark furthermore that g ≡ g˜ k for some fixed k if and only if g(9φ (z)) = eihk,φi g(z). Let now g ∈ L1 (R2n ), and consider its space Fourier transform Z 1 g(z)e−ihs,zi dz. b g (s) := (2π)2n R2n
(3.9)
(3.10)
Given ρ > 0, σ > 0, denoting gb ˜ k (s) the space Fourier transform of g˜ k , define the norm Z X eρ|k| |gb ˜ k (s)|eσ |s| ds. (3.11) kgkρ,σ := k∈Zn
R2n
Definition 3.1. Let ρ > 0, σ > 0. Then Aρ,σ := {g : R2n → C | kgkρ,σ < +∞}.
184
D. Bambusi, S. Graffi, T. Paul
Remarks. 1. If f ∈ Aρ,σ then f is analytic on R2n , and extends to a complex analytic function on a region of the form |Imzi | ≤ ai |Rezi |, with suitable ai . 2. As recalled in the former section, F := Op W (f ) is a trace-class, self-adjoint h¯ 0 pseudodifferential operator in L2 (Rn ) if f ∈ Aρ,σ . Let fb(s) be the Fourier transform of f . Since kfbkL1 ≤ kf kρ,σ , we have Z |fb(s)| ds ≡ kfbkL1 , kF kL2 →L2 ≤ kf kρ,σ . (3.12) kF kL2 →L2 ≤ R2n
Proposition 3.1. Let f : [0, ∗ ] 7 → Aρ,σ for some ρ > 0, σ > 0 and some ∗ > 0. Consider the operator Pf := P0 (h¯ 0 ) + F , F := OpW (f ). Let (1.6) hold. Set: E := sup kf kρ,σ ∈[0,∗ ]
and assume < 0 ρ τ σ 2 , 0 :=
γ . 2τ +5 e2 Eτ τ
(3.13)
Then the operator family Pf := P0 (h¯ 0 ) + F = OpW (p0 (x, ξ ; ω) + f (x, ξ )), D(Pf ) = D(P0 )
(3.14)
is norm-resolvent continuous, self-adjoint with compact resolvents in L2 (Rn ), and there exists a unitary transformation T : L2 (Rn ) → L2 (Rn ) such that Sf (h¯ 0 ) := T Pf (h¯ 0 )T −1 is a h¯ 0 − pseudodifferential operator family with symbol σf (x, ξ, h¯ 0 , ) = p0 (x, ξ, ω) + Z(I (x, ξ ); h¯ 0 , ) + R(x, ξ, h¯ 0 , ),
(3.15)
where Z ∈ Aρ/2,σ/2 and fulfills the estimate kZkρ/2,σ/2 ≤ 2E,
(3.16)
the principal symbol of Z is the normal form of the Hamiltonian p0 (x, ξ ) + f (x, ξ ) 1 computed up to order (0 ρ τ σ 2 /) τ +2 , and R ∈ Aρ/2,σ/2 is exponentially small, namely h i 1 (3.17) kRkρ/2,σ/2 ≤ eE exp −(τ + 2)(0 ρ τ σ 2 /) 2+τ . To prove the result we estimate the quantum normal form, recursively constructed through the quantum Lie transform [DGH,Ba1], and then choose the order of normalization in order to minimize the remainder. The semiclassical pseudodifferential calculus essentially reduces both steps to the classical method. We recall now the procedure of quantum Lie transform. 0 Let W be a bounded self-adjoint operator in L2 and etW/ i h¯ the corresponding unitary 0 group. If G is any bounded operator in L2 its formal image under etW/ i h¯ is 0
0
e−tW/ i h¯ GetW/ i h¯ =
∞ X
t l Gl ,
(3.18)
l=0
where Gl : l = 0, 1, . . . is recursively defined as follows: G0 := G, and 1 1 [W, Gl−1 ], l ≥ 1. l i h¯ 0 In analogy with the classical notion we give the following Gl =
(3.19)
Normal Forms and Quantization Formulae
185
Definition 3.2. Let the bounded self-adjoint operator W admit a smooth Weyl symbol w. Then the quantum Lie transform generated by w is the unitary operator exp (W/ i h¯ 0 ). Recall that if a bounded operator G also admits a smooth Weyl symbol g, then the Weyl symbol of −i[W, G]/h¯ 0 is by definition the Moyal bracket {w, g}M . ∞ X 0 0 gl , where Hence by (3.19) the (formal) Weyl symbol of e−W/ i h¯ GeW/ i h¯ is k=0
g0 := g, gl := {w, gl }M / l, l ≥ 1.
(3.20)
Consider now the operator Pf of symbol p0 + f , and its image under the quantum Lie transform generated by a symbol w of order . Formally the symbol of the image is p0 + {w, p0 }M + f + O( 2 ). For w such that {w, p0 }M + f depends on (x, ξ ) only through the actions Il the transformed operator commutes with P0 up to order 2 . A similar w can be determined and estimated by a purely classical method [Ba2] because, p0 being a quadratic polynomial, the Moyal and Poisson brackets coincide. We recall that in Fourier transform representation, used throughout the paper, the Moyal bracket is (see e.g. [Fo],§3.4): 2 ({g, g }M ) (s) = 0 h¯ 0
∧
Z R2n
h i b g (s 1 )gb0 (s − s 1 ) sin h¯ 0 (s − s 1 ) ∧ s 1 /2 ds 1 .
(3.21)
where, given two vectors s = (v, w) and s 1 = (v 1 , w1 ), s ∧ s 1 := hw, v1 i − hv, w1 i. The main step in proving Proposition 3.1 is: Proposition 3.2 (Iterative Lemma). Let the function H(k) := p0 + Z (k) + f (k)
(3.22)
be such that Z (k) , f (k) ∈ Aρ−kd,σ −kδ with d < ρ/(k + 1) and δ < σ/(k + 1), and let Z (k) be a function of (I1 , . . . , In ) only. Assume that
kZ
(k)
kρ−kd,σ −kδ ≤
E
0 k−1 X l=0
if k = 0 µls if k ≥ 1
kf (k) kρ−kd,σ −kδ ≤ Eµks
(3.23) (3.24)
with µs := 8c9
E , cψ := (τ/e)τ γ −1 . δ2 d τ
(3.25)
For h¯ 0 > 0 fixed, let H (k) be the Weyl quantization of H(k) . If µs < 1/2 there exists a unitary transformation Tk such that Weyl symbol of the transformed operator Tk H (k) Tk−1 := H (k+1) is given by (3.22) with k + 1 in place of k (i.e., H (k+1) = OpW (H(k+1) )), and satisfies (3.23,3.24) with k + 1 in place of k.
186
D. Bambusi, S. Graffi, T. Paul
Remarks. 1. In the notation of Theorem (3.1), OpW (Z (k) ) =
k X
ζs (P0 , h¯ 0 , ) s +
s=0
O( k+1 ). The eigenvalue
ζs ((l1 + 1/2)h¯ 0 ω1 , . . . , (ln + 1/2)h¯ 0 ωn , h¯ 0 , ), li = 0, 1, . . . , i = 1, . . . , n is the k-order coefficient of the quantum perturbation expansion of the eigenvalue of PR (h¯ 0 ) near the eigenvalue (l1 + 1/2)h¯ 0 ω1 + . . . + (ln + 1/2)h¯ 0 ωn of P0 . 2. The extra dependence on of the coefficients ζs is due to the V dependence of the perturbation. The multiplying V is the expansion parameter. The extra dependence disappears when the perturbation is a homogeneous polynomial. The proof of this proposition consists in performing the necessary estimates on the formal algorithm recalled above, to be described in the following lemmas. Let us introduce also the space Aσ of all functions g : R2n → C such that Z |b g (s)|eσ |s| ds < +∞, kgkσ := R2n
and let us estimate the Aσ norm of the Moyal bracket as in Lemma 4.1 of [Ba1]. Lemma 3.1. Let g ∈ Aσ , g 0 ∈ Aσ −δ be smooth. Then, ∀ 0 < δ 0 < σ − δ: k{g, g 0 }M kσ −δ−δ 0 ≤
1 kgkσ kg 0 kσ −δ . e2 δ 0 (δ + δ 0 )
(3.26)
Proof. Since |s ∧ s 1 | ≤ |s| |s 1 |, by definition of Aσ −norm and (3.21), we get k{g, g 0 }M kσ −δ−δ 0 = Z Z 0 e(σ −δ−δ )|s| ds =2
| sin (h¯ 0 (s − s 1 ) ∧ s 1 )/2| 1 ds h¯ 0 R2n R2n Z Z | sin (h¯ 0 s ∧ s 1 )/2| 1 0 1 ≤2 ds e(σ −δ−δ )(|s|+|s |) |b g (s)||b g 0 (s 1 )| ds 2n 2n h¯ 0 R R Z Z 0 0 1 ≤ e(σ −δ−δ )|s| |b g (s)| ds · e(σ −δ−δ )|s | |b g 0 (s 1 )||s ∧ s 1 | ds 1 R2n R2n Z Z 0 1 (σ −δ−δ 0 )|s| ≤ e |b g (s)| |s| ds · e(σ −δ−δ )|s | |b g 0 (s 1 )| |s 1 | ds 1 , R2n
|b g (s)b g 0 (s − s 1 )|
R2n
t whence the assertion because e−δx x ≤ 1/eδ, ∀x > 0, ∀δ > 0. u To estimate the Aρ,σ norm of the Moyal brackets we need to identify their angular Fourier coefficients. Lemma 3.2. For any φ ∈ Tn , and for any smooth symbol g, there exists a unitary operator Xφ : L2 → L2 with the property OpW (g ◦ 9φ ) = Xφ OpW (g)Xφ−1 .
(3.27)
Proof. For any fixed φ the map 9φ is a linear canonical transformation of R2n . Then (3.27) is a characteristic property of Weyl quantization (see e.g. [BS], §5.3). u t
Normal Forms and Quantization Formulae
187
This lemma entails that the k th spatial Fourier coefficient of an h¯ pseudodifferential symbol is also an h¯ pseudodifferential symbol. Lemma 3.3. Let g and g 0 be smooth functions such that g ≡ g˜ k , g 0 ≡ g˜ l0 for some fixed k, l ∈ Zn . Then, denoting h := {g, g 0 }M , one has h ≡ h˜ k+l . Proof. By the former lemma we can write Op W (h ◦ 9φ ) = Xφ OpW (h)Xφ−1 1 [Xφ OpW (g)Xφ−1 , Xφ OpW (g 0 )Xφ−1 ] i h¯ 0 = Op W ({g ◦ 9φ , g 0 ◦ 9φ }M ). =
Moreover, by the bilinearity of the Moyal bracket and (3.9): h ◦ 9φ = {eihk,φi g, eihl,φi g 0 }M = eih(k+l),φi {g, g 0 }M = eih(k+l,)φi h.
t u
We can now estimate the Moyal brackets and the individual terms of the quantum Lie transform expansion. Lemma 3.4. Let g ∈ Aρ,σ and g 0 ∈ Aρ,σ −δ . Then, for any positive δ 0 < σ − δ: k{g, g 0 }M kρ,σ −δ−δ 0 ≤
1 kgkρ,σ kg 0 kρ,σ −δ . e2 δ 0 (δ + δ 0 )
Proof. Denoting once more h = {g, g 0 }M , by Lemmas 3.3 and 3.26 we can write X X kh˜ k kσ −δ−δ 0 eρ|k| ≤ eρ|k+l| k{g˜ k , g˜0 l }M kσ −δ−δ 0 khkρ,σ −δ−δ 0 = k∈Zn
X
≤
k,l∈Zn
eρ|k|+ρ|l|
k,l∈Zn
kgk ˜ ρ,σ kg˜0 kσ −δ kg˜ k kσ kg˜0 l kσ −δ = . 2 0 0 e δ (δ + δ ) e2 δ 0 (δ + δ 0 )
t u
Lemma 3.5. Let g ∈ Aρ,σ and w ∈ Aρ,σ be two smooth symbols, and define gr :=
1 {w, gr−1 }M , r ≥ 1; g0 := g. r
Then gr ∈ Aρ,σ −δ for any 0 < δ < σ , and the following estimate holds: r kgr kρ,σ −δ ≤ δ −2 kwkρ,σ kgkρ,σ . (r)
Proof. Fix r, denote δ˜ := δ/r, and look for a numerical sequence Cl (r)
kgl kρ,σ −δl˜ ≤ Cl . By (3.26) a similar sequence can be recursively defined as follows: (0)
Cl
(r)
= kgkρ,σ , Cl
=
r2 (r) C kwkρ,σ . l 2 δ 2 e2 l−1
such that
(3.28)
188
D. Bambusi, S. Graffi, T. Paul
Therefore we get Cr(r)
1 = (r!)2
r 2 kwkρ,σ δ 2 e2
r kgkρ,σ .
(r)
Denoting Cr := Cr , clearly one has kgr kρ,σ −δ < Cr . Moreover: Cr =
r r −1 (r)
so that the sequence Cl
2r
kwkρ,σ (r − 1)2(r−1) −2r kwkrρ,σ e ≤ Cr−1 , 2 2 2r r [(r − 1)!] δ δ2
majorizes Cr , and this proves the lemma. u t
We now construct and estimate the solution of the “quantum” homological equation. Lemma 3.6. Let g ∈ Aρ,σ . Then the homological equation {p0 , w}M + Z = g
(3.29)
admits analytic solutions w, Z such that Z ◦ 9φ = Z; moreover, for any d < ρ: kZkρ,σ ≤ kgkρ,σ ; kwkρ−d,σ ≤ c9
τ τ 1 kgkρ,σ . ; c := 9 dτ e γ
Proof. Define Z := g˜ 0 ; w :=
X k6 =0
g˜ k . ihω, ki
(3.30)
Equation (3.30) solves (3.29). Since p0 is quadratic in z = (x, ξ ) one has indeed: d w(9ωt (z)) = {p0 , w}M (z) = {p0 , w}(z) = dt t=0 d X g˜ k (z) ihωt,ki X d X g˜ k (9ωt (z)) = e = g˜ k (z). dt t=0 ihω, ki dt t=0 ihω, ki k6=0
k6=0
k6 =0
The estimate of Z is trivial, while that of w is obtained as follows: kwkρ−d,σ =
X kg˜ k kσ e−d|k| eρ|k|−d|k| ≤ kgkρ,σ sup |hω, ki| k6 =0 |hω, ki| k6=0
kgkρ,σ τ τ e−dm mτ = γ γ ed m6=0
≤ kgkρ,σ sup and this proves the lemma. u t
Our final lemma deals with Aρ,σ estimates of the individual terms in the expansion of the quantum Lie transform of p0 (remark that p0 6 ∈ Aρ,σ ).
Normal Forms and Quantization Formulae
189
Lemma 3.7. Let g ∈ Aρ,σ be a smooth symbol, and w be the solution of the homological equation (3.29). Define the sequence pr0 : r = 0, 1, . . . as follows: 1 {w, pr−10 }M , r ≥ 1. r Then, for any 0 < d < ρ, 0 < δ < σ , pr0 ∈ Aρ−d,σ −δ and fulfills the following estimate: r−1 kgkρ−d,σ , k ≥ 1. kpr0 kρ−d,σ −δ ≤ 2 δ −2 kwkρ−d,σ p00 := p0 ; pr0 :=
Proof. It is enough to notice that, by (3.29), {w, p0 }M = Z − g, whence kp10 kσ ≤ t 2kgkσ . Hence the argument of Lemma 3.5 can be taken over to the present case. u Proof of Proposition 3.2. Consider the homological equation {p0 , w}M + Zk = f (k) ,
(3.31)
where Zk depends on z = (x, ξ ) only through I1 , . . . , In . By Lemma 3.6 and (3.24) w and Z (k) exist and fulfill the estimate c9 kwkρ−(k+1)d,σ −kδ ≤ τ Eµks ; kZk kρ−kd,σ −kδ ≤ Eµks . d Define Z (k+1) := Z (k) + Zk and X (k) X (k) X Zl + fl + pl0 ; f (k+1) := l≥1
(k)
l≥1
(k)
Z0 := Z (k) ; Zl
l≥2
1 (k) = {w, Zl−1 }M , l
l≥1
(k)
and analogous definitions for fl , pl0 . Then the symbol of the transformed operator has the form (3.22) with k + 1 in place of k. To get the estimate, let k ≥ 1. Set c9 η := δ −2 kwkρ−(k+1)d,σ −kδ ≤ τ 2 Eµks < 1/2. d δ By Lemmas 3.6, 3.7, 3.5 we can write X X (k) Eµs η kfl kρ−(k+1)d,σ −(k+1)δ ≤ Eµks ≤ 2ηEµks ηl = 1−η l≥1 l≥1 X (k) η ≤ 4Eη kZl kρ−(k+1)d,σ −(k+1)δ ≤ kZ (k) kρ−kd,σ −kδ 1−η l≥1 X kpl0 kρ−(k+1)d,σ −(k+1)δ ≤ 4µks ηE, l≥2
whence the assertion in a straightforward way. This proves Proposition 3.2. u t Proof of Proposition 3.1. Apply k times Proposition 3.2 to the operator Pf (h¯ 0 ) of symbol p0 + f with d := (ρ/2k), δ := (σ/2k). This yields a unitary map T such that the symbol of T Pf (h¯ 0 )T −1 has the form (3.22), Z (k) and f (k) fulfilling the estimates, !k 2τ +5 (k) (k) τ +2 Ec9 k . (3.32) kZ kρ/2,σ/2 ≤ 2E, kf kρ/2,σ/2 ≤ E ρτ σ 2
190
D. Bambusi, S. Graffi, T. Paul
The choice k := (0 ρ τ σ 2 /)1/(τ +2) yields the estimate (3.17); k > 1 by (3.13). This proves Proposition 3.1. u t Let us finally prove Theorem 3.1 approximating a by a function fulfilling the conditions of Proposition 3.1. First we show the existence of a constant µ > 0 such that Z X β β eµ|k| |b a˜ ,k (s)|eµ|s| ds < +∞, (3.33) hka ki := R2n
k∈Zn
where β := (1/2(` + 1)). The proof of this estimate is divided in two lemmas. Lemma 3.8. There exist constants C 1 and C such that: −|j | |j | (|j |!)`+1 . sup ∂ζ j a (9φ (z)) ≤ C 1 C z∈R2n ,φ∈Tn
Here ` is the Gevrey constant of a and ζi denotes a phase variable zi or an angle φi . Proof. Denote h (z, φ) := a (9φ (z)), N = |j | and consider ∂ζN h , where ζ is a fixed variable; we can write ∂ζN h =
N+2 X k=2
CkN Ok ,
(3.34)
where CkN are positive integers, and CkN Ok denotes a sum of CkN terms Ok of the form Ok = (d k−1 a )(9φ (z))G1 (z, φ) · · · Gk−1 (z, φ), k ≤ N. Here (d k a )(9φ (z)) denotes the k th differential of a evaluated at 9φ (z), and Gi : i = 1, . . . , k − 1 is a partial derivative of order at most N of 9φ (z). By (3.7,3.8) one clearly kGi (z, φ)k ≤ R2 for some R2 > 0, whence, by (2.2) has sup |z|≤2R,φ∈Tn
sup z∈R2n ,φ∈Tn
|Ok | ≤
X
sup ∂zkj a (z) R2k−1
|j |=k |z|≤2R
≤ K(R)α −k k 2n R2k−1 (k!)` ≤ K(R)α −k N 2n R2k−1 (k!)` . Hence to estimate (3.34) we just have to estimate the coefficients CkN . To this end, remark that we can symbolically write ∂ζ Ok = Ok+1 + (k − 1)Ok . The l.h.s. is majorized by the r.h.s. via triangular inequality, because in the estimate of Ok each Gj and all its derivatives clearly fulfill the same bound. This easily yields N−1 N −1 N N N (k − 1) + Ck−1 , 3 ≤ k ≤ N. CN +1 = C2 = 1, Ck = Ck
Hence, defining SN :=
N+1 X k=2
CkN one has SN :=
N +1 X k=2
kCkN −1 < N SN −1 whence SN <
N !. This proves the result when all derivatives are performed with respect to the same variable, but the same argument can be applied in general. This proves the lemma. u t
Normal Forms and Quantization Formulae
191
Lemma 3.9. Under the assumption of Theorem 3.1 there exist α˜ and C7 such that h i b 1/2(`+1) 1/2(`+1) ˜ + (α|s|) ˜ . (3.35) a˜ ,k (s) ≤ C7 exp − (α|k|) Proof. Performing l ≥ 1 partial integrations, the above lemma and the uniform Gevrey property (2.2) of a ◦ 9 yield, for some C8 > 0, C9 > 0: b a˜ ,k (s) ≤ C8 (C9 |(k, s)|)−l (l!)`+1 , where: |(k, s)| :=
n X p=1
get
|kp | +
2n X
|sp |. Set l = l(|(k, s)|) := (C9 |(k, s)|/e)1/(`+1) . We
p=1
b a˜ ,k (s) ≤ C10 exp − [C9 |(k, s)|/e]1/(`+1) .
Now remark that |(k, s)| = |k| + |s|; moreover, x 2l + y 2l > (x + y)l ∀ l ∈ N whenever x > 2, y > 2. Therefore, for min(|k|, |s|) > 22(`+1) : 1
1
1
(|k| + |s|) `+1 > |k| 2(`+1) + |s| 2(`+1) .
(3.36)
This yields the lemma for min(|k|, |s|) > 22(`+1) ; its validity when (3.36) does not hold t is ensured by suitably choosing the constants C7 . This proves the lemma. u Taking µ suitably small, it follows immediately from this lemma: Corollary 3.1. There is C11 > 0 independent of a such that hka ki ≤ C11 .
(3.37)
Proof of Theorem 3.1 Assertions 1,2. For 3 > 0 and K > 0 fixed, approximate a by X Z b (3.38) a˜ ,k (s)eihs,zi ds. a3,K (z) := |k|≤K |s|<3
We have: a3,K (z) − a3,∞ (z)
=
X Z |k|>K |s|<3
b a˜ ,k (s)eihs,zi ds.
= OpW (a3,K ), A3,∞ = OpW (a3,∞ ) by (3.12) we can Therefore, denoting A3,K write X Z β 3,∞ − A k ≤ |b a˜ ,k (s)| ds ≤ C11 e−µK , (3.39) kA3,K 2 2 L →L |k|>K |s|<3
β
− A kL2 →L2 ≤ C11 e−µ3 . kA3,∞
(3.40)
Now choose ρ = σ = α , K = 3 = −α with a still undetermined α. On account of (3.37, 3.39, 3.40) this yields: − A kL2 →L2 ≤ 2C11 exp [−µ/ αβ ]. kA3,K
192
D. Bambusi, S. Graffi, T. Paul
Now apply Proposition 3.1 to the operator P0 + A3,K of symbol p0 + a3,K . Then: E := ka 3,K kρ,σ ≤ eρK+σ 3 hkaki = e2 hkaki, and hence, by (3.37): E ≤ eC11 ≤ 0 ρ τ σ 2 = 0 α(τ +2)
(3.41)
holds provided α < 1/(τ + 2) and is suitably small. Hence, considering the symbol of the transformed operator, by (3.16) and (3.12) we have p0 + Z3,K + R3,K kOpW (Z3,K )kL2 →L2 ≤ 2e2 C11 , while by Assertion 1 of Proposition 3.1 the principal symbol of Z3,K coincides with the classical normal form computed up to order 1/(τ +2)
N() := 0
/ τ +2 −α . 1
(3.42)
and P0 +A,R being Moreover, the difference between the cut-off operator P0 +A3,K exponentially small, we can put in the remainder the exponentially small difference between the normal form of P0 + A,R and Z3,K . On the other hand, by (3.17): h i 1 )kL2 →L2 ≤ e3 C11 exp −(τ + 2)0 1/(τ +2) α− τ +2 . kOp W (R3,K We choose α so as to optimize the remainder, i.e. in such a way that αβ =
1 −α τ +2
H⇒
b := αβ =
1 . (τ + 2)(2` + 3)
To conclude the proof of Assertions 1 and 2 of Theorem 3.1 it is enough to set: 1/(τ +2)
∗ := 0
C3 = 2e2 C11 , C4 = (4 + e3 )C11 , C5 = min(µ, (τ + 2)∗ ).
t u
Assertion 3 will be a consequence of the property 3 a (z/) = V (z), ∀|z| ≤ R/2. We actually show that this property holds (mod ((h¯ 0 )∞ ), and on a sphere of radius smaller than R/2) for the symbols recursively defined by the normal form algorithm. Let w , g ∈ 6 1 (R2n ) (notations of [Ro]) be two smooth bounded semiclassical symbols with the property that, for some w, ¯ g¯ (independent of ) and some κ > 0: ¯ h¯ ), 3 g (z/, h¯ / 2 ) = g(z, ¯ h¯ ), |z| ≤ R/2. 3 w (z/, h¯ / 2 ) = w(z, Let M > 0 be fixed. Then the following properties are easily verified: 1. Denote d := {g , w }M . Then for all positive δ < R one has, as h¯ 0 → 0: ¯ h¯ ) + O((h¯ 0 )M ) , ∀|z| < (R − δ)/2. 3 d (z/; h¯ / 2 ) = d(z;
Normal Forms and Quantization Formulae
193
2. Let g˜ ,k be the angular Fourier coefficients of g . Then: 3 g˜ ,k (z/; h¯ / 2 ) = g¯˜ k (z; h¯ ), ∀|z| < R/2. 3. Let ϕt be the flow generated by the Hamiltonian w . Then (as → 0) 3 ϕt (z/; h¯ / 2 ) = ϕ¯t (z; h¯ ) + O((h¯ 0 )M ) , ∀|z| < (R − δ)/2. 0
0
4. Denote W := Op W (w ) and gw the symbol of eiW /h¯ OpW (g )e−iW /h¯ . By the semiclassical Egorov theorem ([Ro], IV.10) one has (as h¯ 0 → 0 and → 0) 3 gw (z/; h¯ / 2 ) = g¯ w (z; h¯ ) + O((h¯ 0 )M ) , ∀|z| < (R − δ). Then Assertion 3 is an immediate consequence of the uniqueness of the normal form of nonresonant systems and of the following lemma. Lemma 3.10. Let M > 0 be fixed and suitably large, and let the function (k)
(k)
H(k) := p0 + Z (k) + k+1 S1 + S2
(3.43)
be such that Z (k) depends on (I1 , . . . , In ) only. Assume that (k)
(k)
3 Z (k) (z/, h¯ / 2 ) = Z¯ (k) (z, h¯ ) , 3 S1 (z/, h¯ / 2 ) = S¯1 (z, h¯ ) (k)
(3.44) (k)
for |z| ≤ (R − δk)/2. Assume furthermore Z (k) , S1 ∈ 6 1 (R2n ) and S2 = Ok ((h¯ 0 )M ). Then there exists a unitary transformation Tk such that the transformed Schrödinger operator Tk H (k) Tk−1 := H (k+1) has a Weyl symbol given by (3.43,3.44) with k + 1 in place of k: H (k+1) = OpW (H(k+1) ). Proof. We construct the generator w of the transformation as in the proof of Lemma 3.2, (k) (k) with k+1 S1 in place of f (k) . Remark that, since S1 ∈ C ∞ (R2n ), its angular Fourier (k) coefficient S˜ are O(|l|−∞ ). Therefore by (1.6) w exists in C ∞ (R2n ), uniformly with 1,l
respect to . Moreover, by Property 2, 3 w(z/, h¯ / 2 ) = w(z, ¯ h¯ ). Then in the notation of Remark 4 above we have (k),w
S1
0 (k),w = S¯1 + O(h¯ ,M ) , Z (k),w = Z¯ (k),w + O(h¯ M ), |z| ≤ (R − (k + 1)δ)/2. 0
Now by the homological equation we easily get: p0w = p¯ 0w + O(h¯ M ). Hence let us set: (k+1)
S1
(k),w (k) := S¯1 − S1 + p¯ 0w − p0 − k+1 {w, p0 }M + Z¯ (k),w − Z (k) , (3.45) (k)
Z (k+1) := Z (k) + k+1 S˜1,0 .
(3.46)
(k),w and the other transformed It is now easy to see, using the explicit formula for S¯1 symbols (see once more [Ro], Theorem IV.10) that the l.h.s. of (3.45) is actually of order . Then the assertion is a direct consequence of the above listed properties. This proves the lemma. u t
194
D. Bambusi, S. Graffi, T. Paul
Proof of Theorem 1.1. By Theorem 2.1 and Theorem 3.1 the eigenvalues λk (h¯ 0 , ), k ∈ (N ∪ {0})n of P (h¯ 0 , ) admit the approximate quantization formula 1 λs (h¯ 0 , ) = ω · (k + )h¯ 0 Z((k1 + 1/2)h¯ 0 , . . . , (kn + 1/2)h¯ 0 , h¯ 0 , ) 2 ˜ + R(, h¯ 0 ), ˜ R(, h¯ ) = O(C4 e 0
−(C5 /)b
) + O(e
−c/h¯ 0 1/`
),
(3.47) (3.48)
where Z() = Z N() is given by (3.2) and the constants have been determined above. On the other hand, Spec[P (h¯ 0 )] = 2 Spec[P (h¯ 0 , )]. Hence, λ ∈ Spec[P (h¯ )] ∩ [0, 2 ] ⇐⇒ λ/ 2 ∈ Spec[P (h¯ 0 , )] ∩ [0, 1]. This yields (1.8), i.e. Assertion (1). To see (1.9), for any fixed M < (∗ /)b write: Z((k + 1/2)h¯ / 2 , h¯ / 2 , ) =
M X
ζs ((k + 1/2)h¯ / 2 , h¯ / 2 ) s
s=1
+
N() X
ζs ((k + 1/2)h¯ / 2 , h¯ / 2 ) s .
s=M
The second addend is O( M ) by construction. By Assertion 3 of 3.1, 3
M X
ζs ((k + 1/2)h¯ / 2 , h¯ / 2 ) s
s=1 ˜M
=Z
((k1 + 1/2)h¯ , . . . , (kn + 1/2)h¯ , h¯ ) + OM ((h¯ / 2 )∞ ).
˜ note that by [Sj], Theorem 1.4 the quantum normal form is constructed To identify Z, as a h pseudodifferential operator commuting with P0 by quantization of the classical Lie transform as in Sect. 3. Hence by Remark 2 after Proposition 3.2 and Assertion 3, formula (3.5), F˜ = F. This concludes the proof of Theorem 1.1. u t Proof of Corollary 1.1. Consider now the scaling parameter , which can be chosen as an arbitrary function (h¯ ) vanishing as h¯ → 0 subject to the only requirement lim h¯ (h¯ )−2 → 0.
h¯ →0
(3.49)
Let
2 = ϕ(h¯ ) :]0, 1[: R+ i be such that ϕ(h¯ ) exp −(C5 /ϕ(h¯ ))b = O(h¯ ∞ ). This condition is satisfied by any h
positive, increasing function ϕ(h¯ ) on ]0, 1] with lim ϕ(h¯ )b ln h¯ → 0. The requirement h¯ →0 h¯ /ϕ(h¯ )b/2+1 → 0 ensures that exp − 2 /h¯ exp − A/ϕ(h¯ )b/2 . This choice of fulfills (3.49) in both cases, and hence (1.11) is proved. u t Acknowledgement. We thank A. Martinez for suggesting the proof of Proposition 2.2.
Normal Forms and Quantization Formulae
195
References [Ba1]
Bambusi, D.: Uniform Nekhoroshev estimates on quantum normal forms. Nonlinearity 8, 93–105 (1995) [Ba2] Bambusi, D.: Exponential Stability of Breathers in Hamiltonian Networks of Weakly Coupled Oscillators. Nonlinearity 9, 433–457 (1996) [BS] Berezin, F.A. and Shubin, M.S.: The Schrödinger Equation. Amsterdam: Kluwer, 1991 [BV] Bellissard, J., Vittot, M.: Heisenberg Picture and non Commutative Geometry of Classical Limit in Quantum Mechanics. Ann. Inst. H. Poincaré 52, 175–235 (1990) [CDS] Combes, J.M., Duclos, P., Seiler, R.: The Born-Oppenheimer Approximation. In: NATO ASI Series B 72, R.: Plenum, New York 1981, 363–391 [DGH] Degli Esposti, M., Graffi, S., Herczynski, J.: Exact quantization of the classical lie algorithm. Ann. Phys. (N.Y.) 208, 363–391 (1991) [Fo] Folland, G.: Harmonic analysis in phase space. Princeton , NJ: Princeton University Press, 1988 [Gi] Giorgilli, A.: Rigorous results on the power expansions for the integrals of a hamiltonian system. Ann. Inst. H. Poincaré, 48, 423–439 (1988) [He] Helffer, B.: h-pseudodifferential operators and applications: An introduction. Lectures at IMA: Quasiclassical methods, 1995 [HS] Helffer, B.,Siöstrand, J.: Multiple wells in the semiclassical limit. Commun.Part. Differ. Eqs. 9, 337– 408 (1984) [Ka] Kato, T.: Perturbation Theory for Linear Operators. Berlin–Heidelberg–New York: Springer-Verlag, 1966–1976 [Ma] Martinez, A.: Microlocal Exponential Estimates and Applications to Tunneling. In: NATO ASI Series C 490, A,sterdam: Kluwer, 1997, pp. 349–378 [Ne] Nekhoroshev, N.N.: Exponential estimate of the stability time of near integrable Hamiltonian systems. Russ. Math. Surveys. 32 (6), 1–65 (1977) [Po] Popov, V.: Invariant tori effective stability and quasimodes with exponentially samll terms. Preprint 1997 [Ro] Robert, D.: Autour de l’approximation semiclassique. Basel: Birkhäuser, 1987 [RS] Reed, M. and Simon, B.: Methods of Modern Mathematical Physics, Vol.IV. New York: Academic Press, 1978 [Si] Simon, B.: Semiclassical analysis of low-lying eigenvalues, I. Ann. Inst. H. Poincaré 38B, 295–307 (1983) [Sj] Sjöstrand, J.: Semi-excited levels in non-degenerate potential wells. Asymptotic Analysis 6, 29–43 (1992) Communicated by B. Simon
Commun. Math. Phys. 207, 197 – 229 (1999)
Communications in
Mathematical Physics
© Springer-Verlag 1999
The Wulff Construction in Three and More Dimensions T. Bodineau CNRS – UMR 7599, Universités Paris 6 and 7, Laboratoire de Probabilités et Modèles Aléatoires, Département de Mathématiques, Case 7012, 2 place Jussieu, 75251 Paris, France Received: 5 January 1999 / Accepted: 27 April 1999
Abstract: In this paper we prove the Wulff construction in three and more dimensions for an Ising model with nearest neighbor interaction.
1. Introduction The problem of phase separation for two dimensional Ising model and the study of the equilibrium shape of crystals (Wulff shape) has been initiated by Dobrushin, Kotecky and Shlosman [DKS]. Among other things, they proved that if at very low temperatures we decrease the averaged magnetization in the + pure phase, we observe the creation of a macroscopic droplet of the − phase which has a deterministic shape on the macroscopic scale. The proof has been first simplified by Pfister [Pf] and then extended to the whole of the phase transition region by Ioffe [I1,I2] (see also [SS] and [PV]). Recently, Ioffe and Schonmann [IS] have completed the DKS theory up to the critical temperature and greatly simplified the original proofs. Moderate deviations in the exact canonical ensemble are also studied in [IS]. In two dimensions, the proofs have been based on duality arguments and on a coarse graining procedure (skeleton). These arguments do not seem to apply in higher dimensions. For more than two dimensions, an alternative procedure has been proposed by Alberti, Bellettini, Cassandro and Presutti [ABCP,BCP] for Ising systems with Kac potentials. They rephrase the whole problem in terms of L1 theory and prove large deviations for the appearance of a droplet of the minority phase in a scaling limit when the size of the domain diverges not much faster than the range of the Kac potentials. This amounts to a weak large deviation principle which is obtained by proving 0-convergence of a functional associated to the spins system [ABCP]. A large deviation principle has then been proved via a tightness property [BCP].
198
T. Bodineau
Their approach has been generalized by Benois, Bodineau, Butta and Presutti [BBBP, BBP] by taking first the thermodynamic limit and then letting the range of interaction go to infinity. The first paper [BBBP] was devoted to the proof of a weak large deviation principle for the macroscopic magnetization which is equivalent to the computation of surface tension. The main idea has been to introduce a coarse graining in order to use the L1 setting. Namely, events in L1 were related to mesoscopic quantities by an argument which we will refer to later as minimal section argument. An exact expression of surface tension was difficult to recover from coarse grained estimates and surface tension was only derived in the Kac limit, i.e. when the range of interactions tends to infinity. The second step [BBP] consisted of proving a tightness property by using the compactness in L1 of the set of functions of bounded variation with finite perimeter. Wulff construction for three dimensional independent percolation has been proven by Cerf [Ce] using a procedure similar to the one of [BBBP] and a novel definition of surface tension. In this case, the dependence on boundary conditions is weaker and, the minimal section argument enables to prove directly a weak large deviation principle by using this appropriate definition of surface tension. As percolation occurs in an infinite volume, there is an extra difficulty and different compactness arguments have been required. In this paper we proceed as in [BBBP]. One of the main difficulies is to recover surface tension from a constraint on the averaged magnetization. The surface tension + is defined as log ZZ+,− , where the partition functions are computed with + boundary conditions and with mixed boundary conditions (+ at the top and − at the bottom), see for instance the paper of Messager, Miracle-Solé and Ruiz [MMR]. To use directly this definition, one would have to find in the bulk surfaces of + spins or of − spins which in fact may not exist. A way to circumvent this problem is to prove that surface tension can be produced by averaging the boundary conditions, choosing the spins with respect to the + pure phase and to the − pure phase. For the Ising model with nearest neighbor interaction, the coarse graining developed by Pisztora [Pi1] will play an analogous role to the one used for the Kac model. Pisztora’s coarse graining is one of the most profound and powerful techniques for the study of the Ising (Potts) model, it provides an accurate description of the Ising model in a non-perturbative regime up to a temperature Tˆc which is conjectured to agree with the critical temperature. In the following, we will mention which of our results hold up to Tˆc . As Pisztora’s coarse graining is defined via the FK representation, several quantities need to be rewritten in terms of the FK representation. In particular, our approach to the surface tension (Sect. 4) is built upon the FK representation and, is motivated by the corresponding construction in [Ce]. This is a key to obtain precise surface order estimates on the logarithmic scale. This is also the only point at which we refer to [Ce], the core philosophy of our proof is based on the renormalization ideas of [BBBP] and [BBP], including the appropriate setup of the geometric measure theory. The coarse graining schemes of the latter works, however, depend on specific properties of Kac potentials and, one of our main technical tasks here is to develop a relevant modification of these renormalization procedures in the nearest neighbor context. A step further in the understanding of the surface tension will be to prove a phase separation theorem for the Kac model with finite range interactions by using a coarse graining defined only in terms of the Gibbs measure [Bo]. After introducing the main notation, we state in Sect. 2 the results and an overview of the paper (see Subsect. 2.3).
Wulff Construction in Three and More Dimensions
199
2. Notation and Results The strategy of the proof can be applied in any dimension larger or equal to 2. Nevertheless, as the proof depends on Pisztora’s coarse graining [Pi1] our results are valid only in dimension larger or equal to 3. In fact, a similar coarse graining should also hold in dimension 2. For simplicity, notation and results are stated in three dimensions.
2.1. Notation. We introduce the following norms on Z3
∀x ∈ Z3 , kxk1 =
3 X
v u 3 uX |xi | and kxk2 = t |xi |2 .
i=1
i=1
Two vertices x, y in Z3 are nearest neighbors if kx − yk1 ≤ 1 and we denote it by x ∼ y. For any finite subset 3 of Z3 , we define its boundary by ∂3 = {x ∈ 3c | y ∈ 3, x ∼ y}, and denote its cardinality by |3|. We consider the Ising model on Z3 with nearest neighbor interaction. Each spin σi , attached at the lattice site i in Z3 , can take values ±1. For any integer N, we set 1N = {1, N}3 and denote the space of configurations in 1N by 61N = {±1}1N . Let σ1N be the spin configuration restricted to 1N . We introduce the Hamiltonian associated to σ1N with boundary conditions σ∂1N , H (σ1N | σ∂1N ) = −
1 X σi σj − 2 i∼j i,j ∈1N
X
σi σj .
i∼j i∈1N ,j ∈∂1N
The Gibbs measure on 61N at inverse temperature β > 0 is µβ,1N (σ1N | σ∂1N ) =
1 exp − βH (σ1N | σ∂1N ) , Zβ (σ∂1N )
where the partition function Zβ (σ∂1N ) is the normalizing factor. When the boundary conditions σ∂1N are identically equal to 1, we simply write µ+ β,1N . There is a critical value βc > 0 and for all β larger than βc , there exists mβ > 0 such that lim µ+ β,1N (σ0 ) = mβ > 0.
N→∞
+ The Gibbs measure µ+ β on 6Z3 obtained by taking the thermodynamic limit of µβ,1N is called the + pure phase and mβ is the equilibrium value of the magnetization.
200
T. Bodineau
2.2. Surface tension. Let us recall the definition of surface tension and related results which can be found in [MMR]. The set of unit vectors in R3 is denoted by S2 . We fix n a vector in S2 and e1 , e2 two vectors orthogonal to n. Let h be a positive constant and N → f (N ) a positive function ¯ which diverges as N goes to infinity. For any integer N, we denote by 3(hN, hN, f (N )) the parallelepiped of R3 centered at 0 with faces parallel to the axis (e1 , e2 , n) such that the lengths of the sides parallel to (e1 , e2 ) are hN, hN and the ones parallel to n is f (N ). ¯ hN, f (N )) ∩ Z3 . The boundary ∂3N is split We introduce 3N the set of vertices 3(hN, into 2 sets ∂ + 3N = {i ∈ ∂3N | i. n ≥ 0}, ∂ − 3N = {i ∈ ∂3N | i. n < 0}. We call ∂ + 3N the upper and ∂ − 3N the lower part of ∂3N . We fix the boundary conditions outside 3N to be equal to 1 on ∂ + 3N and to −1 on ∂ − 3N . The corresponding +,− . partition function on 3N is denoted by Z3 N Definition 2.1. The surface tension in the direction n ∈ S2 is defined by +,− Z3 1 τ (n) = lim − 2 2 log +N . N→∞ h N Z3N
The surface tension defined above coincides with the one defined in [MMR] (see Appendix 8.1). It depends neither on h nor on f as proven in [MMR] (Theorem 2). Let us extend τ by homogeneity v and τ˜ (0) = 0. (2.1) ∀v ∈ R3 − {0}, τ˜ (v) = kvk2 τ kvk2 The pyramidal inequality proven in Theorem 3 of [MMR] ensures that τ˜ is convex. As τ˜ is locally bounded and convex, it is continuous. It was proven by Lebowitz and Pfister [LP] that for all β larger than βc , the surface tension τ (n0 ) in the direction n0 = (1, 0, 0) is positive. From the symmetries and the convexity of τ , we check that τ is uniformly positive on S2 . The spin configuration σ should be seen as a microscopic representation of the system. The macroscopic state of the system is instead determined by the value of an order parameter (the averaged magnetization) which specifies the phase of the system. As β is fixed, it is convenient to replace the order parameter by a parameter u with values ±1. We suppose that the macroscopic region of R3 where our system is confined is T = [0, 1]3 and that outside this region the system is surrounded by the + pure phase. We denote by BV(T , {+1, −1}) the set of functions of bounded variation in ] − 1, 2[3 which are uniformly equal to 1 outside T and which take only values ±1 in T (see [EG] for a review). These functions are simply signed indicators of macroscopic sets of finite perimeter. The fact that ur = 1 for some r in T means that locally at r the system is in equilibrium in the phase +mβ . The precise correspondence between σ and functions on T is described in Sect. 3, where we approximate σ by a coarse graining procedure, introducing a mesoscopic scale. For any function u in BV(T , {+1, −1}), a generalized notion of the boundary of the set {u = −1} can be defined. It is called the reduced boundary and denoted by ∂ ∗ u. If the set {u = −1} is regular then ∂ ∗ u coincides with the usual definition of the boundary.
Wulff Construction in Three and More Dimensions
201
For each x in ∂ ∗ u, there exists a unit vector nx called the measure theoretic unit normal to {u = −1} at x which coincides with the outward normal if the set is regular. Let us introduce the functional F on L1 (T , [− m1β , m1β ]) R F(u) =
∂ ∗ u τ (nx ) dHx , if u ∈ BV(T , {+1, −1}), ∞, otherwise,
(2.2)
where dH is the 2 dimensional Hausdorff measure in R3 . To any subset A of T , we associate the function 1A = 1Ac − 1A and simply write F(A) = F(1A ). An important property is the lower semi-continuity of F with respect to L1 convergence. As τ˜ is convex (see (2.1)), the lower semi-continuity is a consequence of a result by Ambrosio and Braides (see [AmBr] Theorem 2.1 and Example 2.8). The equilibrium crystal shape Wm , called Wulff shape, is a solution of the following isoperimetric problem: Z ur dr ≤ m , (2.3) min F(u) u ∈ BV(T , {+1, −1}), mβ T
where m belongs to ]m∗ , mβ [. We will restrict the parameter m so that, for m in ]m∗ , mβ [ the minimizers of the variational problems in T and R3 are the same. This enables us
to avoid boundary problems. The shape Wm can be explicitly constructed (the Wulff construction) by dilating the set \ x ∈ R3 ; x. n ≤ τ (n)} W= n∈S2
R in order to satisfy the volume constraint mβ T 1Wm (r) dr = m. As Wm = λm W, one has F(Wm ) = λm F(W). Thus F(Wm ) is continuous with respect to m. Taylor [Ta] proved that Wm is a closed convex surface and that all other minimizers of (2.3) are deduced from Wm by shifts (see also [F,FM]). In the following, we suppose that Wm is centered in ( 21 , 21 , 21 ). P 2.3. Heuristics and Results. The total magnetization N13 i∈1N σi will be denoted by M1N . A shift of the magnetization from its equilibrium value leads to large deviations controlled by a surface order Theorem 2.1. There is β0 positive such that for any β larger than β0 and m in ]m∗ , mβ [, lim
N→∞
1 log µ+ β,1N M1N ≤ m = −F(Wm ), 2 N
where m∗ and Wm were defined in (2.3). More precisely, a phase separation occurs on the macroscopic level. In order to describe it, we introduce an intermediate scale called mesoscopic: the magnetization is locally averaged on boxes of size N α with α ∈]0, 1[ (1 N α N ). For any integer L, we define the sub-lattice L L L 3 x = (x1 , x2 , x3 ) ∈ Z . Lx1 + , Lx2 + , Lx3 + (2.4) LL = 2 2 2
202
T. Bodineau
We introduce also B(x, L) the box of length L centered in x in LL , L L . B(x, L) = y ∈ Z3 | ∀i ∈ {1, 2, 3}, − < yi − xi ≤ 2 2
(2.5)
The parameter α will be chosen slightly bigger than 13 (see Subsect. 3.2), this restriction on α is only used because we want to provide a straightforward proof of Theorem 2.2: by tuning properly the parameter α, we avoid many entropic computations in particular in Sect. 7. Nevertheless, notice that an extension of Theorem 2.2 for appropriate finite mesoscopic scales is derived in [BIV]. To simplify the notation, we can choose α in Q −1 and suppose from now that N is of the form 2α k , where k and αk are integers. The set 1N is partitioned into boxes B(x, N α ) of side length N α centered in x in LN α . For general values of N, one would need to partition 1N with boxes which may have different sizes. This is a standard technique and we refer the reader to Pisztora [Pi1]. The local magnetization M is a piecewise constant function on T , ∀r ∈ T , Mr =
1 N 3α
X
σj if ∀i, −
j ∈B(x,N α )
Nα Nα < N ri − xi ≤ . 2 2
As explained in the introduction, it is convenient to formulate the problem of phase separation in terms of L1 theory. For any function u in L1 (T , [− m1β , m1β ]), we denote by V(u, δ) the δ-neighborhood of u, Z 1 1 , ] |ur − vr | dr ≤ δ . V(u, δ) = v ∈ L1 T , [− mβ mβ We can now on phase separation which says that for β large enough with state a theorem M1 ≤ m -probability converging to 1, the function M is close to some . µ+ N β,1N translate of the Wulff shape mβ 1Wm . Theorem 2.2. There is β0 positive such that for any β larger than β0 and m in ]m∗ , mβ [, ∀δ > 0,
lim µ+ β,1N
N→∞
[ M ∈ V(1Wm +r , δ) M1N ≤ m = 1, mβ 0 r∈T
where m∗ , Wm were defined in (2.3) and T 0 = {r ∈ T | Wm + r ⊂ T }. This result is far less sharp than those obtained in the 2 dimensional case (see [IS]). We will follow the scheme of [BBBP,BBP] and deduce Theorems 2.1 and 2.2 from the following statements. Proposition 2.1. Let β be large enough. Then for all u in BV(T , {+1, −1}) such that F(u) is finite, lim lim sup
δ→0 N→∞
1 M log µ+ ∈ V(u, δ) ≤ −F(u). β,1N 2 N mβ
Proposition 2.2. For β large enough and m in ]m∗ , mβ [, lim lim inf
δ→0 N→∞
1 M log µ+ ∈ V(1Wm , δ) ≥ −F(Wm ), β,1N N2 mβ
where m∗ and Wm were defined in (2.3).
Wulff Construction in Three and More Dimensions
203
Since Proposition 2.1 is only a weak large deviation principle, we need to strengthen it by proving an exponential tightness property which is similar to the one in [BBP]. For any a positive, the set Ka = u ∈ BV(T , {+1, −1}) | F(u) ≤ a is compact with respect to convergence in measure: As F is lower semi-continuous, Ka is closed and as the surface tension τ is larger than a positive constant τ0 , the set Ka is included in the compact set of functions of bounded variation in T with perimeter smaller than τa0 (see [EG] Sect. 5.2.3). Proposition 2.3. We fix β large enough. Then there exists a constant Cβ such that for all a and δ positive, lim sup N→∞
1 M log µ+ ∈ V(Ka , δ)c ≤ −Cβ a, β,1N 2 N mβ
where V(Ka , δ) is the δ-neighborhood of Ka in L1 (T , [− m1β , m1β ]). The proofs of Theorems 2.1 and 2.2 are based on well known large deviations arguments (see [DS]). For completeness we prove Theorem 2.2, the proof of Theorem 2.1 is similar. Proof of Theorem 2.2. Throughout this proof, the parameter δ is fixed. We first start by M ∈ F , where F is the closed set computing a logarithmic upper bound of µ+ β,1N mβ Z [ 1 m 1 1 , ] u 6∈ V(1Wm +r , δ) and ur dr ≤ . F = u ∈ L T , [− mβ mβ m β T 0 r∈T
Let a and µ+ β,1N
δ0
be positive constants. We split F into 2 sets
M M M ∈ F ≤ µ+ ∈ F ∩ V(Ka , δ 0 ) + µ+ ∈ V(Ka , δ 0 )c . β,1N β,1N mβ mβ mβ
We choose a such that Cβ a is much larger than F(Wm ), then Proposition 2.3 enables us to bound the last term in the RHS. Let us fix ε positive. Since Ka is compact, we cover it with a finite number ` of neighborhoods V(ui , εi ), where each εi belongs to ]0, ε] and is chosen such that Proposition 2.1 implies lim sup N→∞
M 1 log µ+ ∈ V(ui , εi ) ≤ −F(ui ) + ε. β,1N N2 mβ
δ0
For small enough, we cover F ∩ V(Ka , δ 0 ) with the above neighborhoods which intersect F . Since ui belongs to V(F, ε), we get from Lemma 2.1.2 of [DS], lim sup lim 0
δ →0 N→∞
1 M log µ+ ∈ F ∩ V(Ka , δ 0 ) ≤ − lim inf F(u) β,1N 2 ε→0 u∈V (F,ε) N mβ ≤ − inf F(u). u∈F
Furthermore, results by [F,FM] on the variational problem (2.3) imply that inf F(u) < F(Wm ).
u∈F
204
T. Bodineau
Thus the upper bound is completed. To complete the bound for the probability of the proof it remains to derive a lower M ∈ V(1Wm−ε , δ 0 )} is included event M1N ≤ m . We fix ε positive and check that { m β in {M1N ≤ m} for δ 0 small enough. Proposition 2.2 implies lim inf lim 0
δ →0 N→∞
1 M log µ+ ∈ V(1Wm−ε , δ 0 ) ≥ −F(Wm−ε ). β,1N 2 N mβ
Letting ε go to 0, we obtain lim inf N→∞
1 log µ+ β,1N {M1N ≤ m} ≥ −F(Wm ). 2 N
Combining the upper and the lower bound, we derive Theorem 2.2. u t Let us comment on Propositions 2.1, 2.2 and 2.3.A shift of the averaged magnetization can be realized by 2 competing effects. The first one, which consists of producing a large droplet of − inside the bulk, is controlled by surface tension (Propositions 2.1 and 2.2). The second one consists of increasing homogeneously the number of small − contours. This requires a lot of energy, but may be favored by entropy. This effect is ruled out by Proposition 2.3 which is a combination of an estimate in the phase of small contours with a Peierls type estimate for large contours. In fact, the underlying phenomena are more subtle and it was shown by [IS] in the case of dimension 2, that on the level of moderate deviations the second effect may be the most important. We start by defining a coarse graining on the mesoscopic scale which keeps more details of the microscopic structure than M. This is done, in Sect. 3, via the FK representation by using Pisztora’s results [Pi1]. This coarse graining procedure imposes that β is larger than a critical value β˜c related to the slab percolation threshold (see [Pi1]) and to condition (3.3). It is conjectured that β˜c equals the critical value βc . In Sect. 4, motivated by [Ce], we use an alternative definition of surface tension in terms of the FK representation. We prove the equivalence of several expressions for surface tension which will enable us to compare different boundary conditions. In Sect. 5, Proposition 2.1 is proven along the lines of the argument developed in M ∈ V(u, δ) are those for [BBBP]. It states that the most likely configurations in m β which the + and − phases coexist along the boundary of ∂u, induces Mthis coexistence ∈ V(u, δ) imposed on deviations proportional to a surface order. The L1 constraint m β the magnetization is not strong enough to localize the interface close to ∂u: there might be mesoscopic fingers of one phase percolating into the other. Following [BBBP], we prove by the minimal section argument that one can chop off these fingers without changing too much the probability of the event. The renormalization is an essential feature in the previous procedure. Once the interface is localized on the mesoscopic level, the main problem is to identify surface tension. Notice that in the case of percolation [Ce], the minimal section argument enables to cut the microscopic fingers which connect the domains separated by ∂u. Therefore, one can immediately identify the surface tension factor, because for independent percolation it is defined as the probability that no cluster connects one domain to the other. In the case of spin systems, one would need to find a microscopic surface of + spins on one side of ∂u and another one of − spins on the other side in order to use directly the definition of surface tension. This would seem difficult to achieve because mesoscopic contours enable only to control the averaged magnetization and do not ensure the existence of such microscopic surfaces. We proceed differently
Wulff Construction in Three and More Dimensions
205
and use an alternative definition of surface tension in terms of the FK measure. This requires β to be large. Proposition 2.2 is proven in Sect. 6 under the condition that β is larger than β˜c . The coarse graining is only useful to get the lower bound up to β˜c and it could be avoided if one considers only β large, in which case, a direct proof without using the FK representation is possible. Section 7 is devoted to the proof of Proposition 2.3. Besides its probabilistic interpretation, i.e. the proof of an exponential tightness property, Proposition 2.3 deals with a physical phenomenon of a different nature than the surface tension: it states that the occurrence of many small contours is unlikely. The production of surface tension supposes a balance between energy and entropy. It is a general feature of DKS theory that energy is the dominant factor which rules out the occurrence of small contours. The techniques developed in [I2,SS] and [IS] to control the phase of small contours, for the two dimensional Ising model, are robust enough to be extended to higher dimensions provided Peierls estimate holds. This observation was used in [BBP]. For β large enough, one could have proceeded as in [BBP] and worked only with the Gibbs measure. We use an alternative approach borrowed to [I2] and deduce directly estimates on the phase of small contours by combining estimates of [I2] and Pisztora’s results [Pi1]. Thus Proposition 2.3 holds as soon as β is larger than β˜c . As noticed in the papers on Ising model with Kac potentials, the strategy described above can be applied in any dimension larger or equal to two. As a final remark, we stress the fact that the above results could be easily extended to prove a large deviation principle for the measures µ+ β,1N with action functional F. This setting was developed in [BBP] and also used for percolation [Ce]. This requires a modification of Proposition 2.2 which is described in Remark 8.3 at the end of Subsect. 8.3. 3. Coarse Graining and Mesoscopic Scale 3.1. The FK representation. We describe now the FK representation of Ising model. For a review of FK measures, we refer the reader to [Pi1,Gri] and [ACCN]. The set of edges is E = {x, y} | x ∼ y . For bond percolation, the configurations ω belong to = {0, 1}E . An edge b in E is open if ωb = 1 and closed otherwise. To any subset 3 of Z3 and π included in ∂3, we associate a set of edges [3]πe = {x, y} | x ∼ y, x ∈ 3, y ∈ 3 ∪ π , π
and the space of configurations in 3 is π3 = {0, 1}[3]e . Let ω be a configuration in , an open path (x1 , . . . , xn ) is a finite sequence of distinct nearest neighbors x1 , . . . , xn such that on each edge ω{xi ,xi+1 } = 1. We write {A ↔ B} for the event such that there exists an open path joining a site of A to one of B. A ∗-connected path (x1 ,√. . . , xn ) is a finite sequence of distinct vertices such that kxk − xk+1 k2 is smaller than 3 for all k. The connected components of the set of open edges of ω are called ω-clusters. The ω-cluster associated to the site i is denoted by Ci (ω). Let us now describe the FK representation of the Ising model (see Edwards and Sokal [ES]). Let 3 be a finite subset in Z3 and π a subset of ∂3. The first step is to introduce a measure on π3 . A vertex x of 3 is called π-wired if it is connected by an open path to π. We call π-clusters the clusters defined with respect to the boundary condition π : a π-cluster is a connected set of open edges in π3 and we identify to be the same cluster
206
T. Bodineau
all the clusters which are π-wired, i.e. connected to π. For a given p in [0, 1], we define the FK measure on π3 with boundary conditions π by Y 1 π π,p (1 − p)1−ωb pωb 2c (ω) , 83 (ω) = π,p Z3 b∈[3]π e
π,p
where Z3 is a normalization factor and cπ (ω) is the number of clusters which are not π-wired. If π = ∂3 then the boundary conditions are said to be wired and the corresponding w,p f,p f FK measure on w 3 is denoted by 83 . If π = ∅, we write 83 for the measure on 3 . 1 For any subset 1 of 3, we denote by F3 the σ -field generated by finite dimensional f cylinders associated with configurations in w 3 /1 , then strong FKG property (see [Pi1]) implies that for every increasing function g supported by f1 , 1 ) ≤ 81 (g). 83 − a.s, 81 (g) ≤ 83 (g | F3 w,p
f,p
w,p
w,p
(3.1)
81 (g) ≤ 83 (g) ≤ 83 (g) ≤ 81 (g).
(3.2)
In particular, one has f,p
f,p
w,p
w,p
In order to recover the Gibbs measure µ+ β,3 , we fix the percolation parameter pβ = 1 − exp(−2β) and generate the edges configuration ω in w 3 according to the measure w,pβ 83 . Given ω, we associate to the wired cluster the sign +1 and equip randomly each ωcluster with a color ±1 with probability 21 independently from the others. This amounts to introduce the measure P3ω on {−1, 1}3 such that the spin σi = 1 if Ci (ω) is π -wired and to be the chosen color of Ci (ω) otherwise. The Gibbs measure µ+ β,3 can be viewed w,p
as the first marginal of the coupled measure P3ω (σ )83 β (ω), Z w,p + ∀σ3 ∈ 63 , µβ,3 (σ3 ) = P3ω (σ )83 β (dω). w 3
By abuse of notation, the joint measure will be also denoted by µ+ β,3 . As a consequence of this representation one has mβ = lim µ+ β,1 (σ0 ) = lim 81
w,pβ
1→Zd
1→Zd
({0 ↔ ∂1}) = 2.
In the following, we use mβ or 2 depending on the context. In Theorems 2.1 and 2.2, we consider only the case β large. The first reason to do so is to satisfy the hypothesis of Theorem 5.3 of [Gri] which implies that for β large enough, f,p
w,pβ
lim 81 β ({0 ↔ ∂1}) = lim 81
1→Zd
1→Zd
({0 ↔ ∂1}) = 2.
(3.3)
Throughout the paper we suppose that (3.3) holds. The assumption β large will also be useful for technical reasons in the proof of Lemma 4.3.
Wulff Construction in Three and More Dimensions
207
3.2. Coarse graining. We recall the renormalization procedure introduced by Pisztora [Pi1,DP] for the FK measure. For our purposes, it is preferable to use an alternative construction of the coarse graining [Pi2]. The results of this section hold for β larger than βˆc , where βˆc was defined in [Pi1] in terms of slab percolation threshold. Let β˜c be the smallest value such that (3.3) is satisfied and β˜c ≥ βˆc . It is conjectured that β˜c coincides with the critical value βc . 1 Let γ be in Q ∩]0, 10 [ and α = 13 + γ9 . As in Subsect. 2.3, we partition the domain 1N = {1, N}3 into disjoint boxes B(x, N α ) of length N α centered in x in LN α (see (2.4) and (2.5)). For each x in LN α , we consider also the bigger box B(x, 45 N α ) containing B(x, N α ). Note that if x and y are ∗-neighbors in LN α the boxes B(x, 45 N α ) and B(y, 45 N α ) overlap. Following [Pi1], we introduce events which occur on the box B(x, 45 N α ) for each x in LN α , there is a unique crossing cluster C ∗ in B(x, 5 N α ) . Ux = ω ∈ w 1N 4 A crossing cluster is a cluster which intersects all the faces of the box, \n every open path in B(x, 5 N α ) with diameter ω ∈ w Rx = Ux 1N 4 o larger than N γ is contained in C ∗ , where the diameter of a subset A of Z3 is supx,y∈A kx − yk1 . Notice that the additional parameter γ is needed to control the fluctuations of the magnetization over small clusters (see (3.6)). We also define \n ∗ C crosses every sub-box of side length N γ ω ∈ w O x = Rx 1N o 5 contained in B(x, N α ) . 4 Finally, we consider an event which imposes that the density of the crossing cluster is close to 2 (see (3.3)) in B(x, N α ) with accuracy ζ > 0, \ ∗ α Vxζ = Ux ω ∈ w 1N |C | ∈ [2 − ζ, 2 + ζ ] |B(x, N )| . In the following, parameters α, γ will be fixed, therefore we omit the dependence on these parameters in notation. We will only consider different coarse graining for different values of ζ . ζ Each box B(x, N α ) is labeled by the random variable Yx (ω) depending only on the w configuration ω in 1N , Yxζ (ω) = 1 if ω ∈ Ox ∩ Vxζ ,
Yxζ (ω) = 0 otherwise.
Let {x1 , . . . , x` } be vertices in LN α not ∗-neighbors of x, then [Pi1] implies that there is an integer Nζ such that ∀N ≥ Nζ , 81N β (Yxζ = 0 | Yxζ1 , . . . , Yxζ` ) ≤ exp(−cN γ ) + exp(−cζ0 N α ), w,p
208
T. Bodineau
where cζ0 depends only on ζ and c is a constant. From [LSS] (Theorem 1.3), we deduce ζ
that for N large enough, the random variables {Yx } are dominated by a Bernoulli product measure πρN , πρN (X = 0) = ρN ≤ exp(−cζ N γ ),
(3.4)
where cζ is a positive constant. A similar result was already stated in [Pi1]. The random variables Y ζ are only related to ω, therefore the next step is to define a family of random variables which depend on (σ, ω). We denote by Mx the averaged magnetization in the box B(x, N α ), X 1 σi . (3.5) Mx = M Nx = 3α N α i∈B(x,N )
Pisztora’s results [Pi1] give a control of the deviation of the averaged magnetization from ζ its equilibrium values ±mβ in the boxes B(x, N α ). If Yx = 1, this deviation comes from the random coloring of the small clusters (those of diameter less than N γ ) included in B(x, N α ): this random coloring is independent of the boxes around B(x, N α ). Let ζ be ζ positive and define the new random variables {Zx } which depend on the joint law of (σ, ω), Zxζ (σ, ω) = sign(C ∗ ) if Yxζ (ω) = 1 and |Mx − sign(C ∗ ) mβ | < 2ζ,
Zxζ (σ, ω) = 0
otherwise.
Combining results of [Pi1] and [LSS], we check that there is Nζ such that for all N ζ larger than Nζ the random variables {|Zx |}, taking values in {0, 1}, are dominated by a Bernoulli product measure πρN0 , 0 ≤ exp(−cζ N γ ), πρN0 (X = 0) = ρN
(3.6)
where cζ is a positive constant depending only on ζ . Since the setting is different from [Pi1], we sketch the proof in Appendix 8.2. 3.3. Mesoscopic scale. In Subsect. 2.3, we already used a homogenization procedure on the mesoscopic scale N α . We introduce now a different mesoscopic representation which takes into account more details of the microscopic structure. For a given ζ positive, we associate to any configuration (σ, ω) in 61N × w 1N the ζ piecewise constant function T on T , Nα Nα < N ri − xi ≤ . (3.7) 2 2 If (σ, ω) is close to an equilibrium phase on a mesoscopic scale then T ζ has the sign of this phase. The 2 pure phases are represented by functions T ζ constantly equal to 1 or −1. From (3.6), one knows that for β larger than β˜c , ∀r ∈ T , Trζ (σ, ω) = Zxζ (σ, ω)
if ∀i, −
ζ lim µ+ β,1N ({Tr = 1, ∀r ∈ T }) = 1.
N→∞
The next lemma proves that a knowledge of the asymptotic of T ζ is sufficient to control the local magnetization M. Therefore to prove Propositions 2.1, 2.2 and 2.3, it will be enough to replace M by T ζ . The accuracy of the approximation depends on the parameter ζ which controls the coarse graining.
Wulff Construction in Three and More Dimensions
209
Lemma 3.1. For any δ positive, we set ζ = 41 δ, then Z 1 + mβ T ζ − Mr dr ≥ δ = −∞. log µ lim r β,1N N→∞ N 2 T Proof. One has Z T
mβ T ζ − Mr dr ≤ r
this implies Z T
mβ T ζ − Mr dr ≤ r
Nα N
3 X mβ Z ζ − Mx |, x B(x,N α )
Nα N
3 X B(x,N α )
1Z ζ =0 + 2ζ. x
Since ζ is small enough, Z δ 3(1−α) ζ mβ T ζ − Mr dr ≥ δ ≤ µ+ = 0} ≥ N #{Z , µ+ r x β,1N β,1N 2 T ζ
where #{Zx = 0} is the number of boxes with label 0. Therefore the lemma above will be a consequence of Lemma 3.2. For any δ and ζ positive, 1 log µ+ #{Zxζ = 0} ≥ δN 3(1−α) = −∞. β,1 2 N N→∞ N Proof. One has lim
µ+ β,1N
#{Zxζ
= 0} ≥ δN
3(1−α)
≤
3(1−α) NX
k=δN 3(1−α)
ζ µ+ β,1N #{Zx = 0} = k .
ζ
The random variables |Zx | are dominated by independent variables (3.6), thus for N large enough, 3(1−α) ζ 3(1−α) exp(−cζ δN 3(1−α)+γ ). ≤ 2N µ+ β,1N #{Zx = 0} ≥ δN This implies 2 ζ 3(1−α) ≤ exp ln 2 N 3(1−α) − cζ δN 2+ 3 γ . µ+ β,1N #{Zx = 0} ≥ δN As 3(1 − α) < 2, the entropic factor is negligible and the lemma follows. u t 4. Surface Tension As explained in the introduction, one of the main problems to derive Wulff construction is to recover surface tension from general boundary conditions. In this section we rewrite the surface tension in terms of the FK measure and prove that this new expression depends weakly on boundary conditions. This expression is reminiscent to the one introduced by Cerf [Ce] in the context of percolation. We keep notation of Subsect. 2.2. Throughout this section, we fix the direction n and f (N ) without loss of generality, we set h = 1. We also suppose that log(N ) diverges to infinity as N goes to infinity.
210
T. Bodineau
4.1. First step. The next lemma will be useful to prove Proposition 2.2 | ∂ − 3N } be the event such that there is no open path inside Lemma 4.1. Let {∂ + 3N ↔ + 3 to ∂ − 3 . Then joining ∂ [3N ]w N N e τ (n) = lim − N→∞
1 w,p log 83N β {∂ + 3N ↔ | ∂ − 3N } . 2 N
(4.1)
| ∂ − 3N } takes only into account the paths inside 3N Note that the event {∂ + 3N ↔ and not the identification produced by wired boundary conditions. Proof. We rewrite the quantities in terms of the FK representation. A well known argument implies that for pβ = 1 − exp(−2β), X Y + = exp β(δσx ,σy − 1) Z3 N σ ∈63N <x,y>∈[3N ]w e
Y
X
=
ω∈w 3
N
b∈[3N ]w e
(1 − pβ )1−ωb pβωb 2c
w (ω)
,
where cw (ω) is the number of clusters which are not wired. We prove now an equivalent formula for Y X +,− = exp β(δσx ,σy − 1) , Z3 N σ ∈63N <x,y>∈[3N ]w e
− where boundary conditions are equal to 1 on ∂3+ N and to −1 on ∂3N . We get X Y +,− = 1 − pβ + pβ δσx ,σy , Z3 N σ ∈63N <x,y>∈[3N ]w e
this gives +,− = Z3 N
Therefore +,− = Z3 N
X
X Y (1 − pβ )1−ωb pβωb
σ ∈63N ω∈w 3
N
b
X X Y (1 − pβ )1−ωb pβωb
ω∈w 3N
σ ∈63N
b
Y
δσx ,σy .
b=<x,y> ωb =1
Y
δσx ,σy .
b=<x,y> ωb =1
The boundary conditions imply that configurations ω containing a path joining ∂ + 3N to ∂ − 3N are not taken into account. We keep the definition of wired boundary conditions identifying all the clusters which touch the boundary ∂3N , X Y w +,− = 1{∂ + 3N ↔ (1 − pβ )1−ωb pβωb 2c (ω) . Z3 | ∂ − 3N } (ω) N ω∈w 3
Taking the ratio
+,− Z3 N
+ Z3
N
N
b
w,p , we recover 83N β {∂ + 3N ↔ | ∂ − 3N } . u t
Wulff Construction in Three and More Dimensions
211
¯ 4.2. Second step. In the following, we denote by 30N = 3(N, N, 21 f (N )) ∩ Z3 the parallelepiped included in 3N . Lemma 4.2. One has τ (n) = lim − N→∞
Proof. By definition Lemma 4.1 implies
| {∂ + 30N ↔
1 w,p log 83N β {∂ + 30N ↔ | ∂ − 30N } . 2 N
∂ − 30N } is included in {∂ + 3N ↔ | ∂ − 3N }. Therefore,
τ (n) ≤ lim inf − N→∞
1 w,p log 83N β {∂ + 30N ↔ | ∂ − 30N } . 2 N
| ∂ − 30N } is decreasing and Let us prove the reverse inequality. The event {∂ + 30N ↔ . Thus (3.2) gives supported by [30N ]w e w,pβ w,p | ∂ − 30N } ≥ 830 β {∂ + 30N ↔ | ∂ − 30N } . (4.2) 83N {∂ + 30N ↔ N
Since surface tension does not depend on the function f , Lemma 4.1 implies τ (n) = lim − N→∞
1 w,p log 830 β {∂ + 30N ↔ | ∂ − 30N } . 2 N N
Thus using (4.2), the lemma is proven. u t 4.3. Third step. Let 8f,w 3N be the FK measure with wired boundary conditions on the sides of 3N parallel to n and free on the sides orthogonal to n. Lemma 4.3. There is a constant β0 independent of n and f such that for any β larger than β0 , τ (n) = lim − N→∞
1 + 0 log 8f,w | ∂ − 30N } . 3N {∂ 3N ↔ N2
| ∂ − 30N } is denoted by SN . Applying (3.1), one observes Proof. The event {∂ + 30N ↔ w,pβ f,w that 83N (SN ) ≥ 83N (SN ) so that Lemma 4.2 implies τ (n) ≥ lim sup − N→∞
1 log 8f,w 3N (SN ). N2
(4.3)
To prove the reverse inequality, we introduce the slabs SlN + and SlN − in 3N , 1 3 ¯ N, N, f (N ) + f (N )n ∩ Z3 , SlN + = 3 10 8 3 1 ¯ N, N, f (N ) − f (N )n ∩ Z3 . SlN − = 3 10 8 For any ω in w 3N , we call a vertex x white if ωb = 1 for all edge b incident with − x and black otherwise. Let A+ N (resp AN ) be the event such that there is a surface of white vertices which crosses the slab SlN + (resp SlN − ) and separates the two sides of c the slab orthogonal to n. Equivalently, one can define A+ N as the set of configurations
212
T. Bodineau
ω3N which contain a ∗-connected path of black vertices intersecting the 2 sides of SlN + orthogonal to n. One has f,w f,w + − + − c 8f,w (4.4) 3N (SN ) = 83N SN ∩ AN ∩ AN + 83N SN ∩ (AN ∩ AN ) . First we estimate the last term in the RHS. It is enough to prove an upper bound for c + c ∩ A S . The events SN and A+ 8f,w 3N N N N have distinct supports, so that we can take the conditional expectation with respect to ∂ω, the configuration outside SlN + , + c +c ∂ω = 8f,w 8f,w 3N SN ∩ AN 3N SN 8Sl + (AN ) . N
c
Since A+ N is decreasing, (3.1) implies c
8∂ω (A+ N ) ≤ 8Sl Sl +
f,pβ
N
N
+
c
(A+ N ),
where the free boundary conditions are outside the domain [SlN + ]w e . In order to control this term, we use a Peierls argument (see [Gri], p. 1486). By the comparison result of Aizenman, Chayes, Chayes and Newman [ACCN], the above probability is bounded by p0 pβ ¯ β + with p0 = the percolation (product) measure 8 2, β
SlN
f,pβ
8Sl
N
+
pβ +(1−pβ )
p0
c +c ¯ β (A+ N ) ≤ 8Sl + (AN ). N
We choose β large enough so that pβ0 is close to 1. Then Peierls estimate holds and there is a constant c > 0 such that the probability that a ∗-connected black path joins 2 c f (N )). This comes from the vertices x and y on both sides of SlN + is less than exp(− 10 1 fact that the length of such a path is at least 10 f (N ), p0
¯ β + (A+ c ) ≤ N 2 exp − 8 N Sl N
c f (N ) . 10
One finally obtains c f (N ) 8f,w (4.5) 3N (SN ). 10 + − + We turn now to the estimate of 8f,w 3N SN ∩ AN ∩ AN . For a given ω in AN , we are + going to define the surface S (ω) of white vertices which is the closest to the “upper” side of SlN + . First we construct the black set B+ (ω) as follows: B+ (ω) contains the vertices in the “upper” side of ∂SlN + , i.e. the vertices at distance less than 2 of the hyperplan parallel 17 f (N )n. Furthermore B+ (ω) contains all the black vertices to (e1 , e2 ) and centered in 40 linked by a ∗-connected path of black vertices to the boundary of the “upper” side of SlN + . A vertex x is in S + (ω) if it belongs to the boundary of B+ (ω) and if there is a path of vertices joining x to 0 without crossing B + (ω). By construction the vertices in S + (ω) are white. In the same way we define S − (ω) as the surface of white vertices which is the closest to the “lower” side of SlN − , i.e. to the set of the vertices at distance 17 f (N )n. less than 2 of the hyperplan parallel to (e1 , e2 ) and centered in − 40 + 8f,w 3N SN ∩ AN
c
≤ N 2 exp −
Wulff Construction in Three and More Dimensions
213
The region between the surfaces S + and S − is denoted by SN and by construction is included in SN . Therefore we can consider the conditional expectation of SN with respect to the configurations outside SN (the measurability is discussed in [Gri, p. 1487]). Since S + and S − contain only white vertices, one gets
30N
w,pβ f,w + − 8f,w 3N SN ∩ AN ∩ AN = 83N 1A+ ∩A− 8SN (SN ) . N
N
The event SN is decreasing, thus strong FKG property (3.2) implies w,pβ f,w + − + − w,pβ 8f,w 3N SN ∩ AN ∩ AN ≤ 83N AN ∩ AN 83N (SN ) ≤ 83N (SN ).
(4.6)
Combining (4.4), (4.5) and (4.6), we obtain w,p
β 2 8f,w 3N (SN ) ≤ 83N (SN ) + N exp −
c f (N ) 8f,w 3N (SN ). 10
Applying Lemma 4.2, we get lim inf − N→∞
1 log 8f,w 3N (SN ) ≥ τ (n). N2
The lemma is completed. u t
4.4. Fourth step. Now, we will modify the boundary conditions and prove that the surface tension remains unchanged. The following lemma will be important in the proof of Proposition 2.1. It requires the assumption β large. Notice that the observation used in the proof of Lemma 4.4 that the bonds on the vertical sides can be closed up to a small error has been already made in [Ce]. We denote by ∂ top 30N (resp. ∂ bot 30N ) the face of ∂ + 30N (resp. ∂ − 30N ) orthogonal to | ∂ bot 30N } be the event such that there is no open path inside [30N ]w n. Let {∂ top 30N ↔ e top connecting ∂ 30N to ∂ bot 30N . Finally, we set δ = lim sup N→∞
f (N ) , N
and suppose that δ is finite. Lemma 4.4. There is a constant β0 independent of n and f such that for any β larger than β0 , 1 π,pβ top 0 bot 0 | ∂ 3N } ≤ −τ (n) + cβ δ, lim sup 2 log sup 83N {∂ 3N ↔ π N→∞ N where the constant cβ depends only on β. The above inequality holds uniformly over the boundary conditions π outside [3N ]w e. Notice that in the lemma, π is not a subset of ∂3N but a bond configuration in 3cN .
214
T. Bodineau
Proof. As {∂ top 30N ↔ | ∂ bot 30N } is decreasing, strong FKG property (3.1) implies π,p f,p | ∂ bot 30N } ≤ 83Nβ {∂ top 30N ↔ | ∂ bot 30N } , sup 83N β {∂ top 30N ↔ π
where the free boundary conditions are outside [3N ]w e . Note also that f,p top 0 | ∂ bot 30N } ≤ 24f (N)N 8f,w | ∂ bot 30N } . 83Nβ {∂ top 30N ↔ 3N {∂ 3N ↔ by
(4.7)
| ∂ bot 30N }. The inner boundary of 3N is defined We fix a configuration ω in {∂ top 30N ↔ ∂ ∗ 3N = {x ∈ 3N | ∃y 6∈ 3N , y ∼ x}.
For any vertex x on the sides of ∂ ∗ 3N parallel to n, we modify the edges of ω incident with x into closed edges and denote by ω¯ the new configuration. By construction ω¯ | ∂ − 30N }. Noticing that the number of edges which have been belongs to {∂ + 30N ↔ modified is smaller than 50f (N )N and using (4.7), one has f,p + 0 | ∂ bot 30N } ≤ exp(cβ f (N )N ) 8f,w | ∂ − 30N } , 83Nβ {∂ top 30N ↔ 3N {∂ 3N ↔ where cβ depends only on β. Using Lemma 4.3, we complete the proof. u t 5. Upper Bound: Proposition 2.1 Throughout this section, we fix u in BV(T , {+1, −1}). We recall that F(u) is finite. We split the proof into 3 steps. 5.1. Approximation. First we suppose that ∂ ∗ u is included in the interior of T . The general case will be treated in Subsect. 5.3. We approximate the reduced boundary of u with a finite number of parallelepipeds. Similar theorems were already stated in [ABCP] and [Ce]. The following result is proved in Appendix 8.3. Theorem 5.1. For any δ positive, there exists h positive such that there are ` disjoint parallelepipeds R 1 , . . . , R ` included in T with square basis B 1 , . . . , B ` of size h and height δh. The basis B i divides R i in 2 parallelepipeds R i,+ and R i,− and we denote by ni the normal to B i . Furthermore, the parallelepipeds satisfy the following properties: Z
` Z X |χR i (r) − u(r)| dr ≤ δ vol(R ) and i
Ri
i=1
Bi
τ (ni ) dHx − F(u) ≤ δ,
where χR i = 1R i,+ − 1R i,− and the volume of R i is vol(R i ) = δh3 . The area of B i is h2 .
R Bi
We fix δ positive. The approximation procedure implies lim lim sup
δ 0 →0 N→∞
1 M log µ+ ∈ V(u, δ 0 ) ≤ β,1N N2 mβ lim sup N→∞
` 1 M \ + log µ ∈ V(R i , 2δvol(R i )) , β,1N 2 N mβ i=1
dHx
Wulff Construction in Three and More Dimensions
215
where the ε-neighborhood of R i is Z 1 1 i 1 , ] |v(r) − χR i (r)| dr ≤ ε . V(R , ε) = v ∈ L T , [− mβ mβ Ri According to Lemma 3.1, there is ζ small enough, depending on δ and h, such that lim sup N→∞
` 1 M \ + log µ ∈ V(R i , 2δvol(R i )) ≤ β,1N 2 N mβ i=1
lim sup N→∞
` \ 1 + ζ log µ ∈ V(R i , 3δvol(R i )) . T β,1N 2 N i=1
Therefore to prove Proposition 2.1, it is enough to show that lim sup N→∞
` \ 1 + ζ log µ ∈ V(R i , 3δvol(R i )) ≤ −F(u) + Cβ,u δ, T β,1N 2 N i=1
where the constant Cβ,u depends only on β and u. Each box can be labeled by 3 values 3(1−α) . As 3(1 − α) < 2, this 0, ±1, thus the number of configurations T ζ is less than 3N term has no entropic effect. Thus it remains to check ! 1 + sup µβ,1N {T ζ } ≤ −F(u) + Cβ,u δ, (5.1) lim sup 2 log N →∞ N T ζ ∈ Uδ where U δ denotes which realize T ζ .
T`
i=1 V(R
i , 3δvol(R i ))
and {T ζ } is the set of configurations (σ, ω)
i = 5.2. Minimal section argument. The microscopic domain associated to R i is RN i,+ i,− i /R i,+ . Let Li be the = NR i,+ ∩ 1N and RN = RN N R i ∩ 1N . We also set RN N N i,+ i α is subset of boxes B(x, N ) intersecting RN . The number of boxes intersecting RN
NNi,+ = N 3(1−α)
vol(R i ) + O(N 2(1−α) ) . 2
(5.2)
i,+ .A This error is due to the fact that the partition may not be exact on the sides of RN i,− i,− similar estimate holds for NN , the number of boxes intersecting RN . T To any T ζ in `i=1 V(R i , 3δvol(R i )), we associate the set of bad boxes which are ζ i,+ i,− the boxes in LiN labeled by Zx = 0 and the ones intersecting RN (resp. RN ) labeled ζ ζ by Zx = −1 (resp. Zx = 1). We will now use the L1 -constraint to derive bounds on the number of bad boxes. The i is smaller than 50h2 N 2(1−α) , therefore number of boxes in LiN not included in RN Z X 3(1−α) N |Trζ − 1| dr − |Zxζ − 1| ≤ 100h2 N 2(1−α) . R i,+
i,+ B(x,N α )∩RN 6 =∅
216
T. Bodineau
Since T ζ belongs to V(R i , 3δvol(R i )), one gets from (5.2) that for N large enough the i,+ number of bad boxes in RN is smaller than 10δNNi,+ . In the same way, we check that i,− is smaller than 10δNNi,− . the number of bad boxes in RN i 0 Let R be the parallelepiped included in R i with basis B i and height 2δ h. Its microi 0 . We will apply the minimal section argument introduced in scopic counterpart is RN T i 0↔ i 0 }. | ∂ bot RN [BBBP] and relate the expectation of {T ζ } to the one of `i=1 {∂ top RN i,k k ni . Let BN be the microscopic subset For any integer k, we set B i,k = B i + 10 N 1−α 0
i associated to B i,k , of RN i,k i 0 = j ∈ RN | ∃r ∈ B i,k , kj − N rk1 ≤ 10 . BN
We define Bik as the smallest connected set of mesoscopic boxes containing i,k B(y, N α ) ∈ LiN | B(y, N α ) ∩ BN 6= ∅ . By construction the Bik are disjoint surfaces of boxes. For k positive, let n+ i (k) be the number of bad boxes in Bik and define + δh 1−α N . n+ i = min ni (k); 0 < k < 30 location where the minimum is− achieved and define the minimal Call k + the smallest + i,− and section as Bik . For k non-positive, we denote by Bik the minimal section in RN − k . the number of bad boxes in B n− i i For any {T ζ } in U δ , we will check that the total number of bad boxes is bounded by ` X i=1
− 2(1−α) n+ , i + ni ≤ Cu δN
(5.3)
where Cu is a constant depending only on u. By definition, one has δh 1−α + N ni ≤ 30
X i,+ B(x,N α )∩RN 6=∅
1Z ζ 6=1 ≤ 10δNNi,+ . x
3 2 2(1−α) . Note that h2 is in fact For N large enough, (5.2) implies that n+ i ≤ 10 δh N i the area of B , therefore the approximation procedure implies that `h2 is bounded by a constant depending on the perimeter of {u = −1}. Thus (5.3) holds. We are now going to use all the previous estimates. We define ∃σ such that (σ, ω) ∈ {T ζ } . A = ω ∈ w 1N
Any configuration ω in A will be mapped into ω¯ by the following procedure. For any bad box B(x, N α ) in the minimal sections, we change the open edges of ω located on the sides of the box B(x, 45 N α ) into closed edges. The new configuration ω¯ belongs i 0↔ i 0 }, because any open path of ω which joins ∂ bot R i 0 to ∂ top R i 0 | ∂ bot RN to {∂ top RN N N intersects at least one of the minimal sections on a bad block and therefore is cut by the i 0 to ∂ bot R i 0 and suppose above procedure. Let C be an open path of ω joining ∂ top RN N
Wulff Construction in Three and More Dimensions
217
that C crosses the minimal sections without intersecting a bad box. Then C intersects + − ζ ζ the boxes B(x + , N α ) and B(x − , N α ) in Bik and Bik with labels Yx + = Yx − = 1. This would imply that the crossing clusters of B(x + , N α ) and B(x − , N α ) are connected to ζ ζ C, so that Zx + = Zx − . Therefore one of these boxes has to be a bad box. − 2α edges. From (5.3) the Around the bad boxes, we change at most 20(n+ i + ni )N total number of edges involved in the previous procedure is bounded by 100Cu δN 2 . Therefore we get ` w,pβ \ w,pβ ζ 2 i 0 i 0 ({T }) ≤ 8 δN {∂ top RN ↔ | ∂ bot RN } , A ≤ exp C 8 µ+ β,u 1N 1N β,1N i=1
where the constant Cβ,u depends only on β and u. i and using Lemma 4.4, we derive Conditioning outside each domain RN ! ` Z X 1 w,pβ ζ τ (ni ) dHx + Cβ,u δ + cβ `h2 δ, lim sup 2 log sup 81N {T } ≤ − N ζ δ Bi N→∞ T ∈U i=1
where cβ was defined in Lemma 4.4. Noticing that `h2 is bounded in terms of the perimeter of u and using Theorem 5.1, we derive (5.1). 5.3. Boundary conditions. Let U be the intersection of the reduced boundary ∂ ∗ u and of ∂T . Suppose that U has a positive 2 dimensional Hausdorff measure. In this case we cannot approximate the surface U as in Theorem 5.1 with parallelepipeds included in T . We state a variant of Theorem 5.1 proven in Appendix 8.3. Theorem 5.2. For any δ positive, there exist h positive and ` disjoint squares B 1 , . . . , B ` in ∂T of size h and normal ni such that ` Z X i=1
Bi
Z τ (ni ) dHx −
U
τ (nx ) dHx ≤ δ.
Furthermore, there are ` disjoint parallelepipeds R 1 , . . . , R ` included in T such that one of the faces of R i is B i and the height of R i is δh. The parallelepipeds also satisfy Z |1 + ur | dr ≤ δ vol(R i ). ∀i ≤ `, Ri
The proof of the upper bound is based on local estimates in each parallelepiped, thus we will simply explain how to adapt the previous proof to obtain lim sup N→∞
` \ 1 + ζ log µ ∈ V(R i , δvol(R i )) ≤ − T β,1N 2 N i=1
Z U
τ (nx ) dHx + Cβ,u δ, (5.4)
where Cβ,u is a constant and V(R i , δvol(R i )) is Z 1 1 i i 1 i , ] |vr + 1| dr ≤ δvol(R ) . V(R , δvol(R )) = v ∈ L T , [− mβ mβ Ri
218
T. Bodineau
Combining estimates (5.1) and (5.4), one derives Proposition 2.1 for any function of bounded variation u in T . i be the microscopic set associated to R i . The set of bad boxes is the set of boxes Let RN i and labeled by 0 or 1. Using the L1 constraint, we see that α B(x, N ) intersecting RN the number of bad boxes is smaller than 10δ 2 h3 N 3(1−α) . Let R i 0 be the parallelepiped i included in R i with height δh 2 and such that one of its faces is B . The previous argument implies that the minimal section contains less than 100δh2 N 2(1−α) bad boxes. Therefore we can cut the wired open paths which cross the minimal section and obtain for N large enough, ζ µ+ β,1N T ∈
` \ i=1
` \ i 0 i 0 V(R i , δvol(R i )) ≤ exp Cβ,u δN 2 8w {∂ top RN ↔ | ∂ bot RN } . 1N i=1
0
i is nothing but a subset of ∂1 . Notice that ∂ top RN N i Let R¯ be the union of R i and of the parallelepiped in T c with height δh 2 and one of i i ¯ its faces is equal to B . Denote by RN the corresponding microscopic domain, then by inequality (3.1) one has f,p i 0 i 0 i 0 i 0 ˜ π,w ↔ | ∂ bot RN } ≤ 8 ¯ i β {∂ top RN ↔ | ∂ bot RN } , {∂ top RN 8 i RN
RN
i ˜ π,w where 8 i is the FK measure with wired boundary conditions on the face of RN which RN
coincides with ∂1N and π otherwise. This enables us to apply Lemma 4.4 and to recover (5.4). 6. Lower Bound: Proposition 2.2 The proof rests only on Lemma 4.1, therefore Proposition 2.2 holds as soon as β is larger than β˜c (see Subsect. 3.2). The proof is divided into 2 steps. 6.1. Approximation procedure. We first state an approximation theorem which will be proven in Appendix 8.3. We call a polyhedral set, a set which has a boundary included in the union of a finite number of hyperplans. Theorem 6.1. For any δ positive, there exists a polyhedral set W such that δ δ 1W ∈ V(1Wm , ) and F(W ) − F(Wm ) ≤ . 3 2 For any h small enough there are ` disjoint cubes R 1 , . . . , R ` of size h with basis B 1 , . . . , B ` included in ∂W . Furthermore, the squares B 1 , . . . , B ` cover ∂W up to a S set of measure less than δ denoted by U δ = ∂W/ `i=1 B i and they satisfy ` Z X i=1
Bi
τ (ni ) dHx − F(Wm ) ≤ δ,
where the normal to B i is denoted by ni .
Wulff Construction in Three and More Dimensions
219
We fix δ positive and choose a set W approximating Wm , then δ \ ζ δ M T ∈ V(1W , ) ⊂ ∈ V(T ζ , ) ∈ V(1Wm , δ) . mβ 3 3 mβ M 6∈ V(T ζ , 3δ ) has a Lemma 3.1 implies that there exists ζ such that the event m β probability which vanishes exponentially fast, therefore M
lim inf N→∞
δ M 1 1 ζ log µ+ ∈ V(1Wm , δ) ≥ lim inf 2 log µ+ β,1N β,1N T ∈ V(1W , ) . N →∞ N N2 mβ 3
It remains to find a lower bound for the term in the RHS. For any ε positive, we construct a shell around ∂W which splits T into 2 domains Sε = {r ∈ T | dist(r, ∂W ) ≤ ε}. We set Wε+ = {r ∈ W c | dist(r, ∂W ) ≥ ε} and Wε− = {r ∈ W | dist(r, ∂W ) ≥ ε}. ± be the set of mesoscopic boxes included in N Wε± ∩ 1N . We fix ε such that Let Wε,N δ the volume of Sε is smaller than 10 and choose h smaller than 2ε .
6.2. Exponential bound. The microscopic domain associated to the cube R i is denoted i = N R i ∩ 1 . We set A = T` {∂ + R i ↔ − i by RN N N i=1 N | ∂ RN }. The microscopic domain UNδ = x ∈ 1N ∃r ∈ U δ , kx − N rk1 ≤ 10 is an enlargement of the surface U δ defined in the approximation procedure. We introduce BN , the set of configurations with closed edges in UNδ . Hypotheses on U δ imply that BN is supported by at most 10δN 2 edges. We decompose AN ∩ BN into 2 disjoint sets + + − α ζ µ+ β,1N AN ∩ BN = µβ,1N AN ∩ BN ∩ {∀B(x, N ) ⊂ Wε,N ∪ Wε,N ; |Zx | = 1} + − α ζ + µ+ β,1N AN ∩ BN ∩ {∃B(x, N ) ⊂ Wε,N ∪ Wε,N ; Zx = 0} . We first estimate the last term in the RHS. By definition the events AN ∩ BN and ζ + − ∪ Wε,N ; Zx = 0} have disjoint supports. Taking the {(σ, ω) | ∃B(x, N α ) ⊂ Wε,N conditional expectation with respect to AN ∩ BN and using the stochastic domination (3.6) (see also Remark 8.1), we get + − α ζ µ+ β,1N {∃B(x, N ) ⊂ Wε,N ∪ Wε,N ; Zx = 0} AN ∩ BN ≤ N 3(1−α) exp(−cζ N γ ). Therefore, for N large enough µ+ β,1N AN ∩ BN
+ − α ζ ≤ 2µ+ β,1N AN ∩ BN ∩ {∀B(x, N ) ⊂ Wε,N ∪ Wε,N ; |Zx | = 1} .
220
T. Bodineau
By construction, no configuration ω of AN ∩ BN contains an open path joining S` i S U δ . Therefore any configuration the 2 connected components of 1N / i=1 RN N ζ + − ∪ Wε,N ; |Zx | = 1} contains 2 disconnected in AN ∩ BN ∩ {∀B(x, N α ) ⊂ Wε,N microscopic crossing clusters. The cluster connected to ∂1N is denoted by C + and the other one by C − . The wired constraint imposes the sign 1 to C + . With probability 21 we choose the sign of C − to be −1. We define the event \ \ + ; Zxζ = 1} BN {∀B(x, N α ) ⊂ Wε,N CN (σ, ω) = AN \ − {∀B(x, N α ) ⊂ Wε,N ; Zxζ = −1}, + then µ+ β,1N (AN ∩ BN ) ≤ 4µβ,1N (CN ). Thus for any configuration (σ, ω) in CN , the function T ζ (σ, ω) is equal to 1 on Wε+ and to −1 on Wε− (see (3.7)). Since the volume δ , we have of Sε is less than 10
δ CN ⊂ (σ, ω) T ζ (σ, ω) ∈ V(1W , ) , 3 this leads to δ 1 + ζ µ+ β,1N T (σ, ω) ∈ V(1W , ) ≥ µβ,1N AN ∩ BN . 3 4 As AN ∩ BN depends only on the variable ω, we replace the coupled measure by the w,p FK measure 81N β . The support of BN contains less than 10δN 2 edges, so that ` δ 1 w,pβ \ + i ζ 2 i ) ≥ exp(−c ∈ V(1 , δN )8 {∂ RN ↔ | ∂ − RN } , T µ+ W β 1N β,1N 3 4 i=1
i ↔ i } occur on disjoint | ∂ − RN where cβ is a constant depending on β. The events {∂ + RN supports, taking the conditional expectation with respect to the configuration ∂ωi outside i , we have RN ` δ exp(−cβ δN 2 ) w,pβ Y ∂ωi + i ζ i ) ≥ 8 ∈ V(1 , 8 i ({∂ RN ↔ | ∂ − RN }) . T µ+ W 1N β,1N R 3 4 N i=1
i ↔ i } are non increasing, (3.1) implies | ∂ − RN Since the events {∂ + RN `
δ exp(−cβ δN 2 ) Y w,pβ + i ζ i ) ≥ ∈ V(1 , 8 i {∂ RN ↔ | ∂ − RN } . T µ+ W β,1N R 3 4 N i=1
Taking the limit as N goes to infinity and using Lemma 4.1, we obtain `
lim inf N→∞
X δ 1 + ζ log µ ∈ V(1 , T ) ≥ − W β,1 N N2 3
The proof is completed by letting δ go to 0. u t
i=1
Z Bi
τ (ni )dHx − cβ δ.
Wulff Construction in Three and More Dimensions
221
7. Exponential Tightness: Proposition 2.3 In view of Lemma 3.1, Proposition 2.3 will be a consequence of Lemma 7.1. Let β be larger than β˜c (see Subsect. 3.2), then there is Cβ such that for any a positive, ∀ζ, δ > 0, lim sup N→∞
1 ζ c log µ+ β,1N T ∈ V(Ka , δ) ≤ −Cβ a. 2 N
A high density of small contours can be interpreted on the mesoscopic scale as a high density of random variables Z ζ = 0. Such events are ruled out by Lemma 3.2. If T ζ belongs to V(Ka , δ)c and does not contain many boxes labeled by Z ζ = 0, then there are mainly mesoscopic sets of constant sign. By definition of the variables Z ζ , two sets of different signs are disconnected on the microscopic level and therefore are separated by a layer of boxes with label Z ζ = 0. The problem is that the probability of the event ζ {Zx = 0} may only be of the order of exp(−N γ ) which is much higher than the expected surface order exp(−N 2α ). Thus we introduce another coarse graining in order to recover the surface order. The scheme of the following proof was suggested by D. Ioffe. . Proof. As noticed in Subsect. 5.2, the number of configurations T ζ is less than 3N ζ }). Lemma 3.2 ({T Thus it is enough to fix T ζ in V(Ka , δ)c and to estimate µ+ β,1N enables us to consider only configurations T ζ with a number of boxes labelled by 0 smaller than δN 3(1−α) . This amounts to say that Z 1T ζ =0 dr ≤ δ. (7.1) 3(1−α)
T
r
Each realization of T ζ splits T into T = T+ ∪ T− ∪ T0 , where T ζ is constantly equal to ±1 on T± and to 0 on T0 . From (7.1), the measure of T0 is smaller than δ. The microscopic counterparts of T+ and T− will be denoted by 1N,+ and 1N,− . Moreover as T ζ belongs to V(Ka , δ)c , for any regular set A of T such that T− ⊂ A ⊂ T \ T+ , Z Z dHx ≥ τ (nx ) dHx ≥ a, (7.2) kτ k∞ R
∂A
∂A
where ∂A dHx is the perimeter of ∂A. Note that for each configuration in V(Ka , δ)c the set T− is not empty. Let L be an integer large enough which divides N α and is independent of N . We partition 1N into boxes B(i, L), where i is in LL . We also introduce the boxes B(i, 45 L), and following Subsect. 3.2, define a coarse graining on the scale L. Let {yi } be the family of random variables equal to 1 if the event Oi (for the box B(i, 45 L)) is satisfied and 0 otherwise. We define the microscopic set A˜ N as the union of 1N,− and of the boxes B(i, L) labeled by 1 such that there is a ∗-connected path of boxes B(j, L) labeled by 1 joining B(i, L) to 1N,− . For any configuration (σ, ω) in {T ζ } there is no microscopic path connecting 1N,+ ∪ ∂1N to 1N,− . Therefore there is no ∗-connected path of boxes B(j, L) labeled by 1 connecting 1N,+ ∪ ∂1N to 1N,− . This implies that 1N,− ⊂ A˜ N ⊂ 1N \ 1N,+ .
222
T. Bodineau
S` We define also GN = 1N \ A˜ N . Let GN = i=1 GN,i be the decomposition of GN into maximal connected components composed of boxes B(j, L). The components GN,i such that |GN,i | < N 3α cannot intersect 1N,+ . We set [ GN,i , AN = A˜ N |GN,i |
which satisfies 1N,− ⊂ AN ⊂ 1N \ 1N,+ . To any GN,i with cardinality larger than N 3α , we associate the contour n o ∂ L GN,i = B(j, L) B(j, L) ⊂ GN,i and B(j, 2L) ∩ AN 6= ∅ . By construction ∂ L GN,i is a connected set of boxes B(j, L) such that yj = 0. Furthermore, whenever |GN,i | ≥ N 3α , the number of boxes in the contour ∂ L GN,i (denoted by #∂ L GN,i ) satisfies #∂ L GN,i ≥ c
Nα L
2 ,
(7.3)
where c is a universal constant. Finally, from inequality (7.2), we see that the cardinality of the boundary ∂AN is larger than caN 2 /kτ k∞ , thus 2 X ca N L #∂ GN,i ≥ . (7.4) kτ k L ∞ 3α |GN,i |≥N
Stochastic domination (3.4) implies that for L large enough, a Peierls estimate holds: there is a positive constant CL,β such that the probability that a connected surface S of boxes B(i, L) labeled by 0 contains more than n boxes is bounded by w,p
81N β (#S ≥ n) ≤ N 3 exp(−CL,β n), where #S is the number of boxes in S. By definition the contours ∂ L GN,i are connected surfaces of boxes B(j, L) labeled by 0, so that using (7.3) and (7.4), the proof is concluded by a standard Peierls argument. u t 8. Appendix 8.1. Surface tension. In [MMR], the surface tension τ 0 was defined by tilting the boundary conditions outside the domain 1M = {−M, M}3 , where M is an integer. For a given n in S2 , we denote by DM the intersection of [−M, M]3 with the hyperplan centered in 0 and orthogonal to n. The surface tension τ 0 is τ 0 (n) = lim − M→∞
+,− Z1 1 log +M , SM Z1M
where SM is the area of DM and the mixed boundary conditions in ∂1M are σi = 1 if i.n > 0 and σi = −1 otherwise. The boundary conditions above DM are equal to 1 and to −1 below.
Wulff Construction in Three and More Dimensions
223
This definition coincides with Definition 2.1. This is straightforward from the argument developed by [MMR] (Theorem 2). We recall this argument for completeness. First, we show that τ 0 (n) ≤ lim inf − N→∞
+,− Z3 1 N log + . h2 N 2 Z3 N
(8.1)
Let M, N be 2 integers such that N M. We shall write 0 = − log FM
+,− Z1 M + Z1 M
and F (3N ) = − log
+,− Z3 N + Z3 N
,
where the parallelepiped 3N was introduced in Definition 2.1. We tile DM with kM squares B i of side length hN and at distance 10 from each other. SM i B is smaller These squares are chosen such that the area of DM not covered by ki=1 than C(Mf (N) + kM hN), where C is some positive constant independent of M and N . If the center of B i does not coincide with a site on the lattice, we translate B i at a 0 distance smaller than 1 such that the center of its translate B i belongs to Z3 . Let 3iN 0 be the parallelepiped of basis B i deduced from 3N by translation. There is a choice i of the squares B such that all the parallelepipeds 3iN are included in 1M . Note that F (3iN ) = F (3N ). Inequality (C.2) of [MMR] implies 0 ≤ FM
kM X i=1
F (3iN ) + KC Mf (N ) + kM hN ,
where K is some positive constant. Since |SM − kM h2 N 2 | ≤ C(Mf (N ) + kM hN ), one gets as M goes to infinity, τ 0 (n) ≤
1 h2 N 2
F (3N ) +
KC . hN
Letting N go to infinity, we obtain (8.1). The reverse inequality τ 0 (n) ≥ lim sup − N→∞
+,− Z3 1 N log + h2 N 2 Z3 N
(8.2)
is derived in the same way by choosing M N and by partitioning the basis of 3N with translates of DM . Combining (8.1) and (8.2), we see that τ (n) = τ 0 (n). 8.2. Stochastic domination. The following result is a consequence of [I2] and of the proof of Theorem 1.1 of [Pi1], we sketch the proof for completeness Theorem 8.1. For any ζ positive, there is Nζ such that for all N larger than Nζ , the ζ family of random variables |Zx | is dominated by a product Bernoulli measure πρN0 (see (3.6)).
224
T. Bodineau
Proof. According to [LSS], it is enough to check that ζ ζ ζ γ µ+ β,1N |Zx | = 0 |Zx1 | = ε1 , . . . , |Zx` | = ε` ≤ exp(−cζ N ), where the vertices {x1 , . . . , x` } are not ∗-neighbors of x in LN α and each εi belongs to {0, 1}. Using notation from Subsect. 3.1, one has ζ ζ ζ µ+ β,1N |Zx | = 0, |Zx1 | = ε1 , . . . , |Zx` | = ε` ≤
w,p 81N β Yxζ = 1; P1ωN |Mx − sign(C ∗ )mβ | > 2ζ, |Zxζ1 | = ε1 , . . . , |Zxζ` | = ε` w,p +81N β Yxζ = 0, Yxζ1 = ε1 , . . . , Yxζ` = ε` ; P1ωN |Zxζ1 | = ε1 , . . . , |Zxζ` | = ε` .
Because of stochastic domination (3.4), the last term in the RHS is bounded by ζ ζ exp(−cζ N γ ) µ+ β,1N |Zx1 | = ε1 , . . . , |Zx` | = ε` . In order to control the other term, we have to take into account the deviations occurring from the random coloring of the small clusters, i.e. those of diameter less than N γ . We enumerate the small clusters C1 , . . . , Ck0 included in B(x, N α ). Their cardinals are denoted c1 , . . . , ck0 and their signs s1 , . . . , sk0 . The random variables s1 , . . . , sk0 are iid Bernoulli. For N large enough, one has k0 X ζ 3α N Mx − sign(C ∗ )|C ∗ | − si ci < N 3α . 2 i=1
This comes from the fact that all the clusters intersecting the boundary of B(x, N α ) and ζ distinct from C ∗ have length smaller than N γ when Yx = 1. Thus the total magnetization 2α+γ and does not contribute. By definition of produced by these clusters is less than 6N ζ the event Vx , the unique crossing cluster C ∗ in B(x, N α ) satisfies |mβ N 3α − |C ∗ | | ≤ P0 si ci | > ζ2 N 3α , ζ N 3α . Therefore, we just need to prove large deviations for P1ωN | ki=1 ζ
for configurations ω which satisfy Yx (ω) = 1. By symmetry, it is enough to bound Pk0 ζ ζ 3(α−γ ) (note that k0 is always smaller than P1ωN i=1 si ci > 2 k0 , with k0 ≥ 2 N N 3α ). We follow the argument of [I2] (p. 325). For all t positive, Chebyshev’s inequality implies P1ωN
k0 X i=1
k
0 ζ ζ k0 Y si ci > k0 ≤ exp − t P1ωN exp(tci si ) . 2 2
i=1
As each ci is smaller than N 3γ , one has k0 k0 X ζ ζ 1 X 1 log P1ωN si ci > k0 ≤ −t + log cosh(tci ) k0 2 2 k0 i=1
i=1
ζ ≤ −t + log cosh(tN 3γ ), 2
Wulff Construction in Three and More Dimensions
225
Let 3∗ be the Legendre-Laplace transform of the Bernoulli measure 21 (δ1 + δ−1 ). There exists cζ positive such that for N large enough, k0 X cζ ζ ζ 1 log P1ωN si ci > k0 ≤ −3∗ ≤ − 6γ . k0 2 2N 3γ N i=1
Since k0 ≥
ζ 3(α−γ ) , 2N
P1ωN
this leads to k0 X i=1
si ci >
ζ ζ k0 ≤ exp − cζ N 3α−9γ . 2 2
1 ], we obtain the expected upper bound. u t As γ belongs to ]0, 10 5 α Remark 8.1. Let A be a subset of w 1N with support disjoint from the box B(x, 4 N ). Then the following holds for N large enough: ζ γ µ+ β,1N |Zx | = 0 A ≤ exp(−cζ N ).
This is straightforward from the previous arguments. From [Pi1], we know that ω,p ζ 81N β (Yx = 0) vanishes exponentially fast for arbitrary boundary conditions ω outside the box B(x, 45 N α ). Furthermore, if the magnetization differs from its equilibrium values ±mβ , the deviation occurs from the random coloring of small clusters independent of A. 8.3. Approximation. Before proving Theorems 5.1 and 5.2, let us recall some basic notions of geometric measure theory. Throughout this section, we fix u in BV(T , {+1, −1}) such that F(u) < ∞ and δ in ]0, 1]. We recall that the perimeter of {u = −1}, which is R dH ∗ x , is also finite. The ball of radius r centered in y will be denoted by B(y, r). ∂ u For each y in ∂ ∗ u, we introduce the half-spaces H + (y) = z ∈ R3 | ny .(z − y) ≥ 0 , H − (y) = z ∈ R3 | ny .(z − y) ≤ 0 , where ny is the normal to ∂ ∗ u in y. Let H (y) be the hyperplan H + (y) ∩ H − (y). We fix ζSpositive. According to Theorem 2 (p. 205) of [EG], the reduced boundary ∂ ∗ u equals ni=1 Ki ∪ N, where the 2 dimensional Hausdorff measure of N is less than ζ and each Ki is a compact subset of a C 1 -hypersurface Si . For all x in Ki , the normal nx is also normal to Si and there is r0 > 0 such that uniformly on Ki , (8.3) ∀i, ∀r ≤ r0 , ∀y ∈ Ki , vol B(y, r) ∩ {u = −1} ∩ H + (y) < ζ r 3 , vol B(y, r) ∩ {u = +1} ∩ H − (y) < ζ r 3 . In the decomposition of ∂ ∗ u, one can choose each set Ki such that it is either included in ∂T or at a positive distance from ∂T . Proof of Theorem 5.1. We first approximate the compact sets Ki which do not touch the boundary. The following construction is the same for each Si so it is enough to present it for one Si , that we shall denote by S (with corresponding K ⊂ S).
226
T. Bodineau
As S is C 1 , we can find M pairwise disjoint closed subsets 61 , . . . , 6M of S which cover S up to a set of measure less than ζ and such that each 6i is congruent to the graph of a real function fi : Ui → R of class C 1 , where Ui is a bounded closed set of R2 and fi satisfies the bound |∇fi | ≤ ζ . To any x in Ui , we associate the point gi (x) = (x, fi (x)) of S. Let Ki be the compact subset of Ui such that gi (Ki ) = K ∩ 6i . We choose h in r0 [ (see (8.3)) arbitrarily small and cover Ui with pairwise disjoint cubes C j ⊂ Ui ]0, 10 ζ of side h up to a set of measure less than M . For each cube C j centered in xj and intersecting Ki , we denote by B j the translate of C j centered in gi (xj ). The parallelepiped R j is defined as R j,+ ∪ R j,− , where both parallelepipeds R j,+ and R j,− have a common face B j and height 2δ h (one above and the other below B j ). Let y be in C j ∩ Ki . The parallelepiped R j is included in the ball B(gi (y), 10h). As |∇fi | ≤ ζ , the intersection of the hyperplan H (gi (y)) and R j is contained in {z ∈ R3 | dist(z, B j ) ≤ 2ζ h}. Therefore (8.3) implies that Z ζ |χR j (r) − u(r)| dr ≤ 2ζ h3 + 103 ζ h3 ≤ 104 vol(R i ). (8.4) j δ R The upper bound of τ is denoted by kτ k∞ . It remains to check that Z
k Z X i=1
Bi
τ (ni ) dHx −
τ (nx ) dHx ≤ kτ k∞ CK ζ,
K
(8.5)
where ni is the normal to B i and CK depends on the Hausdorff measure of K. Let C i be the union of cubes C j which intersect Ki , then for h small enough the ζ measure of C i 1Ki (the symmetric difference) is smaller than M and one has Z Z ζ kτ k ∞ , τ (n0 x ) dHx − τ (nx ) dHx ≤ i M gi (C ) gi (Ki ) where n0 is the normal vector to the surface S which coincides with n on K. The normal n0 is uniformly continuous on any compact. Therefore for h small enough, the following holds for any cube C j in Ui : ∀x, y ∈ C j , τ (n0 gi (x) ) − τ (n0 gi (y) ) ≤ ζ. Using the fact that |∇fi | ≤ ζ on each Ui , we derive (8.5). of sets Let us go back to the previous notation and denote by B 1 , . . . B ` the collectionT which approximate the union of sets Ki which are not in ∂ ∗ u. We set also U = ∂ ∗ u ∂T . As τ is bounded, Z τ (nx ) dHx ≤ kτ k∞ ζ. N
We deduce from (8.5) that ` Z X i=1
Bi
Z τ (ni ) dHx −
∂ ∗ u/U
τ (nx ) dHx ≤ Cu ζ,
(8.6)
where Cu depends only on the perimeter of u. From (8.4) and (8.6), we derive Theorem 5.1 for ζ small enough. u t
Wulff Construction in Three and More Dimensions
227
Proof of Theorem 5.2. We are going now to approximate the compact sets Ki included in U = ∂T ∩ ∂ ∗ u. One can also suppose that each Ki is included in one face of ∂T . Notice that H-almost surely for y in Ki , the normal ny is orthogonal to ∂T . r0 [ small enough, there is a covering of Ki with For a given ζ positive and h in ]0, 10 j pairwise disjoint squares B ⊂ ∂T of size h up to a set of measure less than ζ . We denote by R j the parallelepiped in T with one face equal to B j and height δh. Let y be in B j ∩ Ki . The parallelepiped R j is included in the ball B(gi (y), 10h) and (8.3) implies Z ζ |1 + u(r)| dr ≤ 103 ζ h3 ≤ 103 vol(R j ). (8.7) δ Rj Furthermore, for h small enough k Z X
j j =1 B
Z τ (nj ) dHx −
U
τ (nx ) dHx ≤ kτ k∞ ζ,
(8.8)
t where ni is the normal to B i . Combining (8.7) and (8.8), we conclude the proof. u One could have also modified the proof of Lemma 6.4 [Ce] and replaced the approximation in terms of balls, by cubes. Proof of Theorem 6.1. Theorem 6.1 can be viewed as a consequence of a general approximation procedure developed by Alberti and Bellettini [AlBe]. We briefly recall their proof and refer the reader to [AlBe] for details. Let u be a function of bounded variation, then general results of measure theory imply the existence of a sequence {un } of polyhedral functions converging to u in L1 (T ) and such that the vectors measure of the partial first derivatives Dun converge weakly to Du and also that the perimeters of ∂un converge to the one of ∂u. Since τ is continuous, a theorem of Reshetnyak (see [LM]) implies that F(un ) converge to F(u). Therefore for any δ positive, there exists a polyhedral set W such that δ δ 1W ∈ V(1Wm , ) and F(W ) − F(Wm ) ≤ . 3 2 For any h small enough, we approximate the polyhedral set W with disjoint cubes S R 1 , . . . , R ` of size h and basis B 1 , . . . B ` . The set ∂W/ `i=1 B i has arbitrarily small area and is denoted by U δ . As τ is bounded, one has ` Z X i=1
Bi
Z τ (ni ) dHx − F(W ) ≤ kτ k∞
Uδ
dHx ≤ kτ k∞ δ.
This concludes the theorem. u t Remark 8.2. Since we are only interested in approximating the Wulff shape, one could have also used Aleksandrov’s Theorem (see [EG]) which ensures that the boundary of a convex function has almost surely a second derivative. Remark 8.3. As explained previously, Theorem 6.1 holds for any function u of bounded variation. Thus, following the arguments developed in Sect. 6, one can prove the lower bound (Proposition 2.2) for any function u of bounded variation. This implies that a large deviation principle for the measures µ+ β,1N holds.
228
T. Bodineau
Acknowledgements. I am deeply indebted to both D. Ioffe and E. Presutti who have been involved at every stage of this work. This paper would not have been possible without their invaluable suggestions, advice and support. I am also very grateful to A. Chambolle for explaining many results of geometric measure theory and providing advice and references. My gratitude also goes to G. Bellettini and F. Comets for our very useful discussions. I thank S. Shlosman andY.Velenik for useful comments. Finally, I acknowledge the kind hospitality of the Dipartimento di Mathematica di Roma Tor Vergata, where part of this work was completed.
References [AlBe] Alberti, G., Bellettini, G.: Asymptotic behavior of a non local anisotropic model for phase transition. J. Math. Ann. 310 No. 3, 527–560 (1998) [ABCP] Alberti, G., Bellettini, G., Cassandro, M., Presutti E.: Surface tension in Ising system with Kac potentials. J. Stat. Phys. 82, No. 3–4, 743–796 (1996) [AmBr] Ambrosio, L., Braides, A.: Functionals defined on partitions in sets of finite perimeter II: Semicontinuity, relaxation and homogenization. J. Math. pures et appl. 69, 307–333 (1990) [ACCN] Aizenman M., Chayes J., Chayes L., Newman C.: Discontinuity of the magnetization in the onedimensional 1 2 Ising and Potts model. J. Stat. Phys. 50, 1–40 (1988) |x−y|
[BCP]
Bellettini, G., Cassandro, M., Presutti, E.: Constrained minima of non local free energy functionals. J. Stat. Phys. 84, 1337–1349 (1996) [BBBP] Benois, O., Bodineau, T., Butta, P., Presutti, E.: On the validity of van der Waals theory of surface tension. Mark. Proc. and Rel. Fields 3, No. 2, 175–198 (1997) [BBP] Benois, O., Bodineau, T., Presutti, E.: Large deviations in the van der Waals limit. Stoch. Proc. and Appl. 75, 89–104 (1998) [Bo] Bodineau, T.: Wulff construction for Ising model with finite range Kac potentials. In preparation [BIV] Bodineau, T., Ioffe, D., Velenik, Y.: Rigorous probabilistic analysis of equilibrium crystal shapes. Preprint, April (1999) [Ce] Cerf, R.: Large deviations for three dimensional supercritical percolation. Preprint, October (1998) [DKS] Dobrushin, R.L., Kotecky, R., Shlosman, S.: Wulff construction: a global shape from a local interaction. Providence, RI: Am. Math. Soc., 1992 [DP] Deuschel, J. D., Pisztora, A.: Surface order large deviations for high-density percolation. Prob. Theo. Rel. Fields, 104, 467–482 (1996) [DS] Deuschel, J. D., Stroock, D.: Large deviations. London–New York: Academic Press, 1989 [EG] Evans, L., Gariepy, R.: Measure Theory and Fine Properties of Functions. London: CRC Press, 1992 [ES] Edwards, R., Sokal, A.: Generalization of the Fortuin–Kasteleyn–Swendsen–Wang representation and Monte Carlo algorithm. Phys. Rev. D 38 No. 6, 2009–2012 (1988) [F] Fonseca, I.: The Wulff Theorem revisited. Proc. Roy. Soc. London; Ser. A, 432, No 1884, 125–145 (1991) [FM] Fonseca, I., Mueller, S.: A uniqueness proof of the Wulff Theorem. Proc. Roy. Soc. Edinburgh; Sect A, 119, 125–136 (1991) [Gri] Grimmett, G.: The stochastic random cluster process and the uniqueness of random cluster measures. Ann. Prob. 23 No. 4, 1461–1510 (1995) [I1] Ioffe, D.: Large deviations for the 2D Ising model: A lower bound without cluster expansions. J. Stat. Phys. 74, 411–432 (1994) [I2] Ioffe, D.: Exact deviation bounds up to Tc for the Ising model in two dimensions. Prob. Theo. Rel. Fields 102, 313–330 (1995) [IS] Ioffe D., Schonmann R.: Dobrushin-Kotecky-Shlosman Theorem up to the critical temperature. Commun. Math. Phys. 199, 117–149 (1998) [LM] Luckhaus, S., Modica, L.: The Gibbs-Thompson relation within the gradient theory of phase transitions. Arch. Rat. Mech. Anal. 107, 71–83 (1989) [LP] Lebowitz, J., Pfister, C.E.: Surface tension and phase coexistence. Phys. Rev. Lett. 46 No. 15, 1031– 1033 (1981) [LSS] Liggett, T., Schonmann, R., Stacey A.: Domination by product measures. Ann. Prob. 25 No. 1, 71–95 (1997)
Wulff Construction in Three and More Dimensions
229
[MMR] Messager, A., Miracle-Solé, S., Ruiz, J.: Convexity property of the surface tension and equilibrium crystals. J. Stat. Phys. 67 No. 3/4, 449–470 (1992) [Pi1] Pisztora, A.: Surface order large deviations of Ising, Potts and percolation models. Prob. Theo. Rel. Fields 104, 427–466 (1996) [Pi2] Pisztora, A.: Lectures in IHP. Unpublished, Paris, June (1998) [Pf] Pfister, C.E.: Large deviations and phase separation in the two dimensional Ising model. Helvetica Physica Acta 64, 953–1054 (1991) [PV] Pfister, C.E., Velenik Y.: Large deviations and continuum limit in the 2D Ising model. Prob. Theo. Rel. Fields 109 No. 4, 435–506 (1997) [Ta] Taylor, J.: Crystalline variational problems. Bull. Am. Math. Soc. 84 No. 4, 568–588 (1978) [SS] Schonmann, R., Shlosman, S.: Complete analyticity for the 2D Ising model completed. Commun. Math. Phys. 170, 453–482 (1996) Communicated by J. L. Lebowitz
Commun. Math. Phys. 207, 231 – 247 (1999)
Communications in
Mathematical Physics
© Springer-Verlag 1999
An Inverse Scattering Problem with Part of the Fixed-Energy Phase Shifts A. G. Ramm Mathematics Department, Kansas State University, Manhattan, KS 66506, USA. E-mail: [email protected] Received: 3 March 1999 / Accepted: 27 April 1999
Abstract: Assume that q(r) is a real-valued, compactly supported potential, q(r) = 0 Rn , |x| < a}. Let L be for r := |x| > a, q(r) ∈ L2 (Ba ), n = 3, Ba := {x : x ∈P 1 an arbitrary fixed subset of non-negative integers such that ` = ∞, and {δ` } be `6 =0,`∈L
fixed-energy phase shifts corresponding to q(r). The main result is: Theorem. The data {δ` }∀`∈L determine q(r) uniquely. 1. Introduction In this paper a uniqueness theorem is established for inverse scattering with fixed energy data. This theorem says that a real-valued compactly supported spherically symmetric potential q(r) is uniquely determined by the subset of the phase shifts δ` at an arbitrary fixed positive energy. The subset {δ` }`∈L is defined by an arbitrary subset L of integers {0, 1, 2, . . . }, with the property X `∈L,`6=0
1 = ∞. `
(1.1)
No results of this type have been known or conjectured earlier. Condition (1.1) appeared in Müntz’s theorem: it is a necessary and sufficient condition for the completeness of the set {x ` } in L1 (0, a), for an arbitrary fixed a > 0. Such a result gives much deeper understanding of the quantum-mechanical inverse scattering problem with fixed-energy data. It may also be of some practical significance because in some physical experiments the phase shifts can be measured not for all ` and it is important to know what part of the fixed-energy phase shifts is still sufficient for the unique identification of the potential. We now describe the basic ideas of the proof. Our proof is based on two fundamental results.
232
A. G. Ramm
The first result is the uniqueness theorem from [R1] which says that the fixed-energy scattering data A(α 0 , α), ∀α 0 , α ∈ S 2 , determine q(x) ∈ Qa uniquely. Here A(α 0 , α) is the scattering amplitude corresponding to the potential q(x) ∈ Qa := {q : q = q, ¯ q(x) = 0 for |x| > a, q(x) ∈ L2 (Ba )}, where Ba := {x : |x| ≤ a, x ∈ R3 } and the bar stands for complex conjugate. Under this assumption one can check, using Ra Hölder’s inequality, that 0 r|q(r)|dr < ∞. The proof of the above uniqueness theorem is valid in Rn , n ≥ 3, and is not valid for n = 2 (see [R2] for an explanation). An algorithm for inversion of noisy fixed-energy scattering data is developed in [R5] (see also [R2]), where a stability estimate for this algorithm is obtained. In [A] a discussion of the Newton-Sabatier method for inversion of fixed-energy phase shifts is given and an example is constructed of two quite different compactly supported piecewise-constant potentials which produce practically the same ( within the accuracy 10−5 ) fixed-energy phase shifts for all values of the angular momenta. This example illustrates the stability estimate from [R5] and shows that the above estimate is sharp. Actually, property C for a pair of Schrödinger operators is used for a proof of the uniqueness theorem for inverse scattering problem with fixed-energy data. This theorem follows from property C. The notion of property C has been introduced and applied to many inverse problems in a series of papers by the author (see [R2] and references therein, [R3]-[R4]). Let us formulate this notion for a pair {L1 , L2 } of the Schrödinger operators Lj = −O2 + qj − k 2 , j = 1, 2, qj ∈ Qa , k ≥ 0 is a constant, k 2 is the energy of the particle. Let D ⊂ Rn be a bounded domain and Nj := ND (Lj ) := {w : Lj w = 0 in D, w ∈ H 2 (D)}. Let f ∈ L2 (D). Definition 1.1. If Z D
f (x)w1 (x)w2 (x)dx = 0 ∀wj ∈ Nj
⇒ f = 0,
(1.2)
then we say that the pair {L1 , L2 } has property C. Let n = 3 and S 2 be the unit sphere in R3 . Define the scattering solution corresponding to the operator L = −O2 + q − k 2 , and a fixed k > 0, which we take, without loss of generality, to be k = 1 in this paper, as the solution to the equation: (1.3) [O2 + 1 − q(x)]ψ(x, α) = 0 in R3 , ir e x 1 +o , r := |x| → ∞, α 0 := . (1.4) ψ(x, α) = exp(ix · α) + A(α 0 , α) r r r The unit vector α is given, the coefficient A(α 0 , α) is called the scattering amplitude. It is well known that q(x) ∈ Qa determines A(α 0 , α) uniquely. It is proved in [R2] that Z 2 f (x)ψ1 (x, α)ψ2 (x, β)dx = 0 ∀α, β ∈ S ⇒ f = 0, (1.5) D
where ψj (x, α), j = 1, 2, is the scattering solution corresponding to the operator L = −O2 + qj − k 2 , j = 1, 2, and a fixed k = 1.
Inverse Scattering
233
The inverse scattering problem (ISP) with fixed-energy data consists in finding q(x) ∈ Qa given A(α 0 , α) ∀α 0 , α ∈ S 2 . Denote by S˜ 2 an arbitrary small open subset of S 2 . Theorem 1.1 ([R1]). The data A(α 0 , α) ∀α 0 ∈ S˜12 , ∀α ∈ S˜22 , determine q(x) ∈ Qa uniquely. Theorem 1.2 ([R2]). If qj ∈ Qa , j = 1, 2, then (1.2) holds. Theorem 1.2 implies that the pair {L1 , L2 } of Schrödinger’s operators with potentials qj ∈ Qa does have property C. The second result we will use is the uniqueness theorem for analytic functions. Let us assume that h(`) is a holomorphic function in 5+ := {` : Re` > 0}, ` = σ + iτ, σ ≥ 0, and τ are real numbers, h(`) ∈ N (Nevanlinna class in 5+ ), that is Z π 1 − reiϕ + ln h dϕ < ∞, (1.6) sup 1 + reiϕ 0 0, + where ln x = 0 if ln x ≤ 0. Theorem 1.3. If h(`) ∈ N then h(`) = 0, ∀` ∈ L,
(1.7)
h(`) ≡ 0 in 5+ ,
(1.8)
h(`) = 0 ∀` = 0, 1, 2, . . . .
(1.9)
implies
in particular
Theorem 1.3 is a consequence of Theorem 1.4 which is formulated below. This theorem in turn is an immediate corollary to Theorem 15.23 in [Ru, p. 334]. Theorem 1.4. Assume that the following conditions hold: (i) f (z) is holomorphic in the unit disc D1 , (ii) f (z) ∈ N in D1 , that is Z π ln+ |f (reiϕ )|dϕ < ∞, sup 0
(1.10)
(iii) f (zn ) = 0, n = 1, 2, 3, . . . , and ∞ X (1 − |zn |) = ∞.
(1.11)
n=1
Then f (z) ≡ 0 in D1 .
(1.12)
234
A. G. Ramm
Let us explain why Theorem 1.4 implies Theorem 1.3. Note that the function `=
1−w 1+w
(1.13)
1−` 1+`
(1.14)
maps conformally D1 onto 5+ , while w=
is the inverse map 5+ → D 1. The function h(`) = h 1−w 1+w := f (w) is analytic in D1 , f (w) ∈ N in D1 , and f (w` ) = 0. where w` = If ` ∈ L, then X `∈L,`6=0
1−` 1+`
and h(`) = 0. 1 − ` = 1 − 1 + `
X
(1 − |w` |) =
`∈L,`6=0
X `∈L,`6 =0
2 = ∞, `+1
(1.15)
because of assumption (1.1). Thus, f (w) = 0 in D1 by Theorem 1.4, and therefore h(`) = 0 in 5+ . Thus, Theorem 1.3 follows from Theorem 1.4. In order to describe the ideas of our proof, we need to state some known facts from the scattering theory. If q(x) = q(r), r = |x|, that is, the potential is spherically symmetric, and if k = 1, then the scattering solution is: ψ(x, α) =
∞ X `=0
4πi `
x ψ` (r) Y` (x 0 )Y` (α), x 0 := , r r
(1.16)
where ψ` (r) satisfies Eqs. (1.17)–(1.19): `(` + 1) ψ` − q(r)ψ` = 0, r > 0, r2 ψ` (r) = O(r `+1 ) as r → 0, `π + δ` ) + o(1) as r → ∞. ψ` (r) = eiδ` sin(r − 2 ψ`00 + ψ` −
(1.17) (1.18) (1.19)
Here δ` is called the fixed-energy phase shift corresponding to the angular momentum `. The functions Y` (α) = Y`m (α), −` ≤ m ≤ `. in (1.12) are the spherical harmonics orthonormalized in L2 (S 2 ). The summation in (1.16) and in (1.20) below includes the summation with respect to m, −` ≤ m ≤ `, and is not shown for brevity. The corresponding scattering amplitude for the spherically symmetric potential is of the form: A(α 0 , α) = A(α 0 · α) =
∞ X
A` Y` (α 0 )Y` (α),
(1.20)
`=0
where A` = 2πi(1 − e2iδ` ) = 4π eiδ` sin δ` .
(1.21)
Inverse Scattering
235
Recall that we assume k = 1 throughout. Therefore, in the case of spherically symmetric potentials which we consider in this paper, there is a one-to-one correspondence between the scattering amplitude A(α 0 , α) and the set of numbers {A` }`=0,1,2,... . The set {δ` }`=0,1,2,... , −π ≤ δ` < π, determines the set {A` }`=0,1,2,... uniquely, so there is a one-to-one correspondence between the scattering amplitude at a fixed energy and the set of all phase shifts with this energy. Using (1.16) and the orthonormality of the spherical harmonics, one obtains: Z p(r)ψ1 (x, α)ψ2 (x, β)dx ⇐⇒ 0= Ba
0=
∞ Z X
a
`=0 0
p(r)ψ1` (r)ψ2` (r)dr Y` (α)Y` (β), ∀α, β ∈ S 2 ,
(1.22)
where p(r) ∈ Qa is an arbitrary function. Multiplying the second integral in (1.22) by Y` (α), integrating over S 2 and using the orthonormality of the spherical harmonics, one gets that the first equality in (1.22) is equivalent to Z a drp(r)ψ1` (r)ψ2` (r) ∀` = 0, 1, 2, . . . . (1.23) 0= 0
The regular solution ϕ` (r) to Eq. (1.17) is defined uniquely by its behavior near the origin: ϕ` (r) =
r `+1 + o(r `+1 ), r → 0. (2` + 1)!!
(1.24)
This solution is a real-valued function for ` = 0, 1, 2, . . . and r > 0. Its behavior at infinity is known: `π + δ` + o(1), r → +∞, (1.25) ϕ` = |F` | sin r − 2 where δ` is the same as in (1.19) and |F` | is a certain positive constant (the value of the Jost function F` (k) at k = 1). Since ψ` (r) solves (1.17) and satisfies (1.18), it follows that ψ` = c` ϕ` (r), c` = const. Therefore condition (1.23) is equivalent to Z a drp(r)ϕ1` (r)ϕ2` (r) := h(`), ` = 0, 1, 2, . . . , 0=
(1.26)
(1.27)
0
where we have used the real-valuedness of ϕ` (r). Let us now describe the idea of our proof. Step 1. Assuming that the data {δ` }`∈L correspond to two different potentials q1 (r) and q2 (r), qj (r) ∈ Qa , we derive the following orthogonality relation: h(`) = 0 ∀` ∈ L,
p(r) := q1 (r) − q2 (r).
(1.28)
236
A. G. Ramm
Step 2. We prove that the function h1 (`), defined below, in formula (3.65’), is holomorphic in 5+ := {` : ` ∈ C, <` > 0} and belongs to class N defined in (1.6). Condition (1.28) implies h1 (`) = 0 ∀` ∈ L. This implies by Theorem 1.3 that h1 (`) = 0 ∀` ∈ 5+ . Therefore, condition (1.28) implies h(`) = 0, ` = 0, 1, 2, . . . . that is (1.27) holds. This implies, as we have proved above, that Z p(r)ψ1 (x, α)ψ2 (x, β)dx ∀α, β ∈ S 2 . (1.29) 0= Ba
Equation (1.29) and property C for the pair {L1 , L2 } of the Schrödinger operators Lj = t −O2 + qj (x) − 1, qj (x) ∈ Qa , imply p(r) = 0, that is, q1 (r) = q2 (r). u An essential ingredient of our proof of the implication {h1 (`) = 0 ∀` ∈ L} ⇒ {h1 (`) = 0 ∀` = 0, 1, 2, 3, . . . }. The proof of this implication is based on the existence of the transformation operators whose kernel does not depend on `. Existence and uniqueness of such operators as well as the estimate (3.53) (see Sect. 3 below), which we use in the proof of the above implication, are established in Sect. 3. These results, although new and of independent interest, play an auxiliary role in our proof. They are presented in Sects. 3.3 and 3.4 as a part of the proof. The description of the idea of our proof is complete. In Sect. 3 we derive the orthogonality relation (1.28). In the same section we study the analytic properties of h(`) as a function of complex `. In this study there are two basic steps. First, we study the function r Z a πr 2 J 1 (r), drp(r)u` (r), u` (r) := (1.30) h0 (`) := 2 `+ 2 0 where J`+ 1 (r) is the standard Bessel function. 2 Note that ϕ` (r) = u` (r) if q(r) = 0. Define r 1 `+ 1 2 0 2 2 0(` + 1)]2 , H (`) := h0 (`)[ π 2
(1.31)
where 0(z) is the Gamma-function. We prove: Lemma 1.1. The function H (`) is holomorphic in 5+ and H (`) ∈ N . Secondly, we prove the existence of the transformation operator, which sends u` (r) into ϕ` (r). Z r K(r, ρ)u` (ρ)ρ −2 dρ, K(r, 0) = 0. (1.32) ϕ` (r) = u` (r) + 0
It is crucial for our argument that K(r, ρ) does notqdepend on `. Therefore, the analytic 1 properties of h1 (`) := h(`)[0 21 0 ` + 21 2`+ 2 π2 ]2 and H (`), as functions of `, are
Inverse Scattering
237
essentially the same: these two functions are both holomorphic in 5+ and belong to class N . In Sect. 3 we prove some technical estimates for the kernel K(r, ρ) of the transformation operator. Although the transformation operators of the type (1.32) appeared formally earlier in the physical literature [CS, p.185], their existence was not proved. In the literature there exists a construction of the transformation operators whose kernels depend on `, see [V, Le, M]. The difficulty of the existence proof for the transformation operator, whose kernel K(r, ρ) does not depend on `, comes from the fact that the Goursat-type problem which one can derive for K(r, ρ) involves differential operators with variable coefficients which degenerate at the origin. We overcome this difficulty by introducing new variables and reducing the problem to an equivalent Volterra-type integral equation. Existence and uniqueness of the solution to this equation are established in Sect. 3, where some estimates of the solution are given. This concludes the introduction. In Sect. 2 we state the basic uniqueness result. In Sect. 3 proofs are given. 2. Statement of the Basic Result Let us assume that q(x) ∈ Q := {q : q ∈ Qa , q(x) = q(r), r := |x|},
Ra
(2.1)
and let δ` denote the fixed-energy phase shifts. Note that if q ∈ Q, then 0 r|q(r)|dr < c < ∞. Here and below c > 0 stands for various estimation constants. The inverse scattering problem we are interested in can now be formulated: ISP. Given the data {δ` }∀`∈L , where L satisfies condition (1.1), can one recover q(r) ∈ Q uniquely? Our basic result is: Theorem 2.1. Let L be an arbitrary fixed subset of non-negative integers which satisfies condition (1.1). Then the data {δ` }∀`∈L , corresponding to a q(r) ∈ Q, determine q(r) uniquely. This result implies, in particular, that there is no q(r) ∈ Q, q(r) 6 ≡ 0, such that δ2` = 0 ∀` = 0, 1, 2, . . . . It also implies that there is no q(r) ∈ Q, such that δ0 6= 0, δ` = 0, ` = 1, 2, 3, . . . , which means that there are no potentials in Q producing the scattering amplitude A(α, α 0 ) which is constant for all α, α 0 ∈ S 2 at a fixed positive energy, see also [R6] where this has been proved by a different argument for the first time. 3. Proofs 3.1. Proof of the orthogonality relation (1.28). Suppose q1 (r) and q2 (r) generate the same data {δ` }∀`∈L . Subtract from Eq. (1.17) with q = q1 and ψ` = ψ1` similar equation with q = q2 and ψ` = ψ2` to get: ψ`00 + ψ` −
`(` + 1) ψ` − q1 ψ` = p(r)ψ2` , r2
(3.1)
238
A. G. Ramm
where ψ` := ψ1` − ψ2` , p(r) := q1 (r) − q2 (r).
(3.2)
Multiply (3.1) by ψ1` (r), integrate over (0, ∞) and then by parts using (1.18) and (1.19) and the assumption that δ` is the same for ψ1` and ψ2` for ` ∈ L. The result is Z a 0= p(r)ψ1` (r)ψ2` (r)dr ∀` ∈ L. (3.3) 0
This is equivalent to the desired relation (1.28) because of (1.26). u t 3.2. Analytic properties of the function H (`). Proof of Lemma 1.1. Recall the well-known formula [GR, 8.411.8]: r
2 0 π
Z 1 1 `+ 1 (1 − t 2 )` eirt dt. 2 2 0(` + 1)u` (r) = r `+1 2 −1
(3.4)
From (1.30), (1.31), and (3.4) one gets: Z
a
H (`) =
Z drp(r)r
0
2`+2
1
−1
2 ` irt
2
(1 − t ) e dt
.
(3.5)
Let ` = σ + iτ, σ ≥ 0. Then H (`) is a holomorphic function of ` for σ > 0 and Z a dr|p(r)|ra 2σ +1 ≤ ca 2σ . (3.6) |H (`)| ≤ 0
One can always assume a > 1 since σ > 0. Let us check that (3.6) implies that H (`) ∈ N . One has ln+ (ab) ≤ ln+ a + ln+ b
for a, b > 0.
Therefore, using (3.6), one obtains: Z π Z π iϕ 1 − reiϕ 2Re 1−reiϕ + + 1+re |dϕ ln |H ln |ca |dϕ ≤ 1 + reiϕ −π −π Z π 1 − reiϕ Re dϕ ≤ c1 + 2 ln a 1 + reiϕ −π Z π 1 − r2 dϕ ≤ c1 + 2 ln a 2 −π 1 + r + 2r cos ϕ = c1 + 4π ln a < ∞, a > 1. Here we have used the known formula Z π 2π dϕ = , 2 1 − r2 −π 1 + r + 2r cos ϕ
0 < r < 1,
which is easy to check. Estimate (3.7) proves that H (`) ∈ N . Lemma 1.1 is proved. u t
(3.7)
(3.8)
Inverse Scattering
239
3.3. Transformation operators. Define ∂2 Lr ϕ := r 2 2 + r 2 − r 2 q(r) ϕ := L0r ϕ − r 2 q(r)ϕ. ∂r
(3.9)
For the regular solution to (1.17) one has the following differential equation: Lr ϕ` (r) = `(` + 1)ϕ` (r), q and for the function u` (r) = πr 2 J`+ 1 (r) the equation
(3.10)
L0r u` (r) = `(` + 1)u` (r).
(3.11)
2
Let us look for the kernel K(r, ρ) such that formula (1.32) gives the regular solution to Eq. (1.17). Substitute (1.32) into (1.17), drop the index ` for convenience, use (3.10) and (3.11), and get Z r K(r, ρ)uρ −2 dρ 0 = − r 2 q(r)u + (r 2 − r 2 q(r)) 0 Z r Z r −2 2 2 − K(r, ρ)ρ L0ρ udρ + r ∂r K(r, ρ)uρ −2 dρ. (3.12) 0
0
We assume first that K(r, ρ) is twice continuously differentiable with respect to its variables in the region 0 < r < ∞, 0 < ρ ≤ r. This assumption requires extra smoothness of q(r), q(r) ∈ C 1 (0, a). If q(r) satisfies condition (2.1), then Eq. (3.18) below has to be understood in the sense of distributions. Eventually we will work with an integral Eq. (3.45) (see below) for which assumption (2.1) suffices. Note that Z r Z r K(r, ρ)ρ −2 L0ρ udρ = L0ρ K(r, ρ)uρ −2 dρ + K(r, r)ur − Kρ (r, r)u, (3.13) 0
0
provided that K(r, 0) = 0.
(3.14)
dK(r, r) . K˙ := dr
(3.15)
We assume (3.14) to be valid. Denote
Then
Z r 2 ∂r2
0
r
˙ + K(r, r)ur − 2 K(r, r)u K(r, ρ)uρ −2 dρ = Ku r Z r + Kr (r, r)u + r 2 Krr (r, ρ)uρ −2 dρ.
(3.16)
0
Combining (3.12)–(3.16) and writing again u` in place of u, one gets Z r 0= [Lr K(r, ρ) − L0ρ K(r, ρ)]u` (ρ)ρ −2 dρ + u` (r)[−r 2 q(r) + K˙ 0
−
2Kr (r, r) + Kr (r, r) + Kρ (r, r)], ∀r > 0, ` = 0, 1, 2, . . . . r
(3.17)
240
A. G. Ramm
Let us prove that (3.17) implies: Lr K(r, ρ) = L0ρ K(r, ρ), 0 < ρ ≤ r, 2 d K(r, r) 2K(r, r) 2K˙ = . q(r) = 2 − r r r dr r
(3.18) (3.19)
This proof requires a lemma. Lemma 3.1. Assume that ρf (ρ) ∈ L1 (0, r) and ρA(ρ) ∈ L1 (0, r). If Z r f (ρ)u` (ρ)dρ + u` (r)A(r) ∀` = 0, 1, 2, . . . , 0=
(3.20)
0
then f (ρ) ≡ 0 and A(r) = 0.
(3.21)
Proof. Equations (3.20) and (3.4) imply: Z Z 1 d ` r 2 ` dt (1 − t ) dρρf (ρ)eiρt 0= idt −1 0 Z 1 d ` irt + rA(r) (1 − t 2 )` e dt. idt −1 Therefore Z 0=
Z
d ` (t 2 − 1)` dt [ dt ` −1 1
r
dρρf (ρ)eiρt + rA(r)eirt ], l = 0, 1, 2, . . . .
(3.22)
0
Recall that the Legendre polynomials are defined by the formula P` (t) =
1 d` 2 (t − 1)` 2` ! dt `
and they form a complete system in L2 (−1, 1). Therefore (3.22) implies Z r dρρf (ρ)eiρt + rA(r)eirt = 0 ∀t ∈ [−1, 1].
(3.23)
(3.24)
0
Equation (3.24) implies Z
r
dρρf (ρ)eiρt = 0, ∀t ∈ [−1, 1],
(3.25)
0
and rA(r) = 0.
(3.26)
Therefore A(r) = 0. Also f (ρ) = 0 because the left-hand side of (3,25) is an entire function of t, which vanishes on the interval [−1, 1] and, consequently, it vanishes identically, so that ρf (ρ) = 0 and therefore f (ρ) ≡ 0. Lemma 3.1 is proved. u t
Inverse Scattering
241
3.4. Existence and uniqueness of the transformation operators. Let us prove that the problem (3.18), (3.19), (3.14), which is a Goursat-type problem, has a solution and this solution is unique in the class of functions K(r, ρ), which are twice continuously differentiable with respect to ρ and r, 0 < r < ∞, 0 < ρ ≤ r. In this section we assume that q(r) ∈ C 1 (0, a). This assumption implies that K(r, ρ) is twice continuously differentiable. If q(r) ∈ Q, see (2.1), the arguments in this section which deal with the integral equation (3.45) remain valid. Specifically, existence R a and uniqueness of the solution to Eq. (3.45) is proved under the only assumption 0 r|q(r)|dr < ∞ as far as the smoothness of q(r) is concerned. By a limiting argument one can reduce the smoothness requirements on q to the condition (2.1), but in this case Eq. (3.18) has to be understood in the distributional sense. Let us rewrite the problem we want to study: r 2 Krr − ρ 2 Kρρ + [r 2 − r 2 q(r) − ρ 2 ]K(r, ρ) = 0, 0 < ρ ≤ r, Z r r sq(s)ds := g(r), K(r, r) = 2 0 K(r, 0) = 0.
(3.27) (3.28) (3.29)
The difficulty in the study of this Goursat-type problem comes from the fact that the coefficients in front of the second derivatives of the kernel K(r, ρ) are variable. Let us reduce problem (3.27)–(3.29) to the one with constant coefficients. To do this, introduce the new variables: ξ = ln r + ln ρ, η = ln r − ln ρ.
(3.30)
Note that ξ +η
r=e 2 , η ≥ 0,
ξ −η
ρ=e 2 , −∞ < ξ < ∞,
(3.31) (3.32)
and ∂r =
1 1 (∂ξ + ∂η ), ∂ρ = (∂ξ − ∂η ). r ρ
Let
(3.33)
K(r, ρ) := B(ξ, η).
A routine calculation transforms Eqs. (3.27)–(3.29) to the following ones: 1 Bξ η (ξ, η) − Bη (ξ, η) + Q(ξ, η)B = 0, η ≥ 0, −∞ < ξ < ∞, 2 ξ
B(ξ, 0) = g e 2
:= G(ξ ), −∞ < ξ < ∞,
B(−∞, η) = 0, η ≥ 0, where g(r) is defined in (3.28). Here we have defined Q(ξ, η) :=
ξ +η i 1 h ξ +η − eξ +η q e 2 − eξ −η , e 4
(3.34) (3.35) (3.36)
(3.37)
242
A. G. Ramm
and took into account that ρ = r implies η = 0, while ρ = 0 implies, for any fixed η ≥ 0, that ξ = −∞. Note that ξ
sup
−∞<ξ <∞
Z
A
sup
0≤η≤B −∞
e− 2 G(ξ ) < c,
|Q(s, η)|ds ≤ c(A, B),
(3.38) (3.39)
for any A ∈ R and B > 0, where c(A, B) > 0 is a constant. To get rid of the second term on the left-hand side of (3.34), let us introduce the new kernel L(ξ, η) by the formula: ξ
L(ξ, η) := B(ξ, η)e− 2 .
(3.40)
Then (3.34)–(3.36) can be written as: Lηξ (ξ, η) + Q(ξ, η)L(ξ, η) = 0, η ≥ 0, −∞ < ξ < ∞, L(ξ, 0) = e
− ξ2
(3.41)
ξ e2
Z 1 G(ξ ) := b(ξ ) := sq(s)ds, −∞ < ξ < ∞, 2 0 L(−∞, η) = 0, η ≥ 0.
(3.42) (3.43)
We want to prove existence and uniqueness of the solution to (3.41)–(3.43). In order to choose a convenient Banach space in which to work, let us transform problem (3.41)– (3.43) to an equivalent Volterra-type integral equation. Integrate (3.41) with respect to η from 0 to η and use (3.42) to get Z
0
Lξ (ξ, η) − b (ξ ) +
η
Q(ξ, t)L(ξ, t)dt = 0.
(3.44)
0
Integrate (3.44) with respect to ξ from −∞ to ξ and use (3.44) to get Z L(ξ, η) = −
ξ −∞
Z ds
η
dtQ(s, t)L(s, t) + b(ξ ) := V L + b,
(3.45)
0
where Z V L := −
Z
ξ −∞
ds
η
dtQ(s, t)L(s, t).
(3.46)
0
Consider the space X of continuous functions L(ξ, η), defined in the half-plane η ≥ 0, −∞ < ξ < ∞, such that for any B > 0 and any −∞ < A < ∞ one has kLk := kLkAB :=
sup −∞<s≤A t≤B
e−γ t |L(s, t)| < ∞,
(3.47)
Inverse Scattering
243
where γ > 0 is a number which will be chosen later so that the operator V in (3.45) will be a contraction mapping on the Banach space of functions with norm (3.47) for a fixed pair A, B. To choose γ > 0, let us estimate the norm of V . One has: Z kV Lk ≤
sup
≤ kLk
−∞
−∞<ξ ≤A,0≤η≤B
Z
Z
ξ
ξ
sup
Z
ds
−∞<ξ ≤A,0≤η≤B −∞
η
ds
dt|Q(s, t)|e−γ (η−t) e−γ t |L(s, t)| ≤
0
s+t c dt 2es+t +es+t |q e 2 | e−γ (η−t) ≤ kLk, γ (3.48)
η
0
where c > 0 is a constant depending on A, B and Z 2
Z
A −∞
ds
η
dtes+t−γ (η−t) = 2eA
0
A −∞
Z
η
ds
η
η
=2 0
r|q(r)|dr. Indeed, one has:
dtet−γ (η−t) dt ≤ 2eA+B
dtes+t |q(e
s+t 2
s+t 2
)|e−γ (η−t) =
Z
dte−γ (η−t)
η
dte−γ (η−t)
Z
0
Z
e
c1 1 − e−γ B = , γ γ (3.49’)
, one gets:
0
Z
0
0
and, using the substitution σ = e Z
Z
Ra
A+t 2
dσ σ |q(σ )| =
2(1 − e−γ B )
0
γ
A
−∞
Z
a
0
s+t dses+t |q e 2 | dσ σ |q(σ )| :=
c2 . γ (3.49”)
From these estimates inequality (3.48) follows. It follows from (3.48) that V is a contraction mapping in the space XAB of continuous functions in the region −∞ < ξ ≤ A, 0 ≤ η ≤ B, with the norm (3.47) provided that γ > c.
(3.50)
Therefore Eq. (3.45) has a unique solution L(ξ, η) in the region −∞ < ξ < A,
0≤η≤B
(3.51)
for any real A and B > 0 if (3.50) holds. This means that the above solution is defined for any ξ ∈ R and any η ≥ 0. Equation (3.45) is equivalent to problem (3.41)-(3.43) and, by (3.40), one has: ξ
B(ξ, η) = L(ξ, η)e 2 .
(3.52)
Therefore we have proved the existence and uniqueness of B(ξ, η), that is, of the kernel K(r, ρ) = B(ξ, η) of the transformation operator (1.32). Recall that r and ρ are related to ξ and η by formulas (3.31).
244
A. G. Ramm
Let us formulate the result: Lemma 3.2. The kernel of the transformation operator (1.32) solves problem (3.27)(3.29). The solution to this problem does exist and is unique for any potential q(r) ∈ C 1 (0, a) in the class of twice continuously differentiable functions. If q(r) ∈ L∞ (0, a), then K(r, ρ) has first derivatives which are bounded and Eq. (3.27) has to be understood in the sense of distributions. The following estimate holds for any r > 0: Z r |K(r, ρ)|ρ −1 dρ < ∞. (3.53) 0
Proof of Lemma 3.2. We have already proved all the assertions of Lemma 3.2 except estimate (3.53). Let us prove this estimate. Note that Z ∞ Z r η |K(r, ρ)|ρ −1 dρ = r |L(2 ln r − η, η)|e− 2 dη < ∞. (3.54) 0
0
Indeed, if r > 0 is fixed, then, by (3.31), ξ + η = 2 ln r = const. Therefore dξ = −dη, and ρ −1 dρ = 21 (dξ − dη) = −dη, ξ = 2 ln r − η. Thus: Z ∞ Z r 2 ln r−η |K(r, ρ)|ρ −1 dρ = |L(2 ln r − η, η)|e 2 dη 0 0 Z ∞ η |L(2 ln r − η, η)|e− 2 dη. (3.55) =r 0
The following estimate holds: |L(ξ, η)| ≤ ce(2+1 )[ηµ1 (ξ +η)]
1 2 +2
,
(3.56)
where j > 0, j = 1, 2. are arbitrarily small numbers and µ1 is defined in formula (3.60) below, see also formula (3.58) for the definition of µ. Estimate (3.56) is proved below, in Lemma 3.3. From (3.55) and (3.56) estimate (3.53) follows. Lemma 3.2 is proved. u t Lemma 3.3. Estimate (3.56) holds. Proof of Lemma 3.3. From (3.45) one gets: m(ξ, η) ≤ c0 + (W m)(ξ, η), m(ξ, η) := |L(ξ, η)|, Ra where c0 = sup−∞<ξ <∞ |b(ξ )| ≤ 21 0 s|q(s)|ds (see (3.42)), and Z W m :=
ξ −∞
Z ds 0
η
dtµ(s + t)m(s, t), µ(s) :=
s 1 s e 1 + |q(e 2 )| . 2
(3.57)
(3.58)
It is sufficient to consider inequality (3.57) with c0 = 1: if c0 = 1 and the solution m0 (ξ, η) to (3.57) satisfies (3.56) with c = c1 , then the solution m(ξ, η) of (3.57) with any c0 > 0 satisfies (3.56) with c = c0 c1 . Therefore, assume that c0 = 1, then (3.57) reduces to: m(ξ, η) ≤ 1 + (W m)(ξ, η).
(3.59)
Inverse Scattering
245
Inequality (3.56) follows from (3.59) by iterations. Let us give the details. Note that Z η Z η Z ξ Z ξ ds dtµ(s + t) = dt dsµ(s + t) W1 = −∞ 0 0 −∞ Z η dtµ1 (ξ + t) ≤ ηµ1 (ξ + η). = 0
Here we have used the notation
Z µ1 (ξ ) =
ξ
−∞
µ(s)ds,
(3.60)
and the fact that µ1 (s) is a monotonically increasing function, since µ(s) > 0. Note also that µ1 (s) < ∞ for any s, −∞ < s < ∞. Furthermore, Z ξ Z η Z η Z ξ ds dtµ(s + t)tµ1 (s + t) ≤ dtt dsµ(s + t)µ1 (s + t) W 21 ≤ =
−∞ 0 2 2 η µ1 (ξ + η)
2!
2!
0
−∞
.
(3.61)
Let us prove by induction that W n1 ≤
ηn µn1 (ξ + η) . n! n!
(3.62)
For n = 1 and n = 2 we have checked (3.62). Suppose (3.62) holds for some n, then Z η n n Z µn (s + t) η µ1 (ξ + η) tn ξ = dt dsµ(s + t) 1 W n+1 1 ≤ W n! n! n! −∞ n! 0 ≤
ηn+1 µn+1 1 (ξ + η) . (n + 1)! (n + 1)!
(3.63)
By induction, estimate (3.61) is proved for all n = 1, 2, 3, . . . . Therefore (3.59) implies m(ξ, η) ≤ 1 +
∞ 1 X ηn µn1 (ξ + η) 2 +2 ≤ ce(2+1 )[ηµ1 (η+ξ )] , n! n!
(3.64)
n=1
where we have used Theorem 2 from [L, Sect. 1.2], namely the order of the entire P 1 zn function F (z) := 1 + ∞ n=1 (n!)2 is 2 and its type is 2. The constant c > 0 in (3.56) depends on j , j = 1, 2. Recall that the order of an entire function F (z) is the number ρ := lim sup r→∞
ln ln MF (r) , ln r
where MF (r) := max|z|=r |F (z)|. The type of F (z) is the number σ := lim sup r→∞
ln MF (r) . rρ
246
A. G. Ramm
P n It is known [L], that if F (z) = ∞ n=0 cn z is an entire function, then its order ρ and type σ can be calculated by the formulas: ρ
n ln n
lim supn→∞ (n|cn | n ) , σ = . ρ = lim sup 1 eρ n→∞ ln |c | n If cn =
1 , then the above formulas yield ρ (n!)2
=
1 2
and σ = 2. Lemma 3.3 is proved. u t
3.5. Proof of Theorem 2.1. Suppose that there are two potentials which generate the same data {δ` }∀`∈L . In Sect. 3.1 we have proved that this implies (3.3). From (3.3) and (1.26) it follows that (3.3) is equivalent to (1.28). From Lemma 3.2, formula (1.32), Lemma 1.1, and the definition Z a drp(r)ϕ1` (r)ϕ2` (r), (3.65) h(`) = 0
it follows that the function h1 (`) :=
r
2 0 π
1 2 1 0(` + 1)2`+ 2 h(`) ∈ N. 2
(3.65’)
This is checked as in the proof of Lemma 1.1 in Sect. 3.2. There are four terms which one gets from multiplication of ϕ1` (r) by ϕ2` (r), where ϕj ` (r), j = 1, 2. are expressed by formula (1.32) with K(r, ρ) = Kj (r, ρ), j = 1, 2. The first term contains u2` (r) and is identical with (1.30), the second and third terms contain the products of the type u` (r)u` (ρ), while the fourth term contains the term u` (ρ1 )u` (ρ2 ). These terms are treated like in the proof of Lemma 1.1. and estimate (3.53) is used. From (3.65’), Theorem 1.3, and assumption (1.28) it follows that h(`) = 0 for ` = 0, 1, 2, 3, . . . . From this and (1.26) it follows that (1.23) holds for ` = 0, 1, 2, 3, . . . . From (1.23) and (1.22) it follows that Z dxp(r)ψ1 (x, α)ψ2 (x, β) = 0 ∀α, β ∈ S 2 . (3.66) Ba
From (3.66) and Theorem 1.2 one concludes that p(r) = 0. Theorem 2.1 is proved. u t 3.6. Heuristic motivation of the basic result. Here we give an heuristic motivation of the basic result, namely of Theorem 2.1. It is well known that r r 2`+1 2 er πr r 1 J`+ 1 (r) = [1 + o(1)] as ` → ∞. (3.67) u` (r) := √ 2 2 2 2` + 1 2` + 1 One can prove that ϕ` (r) has the same asymptotics r 2`+1 2 er r 1 [1 + o(1)] as ` → ∞. ϕ` (r) = √ 2 2` + 1 2` + 1
(3.68)
Inverse Scattering
247
If one substitutes (3.68) into (1.28), one gets Z a drr 2 p(r)r 2` [1 + o(1)] ∀` ∈ L. 0=
(3.69)
0
If one neglects the term o(1), then one gets Z a drr 2 p(r)r 2` ∀` ∈ L. 0=
(3.70)
0
From (3.70) and the well known Müntz’s theorem [Ru, p.336], it follows that p(r) = 0, which yields the conclusion of Theorem 2.1. This heuristic argument is not a proof because the justification of the passage from (3.69) to (3.70) is not given and is not clear if such a justification can be given directly. Our proof of Theorem 2.1 can be considered as an indirect justification of this heuristic argument. It is known that condition (1.1) is necessary and sufficient for completeness of the set {r ` }`∈L in L1 (0, a) for any fixed a > 0. Therefore one can raise the following question: Is it true that condition (1.1) is necessary for the conclusion of Theorem 2.1 to be valid? This interesting question is open. References [A] [CS] [GR] [L] [Le] [M] [ R1] [R2] [R3] [R4] [R5] [R6] [Ru] [V]
Airapetyan, R., Ramm, A.G. and Smirnova, A.B.: Example of two different potentials which have practical the same fixed-energy phase shifts. Phys. Lett. A 254 3–4, 141–148 (1999) Chadan, K., Sabatier, P.: Inverse problems in Quantum scattering theory. New York: Springer Verlag, 1989 Gradshteyn I., Ryzhik I.: Table of integrals, series and products, Boston: Acad. Press, 1994 Levin, B.: Distribution of zeros of entire functions. AMS Transl. vol. 5 Providence, RI: 1980 Levitan, B.: Inverse Sturm–Liouville problems, Utrecht: VNU Press, 1987 Marchenko, V.: Sturm–Liouville operators and applications. Boston: Birkhäuser Verlag, 1986 Ramm, A.G.: Recovery of the potential from fixed energy scattering data. Inverse Problems 4, 877–886 (1988) Ramm, A.G.: Multidimensional inverse scattering problems. New York: Longman/Wiley, 1992, pp. 1385; Multidimensional inverse scattering problems. Moscow: Mir Publishers, 1994, pp. 1–496 (Russian translation of the expanded monograph) Ramm, A.G.: Completeness of the products of solutions of PDE and inverse problems. Inverse Problems 6, 643–664 (1990) Ramm, A.G.: Necessary and sufficient condition for a PDE to have property C. J. Math. Anal. Appl. 156, 505–509 (1991) Ramm, A.G.: Stability estimates in inverse scattering. Acta Appl. Math. 28 1, 1–42 (1992) Ramm, A.G.: Can a constant be a scattering amplitude? Phys. Lett. A 154A, 35–37 (1991) Rudin, W.: Real and complex analysis. New York: McGraw Hill, 1974 Volk, V.: On inverse formulas for a differential equation with a singularity at x = 0. Uspekhi Math. Nauk 8 4, 1141–151 (1953) (in Russian)
Communicated by B. Simon
Commun. Math. Phys. 207, 249 – 290 (1999)
Communications in
Mathematical Physics
© Springer-Verlag 1999
Spectral Analysis for Systems of Atoms and Molecules Coupled to the Quantized Radiation Field Volker Bach1,?,?? , Jürg Fröhlich2 , Israel Michael Sigal3,??? 1 FB Mathematik MA 7-2, TU Berlin, Str. d. 17. Juni 136, 10623 Berlin, Germany.
E-mail: [email protected]
2 Institut für Theoretische Physik, ETH Hönggerberg, 8093 Zürich, Switzerland.
E-mail: [email protected]
3 Department of Mathematics, University of Toronto, Toronto, M5S 3G3, Canada.
E-mail: [email protected] Received: 3 September 1998 / Accepted: 17 March 1999
Dedicated to Klaus Hepp and Walter Hunziker Abstract: We consider systems of static nuclei and electrons – atoms and molecules – coupled to the quantized radiation field. The interactions between electrons and the soft modes of the quantized electromagnetic field are described by minimal coupling, p → p − eA(x), where A(x) is the electromagnetic vector potential with an ultraviolet cutoff. If the interactions between the electrons and the quantized radiation field are turned off, the atom or molecule is assumed to have at least one bound state. We prove that, for sufficiently small values of the fine structure constant α, the interacting system has a ground state corresponding to the bottom of its energy spectrum. For an atom, we prove that its excited states above the ground state turn into metastable states whose lifetimes we estimate. Furthermore the energy spectrum is absolutely continuous, except, perhaps, in a small interval above the ground state energy and around the threshold energies of the atom or molecule.
1. Introduction and Survey of Results 1.1. The quantum theory of photons and nonrelativistic, bound electrons. In this paper we continue our mathematical analysis of the standard model of nonrelativistic, quantum-mechanical matter in the presence of static nuclei and interacting with the quantized radiation field initiated in [5–7]. The purpose of this paper is to study ground states of atoms and molecules and, for atoms, life-times of their excited states, as described by the standard model, applying and extending the techniques developed [5,6]. In contrast to [5,6] we do not require the assumption of confinement of the interaction, anymore. ? Present address: FB Mathematik, Universität Mainz, 55099 Mainz, Germany.
?? Heisenberg Fellow of the DFG, supported by SFB 288 of the DFG, the TMR-Network on “PDE and
QM”. ??? Supported by NSERC Grant NA 7901
250
V. Bach, J. Fröhlich, I. M. Sigal
The physical system we are analyzing consists of a finite number of nuclei, treated as static sources, and a finite number of nonrelativistic electrons, e.g., atoms, ions, or molecules in a clamped nuclei approximation, interacting with the soft modes of the quantized electromagnetic field which is cut-off in the ultraviolet. The Hilbert space of pure state vectors of the system is given by H := Hel ⊗ F,
(1.1)
where Hel is the Hilbert space of some finite number, N, of electrons, and F is the photon Fock space. Thus, in the Schrödinger (configuration-space) representation, Hel is given by the subspace of totally antisymmetric wave functions in L2 [(R3 × Z2 )N ], where R3 is the configuration space of a single electron, and Z2 describes its spin, i.e., h i (1.2) Hel := AN L2 (R3 × Z2 )N , with AN being the orthogonal projection onto the subspace of totally antisymmetric wave functions, as required by the Pauli principle. The one-photon Hilbert space is given by L2 [R3 × Z2 ], where R3 is the photon momentum space and Z2 describes the two independent transversal polarizations of a photon. (Here and above, the integration measure on R3 is Lebesgue measure.) The photon Fock space is then defined by F :=
∞ M
h i Sn L2 (R3 × Z2 )n ,
(1.3)
n=0
where Sn is the orthogonal projection onto the subspace of totally symmetric n-photon wave functions, in accordance with the fact that photons are bosons. It is convenient to represent the Hilbert space H as the space of antisymmetric, square-integrable wave functions on the N -electron configuration space with values in the photon Fock space F, i.e., i h (1.4) H ∼ = AN L2 (R3 × Z2 )N ; F . The dynamics of the system is generated by the Hamiltonian Hα0 :=
i2 + Vc (x) + Hf , σ j · −i∇ x j − 2π 1/2 α 3/2 Aκ (αx j )
N h X
(1.5)
j =1
where we use units in which h¯ = 1 and the electron mass equals 1/2. In (1.5), σ j = y (σjx , σj , σjz ) denotes the three Pauli matrices associated with the j th electron, xj is its position (in suitable units of length described below), α is the fine structure constant, and Aκ (x) denotes the quantized vector potential of the transverse modes of the electromagnetic field in the Coulomb gauge, i.e., o X Z d 3 k κ(|k|/K) n ελ (k)e−ik·x aλ∗ (k) + ελ (k)∗ eik·x aλ (k) , (1.6) Aκ (x) := √ π 2ω(k) λ=1,2 where κ is an entire function of rapid decrease on the real line, e.g., κ(r) := exp(−r 4 ), cutting off the vector potential in the ultraviolet domain, ω(k) := |k| is the frequency
Spectral Analysis for Systems of Atoms and Molecules
251
of a photon with wave vector k, and ελ (k), λ = 1, 2, are photon polarization vectors satisfying ε λ (k)∗ · εµ (k) = δλµ ,
k · ελ (k) = 0,
for λ, µ = 1, 2.
(1.7)
Moreover, aλ∗ (k), aλ (k) are standard creation- and annihilation operators (see, e.g., [28]) on F obeying the canonical commutation relations [aλ# (k 1 ), aµ# (k 2 )] = 0, [aλ (k 1 ), aµ∗ (k 2 )] = δλµ δ(k 1 − k 2 ),
(1.8)
where a # = a or a ∗ . These objects are densely defined, operator-valued tempered distributions on Fock space F. Fock space F contains a vector , the vacuum vector, uniquely determined, up to a phase, by the properties that kk = 1 and aλ (k) = 0, for all λ and k. A dense set of vectors in F is obtained by applying polynomials in aλ∗ (λ = 1, 2), smeared out with test functions, to the vacuum vector . The term Vc on the right side of Eq. (1.5) defining the Hamiltonian is the Coulomb potential describing the electrostatic interactions between electrons and nuclei. In our units, it is given by Vc (x) :=
M N X X j =1 m=1
−Zm + |x j − R m |
X 1≤i<j ≤N
1 , |x i − x j |
(1.9)
where x ≡ (x 1 , . . . , x N ) ∈ R3N , and R 1 , . . . , R M ∈ R3 are the positions of M static nuclei with atomic numbers Z1 , . . . , ZM . Finally, the Hamiltonian of the free quantized electromagnetic field, Hf , is given by X Z (1.10) d 3 k aλ∗ (k) ω(k) aλ (k). Hf := λ=1,2
The r.s. of Eq. (1.10) defines a densely defined, positive, self-adjoint operator on F with absolutely continuous spectrum on the positive half-axis and a simple eigenvalue at 0 corresponding to the eigenvector . Next, we comment on the units chosen in (1.5). Length (and hence the positions, x j , of the electrons and, R m , of the nuclei) is measured in units of half a Bohr radius, h¯ 2 1 2 rBohr = 2mel e2 , expressed in Gaussian units, where mel is the mass of an electron and −e its electric charge. Photon wave lengths are measured in units of α times half a Bohr radius, α2 rBohr , i.e., the unit for photon wave vectors is 2α −1 rBohr −1 . The energy unit 2
. The ultraviolet cutoff of the radiation is chosen to be 4 Rydberg, with 4 Ry = r2e Bohr field imposed by the function κ(|k|/K) used in the definition of the vector potential Aκ (x) turns off interactions between electrons and photons with energy large compared to K · 4 Ry. The physical value of the fine structure constant α is ≈ 1/137. In this paper, α plays the rôle of a small, dimensionless number. Our results hold for sufficiently small values of α. We shall not verify that the radii of convergence in α of our analytical constructions cover the physical point α = α(phys) ≈ 1/137 (such a verification would presumably require numerical work on a computer). Our main concern in this paper is to analyze properties of the energy spectrum of the Hamiltonian Hα0 and to study resonances for the dynamics generated by Hα0 and estimate their lifetimes. We shall show that, for sufficiently small values of α > 0, Hα0 has a normalizable ground state corresponding to the minimum of its spectrum and
252
V. Bach, J. Fröhlich, I. M. Sigal
that “most” of its spectrum is absolutely continuous. Furthermore, we shall show that the excited bound states of the atom or ion (for a purely technical reason, we exclude molecules here) with electrons decoupled from the radiation field, i.e., the bound states of H0 , turn into metastable states of finite lifetime when that coupling is turned on. We show that Fermi’s golden rule [27] yields an accurate estimate of the lifetimes of metastable states, and we provide a rigorous justification of Bethe’s formula [11,10] for the Lamb shift. Notice that in our proof of a similar statement in [5,6], we required that A(x) be replaced by a confined vector potential κ(|x|/R)Aκ (x), Aκ (x) −→ κ(|x|/R) Aκ (x),
(1.11)
where 1 R < ∞ is large on the atomic length scale, but finite. We do not make this assumption in the present paper, anymore. The standard model of nonrelativistic electrons interacting with the quantized radiation field, cutoff in the ultraviolet, involves passing to the limit R → ∞ in (1.11). In this limit, certain terms in the Hamiltonian Hα0 that describe the coupling of electrons to the quantized radiation field turn out to be what in renormalization jargon is called a marginal perturbation. This makes the problem analyzed in the present paper more subtle and complicated, analytically, than the ones considered in [6,7], where (1.11) was assumed. Because of these analytical complications, we shall not carry out a full-fledged renormalization group analysis of ground states and resonances in the model introduced in (1.5)–(1.10). Instead, we use softer and less quantitative methods to establish existence of ground states, and we do not analyze the exact nature of the atom. The price to be paid is that we lose control over the precise nature of the spectrum of Hα0 in an interval of length O(α 3 ) above the ground state energy, Eo (α), and we do not present a convergent algorithm to calculate E0 (α) and the exact resonance energies to arbitrary accuracy, but merely derive expressions for these quantities which are asymptotic to order O(α 3 ). In this sense, the results of this paper are less precise than the ones in [6, 7]. But, on the positive side, they do not require the full power of the renormalization group methods [7]. In order to describe our results and methods more precisely, we start with a discussion of the spectral properties of the unperturbed Hamiltonian (electrons decoupled from the radiation field), 0 = Hel + Hf , H0 ≡ Hα=0
(1.12)
where Hel is the usual atomic (or molecular) Hamiltonian defined by Hel :=
N X
−1x j + Vc (x).
(1.13)
j =1
We recall some key properties of Hel , viewed as an operator on Hel . For details and proofs, see [27,14] and references given there. We first note that the potential Vc is a P perturbation of the kinetic energy operator 1x := N j =1 1x j with zero relative bound. Thus Hel is a semibounded, self-adjoint operator on the domain D(Hel ) = D(1x ) = Hel ∩ H 2 [(R3 × Z2 )N ], where H 2 is the usual Sobolev space. The essential spectrum of Hel is given by σess (Hel ) = [6, ∞), where 6 is the infimum of the spectrum of Hel , with N replaced by N − 1, as follows from the HVZ theorem (see, e.g., [14]). Thus, 6 is the ionization threshold.
Spectral Analysis for Systems of Atoms and Molecules gr. states
253
(abs.) cont. spectrum exc. states
E0
E1
E2
···
6
Fig. 1. The Spectrum of Hel
In what follows, we shall make the following assumption about the atom, ion, or molecule under consideration, represented by the parameters N, R 1 , . . . , R M , and Z1 , . . . , ZM . For the proof of existence of a ground state of Hα0 in Sect. 2, we assume that Hel has at least one eigenvalue E0 below the ionization threshold 6, i.e., we require that E0 := inf σ (Hel ) < 6 := inf σess (Hel ).
(1.14)
The contents of Sect. 3 on the life-times of metastable states is non-trivial only if, besides E0 , Hel has at least one further eigenvalue E1 , with E0 < E1 < 6, i.e., in Sect. 3 we additionally require that the spectrum σ (Hel ) = {E0 , E1 , . . . } ∪ [6, ∞)
(1.15)
of Hel consists of eigenvalues E0 < E1 < E2 < · · · ≤ 6, of finite multiplicity below 6, possibly with an accumulation point at 6 and (absolutely) continuous spectrum in [6, ∞); see Fig. 1. In Sect. 3 we shall also require an assumption saying, roughly speaking, that there are no accidental selection rules in the system described by Hel , which could prevent excited eigenstates of Hel from decaying radiatively (in 2nd order in the relevant coupling constant). For positive ions and (neutral) atoms or molecules our assumption (1.15) is justified, P as it is known that if N ≤ M m=1 Zm then Hel has infinitely many eigenvalues of finite multiplicity below 6 [36]. In the case of negatively ionized atoms or molecules, i.e., if P N> M has isolated eigenvalues below the ionization m=1 Zm , the question, whether HelP threshold, is more subtle. Indeed, if N ≥ M m=1 (2Zm + 1) then Hel has no eigenvalues at all [26] (see [30,31] for earlier results). Assuming that the atomic numbers of the nuclei are such that E0 < 6, i.e., that Hel has isolated eigenvalues below the ionization threshold, it tends to be an open question whether the ground state corresponding to the energy E0 is unique or not, except when N = 1 or = 2, in which case a standard Perron–Frobenius argument proves uniqueness. (Non-uniqueness for N ≥ 3 may arise as a consequence of the Pauli principle.) The spectrum of the photon Hamiltonian Hf consists of a simple eigenvalue at 0, corresponding to the vacuum vector ∈ F, and absolutely continuous spectrum (of infinite multiplicity) covering the half-axis [0, ∞). Consequently, by separation of variables, the unperturbed Hamiltonian H0 = Hel + Hf on F has spectrum σ (H0 ) = σ (Hel ) + σ (Hf ).
(1.16)
The point spectrum of H0 is the same as the point spectrum of Hel , i.e., it consists of the eigenvalues {Ej }j =0,1,2,... (corresponding to the eigenvectors ϕj,` ⊗ , where {ϕj,` }`=1,2,... ,nj is an orthonormal basis of eigenvectors of Hel corresponding to the eigenvalue Ej of multiplicity nj ). The continuous spectrum of H0 covers the half-axis
254
V. Bach, J. Fröhlich, I. M. Sigal
E0
E1
E2
···
6
Fig. 2. The Spectrum of H0 = Hel ⊗ 1 + 1 ⊗ Hf
[E0 , ∞) and consists of a union of branches [Ej , ∞) starting at the eigenvalues Ej and the branch [6, ∞); as indicated in Fig. 2. We note that the ground state energy E0 = inf σ (Hel ) of the atom or molecule in the absence of the quantized radiation field coincides with the ground state energy E0 ≡ E0 (α = 0) = inf σ (H0 ) of the system of an atom or molecule in the presence of photons, but decoupled from them. But, while E0 is an isolated eigenvalue of Hel , it lies at the tip of a branch of continuous spectrum of the Hamiltonian H0 . Similarly, the energies E1 , E2 , . . . are isolated eigenvalues of Hel ; but they are eigenvalues of H0 imbedded in continuous spectrum of H0 , and each Ej is the threshold of a branch of continuous spectrum of H0 . These spectral properties of H0 make it a difficult problem to analyze, mathematically rigorously, the fate of the eigenvalues Ej of H0 , and the nature of the energy spectrum of the interacting system described by the Hamiltonian Hα0 introduced in (1.5), for α > 0. Although the perturbation, Wα0 := Hα0 − H0 , is a small perturbation of H0 , general analytical methods to deal with this type of problem in perturbation theory do not appear to be available. In [6,7], we have started to develop such methods, tailormade to analyze a class of Hamiltonians describing interactions between nonrelativistic quantum-mechanical matter and the radiation field. In this paper we extend those methods to the Hamiltonian Hα0 (= H0 +Wα0 ) of Eq. (1.5), which describes much of the physics of light atoms or molecules interacting with the quantized electromagnetic field (within the Born–Oppenheimer approximation). For background material concerning the physics described by Hα0 , see e.g., [12,13] and references given there. 1.2. Survey of main results. In the next section, 1.3, we consider the structure and properties of the perturbation Wα0 in the Hamiltonian Hα0 = H0 + Wα0 of (1.5). The strength of the perturbation Wα0 relative to H0 is measured by the parameter g := (α K)3/2 ,
(1.17)
where K is the ultraviolet cutoff in the electromagnetic vector potential A, as described in Eq. (1.6). The parameter K is a “dimensionless energy scale” given by the photon energy above which interactions between electrons and the radiation field are cut off, divided by 4 Ry. In Sect. 1.3 we prove that if g is sufficiently small then the interaction Wg is bounded by H0 , in the sense of Kato [25,28], with relative bound strictly smaller than 1. This proves that, for small α, Hα0 is bounded from below and self-adjoint on the domain of H0 ; see Corollary 1.2. (Under somewhat weaker assumptions one can prove that Hα0 defines a semibounded quadratic form on an appropriate core. It is not easy, however, to characterize the domain of the corresponding self-adjoint operator.) Since g := (α K)3/2 is the relevant coupling parameter, we henceforth write 0 , Hg := Hα=g 2/3 /K
0 Wg := Wα=g 2/3 /K .
(1.18)
Spectral Analysis for Systems of Atoms and Molecules
255
We should emphasize that our bounds on g = (α K)3/2 become poor, as the number of electrons, N, becomes large. This does not mean that the Hamiltonian Hg is ill-defined or unbounded from below, for large values of α, as long as the ultraviolet cutoff K is kept finite. In fact, using path-space methods one can rigorously construct the semi-group exp(−tHg ), for t ≥ 0, and prove that it is self-adjoint, strongly continuous in t > 0 with exp(−tHg ) → 1, as t → 0, for arbitrary values of α and N, as long as K < ∞ (see [18,33] for various ingredients of the proof). However, for the purposes of the analysis presented in Sect. 3, Kato- (or quadratic-form-) boundedness of Wg relative to H0 is an important property. In Sect. 2 we address the question whether Hg has a ground state corresponding to an eigenvalue E0 (g) := inf σ (Hg ). Our main result is Theorem 2.1, which answers this question in the affirmative, as long as the effective coupling constant g is sufficiently small. We remark that the magnitude of the integrand in the definition (1.6) of A(x) is ∼ |k|−1/2 , as k → 0. Hence we are considering a critical case (in the infrared) for which the magnitude of the integrand divided by ω is not square-integrable at k = 0. In case of such a critical behaviour of the coupling function and under the replacement of the photons and the atom by a scalar massless field and a matrix of finite rank, respectively, it is known that ground states (with finite photon number expectation) may not exist, in general [34,1]. This indicates that we are dealing with a borderline case and that the transversality of the photons is crucial for the existence of ground states shown in Theorem 2.1. Indeed, our proof makes explicit use of this property. The method used in Sect. 2 is “non-perturbative” but non-constructive. For this reason, it does not enable us to estimate the multiplicity of the ground state energy E0 (g). However, if the number N of electrons is = 1 or = 2, and if the Zeeman terms in the Hamiltonian Hg are set to 0, then we can construct an L2 -space representation of the photon Fock space F (“electric field” representation) with the property that exp(−tHg ) is positivity-preserving on Hel ⊗F, where Hel is taken in the usual Schrödinger configuration-space representation, and, for N = 2, only spin-singlet states are considered. The ergodicity (positivity-improvement) of exp(−tHg ) (see [27]), which is established in [19], and a Perron–Frobenius argument then establishes uniqueness of the groundstate. It is worthwhile to point out some (fairly standard, but) fundamental consequences of the existence of a ground state of Hg for scattering theory: Using straightforward variants of methods developed in [20–22,3,4], one can construct Möller (wave) operators, ± , such that the range of + (− ) describes asymptotic states of the system consisting of an atom (or molecule) in a ground state accompanied by an outgoing cloud of freely moving photons. The obvious conjecture is that Ran{+ } = Ran{− } =: Hasy , where (0) (0) Hasy is isomorphic to Hg ⊗ F, and where Hg is the space of ground states of Hg . This conjecture, called “asymptotic completeness”, would imply the unitarity of the scattering matrix for the scattering of photons off an atom or molecule below the ionization threshold. We are miles away from proving this conjecture! But in the simplified models of massive photons and confined electrons, it has recently been proven in [17, 15], and for massless photons coupled to confined electrons in the dipole approximation, asymptotic completeness has recently been established in [35]. In Sect. 3, we prove that, outside small neighborhoods of E0 and 6 and below 6, the spectrum of Hg is purely absolutely continuous. (Outside small neighborhoods of the thresholds of Hel and above 6, the spectrum of Hg can be shown to be purely absolutely continuous by using the methods in [6,8].) In particular, Hg does not have any eigenvalues in the vicinity of the imbedded eigenvalues E1 < E2 < · · · < En < 6 − δ,
256
V. Bach, J. Fröhlich, I. M. Sigal E0
E1
E2
···
6
Im(θ) 0 (θ, z) onto the energy plane Fig. 3. A projection of the Riemann surface of z 7 → Fψ,ϕ
of H0 , where δ is a small positive constant depending on g and N. We will, however, make the idea rather precise that imbedded eigenvalues of H0 give rise to metastable states of Hg , and we shall estimate the lifetime, ∝ g 2 , of these metastable states up to an error term O(g 2+ε ), for some ε > 0. Our notion of resonance is based on dilatation analyticity. In order to state our ideas simply, we assume that there is only one atomic nucleus (M = 1 in Eq. (1.9)) of atomic number Z = Z1 ≤ 2N located at R 1 = 0. (The general case of an arbitrary, finite number M of static nuclei can be studied, too, by borrowing ideas developed in [23]. It would be interesting to extend our results to systems of electrons and dynamical nuclei, all coupled to the quantized radiation field. This would require a considerable amount of additional work and will not be done in the present paper.) To describe dilatation analyticity, we start by scaling the positions, x j , of the electrons and the momenta, k, of the photons by x j 7 → eθ x j ,
k 7 → e−θ k .
(1.19)
When θ is real, the transformations (1.19) determine a unitary transformation Uθ on the Hilbert space H defined in Eqs. (1.1)–(1.4). It is easy to see that the subspace, D ⊆ H, of vectors, ψ, with the property that ψ(θ) := Uθ ψ is analytic in θ , for |Imθ | < π/2, is dense in H. Furthermore, one easily checks that Uθ Hf Uθ∗ = e−θ Hf .
(1.20)
These facts, combined with well-known results [2,9] on dilatation analyticity for Schrödinger operators, show that, for arbitrary ψ, ϕ ∈ D, the function
−1 0 (θ, z) := ψθ¯ z − H0 (θ ) ϕ(θ ) , (1.21) Fψ,ϕ where H0 (θ ) = Uθ H0 Uθ−1 , is independent of θ , for |Imθ | < π/2, and, for Imθ =: ϑ 0 (θ, z) is analytic in z in the complement of the shaded region depicted in fixed, Fψ,ϕ Fig. 3. In Sect. 3, we construct the function
−1 g (1.22) Fψ,ϕ (θ, z) := ψθ¯ z − Hg (θ ) ϕ(θ ) , where Hg (θ ) = Uθ Hg Uθ−1 , with Hg as in (1.5), and we show that, for our choice of an g ultraviolet cutoff κ (see (1.6) and below), and for arbitrary ψ and ϕ in D, Fψ,ϕ (θ, z) is g independent of θ, for |Imθ| < π/4 small enough, and, for Imθ =: ϑ fixed, Fψ,ϕ (θ, z) is analytic in z in the complement of the shaded region depicted in Fig. 4. It thus provides g an analytic continuation of the matrix element Fψ,ϕ (0, z) of the resolvent of Hg − z in z from the upper half plane into the lower half plane outside the spectrum of Hg (θ ).
Spectral Analysis for Systems of Atoms and Molecules E0
E1
···
257 Ej
E0 (g)
Ej,` (g) E1 (g) g
Fig. 4. A projection of the Riemann surface of z 7 → Fψ,ϕ (θ, z) onto the energy plane
This implies the absolute continuity of the spectrum of Hg for those energies which are contained in the resolvent set of Hg (θ) (see Corollary 3.3). In Sect. 3, Eqs. (3.8)–(3.14), we introduce a notion of resonance energy Ej (g) corresponding to the energy Ej of the j th excited state of the atom or molecule. We show that the Lamb shift, Re{Ej (g) − Ej }, is given by Bethe’s formula and Im{Ej (g)} is given by Fermi’s Golden Rule, to order g 2 , with error terms that we prove to be O(g 2+ε ), for any 0 < ε < 1/3. If the decay rate for the j th excited state of the atom is non-zero in second order perturbation theory then the lifetime of an initial state close to the j th excited state of the uncoupled atom is shown to be finite. Its inverse is shown to be given by Im{Ej (g)} = 0j g 2 + O(g 2+ε ),
(1.23)
with 0j strictly positive. We remark again that, in contrast to [6, Sect. IV], where similar results have been obtained, we do not require the assumption of confinement, see Eq. (1.11), in the derivation of (1.23), anymore. In [7], by assuming confinement and using a more elaborate renormalization group analysis, we have given a convergent algorithm to compute Im{Ej (g)}, for small values of g > 0. It then follows by standard reasoning that, for g > 0 sufficiently small, the spectrum of Hg is purely absolutely continuous in a neighborhood of every eigenvalue Ej , j ≥ 1, of Hel for which 0j > 0. What we are really looking after, from a physics point of view, is a precise understanding of the decay of the excited states of the atom under the time evolution exp[−it Hg ], as t becomes large. Let δ := dist Ej , σ (Hel ) \ {Ej } > 0, set Ij (δ/2) := (Ej −δ/2, Ej +δ/2), and notice that Ij (δ/2) is an open interval containing Ej and such that dist Ij , σ (Hel ) \ {Ej } = δ/2. Let Fj denote a smooth characteristic function of Ij (δ/2) (see Subsection 3.1). We shall identify a “j th excited state” of the atom with a vector of the form 1/2 (1.24) 9j := Fj (Hg ) exp −g −2 Hf ϕj,` ⊗ η, atomic state to order g 2 described in (3.26) below, and where ϕj,` is an effective excited −2 the operator exp −g Hf essentially eliminates high-energy photons of energy larger than g 2 in the state η ∈ F, which is assumed to be dilatation analytic. An example for such a state is given by η = . We then show that, for any 0 < ε < 1/3, there exists a constant C ≥ 0 and, for any L ∈ N, a constant CL ≥ 0 such that D E 9j exp[−itHg ] 9j i h (1.25) ≤ Bη2 C exp −t g 2 0j − Cg 2+ε ) + CL t −L g 4 ,
258
V. Bach, J. Fröhlich, I. M. Sigal
where Bη := sup|θ|≤θ0 kη(θ)k. This estimate implies that, given ε > 0, there is a finite constant Cε > 0 such that, for t > Cε 0j−1 g −2 , −itH g 9 ≤ ε. 9j e (1.26) j We remark that it is known from other methods (see, e.g., [16]) that, given ε > 0, there is a constant Dε > 0 such that, for 0 ≤ t < Dε g −2 , −itH g 9 ≥ ε. 9j e (1.27) j Estimate (1.26) shows that the state ψj decays, with a life-time bounded above and below by const · g −2 . It is a typical example for the kind of estimates we are able to prove with the help of the methods developed in Sect. 3. We point out that, while the general idea of our method is borrowed from [24], the details of our analysis are more complicated because resonances do not appear as isolated poles of the meromorphic g continuation of Fψ,ϕ (θ, z) in (1.22) onto the second sheet. 1.3. Relative bounds, self-adjointness, and dilatation analyticity. We return to Eq. (1.5), which we write as Hg = H0 + Wg ,
(1.28)
where H0 is defined in (1.12), and we obtain Wg =
N X j =1
−4π 1/2 α 3/2 Aκ (αx j ) · (−i∇ x j ) + 4π α 3 A2κ (αx j ) + 2π 1/2 α 5/2 σ j · ∇ × Aκ (αx j )
(1.29)
α=g 2/3 /K
from expanding the square in (1.5). Our first goal in this subsection is to prove in Corollary 1.2 that Wg is defined on D(H0 ) and obeys the bound
Wg |H0 + iC(N)|−1 ≤ C 0 (N ) (α K)3/2 , (1.30) for some constants C(N), C 0 (N) ≥ 0. This establishes the semiboundedness and selfadjointness of Hg on D(H0 ) for g := (α K)3/2 < C 0 (N )−1 . Our second goal is to establish the dilatation analyticity of Wg (θ) := Uθ Wg Uθ−1 (see Corollary 1.3), where Uθ is the dilatation operator defined in (1.19), i.e., we prove that θ 7 → Wg (θ ) is an analytic function on D(0, θ0 ) := {z ∈ C||z| < θ0 } with values in B(D(H0 ); H), the bounded operators from D(H0 ) to H, for some θ0 > 0. We establish this property by observing that the coupling functions in Wg (θ) are analytic in θ , pointwise in the other parameters, and by verifying a bound similar to (1.30), namely,
Wg (θ) |H0 + iC(N, θ)|−1 ≤ C 0 (N, θ ) g, (1.31) for some constants C(N, θ), C 0 (N, θ) ≥ 0. In fact, Eq. (1.30) is just the special case θ = 0 in Eq. (1.31). Finally, we establish the dilatation analyticity of Hg (θ ) := Uθ Hg Uθ−1 in Corollary 1.4, assuming that Hel = −1x + Vc (x) is the Hamiltonian of an atom, i.e., M = 1. This simplifying assumption could be avoided by using exterior dilatations
Spectral Analysis for Systems of Atoms and Molecules
259
[23,32], at the expense of having to deal with more involved estimates; we do not carry out this analysis here. We characterize the coupling functions in Wg (θ ) in terms of the following functions, √ −θ/2 2e κ(e−θ |k|/K) p exp[−iαk · x] ελ (k) Gx (k, λ; θ) := π K 3 ω(k)
(1.32)
and B x (k, λ; θ ) :=
α
√ −3θ/2 2e κ(e−θ |k|/K) p exp[−iαk · x] k × ελ (k) , i π K 3 ω(k)
(1.33)
where we introduce a dilatation parameter θ ∈ D(0, θ0 ) ⊆ C, for some θ0 > 0 sufficiently small. Note that, pointwise for every x, k ∈ R3 and λ ∈ Z2 , the maps θ 7 → Gx (k, λ; θ) and θ 7 → B x (k, λ; θ) are analytic in D(0, θ0 ). Furthermore, we notice that B x (k, λ; 0) = ∇ × Gx (k, λ; 0). By means of Gx (k, λ; θ ) and B x (k, λ; θ ) we define the following functions on R3 × Z2 with values in the operators on Hel , o −2Gx j (k, λ; θ ) · pj + σ j · B x j (k, λ; θ ) ,
N n X
¯ ∗ := w1,0 (k, λ; θ ) := w0,1 (k, λ; θ)
j =1
(1.34) w2,0 (k 1 , λ1 ; k 2 , λ2 ; θ) := w0,2 (k 1 , λ1 ; k 2 , λ2 ; θ¯ )∗ N n o X := Gx j (k 1 , λ1 ; θ ) · Gx j (k 2 , λ2 ; θ ) ,
(1.35)
j =1
w1,1 (k 1 , λ1 ; k 2 , λ2 ; θ) :=
N n X
Gx j (k 1 , λ1 ; θ¯ )∗ · Gx j (k 2 , λ2 ; θ )
j =1
(1.36)
o + Gx j (k 1 , λ1 ; θ ) · Gx j (k 2 , λ2 ; θ¯ )∗ ,
and these, in turn, serve as coupling functions for the operators defined by Wm,n (θ ) :=
X
Z d 3 k1 · · · d 3 km+n wm,n (k 1 , λ1 ; . . . , k m+n , λm+n ; θ )
λ1 ,... ,λm+n =1,2 aλ∗1 (k 1 ) · · · aλ∗m (k m )
aλm+1 (k m+1 ) · · · aλm+n (k m+n ).
(1.37)
Then we observe that, after normal ordering, the (dilated) interaction Wg (θ ) reads Wg (θ) =
X m+n≤2
g m+n Wmn (θ ) + g 2 Cno ,
(1.38)
260
V. Bach, J. Fröhlich, I. M. Sigal
R where Cno is defined by Cno := 2N |Gx (k, 1; 0)|2 d 3 k, which is independent of x, and g := (α K)3/2 . Henceforth and consistent with our previous definitions, we omit θ in our notation in the undilated case, θ = 0, writing B x (k, λ) := B x (k, λ; 0), Gx (k, λ) := Gx (k, λ; 0) , wm,n (k 1 , λ1 ; . . . , k m+n , λm+n ) := wm,n (k 1 , λ1 ; . . . , k m+n , λm+n ; 0), Wm,n := Wm,n (0).
(1.39) (1.40) (1.41)
Next, we define J : R3 → R+ to be the smallest function such that
w1,0 (k, λ; θ) (−1x + 1)−1/2 , w0,1 (k, λ; θ ) (−1x + 1)−1/2 ≤ J (k) (1.42) holds, for all |θ| ≤ θ0 and (k, λ) ∈ R3 × Z2 , and such that
w2,0 (k 1 , λ1 ; k 2 , λ2 ; θ) , w1,1 (k 1 , λ1 ; k 2 , λ2 ; θ ) , (1.43)
w0,2 (k 1 , λ1 ; k 2 , λ2 ; θ ) ≤ J (k 1 ) J (k 2 ) holds, for all |θ| ≤ θ0 and (k 1 , λ1 ), (k 2 , λ2 ) ∈ R3 × Z2 . Note that, due to ±i∇ x j ≤ (−1x + 1)1/2 , we have that J (k) ≤ C(θ0 ) N K −3/2 |k|−1/2 κ(e−θ |k|/K) + α|k| κ(e−θ |k|/K) , (1.44) for some constant C(θ0 ) > 0. The rapid decay of κ implies that Z 1/2 J (k)2 ω(k)β d 3 k < ∞, 3β :=
(1.45)
for any β > −2. In particular, 3β is uniform in K ≥ 1, for any 1 ≥ β > −2. This uniformity in K ≥ 1 is actually the basic requirement that determines p = 3/2 in the coupling parameter g = α 3/2 K p . The main relative bound that we use is described in the following lemma. Lemma 1.1. For all m, n ∈ N0 with 1 ≤ m + n ≤ 2 and all θ ∈ D(0, θ0 ), the operators Wm,n (θ ) are defined on D(H0 ) and obey the bound
Wm,n (θ) |H0 + iC(N, θ0 )|−1 ≤ 4(1 + 32 + 32 ) (m+n)/2 , (1.46) 0 −1 for some constant C(N, θ) ∈ R+ . Proof. We first note that the canonical commutation relations (1.8) allow us to convert ¯ Indeed, for any ψ ∈ D(H0 ), estimates on W1,0 (θ) into those for W0,1 (θ). X Z ¯ ψk2 + d 3 k kw1,0 (k, λ; θ )ψk2 kW1,0 (θ) ψk2 = kW0,1 (θ) λ=1,2
¯ ψk + 4 320 (−1x + 1)1/2 ψ 2 . ≤ kW0,1 (θ) 2
Furthermore, kW0,1 (θ) ψk ≤
X Z
(1.47)
d 3 k J (k) (−1x + 1)1/2 aλ (k)ψ
λ=1,2
1/2 ≤ 2 3−1 (−1x + 1)1/2 Hf ψ ,
(1.48)
Spectral Analysis for Systems of Atoms and Molecules
and hence kW1,0 (θ ) ψk,
261
kW0,1 (θ) ψk ≤ 2 (30 + 3−1 ) (−1x + Hf + 1)ψ .
(1.49)
Similarly, we convert estimates on W2,0 and W1,1 into those for W0,2 . For W2,0 , for instance, the canonical commutation relations (1.8) imply that ¯ ∗ W0,2 (θ) ¯ (1.50) W2,0 (θ )∗ W2,0 (θ) = W0,2 (θ) Z + 2 dξ1 dξ2 w2,0 (ξ1 ; ξ2 ; θ )∗ w2,0 (ξ1 ; ξ2 ; θ ) Z + 4 dξ1 dξ2 dξ3 w2,0 (ξ1 ; ξ3 ; θ )∗ w2,0 (ξ2 ; ξ3 ; θ ) a ∗ (ξ1 )a(ξ2 ), where we denoted ξ := (k, λ),
R
dξ :=
P
R λ=1,2
d 3 k, and a # (ξ ) := aλ# (k). This yields
¯ ψk + 4320 kψk, kW2,0 (θ) ψk ≤ 3kW0,2 (θ)
(1.51)
and, as in (1.48), we obtain that
kW0,2 (θ) ψk ≤ 32−1 kHf ψk ≤ 32−1 (−1x + Hf + 1)ψ .
(1.52)
The bound (1.46) follows from (1.49), (1.51), (1.52), and the additional use of the fact t that −1x is relatively Hel -bounded with relative bound arbitrarily close to 1. u Now self-adjointness of Hg on D(H0 ) and dilatation analyticity of Wg (θ ) are just two immediate implications of Lemma 1.1. Corollary 1.2. If 0 < g = (α K)3/2 < (1 + 320 + 32−1 )−1/2 /10, then Hg is self-adjoint and semibounded on D(H0 ). Corollary 1.3. The map Wg : D(0, θ0 ) → B(D(H0 ); H), θ 7→ Wg (θ ) is analytic. Finally, we establish the dilatation analyticity of Hg (θ ) := Uθ Hg Uθ−1 , assuming that the potential Vc (x) is dilatation analytic, i.e., D(0, θ0 ) 3 θ → Vc (eθ x) ∈ B(D(H0 ); H) is an analytic function. This property holds in case that Hel = −1x + Vc (x) is the Hamiltonian of an atom, i.e., M = 1, for arbitrary θ . Indeed, in the atomic case we may choose without loss of generality the position of the nucleus to be the origin of the one-electron configuration space, and then we obtain Hel (θ) := Uθ Hel Uθ−1 = e−2θ −1x ) + e−θ Vc (x).
(1.53)
H0 (θ) := Uθ H0 Uθ−1 = Hel (θ ) + e−θ Hf
(1.54)
Therefore,
is an analytic family of type A and, by Corollary 1.3, so is Hg (θ ) = H0 (θ ) + Wg (θ ). We summarize this discussion and a simple consequence of (1.54) in the following corollary. Corollary 1.4. The family {Hg (θ)|θ ∈ D(0, θ0 )} is dilatation analytic, i.e., the map Hg : D(0, θ0 ) → B(D(H0 ); H), θ 7 → Hg (θ ) is analytic. Moreover, there exists a constant b ∈ R+ such that
1Hel (θ) (Hel ± i)−1 ≤ b |θ |, (1.55) where 1Hel (θ) := Hel (θ) − Hel , for all θ ∈ D(0, θ0 ).
262
V. Bach, J. Fröhlich, I. M. Sigal
2. Soft Photon Bound and Existence of a Ground State In this section we derive a new soft photon bound; see Inequalities (2.5)–(2.6) and Theorem 2.3 below. It is tailored for the minimal coupling model, and we use this bound to prove the existence of a ground state. Theorem 2.1. There exists a constant C(N, 6 − E0 ) ≥ 0 such that, for all 0 < g = (α K)3/2 ≤ C(N, 6 − E0 ), the Hamiltonian Hg has a ground state, i.e., E0 (α) := inf σ (Hg ) is an eigenvalue. T Proof. We introduce the notation a # (F ) := a # (F1 ), a # (F2 ), a # (F3 ) , a ∗ (F ) := R 3 R 3 P P ∗ ∗ λ=1,2 d k F (k, λ)aλ (k), and a(F ) := λ=1,2 d k F (k, λ) aλ (k). Here, F is a 3 function on R × Z2 with values in the operators on Hel such that
2 X Z
d 3 k F (k, λ)(−1x + 1)−1/2 < ∞. λ=1,2
We further denote 8(F ) := a ∗ (F )+a(F ), 8(F ) := a ∗ (F )+a(F ), and pj := −i∇ x j . In this notation, the interaction Wg reads Wg =
o −2g8(Gx j ) · p j + g 2 8(Gx j )2 + σ j · g8(B x j ) ,
N n X
(2.1)
j =1
where G and B are defined in Eqs. (1.32), (1.33), and (1.39). Next, we introduce an infrared regularization by switching off the interaction for photons of small momenta. Specifically, we pick a “photon mass”, m > 0, that is, we replace Gx (k, λ), B x (k, λ), and Wg in (1.32), (1.33), and (2.1) by (m)
Gx (k, λ) := χ [ω(k) ≥ m] Gx (k, λ),
(m) Wg
:=
(2.2)
(m) B x (k, λ) := χ [ω(k) ≥ m] B x (k, λ), n PN (m) (m) 2 2 j =1 −2g 8(Gx j ) · p j + g 8(Gx j ) + g σ j (m)
(m)
(m)
(m)
(2.3)
o
(m) · 8(B x j )
, (2.4)
and we denote Hg := H0 + Wg and E0 (g) := inf σ (Hg ). We remark that (m) (0) Hg → Hg = Hg in the norm resolvent sense, as m → 0. This easily follows from an estimate similar to Lemma 1.1 (see [6]). In Theorem 2.2 below we show that, for g = (m) (α K)3/2 sufficiently small, Hg has a ground state, φm , i.e., there exists a normalized (m) (m) solution of Hg φm = E0 (g)φm , for all m > 0. Since kφm k = 1, the family {φm }m>0 contains a weakly convergent subsequence, {φm(n) }n∈N , where limn→∞ m(n) = 0. We put φ0 := w − limn→∞ φm(n) . Then one easily shows [6], that φ0 ∈ D(Hg ) and that Hg φ0 = E0 (α)φ0 . To conclude, it remains to show that φ0 6= 0. To show that φ0 6 = 0, we employ a soft photon bound, as in [6]. There is an important difference, though. The soft photon R in [6] estimated the photon number expectation P bound hφm |Nf φm i, where Nf := λ=1,2 aλ∗ (k)aλ (k)d 3 k, in terms of supx kω−1 Gx k2 . It was derived from a virial type argument, using the commutator of aλ (k) and Hg . This bound does not directly apply to the present problem because kω−1 Gx k2 = ∞, for all x. Modifying the argument slightly by using the commutator of aλ (k) − iFx (k, λ) and
Spectral Analysis for Systems of Atoms and Molecules
263
Hg , for a suitably chosen Fx , we avoid the appearance of supx kω−1 Gx k2 on the right side of the estimate, which we trade for a factor of k |x|φm k. More precisely, in Theorem 2.3 below we show that
2
(2.5) hφm | Nf φm i ≤ C1 (N) g 2 (1 + |x|) φm , for some constant C1 (N) ≥ 0. In [6] we showed that φm is exponentially localized in the electron variables. More generally, there exists an ε > 0 such that
ε|x|
e χ1 (H (m) ) ≤ C2 < ∞, (2.6) g (m)
where χ1 (Hg ) is the spectral projection onto 1 := (−∞ , (6 + E0 )/2), provided that g 2 (1 + 6 − E0 ) is sufficiently small (with C2 and ε independent of g). Since (m) φm ∈ Ran{χ1 (Hg )}, this implies the boundedness of k |x| φm k. Thus there exists a constant, C3 ≡ C3 (N, 6 − E0 ), such that hφm | Nf φm i ≤ C3 g,
(2.7)
for all m > 0. Next, we introduce the projection Pel onto all bound states of Hel below 1 ⊥ 2 (6 + E0 ) < 6. Note that Pel has finite rank and that (6 − E0 )Pel ≤ 2(H0 − E0 ). The latter implies that (m)
hφm | Pel⊥ φm i ≤ 2(6 − E0 )−1 hφm | (E0 (g) − E0 − Wg(m) ) φm i ≤ C3 g 2 .
(2.8)
From (2.7)–(2.8) and Pel P ≥ 1 − Pel⊥ − Nf we draw the important consequence that hφm | Pel P φm i ≥ 1 − 2C3 g 2 ,
(2.9)
where P := |ih| denotes the rank-one projection onto the photon vacuum vector . Thus, if g = (α K)3/2 is sufficiently small then hφm | Pel P φm i ≥ 1/2, for all m > 0. t Since Pel P has finite rank, it follows that φ0 6 = 0. u In the following theorem, we review the proof in [6] of the existence of a ground (m) state of Hg , for m > 0. Theorem 2.2. There exists a constant C ≡ C(N, 6 −E0 ) > 0 such that, for all 0 < g = (m) (m) (m) (α K)3/2 ≤ C, the Hamiltonian Hg has a ground state, i.e., E0 (g) := inf σ (Hg ) is an eigenvalue, for any m > 0. Proof. We only sketch the argument, see [6, Sect. II.2] for details. Alternatively, one may proceed as in [17]. The assertion is proven if we can find some m ˜ > 0 such that (m) (m) (m) ˜ is finite, i.e., Tr [Hg − the sum of the negative eigenvalues of Hg − E0 (g) − m (m) ˜ − > −∞, where the negative part of a real number λ is defined as E0 (g) − m] [λ]− := min{λ, 0}. To this end, we first employ a discretization. Given ε > 0 and a locally integrable function F , we define its ε-average by Z −3 F (k 0 , λ)d 3 k 0 , (2.10) hF (k)iε := ε n(k)+Qε
264
V. Bach, J. Fröhlich, I. M. Sigal
where Qε = [−ε/2, ε/2)3 and n(k) ∈ (εZ)3 is the integer part of k, i.e., k −n(k) ∈ Qε . (m,ε) (m) (m) , byE replacing Gx (k, λ) and B x (k, λ) We define the corresponding interaction, E D Wg D (m)
in (2.2)–(2.4) by Gx (k, λ) (1.46) it follows that
(m)
ε
and B x (k, λ) , respectively. Then, by the bound ε
±(Wg(m,ε) − Wg(m) ) ≤
(m) o(ε 0 ) (Hg(m) − E0 (g) + 1)1/2
1 + |x|
(2.11) (m) (Hg(m) − E0 (g) + 1)1/2 ,
where o(ε 0 ) denotes a function which possibly depends on g, N , κ, m and tends to zero as ε → 0. Here, our original manuscript contained a small mistake in that the factor 1 + |x| was missing, as was kindly pointed out to us by F. Hiroshima. Next, we define (m,ε) (m) by replacing ω(k) in Hf by hω(k)iε . Since |ω(k) − hω(k)iε | = O(ε) and Hf ω(k) ≥ m, we obtain that (m) (m,ε) (m) ≤ 1 + o(ε0 ) Hf . (2.12) 1 − o(ε 0 ) Hf ≤ Hf (m,ε)
Denoting Hg
(m,ε)
:= Hel + Hf
(m,ε)
+ Wg
, we hence obtain that
Hg(m) ≥ 1 − o(ε0 ) Hg(m,ε) (m)
(2.13) (m)
− o(ε 0 ) (Hg(m) − E0 (g) + 1)1/2 (1 + |x|) (Hg(m) − E0 (g) + 1)1/2 . e := (−∞ , E (m) (g) + m) ˜ and observe that Next, we introduce the interval 1 0 (m) (m) (m) (m) (m) ˜ − = χ1 − E0 (g) − m ˜ χ1 [Hg(m) − E0 (g) − m] e (Hg ) Hg e (Hg ).
(2.14)
˜ > 0 sufficiently Furthermore, we note that, thanks to (2.6) and χ1 e = χ1 e χ1 , for m small, we have
χe (H (m) ) (H (m) − E (m) (g) + 1)1/2 (2.15) g g 1 0
(m) (m) 1/2 (m) < ∞. 1 + |x| (Hg − E0 (g) + 1) χ1 e (Hg ) Thus, (m) ˜ − Tr [Hg(m) − E0 (g) − m] (m,ε) (m) (m,ε) (m) − E0 (g) − m ˜ − o(ε0 ) χ1 ≥ Tr χ1 e (Hg ) Hg e (Hg ) (m,ε) (m,ε) (m) (m) − E0 (g) − m ˜ − o(ε0 ) − χ1 ≥ Tr χ1 e (Hg ) Hg e (Hg ) (m,ε) ˜ , ≥ Tr Hg(m,ε) − E0 (g) − 3m/2 −
(2.16)
for ε > 0 sufficiently small, and the finiteness of the right side in (2.16) (and hence the (m,ε) claim) follows if we can show that, for any ε > 0, the discretized Hamiltonian Hg (m,ε) has only finitely many eigenvalues below E0 (g) + 3m/2. ˜ The key point of the discretization by means of the ε-average is the tensor product ⊥ ], where H 3 representation F ∼ = F[Hdisc ]⊗F[Hdisc disc is spanned by χQε +n , n ∈ (εZ)
Spectral Analysis for Systems of Atoms and Molecules
265 (m,ε)
(see [6] for details). Note that, with respect to this representation, we have Hf (m,ε) (m,ε) (m,ε) ∼ (m,ε) ⊗1+1⊗H , Wg ⊗ 1, and hence H = Wg f
∼ =
f
(m,ε) Hg(m,ε) ∼ = Hg(m,ε) ⊗ 1 + 1 ⊗ Hf (m,ε) ≥ E0 (g) + m ⊗ P⊥ + Hg(m,ε) ⊗ P ,
(2.17)
⊥ ] by P . Hence, for 3m denoting the projection onto the vacuum