Commun. Math. Phys. 194, 1 – 45 (1997)
Communications in
Mathematical Physics © Springer-Verlag 1998
¨ Modified Prufer and EFGP Transforms and the Spectral Analysis of One-Dimensional Schr¨odinger Operators Alexander Kiselev1,? , Yoram Last2 , Barry Simon2,?? 1
Department of Mathematics, University of Chicago, Chicago, IL 60637, USA Division of Physics, Mathematics, and Astronomy, California Institute of Technology 253-37, Pasadena, CA 91125, USA
2
Received: 8 April 1997 / Accepted: 19 June 1997
Abstract: Using control of the growth of the transfer matrices, we discuss the spectral analysis of continuum and discrete half-line Schr¨odinger P∞ operators with slowly decaying potentials. Among our results we show if V (x) = n=1 an W (x − xn ), where W has 0, then H has purely a.c. (resp. purely s.c.) spectrum compact support xn /xn+1 →P P and a2n = ∞). For λn−1/2 an potentials, where an are on (0, ∞) if a2n < ∞ (resp. independent, identically distributed random variables with E(an ) = 0, E(a2n ) = 1, and λ < 2, we find singular continuous spectrum with explicitly computable fractional Hausdorff dimension.
1. Introduction In this paper, we will study continuum and discrete Schr¨odinger operators on the halfline (while we don’t always make them explicit, given theory in [10, 26, 32], many of our results extend to suitable whole-line problems). Explicitly, we are interested in the spectral analysis of operators H on L2 (0, ∞; dx) and on `2 ([1, ∞)) given by (Hu)(x) = −
d2 + V (x) dx2
(1.1)
in the continuum case and (Hu)(n) = u(n + 1) + u(n − 1) + V (n)u(n)
(1.2)
in the discrete case. ?
Research supported in part by NSF Grant No. DMS-9022140. This material is based upon work supported by the National Science Foundation under Grant No. DMS9401491. The Government has certain rights in this material. ??
2
A. Kiselev, Y. Last, B. Simon
Suitable boundary conditions are set at x (or n) = 0, so that H is self-adjoint (since in all our examples V is limit point at infinity, a boundary condition is not needed there). We are interested in the spectral properties of such operators in situations where |V (n)| → 0 as n → ∞, but so slowly that the usual scattering theory will not apply. Our main theme in this paper is that there are perturbation techniques of remarkable power for such operators based on two ideas. The first is that one can use the transfer matrix to analyze spectral properties. The transfer or fundamental matrix is a 2 × 2 unimodular matrix defined in the continuum case for any E by 0 a u (x) , (1.3) = TE (x, 0) u(x) b where −u00 + V u = Eu, u0 (0) = a, u(0) = b. In the discrete case a u(n + 1) = , TE (n, 0) b u(n)
(1.4)
where u(1) = a, u(0) = b, and u(n + 1) + u(n − 1) + V (n)u(n) = Eu(n). The second idea is that one can control the transfer matrix by controlling the norms of two solutions of −u00 + V u = Eu and that the critical equations are ones that involve those norms. Two of us [22] have recently found several new criteria for singular or absolutely continuous spectra in terms of transfer matrices, and these criteria will make some of our results here possible. The perturbation equations we will use have not been systematically used in this context except in the paper of Kotani-Ushiroya [21] whose techniques have some overlap with this paper. But they didn’t control the discrete case and their method is so entwined with certain Martingale inequalities that it is unclear how to use them in other contexts. While we were writing up the work for this paper, we received a preprint from Remling [29] that uses this two-pronged approach and has considerable overlap with our Sects. 5 and 6. We will discuss the connection shortly. Here are some of the theorems that we will use that relate spectral properties to behavior of T (n). The first is from [22]: Theorem 1.1. Suppose that there is a fixed sequence ni → ∞ and S is a subset of R so that for a.e. E ∈ S, limi→∞ kTE (ni , 0)k = ∞. Then µac (S) = 0, where µac is the absolutely continuous part of the spectral measure for H. Remarks. 1. The interesting aspect of this theorem is that ni is arbitrary. The result actually allows a more general sequence kTE (ni , mi )k. 2. In typical applications, S is an interval in the essential spectrum. In the other direction, one has the following pair of results: Theorem 1.2. Suppose S is a set so that for a.e. E ∈ S, limx→∞ kTE (x, 0)k < ∞. Then µac (Q) > 0 for any Q ⊂ S with |Q| > 0 where | · | = Lebesgue measure. Rb Theorem 1.3. Suppose there is a sequence ni → ∞ so that a kTE (ni , 0)k4 dE < ∞. Then (a, b) ⊂ spec(H) and the spectral measure is purely absolutely continuous on (a, b) and µac (Q) > 0 for any Q with |Q ∩ (a, b)| > 0.
Modified Pr¨ufer and EFGP Transforms
3
Remarks. 1. That Theorem 1.2 is implied by the Gilbert-Pearson [11] theory was noted by Stolz [36]. A simple proof can be found in [33]. Last-Simon [22] prove a stronger R1 variant in which kTE (x, 0)k is replaced by −1 kTE (x + y, 0)k dy and lim by lim. (The discrete analog holds with lim and without any local integration.) 2. Theorem 1.3 is from Last-Simon [22] although the method used there is not much more than what is in Carmona [1]. As to distinguishing dense pure point from singular continuous spectrum, in one direction we have the following elementary result from Simon-Stolz [35]. R∞ P Theorem 1.4. If n kTE (n, 0)k−2 = ∞ in the discrete case or 0 kTE (x, 0)k−2 dx = ∞ in the continuum case, then Hu = Eu has no solution which is L2 at infinity. The paradigm of results that guarantees a solution L2 at ∞ is Ruelle’s proof [30] of Osceledec’s theorem. His argument is abstracted in [22]. We will need the following in Sect. 8: Theorem 1.5. If limn→∞ [log kTE (n, 0)k/nα ] exists and lies in (0, ∞) for some α > 0, then Hu = Eu has an L2 solution. [22] also has a general abstract result on power decay which, to get an L2 solution, requires 3 log kTE (n, 0)k > . lim n→∞ log n 2 [22] also has an example where the limit is 23 and there is no `2 solution. But there are stronger results that hold a.e. in certain probabilistic situations, so we won’t discuss the power decay result here. In Sect. 8, we will discuss the probabilistic result. As for the technique to control the growth of solutions, in the continuum case we will use modified Pr¨ufer variables defined for E > 0 by √ (1.5a) u0 (x) = E R(x) cos(θ(x)), u(x) = R(x) sin(θ(x)). (1.5b) √ One finds these obey the differential equations (with k = E) V (x) dθ =k− sin2 (θ(x)), dx k
(1.6)
1 d log R(x) = V (x) sin(2θ(x)). dx 2k
(1.7)
Two features of these equations are immediately noteworthy: (a) They separate in the sense that (1.6) does not involve R and after solving it, one obtains R by integration. That R drops out of (1.6) and the right side of (1.7) is an expression of the linearity of the initial equations. (b) If V = 0 in some region (a, b), then in that region R is √ constant and θ(x) = θ(a) + E in (1.5a). The addition k(x − a). It is this fact that leads one to take the factor √ of this E is what distinguishes this from ordinary Pr¨ufer transformations.
4
A. Kiselev, Y. Last, B. Simon
There is a third significant feature which we will turn to momentarily. Given how common these continuum equations are, we would have expected their discrete analogs would have been rediscovered and used many times, but even after some efforts at tracking various literature, we’ve found them in a single chain of four papers (and we are a fifth in this chain, since we learned of them from the fourth paper!). The original discoverer of the correct equation was Thomas Eggarter [9] in 1971. He was not looking Pnat an explicit difference equation but rather a continuum equation with V (x) = V0 i=1 δ(x − xi ). By integrating modified Pr¨ufer variables across the δ-functions, he was led to the transforms (E = 2 cos(k)), u(n) − cos(k)u(n − 1) = R(n) cos(θ(n)), sin(k)u(n − 1) = R(n) sin(θ(n)),
(1.8b) (1.8b)
in which case we have, after some calculation (see Sect. 2), cot(θ(n + 1)) = cot(k + θ(n)) − (sin(k))−1 V (n), V (n)2 V (n) R(n + 1)2 sin(2θ(n) + 2k) + =1− sin2 (θ(n) + k). 2 R(n) sin(k) sin2 (k)
(1.9) (1.10)
Actually, he had only an equation of the form (1.9). The definition of θ(n) and precise (1.9) is in a 1975 paper of Gredeskul-Pastur [13] who followed up on Eggarter’s work. [9, 13] focus on (1.9) because they use the transform to study the integrated density of states. Pastur-Figotin [26] defined R and exploited (1.10) to study the Lyapunov exponent. In recognition of these seminal works, we call (1.8) the EFGP transform. Their approach was further exploited in Chulaevsky-Spencer [2]. It will often be useful to use an equivalent form of (1.9) that appears as (2.14). Notice that (1.9), (1.10) have the two critical properties (a), (b) mentioned for (1.6), (1.7) in the continuum case. In particular, if V (n) = 0 for n in some interval [n0 , n1 ], then in that region R(n) is constant and θ(n) = θ(n0 ) + k(n − n0 ). While the EFGP transform was obtained by integrating a continuum δ-function model, it could also be found by looking for a transform with property (b). We will explain this in Sect. 2. [9, 13, 26, 2] all consider V ’s with no decay as n → ∞ but with a small coupling so that any calculations are only asymptotic in coupling constant. It turns out that the methods are especially well suited when V (n) → 0 at infinity and one obtains results that are exact for a fixed V . For example, in Sect. 8, we will find exact formulas for the local Hausdorff dimensions of certain singular continuous spectral measures. The third critical factor of the modified Pr¨ufer and EFGP transforms is a major theme of this paper, namely, that first-order terms in V are oscillatory while the second-order term has a strong tendency to be strictly positive. This idea is already seen in [26, 2], where γ(E) is O(g 2 ) with g a coupling constant because the first-order terms drop out. Let us be explicit about this idea. In (1.6), one might think the positivity comes via the square in sin2 (θ(x)) but that is wrong! Indeed, in writing sin2 (θ) = 21 − 21 cos(2θ), it is the cos(2θ) that is critical! Formally, (1.6) says θ(x) = kx + θ0 −
V (x) sin2 (kx + θ0 ) + O(V 2 ) ≡ kx + θ0 + δθ + O(V 2 ), k
Modified Pr¨ufer and EFGP Transforms
5
and then using sin(2θ) = sin(2kx + 2θ0 ) + 2 cos(2kx + 2θ0 )δθ + O(V 2 ), we get d log R = t1 + t2 + O(V 3 ), dx where t1 =
1 1 V (x) sin(2(kx + θ0 )) − 2 2k 2k
Z
x
V (x)
V (y) dy cos(2(kx + θ0 ))
x0
is the oscillatory term that is often unimportant, while 1 d t2 = 2 4k dx
Z
2
x
V (y) cos(2ky + 2θ0 ) dy x0
has a positive integral, second order in V . In explicit cases, it is more subtle to prove the second order is strictly positive and, indeed, for examples like V (x) = x−α , α < 21 , where the spectrum is absolutely continuous (by Weidmann [37]), the second-order terms do not cause divergences. This means that results that depend on a finite second-order term should hold more generally than those that depend on an infinite second-order term. Indeed, we Conjecture. If V is bounded and in L2 (R, dx) (or `2 (Z+ )), then the essential support of the a.c. part of the spectrum is all of (0, ∞) (or (−2, 2) in the discrete case). Our idea is that for almost all (but not all; see, e.g., [24, 25, 34]) k, the oscillations should kill the first-order term, and so the L2 condition should suffice to give a bounded transfer matrix for a.e. k and so the stated conclusion about the a.c. spectrum by Theorem 1.2. After discussing the modified Pr¨ufer and EFGP transforms and their relation to the growth of the transfer matrix in Sect. 2, we turn to two warm-up problems in Sects. 3 and 4. In Sect. 3, we show these transforms can replace the Harris-Lutz [15] method in many cases where that method is applicable. In Sect. 4, we look at potentials V with lim x|V (x)| finite and show that for such potentials their positive eigenvalues can only coalesce at E = 0. Since examples are known with countable many eigenvalues embedded in (0, ∞), this result is interesting. In Sects. 5–7, we study sparse potentials. Definition. A Pearson potential is one of the form V (x) =
∞ X
an W (x − xn ),
(1.11)
n=1
where W is a bounded, non-negative function of compact support, an → 0, and 1 ≤ x1 < x2 < x3 < · · ·, xn → 0. (1.12) xn+1
6
A. Kiselev, Y. Last, B. Simon
TheP name is in honor of David Pearson who considered potentials of the form (1.11) ∞ where n=1 a2n = ∞ and xn went to infinity sufficiently fast. To make things precise, think of the example xn = n!. Our major goal in Sects. 5–6 is to prove the following: Theorem 1.6. Let V be a Pearson potential. Then P∞ d2 (1) If n=1 a2n < ∞, the spectrum of − dx 2 + V (x) is purely absolutely continuous on (0, ∞) for any boundary condition at 0. P∞ d2 (2) If n=1 a2n = ∞, the spectrum of − dx 2 + V (x) is purely singular continuous on (0, ∞) for any boundary condition at 0. In Sect. 5, we will actually prove a stronger version of (1): Theorem 1.60 . Let V have the form (1.11) where lim
xn < 1. xn+1
(1.13)
Then (1) holds. P∞ Pearson [27, 28] proved a weak version of (2) in that if n=1 a2n = ∞, there exists some set of xn ’s so that the spectrum is purely singular continuous. In [27], there are hints that a result of type (1) (again with xn sufficiently large) should hold, but nothing explicit. As noted at the end of Sect. 5, for (1) the bumps W (x − xn ) can be n-dependent. At the end of Sect. 6, for [a, b] ≡ S ⊂ (0, ∞), we construct Pearson-like potentials (bumps whose width grows with n) so that there is purely a.c. spectrum on S and purely s.c. spectrum on (0, ∞)\S. In a recent paper, coincident with our work, Remling [29] obtained results related to Theorem 1.6(1) using similar methods. He only obtains the existence of absolutely continuous spectrum (his results are consistent with simultaneous singular continuous spectrum while we prove there is none), and he needs at least exp( 23 n log n) growth on the xn (whereas, if f (n) is a monotone function with f (m) → ∞ no matter how slowly, then xn = exp(nf (n)) obeys (1.12) and xn = exp(an) obeys (1.13)). After this manuscript was completed, we obtained a preliminary version of a preprint of Molchanov [23] with considerable overlap with our results in Sects. 5 and 6. In Sect. 7, we will prove Theorem 1.7. Let xn ∈ Z obey xn /xn+1 → 0. Let V be the potential with V (xn ) = an , V (x) = 0
x 6= xn for any n.
Then, P 2 (1) If an < ∞, the discrete Schr¨odinger operator with potential V has purely a.c. spectrum for (−2, 2). P (2) If a2n = ∞, the operator has purely singular continuous spectrum on (−2, 2).
Modified Pr¨ufer and EFGP Transforms
7
In Sects. 8 and 9, we discuss models with randomness and decay, first studied by Simon [31] and then by Delyon, et al. [7], Delyon [6], and Kotani-Ushiroya [21]. Typical of the models discussed in these sections is (g is positive constant) V (n) = gn−α an , where the an are independent, identically distributed random variables, uniformly distributed in [−1, 1]. We prove (i)
If α > 21 , the spectrum is almost surely purely absolutely continuous in (−2, 2).
(ii) If 0 < α < 21 , the spectrum is almost surely dense pure point in (−2, 2). (iii) If α = 21 , the spectrum is almost surely purely singular continuous in the region |E| < (4 − 13 g 2 )1/2 and dense pure point in the region (4 − 13 g 2 )1/2 ≤ |E| < 2 (if g 2 > 12, interpret (4 − 13 g 2 )1/2 as 0). (iv) In case α =
1 2
and g 2 < 12, in the region |E| < (4 − 13 g 2 )1/2 , the spectrum has
fractional Hausdorff dimension with local dimension (4 − E 2 −
g2 3 )/(4
− E 2 ).
Section 8 handles the discrete case, and Sect. 9 the continuum case. For sparse potentials, we give the details in the continuum case and sketch the discrete case; while for random decaying potentials, we give details in the discrete case and sketch the continuum case. A.K. would like to thank the hospitality of I.H.E.S., and B.S. the hospitality of Hebrew University where some of this work was done.
¨ 2. Modified Prufer and EFGP Transforms We will be interested in solutions of −u00 (x) + V (x)u(x) = k 2 u(x).
(2.1)
u0 (x) = kR(x) cos(θ(x)), u(x) = R(x) sin(θ(x)).
(2.2a) (2.2b)
Change variables to
These are called modified Pr¨ufer variables. The 2π ambiguity in θ is fixed by choosing θ(0) ∈ [0, 2π) and demanding θ(x) be continuous in x. Then a straightforward calculation shows (2.1) is equivalent to the pair of equations V (x) dθ =k− sin2 (θ(x)) dx k 1 d(log R)(x) = V (x) sin(2θ(x)). dx 2k
(2.3) (2.4)
This change of variables is so very useful because if V = 0, then θ(x) = θ0 + kx, R(x) = R0 . We will be able to study V as a perturbation about this solution. As explained in the introduction, one needs to study the asymptotic behavior of the norm of the transfer matrix T (x, 0). For any θ0 in [0, π), let θ(x, θ0 ) solve (2.3) with initial condition θ(0) = θ0 . Then let R(x, θ0 ) solve (2.4) with R(0, θ0 ) = 1. Then
8
A. Kiselev, Y. Last, B. Simon
Theorem 2.1. For any α, β ∈ (0, ∞) and θ1 6= θ2 , there exists non-zero, finite constants C1 and C2 (independent of x and V ) so that C1 max(R(x, θ1 ), R(x, θ2 )) ≤ kT (x, 0)k ≤ C2 max(R(x, θ1 ), R(x, θ2 ))
(2.5)
for all k ∈ (α, β). Proof. Define k(a, b)k2k = (ka)2 + b2 . Then min(1, k)k(a, b)k ≤ k(a, b)kk ≤ max(1, k) k(a, b)k. So defining operator norms in terms of k · kk , we see min(k, k −1 )kT (x, 0)kk ≤ kT (x, 0)k ≤ max(k, k −1 )kT (x, 0)kk , so it suffices to prove (2.5) with k · kk rather than k · k. But kT (x, 0)kk ≥ max(R(x, θ1 ), R(x, θ2 )) is trivial and kT (x, 0)kk ≤ {min[sin( 21 |θ1 − θ2 |), cos( 21 |θ1 − θ2 |)]}−1 max(R(x, θ1 ), R(x, θ2 )) by the lemma below.
If |θ1 − θ2 | ≤ π2 (which can be done by replacing θ1 by π + θ, if need be), then this proof shows we can take C1 = min(α, β −1 ), C2 = max(β, α−1 )[sin( 21 |θ1 − θ2 |)]−1 . Lemma 2.2. Let A be a unimodular matrix. Let uθ = (cos(θ), sin(θ)). Then if |θ1 −θ2 | ≤ π 2, kAk ≤ sin( 21 |θ1 − θ2 |)−1 max(kAuθ1 k, kAuθ2 k). Proof. There exists θ0 so that kAuθ k2 ≥ kAk2 sin2 (θ − θ0 ). If |θ1 − θ2 | < π2 , for any θ0 at least one of | sin(θ0 − θi )| is larger than or equal to | sin( 21 (θ1 − θ2 )|. Remark. One might worry that the lemma involves k · k and not k · kk but kAkk =
k 0 k 0 −1
and this product is also unimodular.
01 A 01 For the discrete case, we are interested in solutions of (0 ≤ k ≤ π) u(n + 1) + u(n − 1) + V (n)u(n) = 2 cos(k)u(n).
(2.6)
EFGP variables R(n), θ(n) are defined by R(n) cos(θ(n)) = u(n) − cos(k)u(n − 1), R(n) sin(θ(n)) = sin(k)u(n − 1).
(2.7a) (2.7b)
A priori θ(n) is only determined mod (2π). We will fix this ambiguity later. Noticing that R(n) sin(k + θ(n)) = sin(k)u(n), (2.8) sin(k + θ(n)) u(n) = . (2.9) u(n − 1) sin(θ(n))
Modified Pr¨ufer and EFGP Transforms
Similarly,
9
R(n) cos(k + θ(n)) = cos(k)u(n) − u(n − 1).
Thus, cot(k + θ(n)) =
(2.10)
cos(k)u(n) − u(n − 1) , sin(k)u(n)
where by definition, − cot(θ(n + 1)) =
cos(k)u(n) − u(n + 1) . sin(k)u(n)
Thus, (2.6) is equivalent to cot(θ(n + 1)) = cot(k + θ(n)) −
V (n) . sin(k)
(2.11)
¯ Writing θ(n) ≡ θ(n) + k, we see, using first (2.7) and then (2.8)/(2.9): R(n + 1)2 = sin2 (k)u(n)2 + (u(n + 1) − cos(k)u(n))2 = sin2 (k)u(n)2 + (u(n − 1) − cos(k)u(n) + V (n)u(n))2 2 V (n) 2 ¯ 2 2 ¯ ¯ sin(θ(n)) = R(n) sin (θ(n)) + R(n) cos(θ(n)) − sin(k) V (n)2 V (n) 2 ¯ 2 ¯ sin(2θ(n)) + = R(n) 1 − sin (θ(n)) . sin(k) sin2 (k) We can summarize with the EFGP equations: V (n) ¯ ; θ(n) = θ(n) + k, sin(k) ¯ cot(θ(n + 1)) = cot(θ(n)) + νk (n),
νk (n) ≡ −
(2.12a) (2.12b)
2
R(n + 1) ¯ ¯ = 1 + νk (n) sin(2θ(n)) + νk (n)2 sin2 (θ). R(n)2
(2.12c)
¯ We will fix the ambiguity in θ by demanding θ(n + 1) − θ(n) ∈ [−π, π). Equation (2.12) can be regarded as analogs of modified Pr¨ufer equations in that if V = 0, R(n) = constant, and θ(n) = θ(0) + kn. As noted in the introduction, Eggarter arrived at the first version of the EFGP transform by looking at continuum models with δ-function potential ((2.12b) is especially transparent in this mode). But one arrived at it by noting that when V (n) ≡ 0, could have 2 cos(k) −1 . This matrix has eigenvalues e±ik and so it the transfer matrix is powers of 0 1 cos(k) sin(k) must be similar to − sin(k) cos(k) . That similarity transformation will make the powers simple. Indeed, 0 sin(k) 2 cos(k) −1 cos(k) sin(k) 0 sin(k) = 1 − cos(k) 1 0 − sin(k) cos(k) 1 − cos(k) so the transform (2.7) precisely realizes the similarity. There is an analog of Theorem 2.1. Define R(n, θ) by requiring R(1) = 1, θ(1) = θ in [0, π). Then
10
A. Kiselev, Y. Last, B. Simon
Theorem 2.3. For any α ∈ (0, π2 ) and θ1 6= θ2 , there exists non-zero, finite constants C1 and C2 (independent of x and V ) so that for all k ∈ (α, π − α), C1 max(R(n, θ1 ), R(n, θ2 )) ≤ kT (n − 1, 0)k ≤ C2 max(R(n, θ1 ), R(n, θ2 )).
(2.13)
Because of the arccot, (2.12b) is somewhat awkward to deal with. Pastur-Figotin [26] have noted an equivalent form of (2.12b) which is straightforward from e2iϕ = 1 + viz., e
2iθ(n+1)
=e
¯ 2iθ(n)
1 1 2 1 + i cot(ϕ)
¯ iνk (n) (e2iθ(n) − 1)2 + . ¯ 2 1 − iνk2(n) (e2iθ(n) − 1)
(2.14)
As an application of (2.14) we have Proposition 2.4. If |νk (n)| < 21 , then ¯ |θ(n + 1) − θ(n)| ≤ π|νk (n)|. Proof. If |νk (n)| <
1 2,
then (2.14) implies that ¯
|e2iθ(n+1) − e2iθ(n) | ≤ Since |eiη − 1| ≥
2|η| π ,
(2.15)
|νk (n)| 4 = 4|νk (n)|. 1 2 2
we get
π ¯ ¯ |θ(n + 1) − θ(n)| ≤ |e2iθ(n+1) − e2iθ(n) |, 4 and so the claimed result.
Note. Kiselev, Remling, and Simon [20] present a way of defining R, θ that makes the analogy to the continuum case transparent, makes (2.14) transparent, improves (2.15), and extends to more general h0 . 3. Conditional Integrals and A.C. Spectrum It follows from [11, 16, 17] that for both continuum and discrete Schr¨odinger operators on [0, ∞), we have (see also [33] for a quick proof): Proposition 3.1. If S is a set of reals so that for each λ ∈ S, supx kTλ (x, 0)k < ∞, then H has purely a.c. spectrum on S in the sense that (i) For any boundary condition θ and any T ⊂ S with |T | > 0, we have ρac θ (T ) > 0. (ii) For any boundary condition θ, ρsing θ (S) = 0. Thus, bounded transfer matrices have important spectral consequences. By Theorems 2.1 and 2.3, if we can show R( · , θ) remains bounded for two initial θ’s, we have boundedness R of T . From this and (2.4), P (2.12c), one easily obtains the well-known result that if |V (x)| dx < ∞, (resp. |V (n)| < ∞), then the spectrum is purely a.c. in (0, ∞) (resp. (−2, 2)). Here is a result allowing more general decay, first in the continuum case.
Modified Pr¨ufer and EFGP Transforms
11
Rβ Theorem 3.2. Fix k 6= 0. Suppose that limβ→∞ x V (y)e2iky dy exists and that Z ∞ V (y)e2iky dy (3.1) Wk (x) = x
obeys
Z
Then
|V (x)Wk (x)| dx < ∞.
(3.2)
lim kT (x, 0)k < ∞.
(3.3)
x→∞
Remarks. 1. This result is not new; it is essentially due to Harris-Lutz [15]. This is a new proof. PN 2. This result implies that if V (x) = m=1 am sin(km x)/xβ , β > 21 , and k 6= ± 21 km for any m, then (3.3) holds, and so by Proposition 3.1, the spectrum is purely a.c. except 2 }. for possible positive eigenvalues of { 41 km 3. In [19], Kiselev proved that if V (x) = O(x− 4 − ), then (3.2) holds off a set of Lebesgue measure zero. 3
Proof. We will show for any θ0 , R(x, θ0 ) is bounded, and then one can appeal to Theorem 2.1 to complete the proof of (3.3). Write θ(x) = kx + ϕ(x), so by (2.3), ϕ obeys V (x) dϕ =− sin2 (kx + ϕ). dx k By (2.4) (and R(0) = 1), Z x 1 dWk 2iϕ Im e dx log R(x) = dx 0 2k Z 2i x dϕ 2iϕ 1 = Im [Wk (x)e2iϕ(x) − Wk (0)e2iθ0 ] − e Wk dx 2k 2k 0 dx if we integrate by parts. By hypothesis, Wk (x) is bdd so using (3.4), Z 1 x | log R(x)| ≤ bdd + |Wk (y)V (y)| dy k 0 is bounded by (3.2).
Remark. A similar argument proves that lim θ − kx −
x→∞
1 2k
Z
x
V (y) dy 0
exists. This in turn lets one prove there are complex solutions η± (k, x) with Z x 1 η± (k, x) exp (∓i kx − V (y) dy → 1, 2k 0 Z x 1 0 η± (k, x) exp ∓i kx − V (y) dy → ik. 2k 0
(3.4)
12
A. Kiselev, Y. Last, B. Simon
Notice that if V ∈ L2 , kx −
Z
1 2k
Z
x 0
x
p
k 2 − V (y) dy + Q(x),
V (y) dy = 0
where limx→∞ Q(x) exists. So if V ∈ L2 , this says that WKB-type solutions exist. This is also what the Harris-Lutz method gives [19]. We are heading toward a proof of Theorem 3.3. Fix k 6= 0, π. Suppose V (n) is a discrete potential with lim
B X
B→∞
exists and that
∞ X
V (m)e2ikm = Wk (n)
m=n
|V (n)Wk (n)| + |V (n)Wk (n + 1)| < ∞.
(3.5)
n=1
Then lim kT (n, 0)k < ∞.
n→∞
Given a function f on {1, 2, . . .}, define (δf )(n) = f (n + 1) − f (n) and note that summation by parts takes the form b X
b X
g(m)(δf )(m) = −
m=a
f (m + 1)(δg)(m) + (f g)(b + 1) − (f g)(a).
m=a
Lemma 3.4. If (3.5) holds for some k, then
P∞ n=1
|V (n)|2 < ∞.
Proof. Since W exists, V → 0 at ∞ and so V is bounded. Thus, writing V (n) = −e−2ikn (δWk )(n), and summing by parts, B X
V (n)2 = bdd +
n=1
B+1 X
V (n)Wk (n)e−2ikn −
n=2
is bounded by (3.5).
B X
V (n)Wk (n + 1)e−2ikn
n=1
Lemma 3.5. Suppose that {an }∞ n=1 is a real sequence so that an → 0 and
Then
N X
QN
n=1 (1
an
as n → ∞
is bounded.
(3.6)
(3.7)
n=1
+ an ) is bounded.
Proof. By (3.6), |an | → 0, so without loss we can suppose that |an | < 1. Then |1+an | ≤ 1 + an ≤ ean and (3.7) implies the result.
Modified Pr¨ufer and EFGP Transforms
13
Proof of Theorem 3.3. By (2.12c), Lemma 3.4, and Lemma 3.5, it suffices to prove that N X
¯
νk (n)e2iθ(n) ≡ G(N )
(3.8)
n=1
is bounded. Define ¯ − kn. ϕ(n) = θ(n) − k(n − 1) = θ(n) Proposition 2.4 and Lemma 3.4 imply that for n large |(δϕ)(n)| ≤ π|νk (n)|.
(3.9)
By the definition (3.8),
G(N ) = −
N X
δWk (n)(sin k)−1 e2iϕ(n)
n=1
= bdd + (sin k)−1
N X
Wk (n + 1)δ(e2iϕ )(n).
n=1
But |δ(e2iϕ )| ≤ 2|δϕ|, so by (3.9) |G(N ) − bdd| ≤ C1
N X
|Wk (n + 1)νk (n)|
n=1
≤ C1
X N
|Wk (n)νk (n)| + |νk (n)|2 < ∞.
n=1
Sometimes it is better to use slightly different Pr¨ufer variables. For example, if R, θ are defined by p u0 (x) = E − V (x) R(x) cos(θ(x)), u(x) = R(x) sin(θ(x)), then d log(R) 1 ∂V = cos2 (θ(x)), dx 2 ∂x 1 from which we see if V (x) → 0 at infinity and ∂V ∂x ∈ L , then solutions are bounded. (This is essentially the proof of Weidmann’s theorem [37] in [33].) If one tries out an 1 2 integration by parts argument, one needs both ∂V ∂x ∈ L and V ∈ L .
14
A. Kiselev, Y. Last, B. Simon
4. Bound States for O(x−1 ) Potentials 2
d If |V (x)| = o(x−1 ), Eastham-Kalf [8] show that − dx 2 +V (x) has no positive eigenvalues; more generally, if lim x|V (x)| = C < ∞, they show any eigenvalue λ must obey λ ≤ C 2 . On the other hand, Naboko [24] and Simon [34] have constructed V (x) decaying arbitrarily slower than x−1 with eigenvalues dense in [0, ∞). In fact, Simon [34] constructed V (x) withP V√ (x) = O(x−1 ) so that there are infinitely many eigenvalues with λi < ∞. In this section, we will handle the borderline case and λi → 0 as long as improve Eastham-Kalf [8] by showing:
Theorem 4.1. Let V (x) obey C = limx→∞ x|V (x)| < ∞. Then there are at most countably many positive eigenvalues λn for which there are solutions un of −u00n + V (x)un = λn un and un ∈ L2 . Moreover,
X
λn ≤
n
C2 . 2
(4.1)
Remarks. 1. We do not specify boundary conditions on V , that is, (4.1) is a bound on all possible boundary conditions at once. P√ 2 λn = ∞ (e.g., λn = π3C 2. There are λn so that (4.1) holds, but 2 n2 ) so there is a gap between Simon’s examples P √and what our bounds allow. We believe the optimal result would be to prove that n λn ≤ C. Without loss of generality by slightly increasing C and looking at [x, ∞), we can suppose that (4.2) |V (x)| ≤ C(1 + |x|)−1 which we henceforth do. The following is standard (see, e.g., Eastham-Kalf [8]): Lemma 4.2. If V is bounded and u solves −u00 + V u = λu and u ∈ L2 , then u0 ∈ L2 . In particular, R(x, θ0 ) ∈ L2 for that θ0 with (u(0), u0 (0)) = (R0 sin(θ0 ), kR0 cos(θ0 )). Proof.
Z
N 0
N Z |u0 |2 dx = u0 u − 0
N Z 0 = u u + 0
N
u00 u dx
0 N
(λ − V )u2 dx, 0
RN so if limN →∞ 0 |u0 |2 dx = ∞, then limN →∞ u0 u = ∞, but that implies u2 (N ) = RN u(0)2 + 2 0 u0 u dx → ∞, contradicting the fact that u ∈ L2 . Lemma 4.3. Let f and g be C 1 functions on [1, ∞) so that |g 0 f | + |f 0 | ∈ L1 . Then
RN 0
f (x)ei(kx+g(x)) dx is bounded as N → ∞ for any k 6= 0.
Modified Pr¨ufer and EFGP Transforms
Proof. Write eikx = Z
1 d ikx ik dx e
N
f (x)e
i(kx+g(x))
1
15
and integrate by parts to see that
Z N |f (N )| |f (1)| 1 + + dx ≤ (|f 0 | + |f g 0 |) dx. |k| |k| |k| 1
Noting that |f (N )| = |f (1)| +
RN 1
|f 0 (y)| dy, we see that the integral is bounded.
Remark. If f (x) → 0 at infinity, this argument shows that limN →∞ exists.
RN f (x)ei(kx+g(x)) dx 1
Lemma 4.4. Let {ei }N i=1 be a set of unit vectors in a Hilbert space H so that α ≡ N suphei , ej i < 1.
(4.3)
i6=k
Then
N X
|hg, ei i|2 ≤ (1 + α)kgk2
(4.4)
i=1
for any g ∈ H. Proof. Let A be the n × n matrix with aij = hei , ej i. Note that the Hilbert-Schmidt P norm of A − 1 is bounded by ( i6=j hei , ej i2 )1/2 ≤ α so (4.3) says that A is invertible. If B is its inverse, then X Bij ej (4.5) fi = obeys hfi , ej i = δij , and thus X hg, ei ifj ≡ Proj of g to the span of the e’s, and so
X
2
hg, e if kgk2 ≥ i i .
By (4.5), hfi , fj i = Bij and since hh, A−1 hiCn ≥ kAk−1 hh, hiCn , we see that n X
|hg, ei i|2 ≤ kAk
X
hg, ei i hfi , fj ihg, ej i
i,j
i=1
≤ kAk kgk2 , which is (4.4).
Proof of Theorem 4.1. It obviously suffices to show for each fixed N < ∞ that N X n=1
λn ≤
C2 . 2
Define Rn (x) to be the R corresponding to the L2 solution u(x, λn ). Normalize u so Rn (0) = 1. By Lemma 4.2,
16
A. Kiselev, Y. Last, B. Simon N X
|Rn (x)|2 ∈ L1
n=1
so lim x
N X
|Rn (x)|2 = 0
n=1
PN
(for if not, eventually n=1 |Rn (x)| ≥ Cx−1 is not L1 ). Thus, we can find Bj → ∞ so that for n = 1, . . . , N , −1/2 Rn (Bj ) ≤ Bj or
Z
2
Bj 0
so by (2.4),
Z
Bj
1 d (log Rn (y)) dy ≤ − ln Bj , dx 2 p
V (x) sin(2θn (y)) dy ≤ −
λn log Bj .
(4.6)
0
Now consider the Hilbert spaces Hj = L2 ((0, Bj ), (1 + x) dx). In Hj , we have Z kV
k2Hj
Bj
≤
C 2 (1 + |x|)−2 (1 + x) dx = C 2 log(Bj ) + O(1).
(4.7)
0
Let e(j) n (y) =
sin(2θn (y)) 1 p χ[0,Bj ] (y), (1 + |y|) Nn(j)
where
Z
Bj
sin2 (2θn (y)) dy. (1 + |y|) 0 √ √ √ Notice that 4θn (y) − 4 λn and 2(θn ± θm ) − 2( λn ± λm ) have derivatives that are O(x−1 ) by (2.3). Thus by Lemma 4.3, Z Bj sin(2θn (y)) sin(2θm (y)) − 21 δnm dy (1 + |y|) 0 Nn(j)
=
are bounded. We conclude that Ni(j) = (j) he(j) i , ek i
1 2
log Bj + O(1),
= O((log Bj )
−1
)
i 6= k.
(4.8) (4.9)
Equations (4.6) and (4.8) imply that p
hV, e(j) n i Hj ≤ −
2λn (log Bj )1/2 + O(1).
(4.10)
Since the number N of eigenfunctions is fixed, but Bj → ∞ for j large, Lemma 4.4 applies and
Modified Pr¨ufer and EFGP Transforms N X
17
2 −1 2 |hV, e(j) n iHj | ≤ (1 + O((log Bj ) )kV kHj .
(4.11)
n=1
But (4.10) and (4.7) then say that 2
X N
λn log(Bj ) ≤ C 2 log(Bj ) + O(1),
n=1
so
N X
λn ≤
n=1
C2 . 2
5. Sparse Potentials: The Continuum, Absolutely Continuous Case Our goal in this section is to prove assertion (1) in Theorem 1.6 and Theorem 1.60 . The idea will be to control kT (x)k4 and then use Theorem 1.3. As explained in Sect. 1, the key is oscillations in sin(2θ(x)) for θ(x) ∼ kxn+1 for x near xn+1 . We will realize this (xn )k . using an integration by parts so we need a priori control on objects like dkTdk Fix a Pearson potential; an is assumed to obey an → 0 and xn+1 > xn + 21. Fix θ0 and solve the modified Pr¨ufer equations for each k ∈ (0, ∞) to get functions θ(x, k) and R(x, k) (with initial conditions θ(x = 0, k) = θ0 , R(x = 0, k) = 1). Fix 1 so supp(W ) ⊂ [−1, 1]. We need two propositions to prepare for bounds in an integration by parts: Proposition 5.1. Suppose that lim xn+1 /xn > 1. For each a, b > 0, there exists a constant C so that for each k ∈ (a, b), ∂θ (5.1) ∂k (xn + 1) ≤ Cxn 2 ∂ θ ≤ Cx2n . (x + 1) n ∂k 2
and
(5.2)
Moreover, uniformly for k ∈ (a, b), 1 ∂θ (x) = 1, x→∞ x ∂k 1 ∂2θ lim 2 (x) = 0. x→∞ x ∂k 2 lim
Proof. Let β = inf n
xn+1 >1 xn
(5.3) (5.4)
(5.5)
by hypothesis. As a preliminary, note that if h, g, f are functions on [a, b], h is C 1 and h0 (x) = f (x) + g(x)h(x).
(5.6)
18
A. Kiselev, Y. Last, B. Simon
Then
|h(b)| ≤ (|h(a)| + (b − a)kf k∞ )e(b−a)kgk∞
as follows from the exact solution of (5.6): Z Rx g(y) dy + h(x) = h(x)e a
Rx
x
f (y)e
y
g(z) dz
(5.7)
dy.
a
Now let h(x) =
∂θ ∂k (x).
From (2.3),
V (x) V (x) ∂h = 1 + 2 sin2 (θ(x)) − sin(2θ(x))h. ∂x k k
(5.8)
This means for x ∈ (xn−1 + 1, xn − 1), we have that ∂h = 1. ∂x
(5.9)
By (5.7) and (5.8), |h(xn + 1)| ≤ e2C|an |1 [|h(xn − 1)| + 21 + 2C|an |1] ≤e
2C|an |1
(5.10)
[|h(xn−1 + 1)| + (xn − xn−1 ) + 2C|an |1],
(5.11)
where we used (5.9) to go from (5.10) to (5.11). In these equations, C is a constant only depending on (a, b). Throughout this proof, C is such a constant whose value can vary from one equation to the next. Let β > 1 be given by (5.5). Pick n0 so large that for n ≥ n0 : β −1 e2|an |C1 ≤ and
1 2
(1 + β −1 )
2C|an |1 2an C1 1 − β −1 . 1+ ≤1+ e xn 2
(5.12)
(5.13)
Since β > 1 and an → 0, such an n0 exists. Next, pick D ≥ 2 so |h(xn0 −1 + 1)| ≤ Dxn0 −1 .
(5.14)
We claim inductively that for n ≥ n0 − 1, we have that |h(xn + 1)| ≤ Dxn
(5.15)
for by (5.14), this holds for n = n0 − 1, and if it holds for n − 1, then by (5.11) and xn−1 ≤ β −1 xn , |h(xn + 1)| ≤ [Dxn−1 + xn − xn−1 + 2C|an |1]e2C|an |1 2C|an |1 2C|an |1 e ≤ xn (D − 1)β −1 + 1 + xn 1 1 − β −1 (1 + β −1 ) + 1 + ≤ xn (D − 1) 2 2 −1 1−β = xn D − (D − 2) ≤ Dxn 2
(by (5.12)/(5.13))
Modified Pr¨ufer and EFGP Transforms
19
since D ≥ 2. Thus, we’ve proven (5.15). Next, let H(x) = h(x) − x, so (5.8) implies that ∂H ∂x ≤ C|an |(1 + |H|)
(5.16)
on (xn − 1, xn + 1). Using (5.7) and (5.15), we conclude (recall the constant C changes from one equation to the next!) |H(xn + 1) − H(xn − 1)| ≤ C|an |xn . Since H(xn−1 + 1) = H(xn − 1), we have that for n ≥ n0 , n X H(xn + 1) xm ≤ C + am xn + 1 xn + 1 (x n + 1) m=n 0
n X C + am β −(n−m) → 0 ≤ xn + 1 m=n 0
as n → ∞ since β > 1 and am → 0. From this and (5.16), we see that | H(x) x | → 0 as x → ∞, which proves (5.1). ∂2θ To prove (5.2), let g = ∂h ∂k = ∂k2 . Then differentiating (5.8) with respect to k, we see that ∂g = 0 on (xn−1 + 1, xn − 1) ∂x ∂g = A(x) + B(x)h(x) + D(x)g(x) + E(x)h2 (x) ∂x
(5.17a) on (xn − 1, xn + 1) (5.17b)
where A, B, D, E are uniformly bounded by Can on this interval with C uniformly bounded as k runs through (a, b). Now use (5.7) and (5.1) to see that |g(xn + 1)| ≤ e2Can 1 [g(xn−1 + 1) + Can x2n 1]. As above, if n is so large that β −2 e2Can 1 ≤
1 2
(1 + β −1 )
and
(Can 1)e2Can 1 ≤
1 2
(1 − β −1 )
then inductively, g(xn + 1) ≤ Cx2n for n large. This is (5.2). Plugging this into (5.17b), we see that g(xn + 1) ≤ C 1 +
n X
am x2m
,
(5.17c)
m=1
which yields limn→∞ g(xn + 1)/x2n = 0 from which (5.4) is immediate.
20
A. Kiselev, Y. Last, B. Simon
Proposition 5.2. For any a, b > 0, there is a C so that for all k ∈ (a, b), log R(xn + 1) ≤ C
n X
|am |,
(5.18)
X ∂ log R (xn + 1) ≤ C |am xm |. ∂k
(5.19)
m=1 n
m=1
Proof. By (2.4), log R(x) is constant for x ∈ (xn−1 + 1, xn − 1) and Z | log R(xn + 1) − log R(xn − 1)| ≤ 2k −1 |an | W (y) dy, so (5.18) holds with C = 2 min(k)−1 From (2.4), we have
R
W (y) dy.
∂θ ∂ ∂ (k log R) = V (x) cos(2θ(x)) , ∂x ∂k ∂k so that the bound (5.1) implies (5.9).
As a final preliminary, we note that Lemma 5.3. Suppose that lim xn+1 /xn > 1. Then for a constant C, ∞ X X
∞
|an am |
n=1 m≤n
X xm ≤C a2n . xn n=1
Proof. Let β = lim xn+1 /xn . Pick 1 < γ < β. Then for m ≤ n, xm /xn ≤ Cγ −|m−n| . Thus, the lemma follows from Young’s inequality that X T (a)n ≡ γ −|m−n| am m
is bounded from `2 to `2 for any γ > 1.
0
Proof of Theorem 1.6 . Let g be a non-negative C ∞ -function compactly supported on (0, ∞). We will prove that Z (5.20) sup g(k)R(k, xn + 1)4 dk < ∞. n
Proving this for two values of θ0 and appealing to Theorem 2.1 gets a uniform bound R on g(k)kT (0, xn + 1)k4 dk. Theorem 1.3 then proves pure absolute continuity of the spectrum on (0, R ∞). Let Bn = g(k)R(xn + 1)4 dk. Notice that by (2.4), R(xn−1 + 1) = R(xn − 1) and (5.21) R(xn + 1)4 = R(xn − 1)4 exp(Qn ), where 2 Qn = k
Z
1 −1
an W (y) sin(2θ(xn + y)) dy.
Since k −1 and an are bounded, Qn is uniformly bounded in n, and so
Modified Pr¨ufer and EFGP Transforms
21
exp(Qn ) ≤ 1 + Qn + CQ2n ≤ 1 + Qn + Ca2n
(5.22)
(where again C is a constant that varies from formula to formula). For y ∈ (−1, 1), we have by (2.3) |θ(xn + y) − θ˜n (y)| ≤ Can , where θ˜n (y) = θ(xn−1 + 1) + k(xn + y − xn−1 − 1), so
Z 1 Qn − 2 ˜ an W (y) sin(2θn (y)) dy ≤ Ca2n . k −1
(5.23)
By (5.21)–(5.23), Bn ≤ Bn−1 (1 + Ca2n ) + En , where
Z E n = an
Z
1
dy −1
(5.24)
2g(k) R(xn−1 + 1, k)4 W (y) sin(θ˜n (y)) dk. k
Notice that we’re implementing our basicQ strategy: We separate out the second-order ∞ terms (which will present no problem since n=1 (1 + Ca2n ) < ∞) and need to control the first-order terms where we have an explicit highly oscillatory factor since θn ∼ kxn . Now 1 ∂θ(xn−1 + 1) ∂ θ˜n (y) = xn + y − xn−1 − 1 + > xn (5.25) ∂k ∂k 2 for n large by the bound (5.3). Thus, we can write sin(θ˜n (y)) =
1 ∂ θ˜n ∂k
∂ (− cos(θ˜n (y))) ∂k
and integrate by parts. After integration by parts, we have three terms En(1) En(2) En(3)
∂[k −1 g(k)] , ∂k ∂R4 , coming from ∂k ∂ 1 . coming from ∂ θ˜ ∂k coming from
∂k
For the En(1) term, we can bound R4 as follows using (5.18) and xn ≥ Cβ n . By (5.10), for n large,
(5.26)
22
A. Kiselev, Y. Last, B. Simon
X n R4 ≤ C exp C am m=1
n ln(β) , ≤ C exp 2 since an → 0. Thus, by (5.19) and (5.26),
En(1) ≤ Cβ n/2 β −n = Cβ −n/2 . For the En(2) term, we use
∂R4 ∂k
(5.27)
R = R4 ∂ log ∂k , (5.19), and (5.25) to see that
En(2) ≤ CBn−1 bn , where bn = Note now that by
P
n−1 X
an am
m=1
a2n
xm . xn
< ∞ and Lemma 5.3, we have ∞ X
bn < ∞.
(5.28)
n=1
For the En(3) term, we use (5.25) and (5.17c) to see that En(3) ≤ Bn−1 cn , where cn = Can As in the proof of Lemma 5.3,
(1 +
X
Pn−1 m=1 x2n
am x2m )
.
cn < ∞.
(5.29)
By (5.24) and the above estimates on En(i) , max(Bn , 1) ≤ (1 + Ca2n + Cbn + Ccn + Cβ −n/2 ) max(Bn−1 , 1). P P By hypothesis, a2n < ∞, and by (5.28–5.29), bn + cn < ∞. Thus N Y
(1 + Ca2n + Cbn + Ccn + Cβ −n/2 )
n=1
is bounded and consequently, so is Bn .
It is easy to see that the methods of this section extend to prove: P Theorem 5.4. Suppose V (x) = Wn (x − xn ), where (i) lim xn /xn+1 < 1, (ii) supp Wn ⊂ [−1, 1] for some fixed 1, P R 2 (iii) n |Wn (y)| dy < ∞. 2
d Then − dx 2 + V (x) has purely a.c. spectrum on (0, ∞).
(5.30)
Modified Pr¨ufer and EFGP Transforms
23
6. Sparse Potentials: The Continuum, Singular Continuous Case In this section, we will prove assertion (2) in Theorem 1.6. The idea will be to force kT (k 2 , xn )k to infinity for almost all k and suitable xn . To do this, we will need to isolate a strictly positive second-order term and show that these second-order terms then dominate the first-order terms because of oscillations. Here is a warm-up problem to show this cancellation mechanism. Let Xn be independent, identically distributed random variables taking the values ±1 with probability 1 2 . Let > 0 and let an be a sequence going to zero as n → ∞. Finally, let Yn =
n X
(a2m + am Xm ).
m=1
Suppose that probability 1,
P
a2n = ∞. We claim there exists a subsequence n(i) → ∞, so with lim Yn(i) = ∞.
i→∞
(6.1) Pn
The reason (6.1) holds is that by the central limit theorem m=1 an Xn is typically pP a2n and, because of the square root, this is smaller not more negative than O − Pn than m=1 a2n . Pn To make a proof, notice that since m=1 a2m → ∞, we can choose n(i) so that Pn(i) 2 2 m=1 am ≥ i . By a Tschbechev inequality, Prob
X n(i) 1
P
1 i2
X 2 am 2 n(i)
am Xm ≥
≤
k
1
Pn(i)
am Xm k2 4 4 1 Pn(i) 2 2 = 2 Pn(i) 2 ≤ 2 2 . i a ) a m m 1 1
1
( 2
< ∞, so by the Borel-Cantelli lemma, with probability 1, eventually n(i) X
a m Xm ≤
1
n(i) X 2 an , 2 1
and thus eventually, Yn(i) ≥
n(i) X 2 am 2 1
diverges. The usual Kolmogorov stopping argument that lets one prove things without subsequences isn’t obviously applicable here in a situation where we assume no regularity on the am ’s (see Sect. 8 for the case am = m−α ). Since a subsequence suffices for our application, we have not tried to push the argument through to get lim Yn = ∞, even in the toy problem. Notice that independence of the Xn ’s was not needed; rather, it suffices to have enough control of E(Xn Xm ) to show that the first-order term is small compared to the second-order term. In the case at hand, we will use integration by parts in k as we did in the last section to get this control. We summarize the key to the above argument with Lemma 6.1. Let Pn , Qn be random variables so that
24
A. Kiselev, Y. Last, B. Simon
(i) Pn (x) ≥ αn > 0 for a.e. x and positive reals αn , P −1 (ii) αn Exp(|Qn |) < ∞, (iii) limn→∞ αn = ∞. Then Pn (x) + Qn (x) → ∞ for a.e. x. If (ii) is replaced with (ii0 ) limn→∞ αn−1 Exp(|Qn |) = 0, then there exists a subsequence n(i) so that Pn(i) (x) + Qn(i) (x) → ∞ for a.e. x. Proof. If (ii0 ) holds, we can find a subsequence so that (ii) holds. Thus, it suffices to prove the result P assuming (ii). By (ii), αn−1 |Qn (x)| < ∞ for a.e. x. In particular, αn−1 Qn (x) → 0 so Pn + Qn ≥ αn [1 − αn−1 |Qn (x)|] → ∞. We will also need the following lemma: Lemma 6.2. Suppose that Bn , αn , βn ≥ 0 are real numbers and that p Bn ≤ Bn−1 + 2αn Bn−1 + βn (n ≥ 1). Then, p
(6.2)
v u n n X p uX Bn ≤ B0 + αk + t βk . k=1
(6.3)
k=1
Proof.PWe give a proof by induction. Equation (6.2) holds for n = 0. Let an = n bn = k=1 βk . By the induction hypothesis, p p p Bn−1 ≤ B0 + an−1 + bn−1 .
Pn k=1
αk ,
(6.4)
Equation (6.2) implies that Bn ≤ So by (6.4),
p 2 Bn−1 + αn + βn .
2 p p B0 + an + bn−1 + βn 2 p p p ≤ B0 + an + bn + 2 bn−1 B0 + a n p p 2 B 0 + a n + bn , ≤
Bn ≤
proving (6.3) inductively.
P So fix a Pearson potential with a2n = ∞. Fix θ0 and let R(x, k) be the solution of (2.3/2.4). Let Yn (k) = log R(xn + 1, k) and δYn (k) = Yn (k) − Yn−1 (k). By (2.4),
Modified Pr¨ufer and EFGP Transforms
δYn (k) =
25
an 2k
Z
1 −1
W (y) sin 2θ(xn + y) dy.
(6.5)
As in Sect. 5, we write θ˜n (y) = θ(xn−1 + 1) + k(xn + y − xn−1 − 1). But we expand θ to the next order by letting Z an y θn(1) (y) = − W (y) sin2 (θ˜n (y)) dy. k −1 Then by (2.3),
(6.6)
θ(xn + y) = θ˜n (y) + θn(1) (y) + O(a2n ),
so by (6.5), δYn (k) = an Xn(1) + a2n Sn + O(a3n ), Z 1 1 W (y) sin(2θ˜n (y)) dy, Xn(1) = 2k −1 (1) Z 1 1 θ (y) W (y) cos(2θ˜n (y)) n . Sn = k −1 an
(6.7a) (6.7b) (6.7c)
In the formula for θn(1) , use sin2 (θ˜n (y)) = 21 (1 − cos(2θ˜n (y))). The cos term from this formula when plugged into (6.7c) gives Z y Z 1 2 1 ˜ ˜ k W (y) cos(2θn (y)) W (s) cos(2θn (s)) dy 2 −1 −1 Z 1 2 1 ˜ = 2 W (y) cos(2θn (y)) dy . 4k −1 lump the contribution of the RWe y W (s) ds, we find −1
1 2
(6.8)
term with the first-order term. Defining X(y) =
δYn (k) = [a2n Zn (k) + an Xn (k)] + O(a3n ),
(6.9)
where Z 1 2 1 ˜n (y)) dy , W (y) cos(2 θ 4k 2 −1 Z 1 1 an W (y)X(y) ˜ ˜ cos(2θn (y)) dy. W (y) sin(2θn (y)) − Xn (k) = 2k −1 2k Zn (k) =
In (6.9), the O(a3n ) means an error bounded by Ca3n , where C is a finite constant for k ∈ [a, b] any compact subinterval of (0, ∞). Define Z 1 f (k) = W (y)e2iky dy. W −1
26
A. Kiselev, Y. Last, B. Simon
Then, Zn (k) =
1 f |W (k)|2 + X˜ n (k), 8k 2
(6.10)
where
1 f X˜ n (k) = 2 |W (k)|2 cos(4(θ˜n (0, k) + ϕ(k)), 8k f (k)). where ϕ(k) = 21 Arg(W f (k) = |W f (k)|e2iϕ(k) , then For let θ˜n (y) = θ˜n (0) + ky. If W
(6.11)
2 Z 1 1 2i(θ˜n (0)+ky) W (y)e dy Zn (k) = 2 Re 4k −1 1 f = 2 |W (k)|2 cos2 (2(θ˜n (0, k) + ϕ(k))). 4k Proof of Theorem 1.6, Part (2). Let X 1 f 2 | W (k)| a2m , 8k 2 n
Pn (k) =
m=1
Qn (k) = Yn (k) − Pn (k); so
δQn (k) = Qn (k) − Qn−1 (k),
δQn (k) = a2n X˜ n (k) + an Xn (k) + O(a3n ). f (k) 6= 0}. Let Let g be a C ∞ -function compactly supported in {k ∈ (0, ∞) | W n 2 Z X am Xm (k) dk, Bn = g(k) Z B˜ n =
m=1
n 2 X 2 g(k) am X˜ m (k) dk. m=1
We will prove that
p
Bn
X n
a2m → 0
(6.12)
m=1
en . Since Pn a3 / Pn a2 → 0 (on account of as n → ∞, P and similarly for B m=1 m m=1 m n an → 0 and m=1 a2m → ∞), (6.12) and the Schwartz inequality imply that X Z n a2m → 0, g(k)|Qn (k)| dk m=1
e > 0 implies that there is a subsequence n(i) so by Lemma 6.1 and inf k∈supp g |W8k(k)| 2 so that Yn(i) (k) → ∞ for a.e. k in supp g. By doing this for two values of θ0 and using Theorem 2.1 and Theorem 1.1, we conclude there is no a.c. spectrum on supp g. f is an entire function, it has isolated zeros and thus, this argument shows σac Since W is empty. By Theorem 1.4, σpp ∩ (0, ∞) is empty, and an elementary argument proves that σ(H) ⊃ [0, ∞). So the spectrum on (0, ∞) is purely singular continuous. It thus suffices to prove (6.12) (the proof for B˜ n is essentially identical).
Modified Pr¨ufer and EFGP Transforms
Let Mn−1 (k) =
27
Pn−1
am Xm (k). Then Z Bn ≤ Bn−1 + g(k)Mn−1 (k)an Xn (k) + Ca2n m=1
for a suitable constant C. Now Xn has cos(2θ˜n (y)) and sin(2θ˜n (y)) terms. As in the last d [. . .] and integrate by parts and section, we write those as a suitable [dθ˜n (y, k)/dk]−1 dk get three terms: −1 One coming from ∂[k ∂kg(k)] ∂k. Noting that |Mn (k)| ≤ Cn, we have that this is bounded by Cn xn . n (k) . Using (5.1), this term is bounded by One coming from ∂M∂k C
n−1 X
an am
m=1
xm . xn
2
n 2 One coming from Ln = [ ∂∂kθ2n ]/[ ∂θ ∂k ] . As in the last section, thisRLn is bounded by Pn−1 C( m=1 am x2m )/x2n . We can use the Schwartz inequality to control g(k)|Mn (k)| dk, p Pn−1 and so bound this term by C Bn−1 an m=1 am x2m /x2n . The net result is the bound p Bn ≤ Bn−1 + 2αn Bn−1 + βn , (6.13)
where αn = C
n−1 X
x2m x2n
|an am |
m=1
and
n−1 n X xm βn = C a2n + + an am . xn xn m=1
By the argument in Lemma 5.3 with xn−1 /xn → 0 and that
n X
αm
m=1
and that
n X
X n
a2m → 0,
m=1
a2n → ∞, we see (6.14)
m=1
n X βm ≤ C 1 + a2m ,
m=1
so
P∞
m=1
v u n X n uX t βm a2m → 0. m=1
(6.15)
m=1
Lemma 6.2 and (6.13–6.15) imply (6.12).
One can modify this construction to make examples of decaying potentials for which the associated Schr¨odinger operator has regions of a.c. spectrum and regions of s.c. specf (k) vanishes in a whole interval so that even though trum. The idea is to arrange that W
28
A. Kiselev, Y. Last, B. Simon
f (k) cannot vanish if W has coman ∈ / `2 , we have a.c. spectrum for those k. Of course, W pact support, so we will take the bump functions of increasing support converging toward a function whose Fourier transform vanishes in an interval. So, let S = [a, b] ⊂ (0, ∞). Let f be an even Schwartz class function that vanishes if k 2 ∈ S and is strictly positive on [0, ∞)\S. P Let an = n−1/2 , xn = (n!)2 , 1n = n−1/12 . Notice that a2n = ∞. Define Z 1 f˜(x) = exp(−2ikx)f (k) dk, 4π Wn (x) = f˜(x)χ(−1n ,1n ) (x) and V (x) =
X
an Wn (x − xn ).
n
We are heading toward: 2
d Theorem 6.3. The half-axis Schr¨odinger operator − dx 2 + V (x) has purely singular spectrum on (0, ∞)\S and purely a.c. spectrum on S.
Lemma 6.4. For any m > 0, there exists a constant Cm with Z f (k) − e2ikx Wn (x) dx ≤ Cm n−m . Proof. Let fn (k) =
R
(6.16)
e−2ikx Wn (x) dx. Then Z 1 sin(1n (k − k 0 )) f (k 0 ) dk 0 , fn (k) = 2π (k − k 0 )
so the left side of (6.16) is Z 1 sin 1n (k − k 0 ) 0 0 [f (k) − f (k )] dk , 0 2π k−k Z g(y, k) sin 1n y dy ,
which has the form
where g(y, k) is Schwartz space in y with bounds (including bounds on derivatives) uniform in k. If we integrate by parts 12m times, we will get (6.16). Proposition 5.1 extends with no change. In the region where f (k) 6= 0, the analysis earlier in this section shows that log R(xn(i) + 1n(i) ) → ∞ for a.e. k and a suitable subsequence xn(i) , so we know the spectrum in (0, ∞)\S is purely singular continuous. On the other hand, if g is C ∞ supported in S, we claim that Z (6.17) sup g(k)R(k, xn + 1n )4 dk < ∞. n
The proof is similar to that in the last section. In place of (5.22), we need to use exp(Qn ) ≤ 1 + Qn + 21 Q2n + O(a3n ).
Modified Pr¨ufer and EFGP Transforms
29
fn (k)|2 /8k 2 and oscillatory terms that we can As in this section, Q2n has a term a2n |W integrate by parts. Noting that n−1 X m=1
X xm an am ≤ n−2 n−1/2 m−1/2 ≤ Cn−2 xn
is still summable and that
n−1
m=1
P
fn (k)|2 is summable by Lemma 6.4, we obtain (6.17). a2n |W
7. Sparse Potentials: The Discrete Case In this section, we will sketch the proof of Theorem 1.7. The proof follows closely that in the last two sections with (2.12) replacing (2.3/2.4). We will make use of (2.14), the Pastur-Figotin form of (2.12b). Fix α > 0 and pick k ∈ (α, π−α) and then N so large that for all such k, |νk (n)| < 21 for n ≥ N0 . Equation (2.14) can then be effectively used to prove the analogs of (5.1/5.2), that is, 2 ∂ θ ∂θ 2 (x ) ≤ Cx (x ) (7.1) n ∂k 2 n ≤ Cxn . ∂k n Equation (2.12c) can be rewritten log R(n + 1) − log R(n) =
1 2
¯ + νk (n)2 sin2 (θ)). ¯ log(1 + νk (n) sin(2θ)
This implies the bound log R(xn ) ≤ C
n X
|am |.
(7.2)
(7.3)
m=1
Next notice that 1 + α sin(2θ) + α2 sin2 (θ) = (1 + 21 α sin(2θ))2 + α2 sin4 (θ). This provides a uniform bound on the argument of the log(·) in (7.2), and so allows one to prove n X ∂ log R(x ) ≤ C a n xm . (7.4) n ∂k m=1
With these tools, the proof of assertion (1) of Theorem 1.6 is similar to that in Sect. 5, only a little simpler since (2.12c) implies ¯ + Cn a2n ). R(n + 1)4 ≤ R(n)4 (1 + νk (n) sin(2θ(n)) The same integration by parts used in Sects. 5 and 6 shows that Z Z g(k)R(n, k)4 νk (n) sin(2θ(n)) dk = C(bn + cn + B −n/2 ) 1 + g(k)R(n, k)4 dk Pn−1 with bn = m−1 an am xm /xn and cn is like bn with x2m /x2n replacing xm /xn . As in Sect. 5, this proves assertion (1) in Theorem 1.7. To prove assertion (2), we must identify a strictly positive second-order term. We write
30
A. Kiselev, Y. Last, B. Simon
log(1 + α sin(2θ) + α2 sin2 (θ)) = α sin(2θ) + α2 (sin2 (θ) −
1 2
sin2 (2θ)) + O(α3 )
(7.5)
= α sin(2θ) + α cos(4θ) − α cos(2θ) + α + O(α ). 1 4
This lets us write
2
1 2
2
log R(n + 1) − log R(n) =
1 4
1 2
2
3
(7.6)
a2n + an Xn ,
and, as in Sect. 6, use the integration by parts machine to prove Z X N
2 an Xn
g(k) dk
1/2 X
a2n → 0
n=1
and complete the proof as there. f (k) since the analog of W here In this case, we don’t need to worry about zeros of W f is δn0 and so W (k) = 1. 8. Random Decaying Potentials: The Discrete Case In this section, we consider discrete situations where the V (n) are independent random variables of zero mean and decaying variance. The results that imply a.c. spectrum require no regularity in E(V (n)2 ), while those for singular spectrum require some kind of regular decay, as we will explain. The results for a.c. spectrum are so general yet so simple to prove that they are a paradigm of the usefulness of the EFGP transform. Theorem 8.1. Suppose Vω (n) are independent random variables with E(Vω (n)) = 0 and X E(Vω (n)2 ) + E(Vω (n)4 ) < ∞. (8.1) n
Then for a.e. ω, hω has purely a.c. spectrum on (−2, 2). Remarks. 1. For E(Vω2 )1/2 ≤ Cn−α with V bounded and α > 21 , we get a.c. spectrum recovering results of Delyon, et al. [7]. bounded, then E(Vω (n)4 ) ≤ CE(Vω (n)2 ) and so (8.1) 2. If the PVω (n) are uniformly 2 becomes n E(Vω (n) ) < ∞; we state the general bound because unbounded V ’s are so easy to accommodate. 3. The case E(Vω (n)2 )1/2 = n−1/2 log(n)−1 is of some interest. This sequence is 2 ` so if V is bounded, the theorem proves a.c. spectrum. Kotani-Ushiroya [21] cannot handle such borderline cases. Proof. Fix θ0 . Then Rω (n) and θω (n) become random variables which are measurable functions of {Vω (j)}j≤n−1 and so independent of {Vω (j)}j≥n . By (2.12c), Vω (n) sin(2θ¯ω (n)) + O(Vω2 + Vω4 ) . R(n + 1)4 = R(n)4 1 + sin k ¯ and R(n), we have Since Vω (n) is independent of θ(n) E(Rω (n)4 Vω (n) sin(2θω (n) )) = E(Vω (n))E(Rω4 (n) sin(2θ¯ω (n))) = 0.
Modified Pr¨ufer and EFGP Transforms
31
Using independence to bound E(R(n)4 Vωj ) by E(R(n)4 )E(Vωj ), we see that E(Rω (n + 1)4 ) ≤ [1 + CE(Vω2 (n) + Vω4 (n))]E(Rω4 (n)), where C is uniformly bounded for k in any (α, π − α) with α > 0. It follows that Z π−α 4 E Rω (n, k) dk < ∞. α
By Fatou’s lemma, for a.e. ω, Z
π−α
Rω (n, k)4 dk < ∞,
lim α
and by Theorem 1.3, the spectrum is purely a.c. on (−2 cos(α), 2 cos(α)). P∞ For the case where n=1 E(V (n)2 ) = ∞, we need some regularity of the fall-off. Rather than try to find complicated general conditions, we consider the case where E(V (n)2 ) ∼ n−2α with α ≤ 21 . The same method can handle a case like E(V (n)2 ) = [n log(n+1)]−1 (which always has singular continuous spectrum of Hausdorff dimension 1) by the kind of arguments we will discuss in the case α = 21 ; in this case for typical energies kT (0, n)k grows like log(n). Explicitly, we suppose 0 < α ≤ 21 ; λ > 0, (i) E(Vω (n)2 )1/2 = λn−α (ii) E(Vω (n)) = 0, (iii) For some > 0, supω |Vω (n)| ≤ Cn−(2α/3)− , (iv) Vω (n) is independent of {Vω (j)}n−1 j=1 . Remarks. 1. Think of the case discussed in [26, 7], where Vω (n) = n−α Xn (ω) with Xn identically distributed bounded, independent random variables. If E(X) = 0 and X is bounded, then (i)–(iv) hold. 2. With some extra effort, we could allow unbounded distributions, and only require that limn→∞ n+α E(Vω (n)2 )1/2 exists and be non-zero. Theorem 8.2. Suppose (i)–(iv) hold. Fix k in (0, π) with k 6=
π 2π 3π 4, 4 4 .
Then for a.e. ω,
λ2 log kT2 cos(k) (n, 0)k Pn −2α = . n→∞ ( j=1 j ) 8 sin2 (k) lim
Remark. In case α < 21 , this says kT (n, 0)k ∼ exp(Cn1−α ) with C = α=
1 2,
this says kT k ∼ n with C = C
λ2 . 8(1−2α) sin2 (k)
If
λ2 . 8 sin2 (k)
Proof. By Theorem 2.3, we need only prove this result with R(n) replacing T for each θ0 . So fix k and θ0 , and let θω (n), Rω (n) solve (2.12). By (2.12c), log R(n + 1) − log R(n) =
1 2
¯ ¯ log(1 + νk (n) sin(2θ(n)) + νk (n)2 sin2 (θ(n))).
(8.2)
Since supω νk (n) → 0 as n → ∞, we can use log(1 + x) = x −
x2 + O(x3 ). 2
(8.3)
32
A. Kiselev, Y. Last, B. Simon
We also use sin2 θ − 21 sin2 (2θ) =
1 4
− 21 cos(2θ) + 41 cos(4θ).
The net result is 1 X E(Vω (n)2 ) + C1 + C2 + C3 + C4 , 8 sin2 (k) n
log R(n) =
j=1
where the corrections have the form X 1 Vω (j) sin(2θ¯ω (j)), 2 sin(k) n
C1 = −
j=1
1 2 ¯ ¯ [Vω (j) − E(Vω (j) )] sin (θω (j)) − sin (2θω (j)) , 2 j=1 n X 1 1 2 1 ¯ ¯ C3 = cos(2 θ cos(4 θ E(V (j) ) (j)) − (j)) , ω ω ω 2 4 2 sin2 (k)
1 C2 = 2 sin2 (k)
n X
2
2
2
j=1
C4 =
n X
O(Vω (j)3 + Vω (j)4 ).
j=1
The theorem follows if we prove that for each q = 1, 2, 3, 4 and a.e. ω, |Cq (ω)| lim Pn −2α = 0. n→∞ j=1 j
(8.4)
Equation (8.4) for q = 4 is an immediate consequence of hypothesis (iii). C1 , C2 clearly have zero expectation values and variances that decay properly for us to hope (8.4) holds; the key to the proof will be a Martingale inequality. C3 will depend on the fact that cos(θ) has zero average and the slow variation of E(Vω (n)2 ). We break the proof to present some needed lemmas. For the first two of these lemmas, let X0 , X1 , . . . , XN be independent random variables, where X0 can be vector valued. Suppose that for j = 1, . . . , N , Zj = Xj fj (X1 , . . . , Xj−1 ; X0 )
(8.5)
with fj a measurable function, and that E(Xj ) = 0.
(8.6)
The following is a variant of a standard Martingale inequality; we provide a proof for the reader’s convenience: Lemma 8.3.
E
sup n=1,2,...,N
|Z1 + · · · + Zn | ≥ r
≤
X N 1 2 E Z j . r2 j=1
(8.7)
Modified Pr¨ufer and EFGP Transforms
33
Proof. Define Yn =
n X
Zj ,
Qn =
j=1
N X
Zj
j=n+1
and let Aj = {ω | |Y1 | ≤ r, |Y2 | ≤ r, . . . , |Yj | > r}. Then χn , the characteristic function of An , is a function only of X0 , X1 , . . . , Xn and thus, if k > n, E(Zk Yn χAn ) = E(Xk )E(fk (X1 , . . . , Xk−1 , X0 )Yn χAn ) = 0. Thus, E(χn Yn2 ) ≤ E(χn (Yn + Qn )2 ), since the cross term has zero expectation when we expand the square. Thus, r2
n X
E(χj ) ≤
j=1
which is (8.7).
N X
E(χj Yj2 ) ≤
j=1
N X
E(χj YN2 ) ≤ E(YN2 ),
j=1
Lemma 8.4. Suppose E(Zn2 ) ≤ Cn−2α . Then for a.e. ω: (1) If α <
1 2
and β > 21 (1 − 2α), then n X −β Zj n = 0. lim n→∞ j=1
(2) If α =
1 2
and β > 21 , then X n Zj (log n)−β = 0. lim n→∞ j=1
(3) If α > 21 , lim
n→∞
n X
Z j = Y∞
j=1
exists, and for any β < α − 21 , X ∞ Zj = 0. lim n+β
n→∞
j=n
Pn Remark. Naively, fluctuations should behave as ( j=1 j −2α )1/2 . This lemma shows they Pn are not much worse. Since we only need that they are small compared to j=1 j −2α , the lemma suffices.
34
A. Kiselev, Y. Last, B. Simon
Proof. (1) Pick β1 so β > β1 ≥ 21 (1 − 2α). By Lemma 8.3, E
sup n−1
j=1,...,2
nβ1 Zk ≥ 2 ≤ C 2−2nβ1 2(n−1) 2−2(n−1)α
2n−1 X+j k=2n−1 +1
(8.8)
≤ C 2−(n−1)(2α+2β1 −1) is summable in n by the choice of β1 . Therefore, by the Borel-Cantelli lemma, for a.e. ω, there is an n0 (ω0 ) so that the sup inside (8.8) is less than 2nβ1 if n ≥ n0 . Let j be larger than 2n0 −1 and pick n so that 2n−1 + 1 ≤ j ≤ 2n . Then |Z1 + · · · + Zj | ≤ |Z1 + · · · + Z2n0 | +
n X
2kβ1
k=1 nβ1
2 2 β1 − 1 2 β1 ≤ |Z1 + · · · + Z2n0 | + β1 j β1 . 2 −1 ≤ |Z1 + · · · + Z2n0 | +
Thus, lim j −β1 |Z1 + · · · + Zj | < ∞. Since β > β1 , the limit for β is 0. (2) Pick β1 with β > β1 > 21 and define Kn =
ω
sup j=1,...,2
X j ≥ nβ1 . Z j n m=1
Then by Lemma 8.3, n
−2β1
E(Kn ) ≤ n
2 X 1 1
j
≤ n−2β1 (1 + n log 2) ≤ Cn1−2β1 ,
Pk
≤ 1 + log k. since Pick an integer m so m(2β1 − 1) > 1. Then 1 1 j
∞ X
E(Knm ) < ∞.
n=1
/ K nm . So by themBorel-Cantelli lemma, for a.e. ω, there is n0 (ω), so if n ≥ n0 , then ω ∈ If j > 2n0 , pick n so that m m 2(n−1) < j ≤ 2n . Then |Z1 + · · · + Zj | ≤ (nm )β1 ≤ 2mβ1 (n − 1)mβ1 ≤ 2mβ1 (log 2)−β1 (log j)β1 . (3) Pick β1 so β < β1 < α − 21 . Then E
2n−1 X+j sup Zk ≥ 2−nβ1 ≤ C 2−2nβ1 2n−1 2−2(n−1)α n−1
j=1,...,2
k=2n−1 +1
≤ C 22β1 2−2(n−1)[α−1/2−β1 ]
Modified Pr¨ufer and EFGP Transforms
35
is summable. Thus, for a.e. ω, there is an n0 (ω) so that for n ≥ n0 (ω), the sup is bounded by 2−nβ1 . Thus, if j1 ≥ j2 ≥ 2n2 −1 ≥ 2n0 −1 , j2 ∞ X X ≤ Z 2−nβ1 → 0 k n=n2
k=j1
as n2 → ∞. So the sum is convergent (i.e., the partial sums are Cauchy). Moreover, if j ≥ 2n0 −1 and n is picked so 2n−1 ≤ j ≤ 2n , then ∞ ∞ X X 2−nβ1 j −β1 Zk ≤ 2−mβ1 = ≤ , 1 − 2−β1 1 − 2−β1 m=n k=j
and thus, if we multiply by j β , the limit is 0.
Lemma 8.5. Suppose that k ∈ R is not in Zπ. Then there exist integers q` → ∞ so that for any θ0 , . . . , θq` , X q` X q` ≤1+ cos(θ ) |θj − θ0 − kj|. j j=1
j=1
Pq
Remark. In essence, we show | j=1 cos(θ0 + kj)| ≤ 1 a stronger result than the ergodic Pq theory result that | q1 j=1 cos(θ0 + kj)| → 0. The weaker ergodic theory result suffices for our application, but the proof of this lemma is easy so we give it. Proof. By general number theory considerations [14], we can find p` , q` so that k − πp` ≤ 1 (8.8) q` q`2 / Z if k ∈ / Zπ. For any p/q ∈ / Z and any θ0 , and p` /q` ∈ q X j=1
Thus
jpπ = 0. cos θ0 + q
q` q` X X jp` π cos(θj ) = cos(θj ) − cos θ0 + q` j=1 j=1 q` X θj − θ0 − jp` π ≤ q` j=1 q` q` X X πp` ≤ |θj − θ0 − kj| + j k − q` j=1
≤
q` (q` + 1) + 2q`2
j=1
q` X j=1
|θj − θ0 − kj|.
(8.9)
36
A. Kiselev, Y. Last, B. Simon
Conclusion of the Proof of Theorem 8.2. We need to verify (8.4) for q = 1, 2, 3. Vω (n) sin(2θ¯ω (n)) ≡ Zn has the form (8.5) and E(Zn2 ) ≤ Cn−2α , so by Lemma 8.4, for a.e. ω, X n −2α j . |C1 (ω)| = o j=1
¯ − 1 sin (2θ)] also has the form of (8.4) since E(Vω2 − [Vω (n)2 − E(Vω (n)2 )][sin (θ) 2 2 E(Vω )) = 0. Since V is bounded, 2
2
E((V 2 − E(V 2 ))2 ) ≤ CE(V 2 ). Thus, for a.e. ω, |C2 (ω)| = o
X n
j −2α
j=1
also. Finally, we will show n X
j −2α cos(4θ¯ω (j)) = o
j=1
X n
j −2α ,
j=1
which proves (8.4) for q = 3. By hypothesis on k, 4k ∈ / Zπ so Lemma 8.5 applies. Let q` be as in that lemma. Note next that by hypothesis (iii) and Proposition 2.4 for j large, |θω (j + 1) − θω (j) − k| ≤ C0 j −2α/3 .
(8.10)
n0 ≥ q`2
(8.11)
Pick n0 so and
−2α/3
4C0 n0
≤ q`−2 .
(8.12)
Suppose N = n0 + Kq` . Then N X q` X −2α K X −2α j cos(4θω (j)) = (n0 + mq` + j) cos(4θω (mq` + j)) j=n0 +1
m=0 j=1
= A 1 + A2 , where A1 is what we get by replacing (n0 + kq` + j)−2α by (n0 + kq` )−2α and A2 is the difference. By Lemma 8.5, (8.10), and (8.12), A1 ≤
K X
(n0 + kq` )−2α [1 + 1],
k=0
while using |(n0 + kq` + j)−2α − (n0 + kq` )−2α | ≤ (n0 + kq` )−2α jn−1 0 and (8.11),
Modified Pr¨ufer and EFGP Transforms
A2 ≤
K X
(n0 +
37
kq` )−2α q`2 n−1 0
≤
k=0
K X
(n0 + kq` )−2α .
k=0
Thus for any N , N N X X −2α −α j cos(4θ (j)) ≤ C + 3q j −2α , ω ` ` j=1
1
and so lim
X N
N →∞
j
−2α
j=1
−1 X N −2α j cos(4θω (j)) ≤ 3q`−α 1
uniformly in ω. Since we can take q` → ∞ by Lemma 8.5, the lim is 0. Theorem 8.6. Suppose that (i)–(iv) hold with α < uous parameter. Then for a.e. ω:
1 2
but we consider V (1) as a contin-
(1) For a dense Gδ of values of V (1), Hω has purely singular continuous spectrum in (−2, 2). (2) For Lebesgue a.e. value of V (1), Hω has dense pure point spectrum in (−2, 2) and the eigenfunctions obey Hω u = 2 cos(km )u with (1 − 2α)λ2 log(|u(n)2 + u(n + 1)2 |1/2 ) = − . n→∞ |n|1−2α 8 sin2 (km ) lim
(8.13)
If we consider a whole-line problem with independent Vω (n), where both {Vω (n)}∞ n=1 and V˜ω (n) ≡ Vω (−n), n = 1, 2, . . . obey hypotheses (i)–(iv) and Vω (0) has a purely a.c. density, then for a.e. ω, Hω has dense pure point spectrum in (−2, 2) and (8.13) holds as |n| → ∞. Remark. This strengthens the result originally proven in [31] and improved in [7] in two ways. First, we get the explicit constant in (8.13). Second, we only require one Vω ( · ) to have an a.e. distribution. Proof. By Theorem 8.2 and Fubini’s theorem for a.e. ω, we have for a.e. k ∈ (0, π), log kT (n)k (1 − 2α)λ2 = . n→∞ n1−2α 8 sin2 (k) lim
Thus by Theorem 8.3 of [22], there is an L2 -solution obeying (8.13). The theorem follows from general principles on rank one perturbations [12, 4, 5, 28]. The case α = 21 has an extra subtlety we will need to deal with, using an argument modeled on Kotani-Ushiroya [21]. The following replaces an explicit but complex formula they use for the projection onto a decaying solution (and fills in a gap in their argument): Lemma 8.7. Let uθ = (cos θ, sin θ) in R2 . For any unimodular matrix A with kAk > 1, let θ(A) be the unique θ ∈ (− π2 , π2 ] with kAuθ k = kAk−1 . Define ρ(A) = kAu0 k/kAuπ/2 k. Let An be a sequence of unimodular matrices with kAn k → ∞ and kAn+1 A−1 n k/kAn kkAn+1 k → 0 as n → ∞. Let ρn = ρ(An ), θn = θ(An ). Then:
38
A. Kiselev, Y. Last, B. Simon
(i) θn has a limit θ∞ if and only if limn→∞ ρn ≡ ρ∞ exists (ρ∞ = ∞ is allowed, but then we only have |θn | → π2 ). (ii) Suppose θn has a limit θ∞ 6= 0, π2 (equivalently, ρ∞ 6= 0, ∞). Then lim
log kAn u∞ k = −1 log kAn k
(8.14)
log |ρn − ρ∞ | ≤ −2. log kAn k
(8.15)
n→∞
if and only if lim
n→∞
Remark. Consider
An =
cosh(n) (−1)n sinh(n)
(−1)n sinh(n) . cosh(n)
Then ρ(An ) ≡ 1 and kAn k → ∞ but θn = (−1)n+1 ( π4 ) does not have a limit. This shows that the condition kAn+1 A−1 n k/kAn kkAn+1 k → 0 is required. Indeed, in this case that limit is 1. Kotani-Ushiroya miss this issue. Proof. (i) Note first that kAn uθ k2 = kAn k2 sin2 (θ − θn ) + kAn k−2 cos2 (θ − θn ). Thus, ρn =
tan2 (θn ) + kAn k−4 . 1 + kAn k−4 tan2 (θn )
(8.16)
(8.17)
It follows that ρn has a finite limit ρ∞ if tan2 (θn ) has a finite limit. By writing ρ−1 n =
cot 2 (θn ) + kAn k−4 , 1 + kAn k−4 cot2 (θn )
this is true also for ρn → ∞ and tan2 (θn ) → ∞. Pick η ∈ [0, π2 ] so tan2 (θn ) → tan2 (η). If η = 0, then θn → 0, and if η → π2 , then |θn | → π2 because of the continuity of tan(θ) on [− π2 , π2 ]. If 0 < η < π2 , we only have |θn | → η and have to worry about the sign (see the remark above). In (8.16), take θ = θn+1 and see that 2 2 sin2 (θn+1 − θn ) ≤ kAn k−2 kAn A−1 n+1 k kAn+1 uθn+1 k 2 = kAn k−2 kAn+1 k−2 kAn+1 A−1 n k −1 −1 since An A−1 n+1 is unimodular, and thus kAn An+1 k = kAn+1 An k. Thus by hypothesis,
sin2 (θn+1 − θn ) → 0. This, together with |θn | → η ∈ (0, π2 ), implies that θn has a limit. (ii) By (8.16), we have that (8.14) holds if and only if lim
n→∞
log |θn − θ∞ | ≤ −2. log kAn k
Since θ∞ 6= 0, π, this is true if and only if
(8.18)
Modified Pr¨ufer and EFGP Transforms
lim
n→∞
39
log | tan2 (θn ) − tan2 (θ∞ )| ≤ −2. log kAn k
By (8.17) and θ∞ 6= π, || tan2 (θn ) − tan2 (θ∞ )| − |ρn − ρ∞ || ≤ CkAn k−4 . Thus, (8.18) holds if and only if (8.15) holds.
Lemma 8.8. Suppose the hypotheses of Theorem 8.2 hold with α = 21 and k 6= is fixed. Then for a.e. ω, there exists an initial condition uθ(ω) so that
π 2π 3π 4, 4 , 4
λ2 log kT2 cos(k) (n, 0)uθ(ω) k =− . n→∞ log(n) 8 sin2 (k) lim
Remark. As noted in [22] (and gotten incorrectly in [21]), Ruelle’s deterministic argument doesn’t ever suffice in this kT k ∼ nγ case. If An is a sequence of unimodular matrices with limn→∞ log kAn k/ log(n) = γ, then [22] has explicit examples (even coming from deterministic Schr¨odinger operators) for each γ > 21 where the decaying solution only obeys limn→∞ log kAn u∞ k/ log(n) = −γ + 1. It also appears one needs γ > 23 to be sure of the existence of decaying solutions. But following [21], the probabilistic argument here can replace Ruelle’s argument. 2
Proof. Let β = 8 sinλ2 (k) . Let R1 (n) and R2 (n) be the R’s associated to θ = 0 and θ = 21 . By the proof of Theorem 8.2 for a.e. ω, lim
n→∞
log kRi (n)k = β. log(n)
(8.19)
Let θi (n) be the corresponding EFGP angles. By (2.7), R1 (n)R2 (n) sin(θ1 (n) − θ2 (n)) = sin(k)[u1 (n)u2 (n − 1) − u1 (n − 1)u2 (n)] = −1 (by the initial conditions R1 (1) = R2 (1) = 0, θ1 (1) = 0, θ2 (1) = Wronskian. Thus by (8.19) for a.e. ω, lim
n→∞
Let ρn =
R1 (n) R2 (n) .
π 2)
and constancy of the
log |θ1 (n) − θ2 (n)| = −2β. log(n)
Then by (2.12c),
Lω (n) ≡ [log ρ(n + 1) − log ρ(n)] = log(1 + A1 (n)) − log(1 + A2 (n)), where Ai (n) = − Define
Vω (n) Vω (n)2 sin(2θi,ω (n)) + sin2 (θi,ω (n)). sin(k) sin2 (k)
F (a, θ) = log(1 − a sin(2θ) + a2 sin2 (θ)).
By a finite Taylor expansion, F (a, θ) =
J−1 X j=1
aj Pj (θ) + O(aJ )
(8.20)
40
A. Kiselev, Y. Last, B. Simon
with P1 (θ) = sin(2θ) and the P ’s, C ∞ in θ. Fix > 0 so for n large, use (8.20) to see that |θ1 − θ2 | = o(n−2β+ ). Choosing J so n−J/3 = o(n−2β−1 ), we see that Lω (n) = −
Vω (n) [sin(2θ1 (n)) − sin(2θ2 (n))] + O(n−2β−1+ ). sin(k)
Since θj (n) depend only on {Vω (k)}k≤n−1 , we can apply part (3) of Lemma 8.4 (with 2α = 1 + 2β − ) to see that for a.e. ω, lim
N X
N →∞
exists and
Lω (n)
(8.21)
1
X ∞ ≤ Cω N −2β+ . L (n) ω
(8.22)
N
By (8.21), lim
R1 (n) R2 (n)
≡ ρ∞ exists and is different from 0 and ∞. Moreover, by (8.22), lim
log |ρ(n) − ρ(∞)| ≤ −2β. log(n)
Lemma 8.7 completes the proof.
Theorem 8.9. Suppose (i)–(iv) hold with α = 21 . Then, (1) For a.e. ω, the essential spectrum of Hω is [−2, 2] and the absolutely continuous spectrum of Hω is empty. (2) If |λ| ≥ 2 and Vω (1) has an absolutely continuous distribution, then for a.e. ω, Hω has dense point spectrum and only dense point spectrum in (−2, 2). (3) If |λ| < 2 and Vω (1) has an absolutely continuous distribution, then for a.e. ω, Hω has purely singular continuous spectrum in {E | |E| < (4 − λ2 )1/2 } and only dense pure point spectrum in {E | (4 − λ2 )1/2 < |E| < 2}. In either case (2) or (3), in the region of point spectrum, there are almost surely eigenvectors of power decay n−β with β=
λ2 . 8 − 2E 2
(8.23)
Remark. This theorem extends results of Delyon, et al. [7], Delyon [6], and KotaniUshiroya [21]. In particular, [7] conjectured that there is a region of point spectrum near E = ±2 no matter how small λ is. Proof. By Theorem 8.2, limn→∞ kTω (0, n)k = ∞ for a.e. E for a.e ω, so by Theorem 1.1, we conclude (3). By Lemma 8.8, for a.e. pairs (ω, E), there is a unique λ2 1 2 decaying solution with rate of decay n−β with β = 8−2E 2 . If β > 2 , this is ` and we 1 have potential point spectrum. If β < 2 , there is no `2 solution. The general theory of rank one perturbations ([32, 5]) then yields (2) and (3). We can compute the precise Hausdorff dimension of the singular continuous spectral measures in this case:
Modified Pr¨ufer and EFGP Transforms
41
Theorem 8.10. Fix λ < 2 and a model obeying (i)–(iv) with α = 21 . In the region |E| ≤ (4 − λ2 )1/2 , define 4 − E 2 − λ2 d(E, λ) = . 4 − E2 Suppose Vω (1) has an absolutely continuous density. Then for a.e. ω, the spectral measure, µ, has dimension d(E, λ) at E in the sense that for any , there is a δ so that µ(A) = 0 if A is a subset of (E − δ, E + δ) of Hausdorff dimenion less than (d − ), and there is a subset B of Hausdorff dimension less than (d + ), so µ((E − δ, E + δ)\B) = 0. PL Proof. Let kukL = ( j=1 u(j)2 )1/2 . By the general theory of rank one perturbations, Theorem 8.2, Lemma 8.8, and the assumption of Vω (1) for a.e. ω, µ is supported on the set of energies where most solutions grow as nβ and one decays as n−β , where β(E, λ) is given by (8.23). The hypothesis for singular spectrum is precisely β < 21 . Since β < 21 , ku1 kL ∼ L−β L1/2 while ku2 kL ∼ Lβ L1/2 , where a ∼ b is shorthand for lim log(a) log(b) = 1. The Jitomirskaya-Last version [16, 17] of the Gilbert-Pearson [11] theory says that the Borel transform of the spectral measure is supported on the set of E’s, where ku2 kL , (8.26) |m(E + i)| ∼ ku1 kL and E is given by ku1 kL ku2 kL =
1 2
(the ∼ in (8.26) holds in the strong sense that the ratio lies in the interval (5 − √ 24 )). Thus, ∼ L−1 and (8.26) says that
(8.27) √
24 , 5 +
|m(E + i)| ∼ −2β . Since β is continuous, the theory in [3] then says that the local dimension is given by 1 − 2β as claimed. 9. Random Decaying Potentials: The Continuum Case Having done the discrete random case, we will only sketch the continuum case. We will specialize to a situation where {V (x)}n≤x
n+1
|Vω (y)| dy.
an (ω) = n
Suppose (i) E(Vω (x)) = 0 for each x, P 2 Can ) < ∞ for all C > 0, (ii) n E(an e (iii) {V (x)}n≤x
(9.1)
42
A. Kiselev, Y. Last, B. Simon 2
d 2 Then for a.e. ω, − dx 2 + Vω (x) on L (0, ∞) has purely absolutely continuous spectrum on (0, ∞) for any boundary condition.
Remarks. 1. Our methods imply for all E > 0 and a.e. ω, TE (n) is bounded, and that d2 implies − dx 2 +Vω (x) is limit point at infinity, so we need not worry about self-adjointness issues. R∞ 2. A simple example where (ii) holds is if supω,x |Vω (x)| < ∞ and E( 0 V (y)2 dy) < ∞. Proof. By (2.4), R4 (n + 1) = R4 (n) exp(Bn (ω)), where Bn (ω) =
2 k
Z
(9.2)
n+1
Vω (x) sin(2θω (x)) dx.
(9.3)
2 an (ω). k
(9.4)
n
By (2.3), |θω (x) − θω (n) − k(x − n)| ≤ Using
|ex − 1 − x| ≤ 21 x2 ex ,
we obtain from (9.2)–(9.4), R4 (n + 1) ≤ R4 (n)(1 + Ca2n eCan ) + Qn , Z n+1 2 Qn = R4 (n) Vω (x) sin(2θω (n) + 2k(x − n)) dx k n
(9.5a) (9.5b)
for some constant C uniformly bounded for k in any compact of (0, ∞). Since Vω (x) is independent of {Vω (y)}y≤n , it is independent of R(n) and θω (n), and so E(Qn ) = 0. Moreover, an is independent of R(n), so (9.5) implies that E(R4 (n + 1)) ≤ E(R4 (n))E(1 + Ca2n eCan ). By condition (ii), we see that lim E(R4 (n)) < ∞
n→∞
with bounds uniform in k on compacts of (0, ∞). Thus by Fatou’s lemma, for a.e. ω, Rb lim a Rω4 (n, k) dk < ∞ and so the spectrum is purely absolutely continuous by Theorem 1.3. Theorem 9.2. Let f be supported on (0, 1) and let Vω (x) =
∞ X
(n + 1)−α Xn (ω)f (x − n),
n=0
where {Xn (ω)} are independent, identically distributed bounded variables of mean zero / Zπ, and 0 < α ≤ 21 . Then for 4k ∈ 2 Z log kT (n, 0)k E(Xn2 ) 1 iky Pn −α = f (y)e dy . (9.6) lim 2 n→∞ 8k 0 j=1 j
Modified Pr¨ufer and EFGP Transforms
43
Remarks. 1. This implies pure point spectrum for a.e. ω if α < 21 . 2. If α = 21 , we get singular continuous spectrum for large E and pure point spectrum R1 R1 for small E (assuming 0 f (y) dy 6= 0 or 0 yf (y) dy 6= 0) and no a.c. spectrum. Sketch. Define θn (y) = θ(n) + ky and δθn (y) = −(n + 1)
Z
−α
y
f (y) sin2 (θn (y)) dy.
Xn
(9.7)
0
By (2.3),
|θn (n + y) − θn (y) − δθn (y)| = O(n−2α )
for y ∈ (0, 1). Plugging this into (2.4), we find that log R(n + 1) − log R(n) = Yn(1) + Yn(2) + O(n−2α ), where Yn(1) = and Yn(2) =
(n + 1)−α Xn (ω) 2k
(n + 1)−α Xn (ω) 2k
Z
Z
1
f (y) sin(2θn (y)) dy 0
1
2f (y) cos(2θn (y))(δθn )(y) dy. 0
By using Lemmas 8.2 and 8.3, one sees that X X n n (1) −2α (j + 1) Yj → 0 j=0
j=0
for a.e. ω. The same lemmas replace Xn (ω)2 by E(Xn2 (ω)) in Yn(2) . So if we let $ Pn let us −2α ) terms, we see that indicate equal up to o( j=0 (j + 1) log R(n) $
n−1 X
(Yn(3) + Yn(4) ),
j=0
where we use sin2 (θn (y)) = 21 − 21 cos(2θn (y)) and let Yn(3) indicate the − 21 cos(2θ) terms and Yn(4) the 21 terms. By an argument analogous to the one in the proof of Theorem 9.2 P that used Lemma 8.5, Yn(4) $ 0 because k ∈ / Zπ. As in (6.5), we get log R(n) $
n−1 X j=0
(j + 1)−2α E(Xn (ω)2 ) 4k 2
Z
2
1
f (y) cos(2θj (y)) dy
.
0
As in the proof of Lemma 6.2, this last square is Z 2 1 1 iky f (y)e 2 0 plus a term that has cos(4θj (y)), which we can handle using Lemma 8.5.
44
A. Kiselev, Y. Last, B. Simon
References 1. Carmona, R.: One-dimensional Schr¨odinger operators with random or deterministic potentials, New spectral types. J. Funct. Anal. 51, 229–258 (1983) 2. Chulaevsky, V. and Spencer, T.: Positive Lyapunov exponents for a class of deterministic potentials. Commun. Math. Phys. 168, 455–466 (1995) 3. del Rio, R., Jitomirskaya, S., Last, Y. and Simon,B.: Operators with singular continuous spectrum, IV. Hausdorff dimensions, rank one perturbations, and localization. J. d’Analyse Math. 69, 153–200 (1996) 4. del Rio, R., Makarov, N. and Simon, B.: Operators with singular continuous spectrum, II. Rank one operators. Commun. Math. Phys. 165, 59–67 (1994) 5. del Rio, R., Simon, B. and Stolz, G.: Stability of spectral types for Sturm-Liouville operators. Math. Research Lett. 1, 437–450 (1994) 6. Delyon, F.: Apparition of purely singular continuous spectrum in a class of random Schr¨odinger operators. J. Statist. Phys. 40, 621–630 (1985) 7. Delyon, F., Simon, B. and Souillard, B.: From power pure point to continuous spectrum in disordered systems. Ann. Inst. H. Poincar´e 42, 283–309 (1985) 8. Eastham, M.S.P. and Kalf, H.: Schr¨odinger-type Operators With Continuous Spectra. London: Pitman Books, 1982 9. Eggarter, T.: Some exact results on electron energy levels in certain one-dimensional random potentials. Phys. Rev. B5, 3863–3865 (1972) 10. Gilbert, D.J.: On subordinacy and analysis of the spectrum of Schr¨odinger operators with two singular endpoints. Proc. Roy. Soc. Edin. 112A, 213–229 (1989) 11. Gilbert, D.J. and Pearson, D.B.: On subordinacy and analysis of the spectrum of one-dimensional Schr¨odinger operators. J. Math. Anal. 128, 30–56 (1987) 12. Gordon, A.Ya.: Pure point spectrum under 1-parameter perturbations and instability of Anderson localization. Commun. Math. Phys. 164, 489–505 (1994) 13. Gredeskul, S.A. and Pastur, L.A.: Behavior of the density of states in one-dimensional disordered systems near the edges of the spectrum. Theor. Math. Phys. 23, 132–139 (1975) 14. Hardy, G.H. and Wright, E.M.: An Introduction to the Theory of Numbers, 5th ed., Oxford: Oxford University Press, 1979 15. Harris W.A. and Lutz, D.A.: Asymptotic integration of adiabatic oscillator. J. Math. Anal. Appl. 51, 76–93 (1975) 16. Jitomirskaya, S. and Last, Y.: Dimensional Hausdorff properties of singular continuous spectra. Phys. Rev. Lett. 76, 1765–1769 (1996) 17. Jitomirskaya, S. and Last, Y.: Power law subordinacy and singular spectra I. Half-line operators. In preparation 18. Kahane, J.P.: Some Random Series of Functions. Cambridge: Cambridge University Press, 1985 19. Kiselev, A.A.: Absolutely continuous spectrum of one-dimensional Schr¨odinger operators and Jacobi matrices with slowly decreasing potentials. Commun. Math. Phys. 179, 377–400 (1996) 20. Kiselev, A., Remling, C. and Simon, B.: Effective perturbation methods for one-dimensional Schr¨odinger operators. Preprint 21. Kotani, S. and Ushiroya, N.: One-dimensional Schr¨odinger operators with random decaying potentials. Commun. Math. Phys. 115, 247–266 (1988) 22. Last, Y. and Simon, B.: Eigenfunctions, transfer matrices, and absolutely continuous spectrum of onedimensional Schr¨odinger operators. To appear in Invent. math. 23. Molchanov, S.: One-dimensional Schr¨odinger operators with sparse potentials. Preprint 24. Naboko, S.: Dense point spectra of Schr¨odinger and Dirac operators. Theor. Math. Phys. 68, 18–28 (1986) 25. Naboko, S. and Yakovlev, S.I.: On the point spectrum of discrete Schr¨odinger operators. Func. Anal. Appl. 26, 85–88 (1992) 26. Pastur, L. and Figotin, A.: Spectra of Random and Almost-Periodic Operators. Berlin: Springer, 1992 27. Pearson, D.B.: Singular continuous measures in scattering theory. Commun. Math. Phys. 60, 13–36 (1978) 28. Pearson, D.B.: Pathological spectral properties. Mathematical Problems in Theoretical Physics, Proc. Internat. Conf. Math. Phys., Lausanne, Lecture Notes in Phys. 116, 1979, Berlin–New York: Springer, 1980, pp. 49–51 29. Remling, C.: A probabilistic approach to one-dimensional Schr¨odinger operators with sparse potentials. Commun. Math. Phys. 185, 313–323 (1997)
Modified Pr¨ufer and EFGP Transforms
45
30. Ruelle, D.: Ergodic theory of differentiable dynamical systems. Publ. Math. IHES 50, 275–306 (1979) 31. Simon, B.: Some Jacobi matrices with decaying potentials and dense point spectrum. Commun. Math. Phys. 87, 253–258 (1982) 32. Simon, B.: Spectral analysis and rank one perturbations and applications. CRM Lecture Notes Vol. 8, J. Feldman, R. Froese, L. Rosen, eds., Providence, RI: Amer. Math. Soc., 1995, pp. 109–149 33. Simon, B.: Bounded eigenfunctions and absolutely continuous spectra for onedimensional Schr¨odinger operators. Proc. Amer. Math. Soc. 124, 3361–3369 (1996) 34. Simon, B.: Some Schr¨odinger operators with dense point spectrum. Proc. Amer. Math. Soc. 125, 203–208 (1997) 35. Simon, B. and Stolz, G.: Operators with singular continuous spectrum, V. Sparse potentials. Proc. Amer. Math. Soc. 124, 2073–2080 (1996) 36. Stolz, G.: Bounded solutions and absolute continuity of Sturn-Liouville operators. J. Math. Anal. Appl. 169, 210–228 (1992) 37. Weidmann, J.: Zur Spektral theorie von Sturm-Liouville Operatoren. Math. Z. 98, 268–302 (1967) Communicated by D. C. Brydges
Commun. Math. Phys. 194, 47 – 60 (1998)
Communications in
Mathematical Physics © Springer-Verlag 1998
Conformally Symplectic Dynamics and Symmetry of the Lyapunov Spectrum Maciej P. Wojtkowski1 , Carlangelo Liverani2 1 Department of Mathematics, University of Arizona, Tuscon, Arizona 85721, USA. E-mail:
[email protected] 2 Mathematics Department, University of Rome “Tor Vergata”, 00133 Rome, Italy. E-mail:
[email protected]
Received: 18 March 1997 / Accepted: 5 June 1997
Abstract: A generalization of the Hamiltonian formalism is studied and the symmetry of the Lyapunov spectrum established for the resulting systems. The formalism is applied to the Gausssian isokinetic dynamics of interacting particles with hard core collisions and other systems.
0. Introduction We study the symmetry of the Lyapunov spectrum in systems more general than Hamiltonian but closely related to the symplectic formalism. We call these systems conformally Hamiltonian. They are determined by a non-degenerate 2-form 2 on the phase space and a function H, called again a Hamiltonian. The form 2 is not assumed to be closed but it satisfies the following basic condition d2 = γ ∧ 2, for some closed 1-form γ. This condition guarantees that, at least locally, the form 2 can be multiplied by a nonzero function to give a bona fide symplectic structure. The skew-orthogonality of tangent vectors is preserved under multiplication of the form by any nonzero function, hence the name conformally symplectic structure. These ideas were known to geometers for a long time, see for example the paper of Vaisman [V]. The conformally Hamiltonian (with respect to the form 2) vector field ∇2 H is defined by the usual relation 2 (·, ∇2 H) = dH(·). The Hamiltonian function H is again a first integral of the system. In Sect. 2 we prove that for any conformally Hamiltonian system restricted to a smooth level set of the Hamiltonian the Lyapunov spectrum is symmetric, with symmetric exponents adding up to a constant. More precisely the direction of the flow has to be factored out. In Sect. 3 we extend this formalism to flows with collisions.
48
M. P. Wojtkowski, C. Liverani
In Sect. 1 we give an independent proof of the fact that for any conformally symplectic cocycle we have the symmetry of the Lyapunov spectrum. The first proof of this fact in the symplectic case goes back to Benettin et al. [B-G-G-S]. Our proof is based on an alternative description of Lyapunov exponents and it borrows an idea from [W1] (Lemma 1.2). In Sect. 4 we present examples which were recently the subject of several papers. We show that the Gaussian isokinetic dynamics can be viewed as a conformally Hamiltonian system, by which we immediately recover the results of Dettmann and Morriss [D-M 1, D-M 2], on the symmetry of the Lyapunov spectrum and the Hamiltonian character of the dynamics. We extend these results to systems with collisions, taking advantage of the fact that our formalism works equally well for collisions as it does for flows. Such systems were studied in the papers of Garrido and Gallavotti, [G-G], Dellago, Posch and Hoover, [D-P-H], and Bonetto, Gallavotti and Garrido, [B-G-G], and the symmetry of the Lyapunov spectra was demonstrated numerically. In such systems the symmetry of the Lyapunov spectrum was discovered as a pairing rule by Evans, Cohen and Morriss, [E-C-M, E-M]. Chernov et al, [Ch-E-L-S], studied rigorously the Lorentz gas of periodic scatterers with an electric external field in dimension 2. Latz, van Beijeren and Dorfman, [L-B-D], considered a thermostated random Lorentz gas in 3-dimensions and found the symmetry there. Let us note that we prove the symmetry of the Lyapunov spectrum for any invariant (ergodic) measure and not only for the SRB measure, which is the easiest to access numerically. We also show that the Gaussian isokinetic dynamics on a Riemannian manifold can be given the same treatment. The last application is to Nos´e–Hoover dynamics. We show that the Hoover equations can be naturally viewed as a conformally Hamiltonian system, thus giving another proof of the symmetry of Lyapunov spectra for this system. It was originally proven by Dettmann and Morriss, [D-M 3]. Finally let us note that our approach is one of several possible. Recently Choquard, [Ch], showed that the isokinetic and Nos´e-Hoover dynamics can be considered as Lagrangian systems. Even in the conformally symplectic framework one can keep the Hamiltonian unchanged and modify the form 2, or keep the form unchanged and modify the Hamiltonian, or keep both the form and the Hamiltonian unchanged, but change time on a level set of the Hamiltonian. We elaborate on that in Remark 2.1. We believe that our approach sheds new light on these issues. 1. Conformally Symplectic Group Pn Let ω = i=1 dpi ∧ dqi be the standard linear symplectic form in Rn × Rn . Proposition 1.1. For an invertible linear mapping S acting on R2n = Rn × Rn , the following are equivalent: (a) ω(Su, Sv) = βω(u, v) for some scalar β and all u, v ∈ R2n ; (b) ω(Su, Sv) = 0 if and only if ω(u, v) = 0, i.e., S preserves skew-orthogonality of vectors; (c) S maps Lagrangian subspaces of R2n onto Lagrangian subspaces. Proof. It is apparent that (a) implies (b) and that (b) is equivalent to (c). It remains to prove that (c) implies (a). By composing S with an appropriate linear symplectic map
Conformally Symplectic Dynamics
49
we can assume without loss of generality that S preserves the Lagrangian subspaces Rn × {0} and {0} × Rn , i.e., S is block diagonal. Moreover, again by multiplying by an appropriate linear symplectic map, we can assume that S is equal to identity on Rn × {0}. By (b) we conclude that S is diagonal on {0} × Rn . A simple calculation shows that to satisfy (b) this diagonal matrix must be a multiple of identity, which gives us (a). We will call a linear map from GL(R2n ) conformally symplectic if it satisfies one of the properties in Proposition 1.1. The group of all conformally symplectic maps will be denoted by CSp(R2n ). Let X be a measurable space with probabilistic measure µ and let T : X → X be an ergodic map. Let further A : X → GL(R2n ) be a measurable map such that Z log+ kA(x)kdµ(x) < +∞. (1.1) X
We define the matrix valued cocycle Am (x) = A(T m−1 x) . . . A(x). By the Oseledets Multiplicative Ergodic Theorem, [O], which in this generality was first proven by Ruelle, [R], there are numbers λ1 < . . . < λs , called the Lyapunov exponents of the measurable cocycle A(x), x ∈ X, and for almost all x a flag of subspaces {0} = V0 ⊂ V1 (x) ⊂ . . . ⊂ Vs−1 (x) ⊂ Vs = R2n , such that for all vectors v ∈ Vk (x) \ Vk−1 (x), lim
m→+∞
1 log kAm (x)vk := λ(v) = λk . m
In addition, denoting by dk the difference between the dimensions of Vk and Vk−1 (dk is called the multiplicity of the k th Lyapunov exponent), the following holds: Z s X d k λk = log |det(A(x))|dµ(x), (1.2) k=1
X
i.e., the sum of all Lyapunov exponents is equal to the average exponential rate of volume growth. Given a measurable cocycle A(x), x ∈ X, satisfying (1.1) and with values in the conformally symplectic group CSp(R2n ) we obtain a measurable function β = β(x) such that ω(A(x)u, A(x)v) = β(x)ω(u, v), (1.3) for all vectors u, v ∈ R2n . Let us define Z b := log |β(x)|dµ(x).
(1.4)
X
Lemma 1.1. If a measurable cocycle A(x), x ∈ X, satisfies (1.1) and it has values in the conformally symplectic group CSp(R2n ), then s X k=1
dk λk = nb.
50
M. P. Wojtkowski, C. Liverani
Proof. Since ω n is the volume form it follows from (1.3) that the determinant of A(x) is det A(x) = β(x)n . The lemma follows by applying (1.2).
Lemma 1.2. Given a measurable cocycle A(x), x ∈ X, satisfying (1.1) and with values in the conformally symplectic group CSp(R2n ), for each two non skew-orthogonal vectors u, v ∈ R2n , i.e., ω(u, v) 6= 0, we have λ(u) + λ(v) ≥ b. Proof. For the standard Euclidean norm k · k we have |ω(u, v)| ≤ kukkvk. From (1.3) we obtain m
m
ω(A (x)u, A (x)v) = ω(u, v)
m−1 Y
β(T i x).
i=0
Therefore, log |ω(Am (x)u, Am (x)v)| = log |ω(u, v)| +
m−1 X
log |β(T i x)|,
i=0
and 1 1 1 log |ω(Am (x)u, Am (x)v)| ≤ log kAm (x)uk + log kAm (x)vk. m m m Putting these relations together and using the Birkhoff Ergodic Theorem we conclude that Z log |β(x)|dµ(x) ≤ λ(u) + λ(v). b= X
The following lemma is obvious. We formulate it to streamline the proof of Theorem 1.4 where it is used twice. For a linear subspace X ⊂ R2n we denote by X ∠ the skew-orthogonal complement of X, i.e, X ∠ ⊂ R2n is the linear subspace containing all vectors v such that ω(u, v) = 0 for all u ∈ X. Since ω is assumed to be nondegenerate we have dim X + dim X ∠ = 2n
and
(X ∠ )∠ = X.
Lemma 1.3. Let U, V ⊂ R2n be two linear subspaces. If ω(u, v) = 0 for all u ∈ U and v ∈ V , then U ⊂ V ∠ , V ⊂ U ∠ and dim U + dim V ≤ 2n. Theorem 1.4. If a measurable cocycle A(x), x ∈ X, satisfies (1.1) and it has values in the conformally symplectic group CSp(R2n ), then we have the following symmetry of the Lyapunov spectrum: λk + λs−k+1 = b, where b is given by (1.4), and the multiplicities of λk and λs−k+1 are equal, for k = 1, 2, . . . , s. Moreover the subspace Vs−k is the skew-orthogonal complement of Vk .
Conformally Symplectic Dynamics
51
Proof. Let µ1 ≤ µ2 ≤ . . . ≤ µ2n be the Lyapunov exponents taken with repetitions according to their multiplicities. By Lemma 1.1 µ1 + µ2 + . . . + µ2n = nb. We can choose a flag of subspaces {0} = W0 ⊂ W1 (x) ⊂ . . . ⊂ W2n−1 (x) ⊂ W2n = R2n , such that dim Wl = l and for all vectors v ∈ Wl (x) \ Wl−1 (x) the Lyapunov exponent λ(v) = µl , for l = 1, 2, . . . , 2n. (Note that except in the case of all multiplicities equal to 1 there is a continuum of such flags.) Since for any l ≤ n, dim Wl + dim W2n−l+1 = 2n + 1, by Lemma 1.3 there are vectors u ∈ Wl and v ∈ W2n−l+1 such that ω(u, v) 6= 0. By continuity there must be ˜ v) ˜ 6= 0. It follows also vectors u˜ ∈ Wl \ Wl−1 and v˜ ∈ W2n−l+1 \ W2n−l such that ω(u, from Lemma 1.2 that µl + µ2n−l+1 ≥ b, for l = 1, 2, . . . , n. Adding these inequalities together, we get nb =
n X
(µl + µ2n−l+1 ) ≥ nb,
l=1
which shows that all the inequalities must be actually equalities. It follows immediately that for any k = 1, . . . , s, the multiplicities of λk and λs−k+1 are equal and λk +λs−k+1 = b. To show that the subspace Vs−k is the skew-orthogonal complement of the subspace Vk we observe that ω(u, v) = 0 for any u ∈ Vk and v ∈ Vs−k . Indeed, if this is not the case we could use Lemma 1.2 to claim that λk + λs−k ≥ b, which leads to the contradiction b = λk + λs−k+1 > λk + λs−k ≥ b. We can now apply Lemma 1.3 and we obtain Vs−k ⊂ V∠ k . Since the dimensions of these . subspaces are equal we must have Vs−k = V∠ k 2. Conformally Symplectic Manifolds and Conformal Hamiltonian Flows Let M be a smooth manifold of even dimension. A conformally symplectic structure on M is a differentiable 2-form 2 which is non-degenerate and has the following basic property: d2 = γ ∧ 2, (2.1) for some closed 1-form γ. A manifold with such a form 2 is called conformally symplectic. The origin of this name becomes clear when one observes that locally γ = dU for some smooth function U and d(e−U 2) = 0, i.e. e−U 2 defines a bona fide symplectic structure. For a given function H : M → R, called a Hamiltonian, let us consider a vector field ∇2 H defined by the usual relation 2(·, ∇2 H) = dH.
(2.2)
We will call it the conformally Hamiltonian vector field, or conformally symplectic, or simply a Hamiltonian vector field when the conformally symplectic structure is clearly
52
M. P. Wojtkowski, C. Liverani
chosen. Note that our definition does not coincide with the definition of a Hamiltonian vector field from [V]. Let 8t denote the flow defined by the vector field F = ∇2 H. The Hamiltonian function H is a first integral of the system. Indeed we have d H = dH(∇2 H) = 2(∇2 H, ∇2 H) = 0. dt Let us consider the Lie derivative of the form 2 in the direction of vector field F , i.e., (LF 2) (ξ, η) :=
d 2(D8u ξ, D8u η)|u=0 . du
Theorem 2.1. For a Hamiltonian vector field F = ∇2 H we have LF 2 = γ(F )2 + γ ∧ dH.
(2.3)
Proof. We will use the Cartan formula ([A-M-R]) LF = iF d + diF , where iF is the interior and d the exterior derivative. (For a differential m-form ζ the interior derivative iF ζ is the differential (m − 1)-form obtained by substituting F as the first vector argument of ζ.) We have iF 2 = −dH and we get immediately LF 2 = iF d2 − d2 H = iF (γ ∧ 2) = γ(F )2 + γ ∧ dH. Let us restrict the flow 8t to one smooth level set of the Hamiltonian, M c = {z ∈ M |H(z) = c}. In particular we assume that on M c the differential dH and the vector field F do not vanish. For two vectors ξ, η from the tangent space Tz M c , we introduce w(t) = 2(Dz 8t ξ, Dz 8t η). By (2.3) we get d w(t) = γ(F (8t z))w(t), dt since dH vanishes on the tangent space Tz M c . We conclude that 2(Dz 8t ξ, Dz 8t η) = β(t)2(ξ, η),
(2.4)
for every ξ, η, from Tz M c and Z
t
β(t) = exp
u
γ(F (8 z))du . 0
(2.5)
Conformally Symplectic Dynamics
53
Remark 2.1. Let us note that under a non-degenerate time change a conformally Hamiltonian vector field is still conformally Hamiltonian with the same Hamiltonian function but with respect to a modified conformally symplectic form. More precisely if F = ∇2 H is a Hamiltonian vector field, then if the new time τ is related to the original time t by dτ = f, dt for some function f of the phase point, we get that the vector field f1 F is conformally e = f 2. Indeed symplectic with respect to the form 2 1 e = (d ln f + γ) ∧ 2 e and 2(·, e F ) = dH. d2 f Alternatively we can keep the same conformally symplectic form and modify the Hamiltonian separately on each level set. Indeed we have 1 1 1 (H − c) , 2(·, F ) = dH = d f f f where the last equality is valid only on the level set {H = c}. Finally, let us consider the symplectic form e−U 2. On the level set {H = c} we have d(e−U (H − c)) = e−U dH. It follows that on this level set e−U 2(·, F ) = d e−U (H − c) , and, as a result, the vector field F coincides locally with the Hamiltonian vector field given by the Hamiltonian e−U (H − c) (with respect to the symplectic form e−U 2). This observation provides an alternate way to derive (2.4) by using the preservation of the symplectic form e−U 2 by any Hamiltonian flow (with respect to this symplectic form). For a fixed level set M c we introduce the quotient of the tangent bundle T M c of M by the vector field F = ∇2 H, i.e., by the one dimensional subspace spanned by F . Let us denote the quotient bundle by TbM c . The form 2 factors naturally from T M c to TbM c , in view of (2.2). The factor form defines in each of the quotient tangent spaces Tbz M c , z ∈ M c , a linear symplectic form. The derivative of the flow preserves the vector field F , i.e., c
Dz 8t (F (z)) = F (8t z). As a result the derivative can be also factored on the quotient bundle and we call it the transversal derivative cocyle and denote it by At (z) : Tbz M c → Tb8t z M c . It follows immediately from (2.4) that the transversal derivative cocycle is conformally symplectic with respect to 2 (or more precisely the linear symplectic form it defines in the quotient tangent spaces). We fix an invariant probability measure µc on M c and assume that
54
M. P. Wojtkowski, C. Liverani
Z Mc
kDz 8t kdµc (z) < +∞.
Under this assumption the derivative cocycle has well defined Lyapunov exponents, cf. [O, R]. Then the transversal derivative cocycle has also well defined Lyapunov exponents which coincide with the former except that one zero Lyapunov exponent is skipped. We can immediately apply Theorem 1.4 to the transversal derivative cocycle and we get the following. Theorem 2.2. For a Hamiltonian flow 8t , defined by the vector field F = ∇2 H, restricted to one level set M c we have the following symmetry of the Lyapunov spectrum of the transversal derivative cocyle with respect to an invariant ergodic probability measure µ. Let {0} ⊂ V0 (z) ⊂ V1 (z) ⊂ . . . ⊂ Vs−1 (z) ⊂ Vs = Tbz M c be the flag of subspaces at z associated with the Lyapunov spectrum λ1 < λ2 < . . . < λs−1 < λs of the transversal derivative cocycle At (z), z ∈ M c . Then the multiplicities of λk and λs−k+1 are equal and R
λk + λs−k+1 = a, for k = 1, 2, . . . , s,
where a = M c γ(F (z))dµc (z). Moreover the subspace Vs−k is the skew-orthogonal complement of the subspace Vk . Note that the invariant measure µc can be supported on a single periodic orbit, so that Theorem 2.2 applies as well to the real parts of the Floquet exponents. To apply Theorem 1.4 it is enough to have linear symplectic forms in each of the quotient tangent spaces (to the level set), not necessarily coming from a conformally symplectic structure on the phase space. But then one needs to check directly how the transversal derivative cocycle acts on these forms, because we do not have the advantage of Theorem 2.1. This is essentially the line of argument in [D-M 1] and [D-M 3]. 3. Conformally Symplectic Flows with Collisions Let M be a smooth manifold with piecewise smooth boundary ∂M . We assume that the manifold M is equipped with a conformally symplectic structure 2, as defined in Sect. 2. Given a smooth function H on M with non vanishing differential we obtain the non vanishing conformally Hamiltonian vector field F = ∇2 H on M . The vector field F is tangent to the level sets of the Hamiltonian M c = {z ∈ M |H(z) = c}. We distinguish in the boundary ∂M the regular part, ∂Mr , consisting of the points which do not belong to more than one smooth piece of the boundary and where the vector field F is transversal to the boundary. The regular part of the boundary is further split into “outgoing” part, ∂M− , where the vector field F points outside the manifold M and the “incoming” part, ∂M+ , where the vector field is directed inside the manifold. Suppose that additionally we have a piecewise smooth mapping 0 : ∂M− → ∂M+ , called the collision map. We assume that the mapping 0 preserves the Hamiltonian, H ◦ 0 = H, and so it can be restricted to each level set of the Hamiltonian. We assume that all the integral curves of the vector field F that end (or begin) in the singular part of the boundary lie in a codimension 1 submanifold of M . We can now define a flow 9t : M → M , called a flow with collisions, which is a concatenation of the continuous time dynamics 8t given by the vector field F , and the
Conformally Symplectic Dynamics
55
collision map 0. More precisely a trajectory of the flow with collisions, 9t (x), x ∈ M , coincides with the trajectory of the flow 8t until it gets to the boundary of M at time tc (x), the collision time. If the point on the boundary lies in the singular part then the flow is not defined for times t > tc (x) (the trajectory “dies” there). Otherwise tc the trajectory is continued at the point 0(9 x) until the next collision time, i.e., for 0 ≤ t ≤ tc 0(9tc (x) x) , 9tc +t x = 8t 09tc x. We define a flow with collisions to be conformally symplectic, if for the collision map 0 restricted to any level set M c of the Hamiltonian we have 0∗ 2 = β2,
(3.1)
for some non vanishing function β defined on the boundary. More explicitly we assume that for every vector ξ and η from the tangent space Tz ∂M c to the boundary of the level set M c we have 2(Dz 0ξ, Dz 0η) = β2(ξ, η). We restrict the flow with collisions to one level set M c of the Hamiltonian and we denote the resulting flow by 9tc . This flow is very likely to be badly discontinuous but we can expect that for a fixed time t the mapping 9tc is piecewise smooth, so that the derivative D9tc is well defined except for a finite union of codimension one submanifolds of M c . We will consider only such cases. We choose an invariant measure in our system which satisfies the condition that all the trajectories that begin (or end) in the singular part of the boundary have measure zero. Usually there are many natural invariant measures satisfying this property. For instance we get one by taking a Lebesgue measure ν in RT M c and averaging it over increasing time intervals ( T1 0 9tc∗ νdt as T → +∞). Let us denote the chosen invariant measure by µc . This measure µc defines the measure µcb on the boundary ∂M c , which is an invariant measure for the section of the flow (Poincar´e map of the flow). With respect to the measure µc the flow 9tc is a measurable flow in the sense of the Ergodic Theory and we obtain a measurable derivative cocycle D9tc : Tx M c → T9tc x M c . We can define Lyapunov exponents of the flow 9tc with respect to the measure µc , if we assume that Z Z t log+ ||Dx 9c ||dµc (x) < +∞ and log+ ||Dy 0||dµcb (y) < +∞ Mc
c ∂M−
(cf. [O, R]). The derivative of the flow with collisions can be also naturally factored onto the quotient of the tangent bundle T M c of M c by the vector field F , which we denote by TbM c . Note that for a point z ∈ ∂M c the tangent to the boundary at z can be naturally identified with the quotient space. We will again denote the factor of the derivative cocycle by At (x) : Tbx M c → Tb9tc x M c . We will call it the transversal derivative cocycle. If the derivative cocycle has well defined Lyapunov exponents then the transversal derivative cocycle has also well defined Lyapunov exponents which coincide with the former ones except that one zero Lyapunov exponent is skipped. For a conformally symplectic flow with collisions the factor At (x) of the derivative cocycle on one level set changes the form 2 by a scalar, (2.3) and (3.1), so that we can immediately apply Theorem 1.4 and we get
56
M. P. Wojtkowski, C. Liverani
Theorem 3.1. For a conformally symplectic flow with collisions 9tc we have the following symmetry of the Lyapunov exponents for a given ergodic invariant probability measure µc . Let {0} ⊂ V0 (x) ⊂ V1 (x) ⊂ . . . Vs−1 (x) ⊂ Vs = Tbx M c be the flag of subspaces at x associated with the Lyapunov spectrum λ1 < λ2 < . . . < λs−1 < λs of the transversal derivative cocycle At (x), x ∈ M c . Then the multiplicities of λk and λs−k+1 are equal and λk + λs−k+1 = a + b, for k = 1, 2, . . . , s, R R where a = M c γ(F )dµc and b = τ1 ∂M c log |β(y)|dµcb (y). τ = ∂M c tc (y)dµcb (y) is − − the average collision time on the section of the flow. Moreover the subspace Vs−k is the skew-orthogonal complement to Vk . R
4. Applications A. Gaussian isokinetic dynamics. The equations of the system are (cf. [D-M 1]) q˙ = p, p˙ = E − αp, where α =
hE, pi . hp, pi
(4.1)
In these equations q describes a point in the multidimensional configuration space RN , p is the momentum (velocity) also in RN and h·, ·i is the arithmetic scalar product in RN . The field of force E = E(q) is assumed to be irrotational, i.e., it has locally a potential . function U = U (q), E = − ∂U ∂q P Let us denote by pdq the 1-form which defines the standard symplectic Pκ = structure ω = dκ = dp ∧ dq. We introduce the following 2-form 2=ω+
hE, dqi ∧ κ. hp, pi
We choose the Hamiltonian to be H = 21 hp, pi and we denote by F the vector field defined by (4.1). We have 2(·, F ) = dH, (4.2) but the form 2 does not give us a conformally symplectic structure because the relation (2.1) fails. To correct this setback we fix one level set of the Hamiltonian M c = {H = 1 2 hp, pi = c} and define another 2-form 2c = ω +
hE, dqi ∧ κ. 2c
Now we get a conformally symplectic structure. Indeed hE, dqi hE, dqi d2c = − ∧ 2c , and locally =d 2c 2c
−U 2c
.
Moreover on M c we still have 2c (·, F ) = dH so that the restriction of (4.1) to M c coincides with a conformally Hamiltonian system with respect to the 2-form 2c and with the Hamiltonian H = 21 hp, pi. We can immediately apply Theorem 2.1 and we obtain
Conformally Symplectic Dynamics
57
that for any invariant ergodic probability measure µc on M c the Lyapunov exponents λ1 < . . . < λs satisfy Z Z 1 λk + λs−k+1 = − hE, pidµc = αdµc . 2c M c Mc Note that if the vector field of force has a global potential, E = − ∂U ∂q , then by the R R 1 1 Birkhoff Ergodic Theorem the integral − 2c M c hE, pidµc = 2c M c dU (F )dµc is equal −U to the time average of dU 2c dt and so it must vanish. Another way to see it is that e defines a global symplectic structure and on Mc our flow is Hamiltonian with respect to this symplectic structure and a modified Hamiltonian e = e−U ( 1 hp, pi − c). H 2 Indeed as discussed in Remark 2.1 on M c we have e e−U 2c (·, F ) = dH. For a Hamiltonian flow the symmetric Lyapunov exponents must add up to zero. B. Gaussian isokinetic dynamics on a Riemannian manifold. For a given Riemannian P gij dqi dqj we canP naturally generalize the manifold N with the metric tensor ds2 = pdq is independent of form 2 to the cotangent bundle T ∗ N . Indeed the 1-form κ = the coordinate system, cf. [A], and for a given closed 1-form γ we put 2c = dκ −
1 γ ∧ κ. 2c
1 We get d2c = − 2c γ ∧ 2c . Taking γ = dU for some potential function (single or P ij multi-valued) and the Hamiltonian H = 21 g pi pj we obtain the Gaussian isokinetic dynamics, [Ch], on the level set H = c by the relation (4.2). We can repeat the discussion in part A and we conclude again that the Lyapunov exponents must be symmetric and they add up to zero, if the potential U is single-valued.
C. The Gaussian isokinetic dynamics with collisions. Let us consider n spherical particles in a finite box B contained in Rd or the torus Td . We assume that the particles interact with each other by the potential V (q1 , q2 , . . . , qn ) (qk ∈ B, k = 1, . . . , n denote the positions of the particles) and that they are subjected to the external fields given by the potentials Vk (qk ), k = 1, . . . , n. Further we assume that the particles have the radii r1 , . . . , rn , the masses m1 , . . . , mn , and that they collide elastically with each other and the sides of the box, which can be flat or curved. The last element in the description of the system is the Gaussian isokinetic thermostat. As described in Part A and B the Gaussian isokinetic thermostat gives rise to a conformally Hamiltonian flow with the Pn p2 Hamiltonian H = k=1 2mkk and an appropriate conformally symplectic structure. We will check below that the collisions in this system preserve the form 2c giving rise to a conformally symplectic flow with collisions. Theorem 3.1 can be thus applied to our system giving us the symmetry of the Lyapunov spectrum. We introduce the canonical change of variables which brings the kinetic energy into the standard form,
58
M. P. Wojtkowski, C. Liverani
xk =
√
mk qk , pk vk = √ . mk
The advantage of these coordinates is that although the collision manifolds in the configuration space become less natural, the collisions between particles (and the walls of the box) are given by the billiard rule in the configuration space. The equations of motions in the (x, v) coordinates are x˙ = v, (4.3) ∂U − αv, v˙ = − ∂x Pn (v) where U = U (x) = V + k=1 Vk is the total potential of the system and α = − dU hv,vi . We introduce the differential 2-form X 1 dv ∧ dx − dU ∧ hv, dxi. 2c = 2c As in part A we conclude that the form satisfies (2.1) and the system (4.3) restricted to M c coincides with the conformally Hamiltonian system defined by this form and the Hamiltonian H = 21 hv, vi. Proposition 4.1. The collision maps preserve the form 2c . Proof. A collision manifold is locally given by an equation of the form g(x) = 0, where g is some differentiable function on Rnd . Note that the general form of the collision map is the same for collisions of particles and the collisions with the sides of the box. Let n(x), for x ∈ {x ∈ Rnd |g(x) = 0}, denote the unit normal vector to the collision manifold in the configuration space. The collision map is defined as x + = x− , v + = v − − 2hv − , n(x− )in(x− ),
(4.4)
where the index + corresponds to the values of x and v after the collision and the index − to the values before the collision. As a result of these formulas we get immediately that (4.5) δx+ = δx− . It is well known, [W1, W2], that in an elastic collision the symplectic form ω is preserved. It remains to show the preservation of the second term in 2c . It follows immediately from (4.4) and (4.5), because hv + , δx+ i = hv − , δx− i − 2hv − , n(x− )ihn(x− ), δx− i, and the last term is zero since we only take the variations (δx− , δv − ) tangent to the collision manifold, i.e., δx− is orthogonal to n(x). The proposition is proven. It follows from Proposition 4.1 that also the form e−U 2c is preserved under collisions. Hence, as remarked in Parts A and B, if the potential U is singlevalued then the system restricted to one energy level coincides with a globally Hamiltonian system (with collisions) with respect to the symplectic form e−U 2c with the Hamiltonian function e = e−U ( 1 hp, pi − c). We conclude that the occurrence of dissipation in such equal to H 2
Conformally Symplectic Dynamics
59
systems is related to the topology of the configuration space (the multivaluedness of the potential U ). D. Nos´e-Hoover dynamics. The Nos´e Hamiltonian is, cf. [D-M 3], H(q, s; π, ps ) =
N X i=1
p2 πi2 + ϕ(q) + s + C ln s, 2 2mi s 2
withP a non-physical time denoted by λ and some constant C. The symplectic form is ω = dπ∧dq+dps ∧ds. Changing the variables as π = sp and σ = ln s the Hamiltonian becomes N X p2 p2i (4.6) + ϕ(q) + s + Cσ, H(q, σ; p, ps ) = 2mi 2 i=1 P P and the symplectic form is ω = eσ i dpi ∧ dqi + dps ∧ dσ + dσ ∧ ( i pi dqi ) . Note that now in the Hamiltonian the thermostat (σ, ps ) is decoupled from the system but the σ coupling is shifted to the symplectic form. We make finally the time change dλ dt = e . We choose not to change the Hamiltonian but rather to modify the 2-form, e−σ ω(·, eσ ∇ω H) = dH. We end up with the Hamiltonian (4.6) and the conformally symplectic structure X X dpi ∧ dqi + dps ∧ dσ + dσ ∧ ( pi dqi + ps dσ). 2 = e−σ ω = i
We have d2 = dσ ∧ 2. Note the similarity of 2 with the form used in the discussion of the isokinetic dynamics above. This form and the Hamiltonian give us the Hoover equations pi , q˙i = mi ∂ϕ p˙i = − − ps pi , ∂qi σ˙ = ps , X p2 i p˙s = − C. m i i On any level set we can drop the equation for σ since σ can be trivially obtained from other variables using the constancy of the Hamiltonian. By Theorem 2.1 we have the symmetry of the Lyapunov spectrum for this system reduced to one level of the Hamiltonian. Moreover the Lyapunov exponents add up to the time average of σ. ˙ This average must be zero, unless σ grows linearly, which is unlikely. Note that the Nos´e–Hoover system is open in the sense that arbitrarily large fluctuations of σ cannot be ruled out. Acknowledgement. We thank Dmitri Alexeevski, Federico Bonetto and Philippe Choquard for many enlightening discussions during our stay at the ESI in December of 1996. In addition, we thank David Ruelle for valuable comments. We are also grateful for the opportunities provided by the hospitality of the Erwin Schr¨odinger Institute in Vienna where this work was done. M.P.Wojtkowski has been partially supported by NSF Grant DMS-9404420.
60
M. P. Wojtkowski, C. Liverani
References [A]
Arnold, V.I.: Mathematical Methods in Classical Mechanics. Berlin–Heidelberg–New York: Springer Verlag, 1978 [A-M-R] Abraham, R., Marsden, J.E., Ratiu, T.: Manifolds, tensor analysis and applications. Berlin– Heidelberg–New York: Springer Verlag, 1988 [B-G-G-S] Benettin, G., Galgani, I., Giorgilli, A., Strelcyn, J.-M.: Lyapunov characteristic exponents for smooth dynamical systems and for Hamiltonian systems; a method for computing all of them. Meccanica 15, 9–20 (1980) [B-G-G] Bonetto, F., Gallavotti, G., Garrido, P.L.: Chaotic principle: An experimental test. Preprint (1996) [Ch] Choquard, Ph.: Lagrangian formulation of Nos´e–Hoover and of isokinetic dynamics. ESI report (1996) [Ch-E-L-S] Chernov, N.I., Eyink, G.L., Lebowitz, J.L., Sinai, Ya.G.: Steady-state electric conduction in the periodic Lorentz gas. Commun. Math. Phys. 154, 569–601 (1993) [D-P-H] Dellago, C.P., Posch, H.A., Hoover, W.G.: Lyapunov instability in a system of hard disks in equilibrium and nonequilibrium steady states. Phys. Rev. E 53, n.2, 1485–1501 (1996) [D-M 1] Dettmann, C.P., Morriss, G.P.: Proof of Lyapunov exponent pairing for systems at constant kinetic energy. Phys. Rev. E 53, 5541 (1996) [D-M 2] Dettmann, C.P., Morriss, G.P.: Hamiltonian formulation of the Gaussian isokinetic thermostat. Phys. Rev. E 54, 2495 (1996) [D-M 3] Dettmann, C.P., Morriss, G.P.: Hamiltonian reformulation and pairing of Lyapunov exponents for Nos´e–Hoover dynamics. Phys. Rev. E 55, 3693 (1997) [E-C-M] Evans, D.J., Cohen, E.G.D., Morriss, G.P.: Viscosity of a simple fluid from its maximal Lyapunov exponents. Physical Review 42A, 5990–5997 (1990) [E-M] Evans, D.J., Morriss, G.P.: Statistical Mechanics of nonequilibrium liquids. New York: Academic Press, 1990 [G-G] Garrido, P.L., Gallavotti, G.: Billiards correlation functions. J. Stat. Phys. 76, 549–586 (1994) [L-B-D] Latz, A., van Beijeren, H., Dorfman, J.R.: Lyapunov spectrum and the conjugate pairing rule for a thermostatted random Lorentz gas: Kinetic theory. Preprint (1996) [O] Oseledets, V.I.: A multiplicative ergodic theorem: Characteristic Lyapunov exponents of dynamical systems. Trans. Moscow Math. Soc. 19, 197–231 (1968) [R] Ruelle, D.: Ergodic theory of differentiable dynamical systems. Publ. Math. IHES 50, 27–58 (1979) [V] Vaisman, I.: Locally conformal symplectic manifolds. Int. J. Math.–Math. Sci. 8, 521–536 (1985) [W1] Wojtkowski, M.P.: Measure Theoretic Entropy of the system of hard spheres. Ergodic Theory and Dynamical Systems. 8, 133–153 (1988) [W2] Wojtkowski, M.P.: Systems of classical interacting particles with nonvanishing Lyapunov exponents. In: Lyapunov Exponents, Proceedings, Oberwolfach 1990, L. Arnold, H. Crauel, J.-P. Eckmann (Eds.), Lecture Notes in Math. Vol. 1486, 1991, pp. 243–262 Communicated by J. L. Lebowitz
Commun. Math. Phys. 194, 61 – 70 (1998)
Communications in
Mathematical Physics © Springer-Verlag 1998
The Pair Correlation Function of Fractional Parts of Polynomials? Ze´ev Rudnick1 , Peter Sarnak2 1 Raymond and Beverly Sackler School of Mathematical Sciences, Tel Aviv University, Tel Aviv 69978, Israel 2 Department of Mathematics, Princeton University, Princeton NJ 08544, USA
Received: 22 July 1997 / Accepted: 24 September 1997
Abstract: We investigate the pair correlation function of the sequence of fractional parts of αnd , n = 1, 2, . . . , N , where d ≥ 2 is an integer and α an irrational. We conjecture that for badly approximable α, the normalized spacings between elements of this sequence have Poisson statistics as N → ∞. We show that for almost all α (in the sense of measure theory), the pair correlation of this sequence is Poissonian. In the quadratic case d = 2, this implies a similar result for the energy levels of the “boxed oscillator” in the high-energy limit. This is a simple integrable system in 2 degrees of freedom studied by Berry and Tabor as an example for their conjecture that the energy levels of generic completely integrable systems have Poisson spacing statistics.
1. Introduction Hermann Weyl [11] proved that for an integer d ≥ 1 and an irrational α, the sequence of fractional parts αnd mod 1, n = 1, 2, . . . is equidistributed in the unit interval. A different aspect of the random behavior of the sequence has attracted attention recently: Are the spacings between members of the sequence distributed like those between members of a sequence of random numbers in the unit interval (or as some would say, do they have a “Poissonian” distribution)? This issue came up in the context of the distribution of spacings of the energy levels of integrable systems [1]. For the case d = 1 the spacings between the fractional parts of αn are essentially those of the energy levels of a two-dimensional harmonic oscillator [4, 2, 3]. For d = 2 the spacings are related to the spacings between the energy levels of the “boxed oscillator” [1], a particle in a ? Supported in part by grants from the U.S.-Israel Binational Science Foundation, the Israel Science Foundation, and the NSF.
62
Z. Rudnick, P. Sarnak
2-dimensional potential well with hard walls in one direction and harmonic binding in the other. The spacings of αn2 mod 1 were also investigated numerically in [5]. If d = 1 it is elementary that the consecutive spacings have at most 3 values [9, 10]. Hence the sequence is not random in this case. For d ≥ 2 the picture is very different. To explain it, we recall a basic classification of real numbers with regards to their Diophantine approximation properties: We say α is of type κ if there is c = c(α) > 0 so that |α − p/q| > c/q κ for all integers p, q. For rational α, κ = 1 and α is irrational if and only if κ ≥ 2. It is well known that almost all α (in the sense of measure theory) are of type κ = 2 + for all > 0. We will call such α “Diophantine”. For instance, algebraic irrationals are of this type (Roth’s theorem). In [7] we establish some results towards the conjecture that αnd mod 1 is Poissonian for any α of Diophantine type. In this note we examine the behavior for almost all α, which according to the above should be Poissonian. The statistic we examine is the pair correlation: The pair correlation density for a sequence of N numbers θ1 , . . . , θN ∈ [0, 1] which are equidistributed as N → ∞, measures the distribution of spacings between the θj at distances of order of the mean spacing 1/N . Precisely, if kxk = distance(x, Z) then for any interval [−s, s] set 1 n so . (1.1) R2 ([−s, s], {θn }, N ) = # 1 ≤ j 6= k ≤ N : kθj − θk k ≤ N N For random numbers θj chosen uniformly and independently, R2 ([−s, s], {θn }, N ) → 2s with probability tending to 1 as N → ∞. Our main result is that this holds for the sequence of fractional parts {αnd mod 1} for almost every α: Denoting by R2 ([−s, s], α, N ) the pair correlation sum (1.1) for this sequence, we show Theorem 1. For d ≥ 2, there is a set P ⊂ R of full Lebesgue measure such that for any α ∈ P , and any s ≥ 0, R2 ([−s, s], α, N ) → 2s,
N → ∞.
Remark 1.1. The proof given below does not provide (and we do not know of) any specific α which is provably in P . Remark 1.2. Already with the pair correlation we see the necessity of a condition on the type of α. For if there are arbitrarily large integers p, q so that 1 p , |α − | ≤ q 10q d+1 then R2 ([−s, s], α, N ) 6→ 2s. Indeed if we choose N = q, then for m 6= n ≤ N ,
d d d
d
d
n α − md α = (n − m )p + t(n − m )
q 10q d+1
d with Hence either n α − md α ≤ 1/10q = 1/10N if q divides nd − md , or
d |t| ≤ 1.
n α − md α ≥ 9/10q = 9/10N otherwise. Thus there are no normalized differences
N nd α − md α in the interval (1/10, 9/10) for this sequence of N = q.
Pair Correlation of Fractional Parts
63
The proof of the theorem follows the steps in [8] (where a similar assertion is proven for the values of binary quadratic forms). We first establish that as a function of α ∈ [0, 1], R2 ([−s, s], α, N )P→ 2s in L2 (0, 1). This together with standard bounds on the Weyl sums S(n, N ) = x≤N e(nαxd ) allows us to pass to almost everywhere convergence. In Sect. 4 we briefly discuss higher correlations and show that they do not converge in L2 to the expected value. Thus our approach does not lend itself directly to establishing almost everywhere convergence of the higher correlations. 2. Bounding the Variance Let f ∈ Cc∞ (R) be a test function and set R2 (f, {θn }, N ) := where FN (y) =
X
1 N
FN (θj − θk ),
(2.1)
1≤j6=k≤N
X
f (N (y + m)).
(2.2)
m∈Z
The function FN (y) is periodic and has a Fourier expansion 1 X b n e(ny). f FN (y) = N N
(2.3)
n∈Z
Hence R2 (f, {θn }, N ) =
1 X b n f N2 N n∈Z
X
e(n(θj − θk )).
(2.4)
1≤j6=k≤N
In particular, if θn = αnd mod 1, then the pair correlation function is given by 1 X b n sof f (n, N ), f R2 (f, α, N ) = 2 N N
(2.5)
n∈Z
where sof f (n, N ) :=
X
e nα(xd − y d ) .
(2.6)
1≤x6=y≤N
As a function of α, R2 (f, α, N ) is periodic and from (2.5) its Fourier expansion is X R2 (f, α, N ) = bl (N )e(lα), (2.7) l∈Z
where for l 6= 0, bl (N ) =
1 X N2 n6=0
X 1≤x6=y≤N n(xd −y d )=l
The mean of R2 (f, α, N ) over α ∈ [0, 1] is
n . fb N
(2.8)
64
Z. Rudnick, P. Sarnak
1 hR2 i = b0 (N ) = 2 N
X
fb(0) =
1≤x6=y≤N
1 1− N
fb(0),
(2.9)
so that Z hR2 i =
∞
−∞
f (x)dx + O
1 N
,
(2.10)
which is the expected value for a random sequence. We next estimate the variance of R2 (f, α, N ) as a function of α: Proposition 2. As a function of α ∈ [0, 1],
R2 (f, α, N ) − fb(0) N −1/2+
(2.11)
2
for any > 0, the implied constants depending on and f . Proof. It is easy to see from (2.8) that since f ∈ Cc∞ (R), the Fourier coefficients bl (N ) are negligable for l ≥ N d+1+δ for any fixed δ > 0. Also from (2.8) we have for l 6= 0, τ (|l|)2 , N2
bl (N )
(2.12)
where τ (|l|) is the numbers of divisors of |l|. This is because the factors of l determine n, x, y. We will use the well-known estimate τ (m) m ,
for any > 0.
Thus by Parseval
2
R2 (f, α, N ) − fb(0) =
fb(0) N
2
!2 +
X N l6=0
N2
|bl (N )|
06=|l|≤N d+1+δ X
N N2 N N2
|bl (N )|2
l6=0
X
=
X
N |bl (N )| + smaller order term N2
|bl (N )|
l6=0
X
1≤x6=y≤N n∈Z
1 b n | N −1+ . |f N2 N
(2.13)
Pair Correlation of Fractional Parts
65
3. Almost-Everywhere Convergence 3.1. Overview of the argument for Theorem 1. In order to prove Theorem 1 from the decay of the variance of the pair correlation, we first show that for each f ∈ Cc∞ (R), there is a set of full measure P (f ) ⊂ R so that for all α ∈ P (f ), R2 (f, α, Nm ) → fb(0)
(3.1)
for a subsequence Nm which grows faster than m. Indeed, fix δ > 0, and let {Nm } be a sequence of integers with Nm ∼ m1+δ . Set XN (α) = R2 (f, α, N ) − fb(0). By Proposition 2, kXN k2 N −1+ for all > 0 and so 2
∞ Z X m=1
1
|XNm (α)2 |dα < ∞. 0
Therefore (since |XNm |2 ≥ 0) Z 0
and so
P m
1
X
|XNm (α)| dα = 2
m
XZ m
1
|XNm (α)|2 dα < ∞, 0
|XNm |2 ∈ L1 (0, 1). Thus the sum is finite almost everywhere: X
|XNm (α)|2 < ∞,
for almost all α.
m
Therefore, XNm (α) → 0 as m → ∞ for almost all α, that is we have (3.1) on a set P (f ) of α’s which we may assume consists only of Diophantine numbers. To go from almost everywhere convergence along a subsequence to almost everywhere convergence, we will show that as a function of N , R2 (f, N, α) does not oscillate much for Diophantine α. More precisely, there is some ν > 0 so that if Nm ≤ n < Nm+1 then for Diophantine α, there is c(f, α) > 0 so that −ν |Xn (α) − XNm (α)| c(f, α)Nm . δ , this estimate in turn will follow from: Because 0 ≤ n − Nm ≤ Nm+1 − Nm Nm
Proposition 3. Let 0 < δ < 1/2d−1 . Then for all f ∈ Cc (R) and all α of Diophantine type, there is some c(f, α) > 0 so that for all 0 ≤ k ≤ N δ , |XN +k (α) − XN (α)| ≤ c(f, α)N −ν , where ν < 1/2d−1 − δ.
66
Z. Rudnick, P. Sarnak
Since XNm (α) → 0 for all α ∈ P (f ), which by throwing out a measure-zero subset we assumed consisted only of Diophantine α’s, Proposition 3 implies Xn (α) → 0 for all α ∈ P (f ). We will prove this proposition after finishing the proof of Theorem 1. What remains to do is to find one subset P ⊂ R of full measure for which R2 (f, α, N ) → R∞ f (x)dx for all α ∈ P and all f which are characteristic function of intervals [−s, s] −∞ (or in Cc∞ (R)). To do this, pick a (countable) sequence of positive fi ∈ Cc∞ (R) so that for eachRf ≥ 0 as above, there are subsequences {fi± } ⊂ {fi } which satisfy fi− ≤ f ≤ fi+ ∞ and −∞ (fi+ − fi− )(x)dx → 0. Take P := ∩i P (fi ) which is still of full measure. For every α we have R2 (fi− , α, N ) ≤ R2 (f, α, N ) ≤ R2 (fi+ , α, N ), R∞ R∞ and in addition for α ∈ P , we have R2 (fi± , α, N ) → −∞ fi± . Since −∞ fi± → R∞ R∞ f , this shows that R2 (f, α, N ) → −∞ f for α ∈ P and gives Theorem 1. −∞ The proof of Proposition 3 will occupy the rest of this section. 3.2. Estimates for Weyl sums. We Pbegin with some consequences of Weyl’s estimates for the “Weyl sums” S(n, N ) = x≤N e(nαxd ) which we will need. Throughout the remainder of this section, we set D = 2d−1 . Lemma 4. For α Diophantine, and M ≥ 1, we have X |S(n, N )|D M 1+ N D−1+ 1≤n≤M
for all > 0 (D = 2d−1 ). Proof. This follows from proof of Weyl’s inequality (see [6], Lemma 3). We will outline the steps. By repeated squaring, one finds that |S(n, N )|
D
N
D−1
+N
D−d
N X y1 ,...,yd−1 =1
1 min N, kd!nαy1 . . . yd−1 k
,
where k·k denotes the distance to the nearest integer. Now sum over n ≤ M , collecting together terms with the product d!ny1 . . . yd−1 having a given value m. The number of such terms is at most the divisor function τ (m) m . Since the maximal value of m is d!M N d−1 , we find X X 1 . |S(n, N )|D M N D−1 + M N D−d+ min N, kmαk (3.2) 1≤n≤M m≤d!M N d−1 Proceeding as in [6], we replace α by a rational approximation a/q with |α−a/q| ≤ 1/q 2 , and divide the range of summation into consecutive blocks of length q. This will give X 1 M N d−1 + 1 · (N + q log q). min N, kmαk q d−1 m≤d!M N
Inserting into (3.2) we get
Pair Correlation of Fractional Parts
X
67
|S(n, N )|
D
MN
D−1
+M N
1≤n≤M
D−d+
M N d−1 + 1 · (N + q log q). q (3.3)
Now choose q ≤ M N d−1 with |α − a/q| ≤ 1/qM N d−1 (so certainly |α − a/q| ≤ 1/q 2 so (3.3) holds). Since α is Diophantine, |α − a/q| 1/q 2+ which gives q (M N d−1 )1− . Therefore M N d−1 + 1 · (N + q log q) (M N d−1 )1+ , q and consequently
X
|S(n, N )|D M 1+ N D−1+
1≤n≤M
as required.
As an immediate consequence of this lemma, we get on repeatedly using the CauchySchwarz inequality that Corollary 5. For α Diophantine, and M ≥ 1, X |S(n, N )|2 M 1+ N 2−2/D+
(3.4)
1≤n≤M
and
X
|S(n, N )| M 1+ N 1−1/D+ .
(3.5)
1≤n≤M
3.3. Proof of Proposition 3. We first show XN +k (α) − XN (α) =
1 N2
X 0<|n|≤M
n {sof f (n, N + k) − sof f (n, N )} fb N + O M 2+ N −2+δ−2/D .
(3.6)
We use the representation (2.5), XN (α) =
1 X b n sof f (n, N ). f N2 N n6=0
Since f ∈ Cc∞ (R), its Fourier transform fb is rapidly decreasing and so on using the trivial estimate |sof f (n, N )| ≤ N 2 we see that for any b > 0, M = N 1+b , n X 1 sof f (n, N ) + rapidly decaying term. XN (α) = 2 fb N N 06=|n|≤M
Next we use |sof f (n, N )| ≤ N + |S(n, N )|2 and Corollary 5 to deduce that X X |sof f (n, N + k)| ≤ M N + |S(n, N + k)|2 M 1+ N 2−2/D . (3.7) 06=|n|≤M 06=|n|≤M
68
Z. Rudnick, P. Sarnak
Next we claim that n X 1 1 sof f (n, N + k) = 2 fb 2 (N + k) N +k N 06=|n|≤M
n sof f (n, N + k) fb N
X 06=|n|≤M
+ O(M 2+ N −2+δ−2/D ). Indeed, write 1 1 = 2 +O 2 (N + k) N
k N3
1 + O(N −3+δ ) N2
=
and nk n n n = + O( 2 ) = +O N +k N N N
M N 2−δ
,
so that for |n| ≤ M , k < N δ , fb
n b n =f +O N +k N
M N 2−δ
.
Therefore n sof f (n, N + k) N +k 06=|n|≤M n X 1 sof f (n, N + k) fb − 2 N N 06=|n|≤M X 1 M 1 b n +O = f + O( ) sof f (n, N + k) N2 N 3−δ N N 2−δ 06=|n|≤M n X 1 sof f (n, N + k) fb − 2 N N 06=|n|≤M X 1 M + |sof f (n, N + k)| N 4−δ N 3−δ X
1 (N + k)2
fb
06=|n|≤M
M M 1+ N 2−2/D = M 2+ N −2+δ−2/D N 4−δ
by (3.7)
as required. This proves (3.6). Next we express the difference sof f (n, N + k) − sof f (n, N ) as X X e(−nαy d ) e nαxd sof f (n, N + k) − sof f (n, N ) = 2 Re N +1≤y≤N +k
+
X
1≤x≤N
e nα(xd − y d ) .
N +1≤x6=y≤N +k
We estimate the second term trivially by k 2 : |sof f (n, N + k) − sof f (n, N )| ≤ k|S(n, N + k)| + k 2 .
(3.8)
Pair Correlation of Fractional Parts
69
Then inserting this into (3.6) we get XN +k (α) − XN (α)
1 N2 k N2
X
k|S(n, N + k)| + k 2 + M 2+ N −2+δ−2/D
0<|n|≤M
X
|S(n, N )| +
0<|n|≤M
M k2 + M 2+ N −2+δ−2/D N2
k M k2 1+ 1−1/D M N + + M 2+ N −2+δ−2/D N2 N2
by (3.5)
M 1+ kN −1−1/D N b+δ−1/D+ . Since b > 0 can be made arbitrarily small, this proves our proposition.
4. Triple and Higher Correlations The higher correlations run into some basic difficulties. For example, consider the triple correlation for αn2 mod 1. For a test function f ∈ Cc∞ (R2 ), let X FN (y1 , y2 ) = f (N (y1 + m1 ), N (y2 + m2 )) . (4.1) (m1 ,m2 )∈Z2
This function is periodic and has a Fourier expansion 1 X b k e(k · y). f FN (y) = 2 N N 2
(4.2)
k∈Z
The triple correlation function of the sequence αn2 mod 1 and for the test function f is R3 (f, α, N ) = where the sum of R3 is
P0
1 N
0 X
FN α(x2 − y 2 ), α(y 2 − z 2 ) ,
(4.3)
1≤x,y,z≤N
is over all triples of distinct integers x, y, z. The Fourier expansion R3 (f, α, N ) =
X
cl (N )e(lα)
(4.4)
l
with 1 cl (N ) = 3 N
0 X 1,x,y,z≤N, k1 ,k2 k1 (x2 −y 2 )+k2 (y 2 −z 2 )=l
k 1 k1 b , . f N N
(4.5)
There is no doubt that the mean hR3 i = c0 (N ) → fb(0, 0), as N → ∞, the expected answer for random sequence, and that more generally cl (N ) → 0 if l 6= 0. That is to say that R3 (f, α, N ) → fb(0, 0) in the weak sense. This can probably be proven. However, a much greater difficulty appears and that is that if f (0) 6= 0 then
70
Z. Rudnick, P. Sarnak
2
R3 (f, N ) − fb(0) N.
(4.6)
2
This shows that the L2 approach to almost-everywhere convergence is problematic in this case. In fact, this feature of the L2 -norm being as large a manifestation of R3 being very large at rational α’s. For almost all α we still expect that R3 (f, α, N ) → fb(0, 0). To prove (4.6) note that as N → ∞, X l
1 cl (N ) = 3 N
0 X
X
1≤x,y,x≤N k1 ,k2
k 1 k1 b , ∼ N 2 f (0). f N N
Hence if f (0) 6= 0 then N | 2
X
cl (N )| ≤
l
X
!1/2 |cl (N )|
!1/2 1
lN 3
l
= N 3/2
X
2
X
!1/2 |cl (N )|2
,
l
which gives (4.6). References 1. Berry, M.V. and Tabor, M.: Level clustering in the regular spectrum Proc. R. Soc. London A356, 375–394 (1977) 2. Bleher, P.M. The energy level spacing for two harmonic oscillators with golden mean ratio of frequencies. J. Stat. Phys. 61, 869–876 (1990) 3. Bleher, P.M.: The energy level spacing for two harmonic oscillators with generic ratio of frequencies. J. Stat. Phys. 63, 261–283 (1991) 4. Pandey, A., Bohigas, O. and Giannoni, M.J.: Level repulsion in the spectrum of two-dimensional harmonic oscillators. J. Phys. A: Math. Gen. 22, 4083–4088 (1989) 5. Casati, G., Guarneri, I. and Izrailev, F.M.: Statistical properties of the quasi-energy spectrum of a simple integrable system. Phys. Lett. A 124, 263–266 (1987) 6. Davenport, H.: Analytic Methods for Diophantine Equations and Diophantine Inequalities. Ann Arbor, Michigan: Ann Arbor Publishers, 1962 7. Rudnick, Z., Sarnak, P. and Zaharescu, A.: In preparation 8. Sarnak, P.: Values at Integers of Binary Quadratic Forms. To appear in a volume in memory of C. Herz, edited by S. Drury 9. S´os, V.: On the distribution mod 1 of the sequence nα. Ann. Univ. Sci. Budapest. E¨otv¨os Sect. Math. 1, 127–134 (1958) 10. Swierczkowski, S.: On succesive settings of an arc on the circumference of a circle Fund. Math. 46, 187–189 (1958) ¨ 11. H. Weyl: Uber die Gleichverteilung von Zahlen mod. Eins. Math. Ann. 77, 313–352 (1916) Communicated by Ya. G. Sinai
Commun. Math. Phys. 194, 71 – 86 (1998)
Communications in
Mathematical Physics © Springer-Verlag 1998
Classification of Bicovariant Differential Calculi on Quantum Groups (a Representation-Theoretic Approach) Pierre Baumann, Fr´ed´eric Schmitt U.F.R. de Math´ematique, Universit´e Louis Pasteur, 7 rue Ren´e Descartes, F-67084 Strasbourg Cedex, France Received: 1 December 1996 / Accepted: 29 September 1997
Abstract: The restricted dual of a quantized enveloping algebra can be viewed as the algebra of functions on a quantum group. According to Woronowicz, there is a general notion of bicovariant differential calculus on such an algebra. We give a classification theorem of these calculi. The proof uses the notion (due to Reshetikhin and SemenovTian-Shansky) of a factorizable quasi-triangular Hopf algebra and relies on results of Joseph and Letzter. On the way, we also give a new formula for Rosso’s bilinear form.
Introduction Let G be a semi-simple connected simply-connected complex Lie group, g its Lie algebra, Uq g the quantized enveloping algebra of g. Uq g is a Hopf algebra. The associated quantum group is an object of non-commutative geometry. According to a point of view due to Woronowicz and developed by Faddeev, Reshetikhin and Takhtadzhyan [F–R–T], one may view the restricted (Hopf) dual (Uq g)∗ res as the algebra Aq G of functions on this quantum group. In this way, the Peter–Weyl theorem becomes a definition: the rational representations of the quantum group are the finite-dimensional representations of Uq g. In order to study the differential geometry of quantum groups, Woronowicz [Wo] defined the notion of bicovariant differential calculus. As in the classical case, one needs only to define the differential of functions at the unity point of the quantum group. If ε : Aq G → C(q) is the augmentation map, this amounts to take the residual class of functions belonging to ker ε modulo a right ideal R ⊆ ker ε. In the classical case, one takes R = (ker ε)2 . As for quantum groups, it is more important to preserve the group structure than the infinitesimal structure, and one is led to select ideals R as above by the requirement of a certain invariance condition. In this article, we solve the classification problem for these ideals R, and we give a picture of what they look like. We now compare our results with previous ones. Rosso [Ro3] showed how to use the quasi-triangular structure of Uq g in order to construct left covariant differential calculi
72
P. Baumann, F. Schmitt
on the quantum group. Modifying this construction, Jurˇco [Ju] used the R-matrix in the natural representation of Uq g (and in the dual of it) so as to construct bicovariant differential calculi: he obtained particular cases (when M is the natural g-module or its dual) of our Theorem 2. (In this spirit, see also [F–P].) As regards classification results, Schm¨udgen and Sch¨uler have classified the ideals R as above, but only when g is of classical type, and under restrictive assumptions on R. Most of the results in [S–S1, S–S2] are particular cases of our Theorem 1. For instance, the classification given in Theorem 2.1 of [S–S1] corresponds (in the wording of our theorem) to the ideals R constructed (up to a twisting character χ : 2X/2Q → C× , as explained in Sect. 3.3) from the natural Uq sln -module or its dual. Let us explain our proof and the contents of our article. Our proof relies on the quasi-triangular structure of Uq g. Since the formalism of R-matrices may be justified only for finite dimensional Hopf algebras, we will employ the dual notion of a co-quasitriangular (c.q.t.) Hopf algebra (see [L–T]): the algebra Aq G is c.q.t. We use then a bilinear form on Aq G, introduced by Reshetikhin and Semenov-Tian-Shansky. As Uq g is a factorizable quasi-triangular Hopf algebra (in the terminology of [R–S]), this pairing is non-degenerate and gives a linear injection Aq G ,→ Uq g ⊆ (Aq G)∗ . The image of R under this map is nearly the annihilator of a Uq g-module. It is then easier to discuss what R may be. The definitions and the proofs of these assertions are given in Sects. 1 and 2. In Sect. 3, we present a contruction of bicovariant differential calculi valid for any factorizable c.q.t. Hopf algebra. Finally we link, in the case of Aq G, these constructions with our classification result. Notations. • Let A be a k-algebra. If M is an A-module, its annihilator is denoted by annA M. If m ∈ M and m∗ ∈ M∗ (the k-dual of M), we denote by θM (m, m∗ ) the matrix coefficient (A → k, a 7→ hm∗ , a · mi). P • For a Hopf algebra H, we will use Sweedler’s notation for coproduct (1(a) = a(1) ⊗ a(2) ) and for coaction on comodules. The sum sign will generally be omitted. We will denote the augmentation and the antipode of H by ε and S respectively. • Let H be a Hopf algebra, and H∗ res be the restricted (Hopf) dual of H. A finitedimensional left H-module M (with a basis (mi ) and the dual basis (m∗i ) of M∗ ) can be H∗ res -comodule with structure map δR : (M → M ⊗ H∗ res , m 7→ P viewed as a right ∗ mi ⊗ θM (m, mi ) ). 1. Co-Quasi-Triangular Hopf Algebras
1.1. Some definitions. Let H be a Hopf algebra over a field k. A right crossed bimodule over H (in the sense of Yetter [Ye]) is a k-vector space M, which is also P a right H-module, : (M → M ⊗ H, m 7→ m(0) ⊗ m(1) )), both a right H-comodule (with structure map δR P m(0) · a(2) ⊗ S(a(1) )m(1) a(3) (for m ∈ M, structures being compatible: δR (m · a) = a ∈ H). When M and N are right crossed bimodules over H, M ⊗ N becomes a right crossed bimodule for the action (m ⊗ n) · a = m · a(1) ⊗ n · a(2) and the coaction δR (m ⊗ n) = (m(0) ⊗ n(0) ) ⊗ m(1) n(1) . There are two easy examples: we can endow H with the structures: a · b = ab and δR : (H → H ⊗ H, a 7→ a(2) ⊗ S(a(1) )a(3) ). Alternatively, we can put on H the structures a · b = S(b(1) )ab(2) (right adjoint action) and δR : (H → H ⊗ H, a 7→ a(1) ⊗ a(2) ).
Classification of Bicovariant Differential Calculi on Quantum Groups
73
When 0 is a bicovariant bimodule (see [Wo]), the space 0L of left coinvariants is a right crossed bimodule over H. Conversely, any right crossed bimodule over H is the space of left coinvariants of a bicovariant bimodule. Finally (H still being a Hopf algebra), we endow the tensor product coalgebra H∗ res ⊗ H with the product (f ⊗ a)(g ⊗ b) = hg(3) , a(3) ihg(1) , S(a(1) )i(g(2) f ⊗ a(2) b). We obtain a bialgebra, called Drinfel 0d’s double of H and denoted by D(H). (Here H∗ res is the standard dual of H, the coproduct is not brought into its opposite.) When M is a right crossed bimodule over H, it is a right D(H)-module for the actions: m · (f ⊗ 1) = hf, m(1) im(0) , m · (1 ⊗ a) = m · a. 1.2. Definition of a co-quasi-triangular Hopf algebra. We give the definition of c.q.t. Hopf algebras, by now usual (see [L–T] for historical notes): Definition 1. A co-quasi-triangular Hopf algebra is a pair (A, γ), where A is a Hopf algebra and γ : A → A∗ res is a coalgebra morphism and an algebra antimorphism such that we have the Yang–Baxter equation (or rather the Baxter commutation relations): a(1) b(1) hγa(2) , b(2) i = hγa(1) , b(1) ib(2) a(2) for all a, b ∈ A. That γ is a coalgebra morphism and an algebra antimorphism gives us that for all a, b ∈ A, hγa, bi = hγS(a), S(b)i. We call δ : A → A∗ the map such that hδa, bi = hγb, S(a)i, for all a, b ∈ A. Hence we have hγa, bi = hδb, S(a)i. We verify easily that δ takes its values in A∗ res and that (A, δ) is a c.q.t. Hopf algebra. If U is a Hopf algebra quasi-triangular for an R-matrix R12 , then U∗ res becomes a c.q.t. Hopf algebra for the map γ given by: for a, b ∈ U∗ res , hγ(a), bi = hb ⊗ a, R12 i, −1 and then hδ(a), bi = hb ⊗ a, R21 i. This follows from Drinfel 0d’s classical axioms. For instance, let H be a finite-dimensional Hopf algebra, and U = D(H): the dual vector space of the restricted dual of U. If (ei ) is a basis for space H ⊗ H∗ of U is the underlying P H, the canonical R-matrix is (e∗i ⊗ 1) ⊗ (1 ⊗ ei ) ∈ U ⊗ U. It corresponds to the maps γ : (H ⊗ H∗ → U, a ⊗ f 7→ ε(a)f ⊗ 1) and δ : (H ⊗ H∗ → U, b ⊗ g 7→ g(1)ε ⊗ S−1 (b)) (the antipode of a finite-dimensional Hopf algebra being invertible). The category of left modules over a quasi-triangular Hopf algebra is braided. The translation in the present formalism is given by the following proposition: Proposition 1. Let (A, γ) be a c.q.t. Hopf algebra. If M is a right A-comodule, it becomes a right crossed bimodule over A when endowed with the right module structure given by: for m ∈ M and a ∈ A, m · a = hγa, m(1) im(0) . This extra structure is compatible with tensor products of comodules and crossed bimodules. Proof. Let δR : (M → M ⊗ A, m 7→ m(0) ⊗ m(1) ) the structure map for M. Then we have: m(0) · a(2) ⊗ S(a(1) )m(1) a(3) = m(0) ⊗ hγa(2) , m(1) iS(a(1) )m(2) a(3) = m(0) ⊗ S(a(1) )a(2) m(1) hγa(3) , m(2) i = m(0) ⊗ m(1) hγa, m(2) i = δR (m · a). The compatibility with tensor products is a consequence of γ being a coalgebra homomorphism. We also note that the antipode of a c.q.t. Hopf algebra is always invertible, the square of its transpose being an inner automorphism of the algebra A∗ (see [Dr2]).
74
P. Baumann, F. Schmitt
Finally, when (A, γ) is a c.q.t. Hopf algebra, we have the maps γ and δ, and Radford [Ra] has shown that (im γ)(im δ) = (im δ)(im γ) is a sub-Hopf-algebra of A∗ res . This was shown in the early [R–S]: there is a Hopf algebra structure (with invertible antipode) on the tensor product coalgebra A ⊗ A such that the map (A ⊗ A → A∗ res , a ⊗ b 7→ γb · δa) is a coalgebra morphism and an algebra antimorphism. Example. In the F.R.T. construction [F–R–T], one considers matrices L+ and L− , whose elements lie in im γ and im δ respectively. Then Faddeev, Reshetikhin and Takhtadzhyan defined Uq g to be the algebra (im γ)(im δ). 1.3. The maps I and J. We fix in this subsection a c.q.t. Hopf algebra (A, γ) over the field k, and note δ the associated map. We define two maps I : (A → A∗ res , a 7→ γ(a(1) ) Sδ(a(2) )) and J : (A → A∗ res , a 7→ Sδ(a(1) ) γ(a(2) )). Equivalently, we may consider the pairing of two elements a, b ∈ A: hI(a), bi = hJ(b), ai. (When A is the dual of a quasi-triangular Hopf algebra, this pairing is ha ⊗ b, R21 R12 i.) We have I = S ◦ J ◦ S and J = S ◦ I ◦ S. We will now state an important property of the map I. A∗ res is a left A∗ res ⊗ A∗ res module for the law (x ⊗ y) · z = xz S(y). A is a right crossed bimodule over A for the structures: a·b = ab and δR : (A → A⊗A, a 7→ a(2) ⊗S(a(1) )a(3) ), so A is a right D(A)module. Let 5 : (D(A) ≡ A∗ res ⊗A → A∗ res ⊗A∗ res , x⊗b 7→ γ(b(1) )x(1) ⊗δ(b(2) )x(2) ). Proposition 2. In the set-up above, 5 is an algebra antimorphism. If ξ ∈ D(A) and a ∈ A, then I(a · ξ) = 5(ξ) · I(a). Proof. That 5 is an antimorphism is already in [R–S]. Then, as a consequence of the Yang–Baxter equation, we may write, for x ∈ A∗ res and a ∈ A, that Sγ(a(1) )hx, a(2) i = hx(2) , a(1) ix(1) Sγ(a(2) )S(x(3) ). Then we compute, for ξ = x ⊗ b ∈ D(A): I(a · ξ) = hx, S(a(1) )a(3) i I(a(2) b) = γ(b(1) ) hx, S(a(1) )a(4) i γ(a(2) ) Sδ(a(3) ) Sδ(b(2) ) = γ(b(1) ) hx(1) , S(a(1) )i SγS(a(2) ) hx(2) , a(4) i Sδ(a(3) ) Sδ(b(2) ) = γ(b(1) ) hx(2) , S(a(2) )ix(1) SγS(a(1) )S(x(3) )hx(5) , a(3) ix(4) Sδ(a(4) )S(x(6) )Sδ(b(2) ) = γ(b(1) ) hx(2) , S(a(2) )i x(1) SγS(a(1) ) hx(3) , a(3) i Sδ(a(4) ) S(x(4) ) Sδ(b(2) ) = γ(b(1) ) x(1) γ(a(1) ) Sδ(a(2) ) S(x(2) ) Sδ(b(2) ) = 5(ξ) · I(a). We single out the particular case b = 1: Proposition 3. We consider A and A∗ res as left A∗ res -modules for the adjoint action: if x, y ∈ A∗ res and a ∈ A, x · a = hx, S(a(1) )a(3) ia(2) and x · y = x(1) y S(x(2) ). Then I : A → A∗ res is a morphism of A∗ res -modules. Finally, we give the definition, originally due to Reshetikhin and Semenov-TianShansky [R–S]: Definition 2. One says that (A, γ) is factorizable if the pairing (A × A → k, (a, b) 7→ hI(a), bi) is non-degenerate. Thus (A, γ) is factorizable iff the maps I and J are injective. It is possible to show that (A, γ) is factorizable iff (A, δ) is so. 1.4. A related construction. First, let U be a Hopf algebra. It is a left U-module for the adjoint action: x · y = x(1) y S(x(2) ). We let F` (U) be the sum of all finite-dimensional
Classification of Bicovariant Differential Calculi on Quantum Groups
75
U-submodules of U. It is known [J–L1] that F` (U) is a subalgebra of U, a left coideal in U, and a U-submodule for the left adjoint action. The multiplication in U defines a morphism of left U-modules F` (U) ⊗ F` (U) → F` (U). We can then do the semi-direct product F` (U) ⊗ U: we obtain an algebra U ⊗ U denoting the ordinary tensor product algebra, there is an algebra morphism (F` (U) ⊗ U → U ⊗ U, x ⊗ y 7→ xy(1) ⊗ y(2) ). We can make the same constructions on the right: we obtain an algebra Fr (U). If the antipode of U is invertible, the algebra morphism (U ⊗ Fr (U) → U ⊗ U, x ⊗ y 7→ x(1) ⊗ x(2) y) has the same image as the previous one. Hence this image contains F` (U) ⊗ Fr (U) ⊆ U ⊗ U. We take now a c.q.t. Hopf algebra (A, γ), with δ, I and J as in the preceding subsection. Let U = (im γ)(im δ) be the minimal sub-Hopf-algebra of A∗ res in which γ and δ take their values. We consider on A and A∗ res the A∗ res -module structures of Proposition 3. By restriction, A and A∗ res are U-modules, and I : A → A∗ res is a morphism of Umodules. We can see that I takes its values in U, which is a U-submodule of A∗ res . Further, A is the sum of its finite-dimensional U-submodules, hence im I ⊆ F` (U). Proposition 4. Let (A, γ) be a c.q.t. factorizable Hopf algebra, and I be the associated map. Let U be the sub-Hopf-algebra (im γ)(im δ) ⊆ A∗ res . We suppose that im I = F` (U). Then I induces a bijection between: • the set of right ideals R of A, which are subcomodules for the right coaction δR : (A → A ⊗ A, a 7→ a(2) ⊗ S(a(1) )a(3) ). • the set of two-sided ideals I of F` (U), which are U-submodules for the adjoint action. This bijection preserves dimensions, codimensions, and the inclusion ordering in both sets. Proof. By assumption, I : A → F` (U) is a U-module isomorphism. We adopt the notations of Proposition 2. A is a D(A)-module, and U ⊗ A is (the underlying space of) a sub-Hopf-algebra of D(A), so we will view A as a right U ⊗ A-module: 1 ⊗ A acts on A by right multiplication, Uop ⊗ 1 acts on A by the left adjoint action. The injectivity of I implies that im J ⊆ U separates the points of A: hence the sub-U ⊗ A-modules of A are the right ideals which are subcomodules for the right coaction δR . On the other hand, we let E be the image of the morphism (F` (U) ⊗ U → U ⊗ U, x ⊗ y 7→ xy(1) ⊗ y(2) ). U is a U ⊗ U-module, so is an E-module, and F` (U) is a sub-E-module of U. E contains F` (U) ⊗ Fr (U), with S(Fr (U)) = F` (U). Therefore, the sub-E-modules of F` (U) are the two-sided ideals I which are U-submodules for the adjoint action. Now the proposition is a consequence of Proposition 2: writing 5 as the composition (F` (U)⊗A∗ res → A∗ res ⊗A∗ res , x⊗y 7→ xy(1) ⊗y(2) )◦(A∗ res ⊗A → F` (U)⊗A∗ res , x⊗ a 7→ I(a(1) ) ⊗ δ(a(2) )x), and using the assumption im I = F` (U), we can see that E is the image of U ⊗ A through 5. 2. The Case of the Quantum Coordinate Algebra
2.1. Notations. In this section, we study the preceding constructions in the case where A is the algebra Aq G of regular functions on a quantum group. Let g be a finite-dimensional semi-simple split Lie algebra, h a splitting Cartan subalgebra, {α1 , . . . , α` } ⊆ h∗ a basis for the root system, {α1∨ , . . . , α`∨ } ⊆ h the inverse roots, P ⊆ h∗ and Q ⊆ h∗ the weight and the root lattices. The choice of an invariant (under Weyl group action) scalar product (·|·) allows us to identify h and h∗ ,
76
P. Baumann, F. Schmitt
with αi = di αi∨ , di = (αi2|αi ) . We choose the normalization of (·|·) so that (λ | µ) ∈ Z whenever λ and µ belong to P. We denote by ρ half the sum of the positive roots, by P+ the set of dominant weights, and by w0 the longest element in the Weyl group. We now choose the following version of Uq g: this is a C(q)-algebra (q is generic) generated by Ei , Fi and Kλ (λ ∈ P). The relations are the usual ones among which: K −K−αi Kλ Ei = q (λ|αi ) Ei Kλ , Kλ Fi = q −(λ|αi ) Fi Kλ , Ei Fj − Fj Ei = δij qαdii −q−d . The i coproduct is given by: 1Kλ = Kλ ⊗Kλ , 1Ei = Ei ⊗1+Kαi ⊗Ei , 1Fi = 1⊗Fi +Fi ⊗ K−αi . We note S the antipode of Uq g. If one chooses a dominant weight λ and a character χ : P/2Q → C× , one knows how to construct a simple finite-dimensional Uq g-module, in which there is a highest weight vector mλ such that Kµ ·mλ = χ(µ mod 2Q)q (µ|λ) mλ . We note Lχ (λ) such a Uq g-module ; when χ is the trivial character, we simply write L(λ), and then Lχ (λ) = L(λ) ⊗ Lχ (0). The matrix coefficients of the representation L L(λ) span a linear subspace C(λ) of the restricted dual of Uq g, and we let Aq G = λ∈P+ C(λ). This is a Hopf subalgebra of (Uq g)∗ res . The elements of Aq G separate the points of Uq g [J–L1], so that there is an inclusion of Uq g into the dual of Aq G, actually into the restricted dual of Aq G. We note S the antipode of Aq G, which is the restriction to Aq G of the transpose of the antipode of Uq g. ThereP is an R-matrix for Uq g [Dr1, Ta, Ga]. We choose the R-matrix with the structure (diagonal part)(monomial in F ) ⊗ (monomial in E). If a and b belong to Aq G, the number hR12 , b ⊗ ai ∈ C(q) is well-defined (thanks to the weight graduation of Uq g and of any finite-dimensional Uq g-module), and we can define γ, δ : Aq G → (Aq G)∗ such that hR12 , b⊗ai = hγ(a), bi = hδ(b), S(a)i. (Aq G, γ) and (Aq G, δ) are c.q.t. Hopf algebras, im(γ) and im(δ) are the sub-Hopf-algebras U− U0 and U0 U+ of Uq g ⊆ (Aq G)∗ res respectively, and Uq g is the sub-Hopf-algebra (im γ)(im δ) = (im δ)(im γ) of (Aq G)∗ res . 2.2. Factorizability of Aq G. Let (Aq G, γ) be the c.q.t. algebra presented above, and δ be the associated map. For all the section, we endow Aq G and Uq g with the left adjoint action of Uq g, as in Sect. 1.4: in particular, the map I : Aq G → F` (Uq g) is a morphism of left Uq g-modules. Joseph and Letzter [J–L1, J–L2] have studied the structure of F` (Uq g), and we need the following results: • L If λ ∈ P+ , K−2λ generates a finite dimensional Uq g-submodule of Uq g, and F` (Uq g) = λ∈P+ (Uq g · K−2λ ). • Each block Uq g · K−2λ contains a unique one-dimensional Uq g-submodule; it defines a unique (up to scalars) element zλ of the center of Uq g. • F` (Uq g) ⊆ (Aq G)∗ separates the points of Aq G. The next assertion has been stated in [R–S]: Proposition 5. (Aq G, γ) is a factorizable c.q.t. Hopf algebra, and im I = F` (Uq g). Proof. Let λ ∈ P+ , L(λ) the standard Uq g-module, mλ a highest weight vector, mw0 λ a lowest weight vector, (mi ) a basis for L(λ) composed of weight vectors, (m∗i ) the dual basis. We have: • The matrix element θL(λ) (mw0 λ , m∗w0 λ ) is the linear form on Uq g given by (in the triangular decomposition U+ ⊗ U0 ⊗ U− of Uq g): EKµ F 7→ ε(E)q (w0 λ|µ) ε(F ). • On this element, γ takes the value Kw0 λ and δ the value K−w0 λ .
Classification of Bicovariant Differential Calculi on Quantum Groups
77
• The image by γ (respectively δ) of the matrix element θL(λ) (mi , m∗w0 λ ) (respectively θL(λ) (mw0 λ , m∗i )) is zero if i 6= w0 λ. So we have: I(θL(λ) (mw0 λ , m∗w0 λ )) = γ((θL(λ) (mw0 λ , m∗w0 λ ))(1) ) S(δ((θL(λ) (mw0 λ , m∗w0 λ ))(2) )) P = γ(θL(λ) (mi , m∗w0 λ )) S(δ(θL(λ) (mw0 λ , m∗i ))) = γ(θL(λ) (mw0 λ , m∗w0 λ )) S(δ(θL(λ) (mw0 λ , m∗w0 λ ))) = K2w0 λ .
Hence im I is a Uq g-submodule of F` (Uq g) which contains all the K2w0 λ (λ ∈ P+ ), so im I = F` (Uq g). We now want to show that J is injective. If b ∈ ker J, then for all a ∈ Aq G, hI(a), bi = hJ(b), ai = 0, so b is null when viewed as a linear form on im I = F` (Uq g). Then b = 0, because F` (Uq g) separates the points of Aq G. Finally, owing to the formula J = S ◦ I ◦ S and to the invertibility of S, I is also injective. This concludes the proof of the proposition. There is another way to present this result. Rosso [Ro1] introduced a bilinear non-degenerate ad-invariant form on Uq g, that Caldero [Ca] writes (Uq g × Uq g → C(q 1/2 ), (x, y) 7→ hζ(x), S−1 (y)i), where ζ : Uq g → (Uq g)∗ . Rosso’s non-degeneracy result is that ζ is injective; Caldero’s theorem states that ζ maps F` (Uq g) onto Aq G ⊆ (Uq g)∗ res . The triangular behaviour of Rosso’s form gives us that ζ(K2w0 λ ) = θL(λ) (mw0 λ , m∗w0 λ ). The ad-invariance of Rosso’s form can be translated for ζ: when we restrict ζ to F` (Uq g) and Aq G, ζ is a morphism of Uq g-modules for the adjoint structures. Now I ◦ ζ : F` (Uq g) → F` (Uq g) and ζ ◦ I : Aq G → Aq G are morphisms of Uq g-modules and fix the respective generators K2w0 λ and θL(λ) (mw0 λ , m∗w0 λ ) of these modules. (The fact that θL(λ) (mw0 λ , m∗w0 λ ) generates the Uq g-submodule C(λ) of Aq G is equivalent to the fact that m∗w0 λ ⊗ mw0 λ generates the Uq g-module L(λ)∗ ⊗ L(λ).) So we conclude that ζ and I are mutually inverse isomorphisms, and that I is a bijection between C(λ) and Uq g · K2w0 λ . The analysis also shows the amusing side-result: Proposition 6. If x ∈ F` (Uq g), y ∈ Uq g, then the Rosso form on (x, y) is given by hI−1 (x), S−1 (y)i, where I : (Aq G → F` (Uq g), a 7→ ha ⊗ idUq g , R21 R12 i) is related to the universal R-matrix and S is the antipode of Uq g. Remarks. 1. It is also possible to give an heuristic proof of this result, using the canonical R-matrix for Drinfel 0d’s double and using Rosso’s formula for his form [Ro2]. 2. In the preceding discussion, we were lying a bit. Caldero’s map ζ does not give exactly Rosso’s bilinear form, but our formula connecting I and Rosso’s form is correct as stated. In our notations, Caldero’s map ζ is the inverse of the map (Aq G → F` (Uq g), a 7→ δ(a(1) ) Sγ(a(2) )). Later, we will need to know the relations between the central elements zλ defined above. To this aim, we recall Drinfel 0d’s construction of the center of Uq g [Dr2]. Let λ ∈ P+ and t ∈ Aq G be the quantum trace in L(λ): for x ∈ Uq g, ht, xi = Tr L(λ) (K2ρ x). t is an invariant element for the adjoint action of Uq g in Aq G, so I(t) is central, and belongs to Uq g · K2w0 λ . We choose the normalization of z−w0 λ by letting z−w0 λ = I(t). We then have a Mackey-like theorem (which is implicit in [Dr2] and in the thesis of Caldero, Chap. II, 1.2): L Proposition 7. Let cνλµ be the fusion coefficients for g: L(λ) ⊗ L(µ) ' ν cνλµ L(ν). P ν Then zλ zµ = ν cλµ zν .
78
P. Baumann, F. Schmitt
Proof. Let µ ∈ P+ . We compute J(θL(µ) (mµ , m∗µ )) = K2µ (with the help of the formulas J = S ◦ I ◦ S and S(θL(µ) (mµ , m∗µ )) = θL(−w0 µ) (m−µ , m∗−µ )). Now let λ ∈ P+ and let t be the quantum trace in L(λ). Let 9 be the Harish-Chandra morphism from the center of Uq g to U0 [Ro1]. We want to compute 9(z−w0 λ ) on µ + ρ. (Evaluation on µ + ρ means the algebra homomorphism (U0 → C(q), Kλ 7→ q (λ|µ+ρ) ).) The result will be the image of z−w0 λ by the central character of L(µ). So it is hI(t), θL(µ) (mµ , m∗µ )i = hJθL(µ) (mµ , m∗µ ), ti = hK2µ , ti = Tr L(λ) (K2µ K2ρ ) = Tr L(λ) (K2(µ+ρ) ). Thus 9(z−w0 λ ) equals the sum of K2ν for ν in the set of weights (with multiplicities) of L(λ). We then use the fact that 9 is an injective algebra homomorphism. We denote by G the Grothendieck ring of the category of finite-dimensional Uq g-modules whose components are modules L(λ), without any twisting character χ : P/2Q → C× . Let Z(Uq g) be the center of Uq g, and Z[P] the group algebra of P (with the standard Z-basis denoted by (eν )ν∈P ). The map (G → Aq G, [M] 7→ Tr M (K2ρ )) is a ring homomorphism. If a, b ∈ Aq G are such that I(a) belongs to the center of Uq g, then I(ab) = I(a) I(b). As a consequence, the map G → Z(Uq g), [M] 7→ I(Tr M (K2ρ )) is a ring homomorphism. This shows again the statement in Proposition 7, and we can paraphrase the above proof by saying that the following diagram is commutative: G ch
? Z[P]
- Aq G
I
- Z(Uq g) 9
? - U0
Here ch : G → Z[P] is the ring homomorphism which maps a module to its formal character, and the bottom arrow is the map (Z[P] → U0 , eν 7→ K2ν ). 2.3. A technical result on the representation ring. We have just introduced a Grothendieck ring G: by the classical results of Lusztig and Rosso, G is naturally isomorphic to the representation ring of g. The elements [L(λ)] (λ ∈ P+ ) form a Z-basis of G and a Q-basis of G ⊗Z Q. Proposition 8. Let λ ∈ P+ . Then the ideal of R ⊗Z Q generated by the elements [L(λ + $)] ($ ∈ P+ ) is the whole algebra G ⊗Z Q. The proof of this proposition can be skipped without any drawback. Before we give it, we have to state an elementary lemma: ` Lemma. Let (µ(1) , . . . , µ(k) ) ∈ (C` )k be such that their image in C/Z are all P (m) different, and let (P (1) , . . . , P (k) ) ∈ (C[X1 , . . . , X` ])k . If (n1 , . . . , n` ) mP P (m) ` exp(2πi j nj µj ) = 0 holds for all (n1 , . . . , n` ) ∈ N , then the polynomials P (1) , . . . , P (k) are all equal to zero. For ` = 1, this lemma states linear independence of elementary solutions of a linear difference equation. The general proof is by induction on `. Proof of Proposition 8. In this proof, we are in a classical context and we do not identify h and h∗ . Let R ⊆ h∗ and R∨ ⊆ h be the direct and inverse root systems, (α 7→ α∨ ) the canonical bijection between R and R∨ , and Q(R∨ ) ⊆ h the root lattice. P = P(R) ⊆ h∗ is still the weight lattice; we denote by {α1∨ , . . . , α`∨ } the set of inverse simple roots,
Classification of Bicovariant Differential Calculi on Quantum Groups
79
and by {$1 , . . . , $` } the set of fundamental weights. R∨ and R define Q-structures on h and h∗ , and we can define hR and hC . The Weyl group W operates on h and h∗ , and the affine Weyl group Wa = W n Q(R∨ ) operates on h. Let Z[P] be the Z-algebra of the group P, Z[P]W be the set of elements which are invariant under Weyl group action, ∼ ch : (G →Z[P]W ) be the ring isomorphism “formal character”. Finally, we denote by ε(w) = ±1 the determinant of an element w of the Weyl group. For µ ∈ hC , let evµ : (Z[P] → C) be the ring morphism which sends a basic element eν (ν ∈ P) to exp(2πihµ, νi), where exp is the complex exponential. This extends to an algebra morphism evµ : (C[P] → C). If ν ∈ P+ , let fν be the map hC → C, µ 7→ evµ (ch L(ν)) . We first assert that given any (x1 , . . . , x` ) ∈ C` , there exists µ ∈ hC such that for all i ∈ {1, . . . , `}, f$i (µ) = xi . We view C[P] Pas the µi αi∨ coordinate ring of the affine variety (C× )` , and we view an element µ = 2πiµ1 2πiµ` × ` (µi ∈ C) as the point (e ,... ,e ) ∈ (C ) . By the Nullstellensatz, it is sufficient to prove that the elements (ch L($i ) − xi e0 ) (i = 1, . . . , `) generate a proper ideal in C[P]. This is already true in C[P]W by [Bo], Ch. VI, § 3, Th´eor`eme 1. The case of C[P] is given by a standard trick: let \ : (C[P] → C[P]W ) be the projection onto the trivial homogeneous component of W; \ is a morphism of C[P]W P in C[P] for the action 0 modules, and thus a relation Qi · (ch L($i ) − xi e ) = 1 in C[P] would give a relation P \ Qi · (ch L($i ) − xi e0 ) = 1 in C[P]W , which is impossible. We now want to prove a formula for the character fν (µ) = evµ (ch L(ν)). We first remark that fν is invariant under the action of the affine Weyl group Wa in hC . If the real part Re(µ) of µ lies in an open alcove of hR , our formula will just be Weyl’s character formula: P w∈W ε(w) exp(2πihwµ, ν + ρi) . fν (µ) = P w∈W ε(w) exp(2πihwµ, ρi) Writing the denominator as a product over the positive roots: Q exp(2πihµ, ρi) α∈R,α≥0 (1 − exp(−2πihµ, αi)), we can see that it is a non-zero complex number. In the general case, we let T = {α ∈ R | Re(hµ, αi) ∈ Z}: this is a closed symmetric subset of R ([Bo], Ch. VI, § 1, D´efinition 4), thus T is a root system in the vector space V1 ⊆ h∗R that it spans ([Bo], Ch. VI, § 1, Proposition 23). The stabilizer of µ in Wa is generated by the reflections across the affine hyperplanes in which Re(µ) lies ([Bo], Ch. V, § 3, Proposition 2), thus W1 := {w ∈ W | µ − wµ ∈ Q(R∨ )} is precisely the subgroup generated by reflections along α∨ (α ∈ T), and its restriction to V1 isPthe Weyl group of T. Let σ be half the sum of the inverse positive roots of T: σ = 21 α∈T,α≥0 α∨ . In restriction to V1 , σ is the sum of the fundamental weights of the root system T∨ of V1∗ . Let h be a small real parameter: Re(µ) + hσ then lies in an open alcove of hR and we can compute (with a little abuse): fν (µ) = lim fν (µ + hσ) h→0 P P = lim
h→0
w∈W/W1
P
w1 ∈W1
w∈W/W1
P
ε(ww1 ) exp(2πihwµ, ν + ρi) exp(2πihhw1 σ, w−1 (ν + ρ)i)
w1 ∈W1
ε(ww1 ) exp(2πihwµ, ρi) exp(2πihhw1 σ, w−1 ρi)
.
In the sums, we fix w ∈ W/W1 and compute the sums on w1 : in the numerator for instance, we have an alternating sum of exp(2πihhw1 σ, w−1 (ν + ρ)i), where w−1 (ν + ρ) ∈ P(R) has to be projected on V1 , as in [Bo], Ch. VI, § 1,P Proposition 28. The formula (valid in the group algebra of the weight lattice of T∨ ): w1 ∈W1 ε(w1 )ew1 σ = Q ∨ eσ α∈T,α≥0 (1 − e−α ) then gives:
80
P. Baumann, F. Schmitt
P fν (µ) =
w∈W/W1
P
ε(w) exp(2πihwµ, ν + ρi)
w∈W/W1
Q
ε(w) exp(2πihwµ, ρi)
α∈T,α≥0 hα
Q
∨
α∈T,α≥0 hα
, w−1 (ν + ρ)i
∨ , w −1 ρi
.
As ν + ρ and ρ are regular, neither of the products occurring here can be zero. (We will see soon that the denominator cannot be zero.) We now prove that the ideal of G ⊗Z C generated by the elements [L(λ+$)] ($ ∈ P+ ) is the whole algebra G ⊗Z C. We consider again [Bo], Ch. VI, § 3, Th´eor`eme 1: this time, the isomorphism ϕ : C[X1 , . . . , X` ] → C[P]W is given by ϕ(Xi ) = ch L($i ). Composing with the isomorphism ch : G → Z[P]W , we can see that G ⊗Z C is a polynomial algebra over C. We suppose by the way of contradiction that the elements [L(λ+$)] ($ ∈ P+ ) all belong to some maximal ideal of G ⊗Z C. Then, by the Nullstellensatz, there exists a point (x1 , . . . , x` ) ∈ C` such that for all $ ∈ P+ , ϕ−1 (ch L(λ + $))(x1 , . . . , x` ) = 0. We can find µ ∈ hC such that f$i (µ) = xi (i = 1, . . . , `): then fλ+$ (µ) = 0 for all $ ∈ P+ . We next use the formula: fλ+$ (µ) (denominator) X ε(w) exp(2πihwµ, λ + $ + ρi) = w∈W/W1
Y
hα∨ , w−1 (λ + $ + ρ)i,
α∈T,α≥0
P
any integers. The wµ (w ∈ W/W1 ) are all and write $ = ni $i , where (ni ) ∈ N` are Q distinct modulo Q(R∨ ), and the expressions α∈T,α≥0 hα∨ , w−1 (λ + $ + ρ)i are nonzero polynomials in (n1 , . . . , n` ) (they never vanish indeed). Then the above lemma states that the right-hand side cannot vanish for all (ni ) ∈ N` . This proves first that the denominator is not null, and second that fλ+P ni $i (µ) cannot vanish for all (ni ) ∈ N` . We have reached a contradiction. is then easy: we have shown that we can express in To go down to the case of G ⊗Z QP G ⊗Z C the unity as a finite sum 1 = xi [L(τi )][L(νi )], where τi ∈ P+ , νi ∈ λ + P+ and xi ∈ C. As the structure constants of G ⊗Z C are integer-valued, this system, viewed as linear equations in (xi ), has a solution in C, so has a solution in Q. 2.4. Classification of some ideals of F` (Uq g). In order to achieve our classification of ideals R ⊆ Aq G in the next section, we must study the ideals I ⊆ F` (Uq g) which are stable by the adjoint action of Uq g. The analysis requires the use of the subalgebra V of Uq g generated by F` (Uq g) and by the elements K2λ (λ ∈ P+ ). Joseph and Letzter [J–L1] have shown that V is the subalgebra generated by the elements Ei , Fi Kαi and K2λ (λ ∈ P). As it is such a “big” subalgebra of Uq g, its representation theory is similar to that of Uq g. We will describe it in the next subsection, but in the following proof, we need to know that the annihilator of a finite-dimensional V-module is homogeneous with respect to the Q-graduation of V. Proposition 9. The following two properties for a subspace I ⊆ F` (Uq g) are equivalent: (1) I is the annihilator in F` (Uq g) of a finite-dimensional V-module; (2) I is a finite-codimensional two-sided ideal of F` (Uq g) and a Uq g-submodule of F` (Uq g) for the left adjoint action. Proof. We first show that (1) ⇒ (2). If M is a finite-dimensional V-module, its annihilator in V is a finite-codimensional two-sided ideal of V, and is homogeneous w.r.t. the Q-graduation of V. It is then easy to see that annV M is a Uq g-submodule of V for
Classification of Bicovariant Differential Calculi on Quantum Groups
81
the left adjoint action. The annihilator I = (annV M) ∩ F` (Uq g) of M in F` (Uq g) thus satisfies the property (2). Conversely, let I ⊆ F` (Uq g), satisfying the property (2). We consider the left regular F` (Uq g)-module M = F` (Uq g)/I . I is its annihilator, so it is sufficient to show that M extends to a V-module. We thus want to show that the elements K−2λ ∈ F` (Uq g) (λ ∈ P+ ) map to invertible operators in End(M). 1) M is a finite-dimensional algebra, and is also a left Uq g-module (for the adjoint action). The multiplication in M defines a morphism of left Uq g-modules: M ⊗ M → M. Thus the Q-graduation of M (defined by the structure of Uq g-module) is an algebra grading. 2) We fix λ ∈ P+ . We can write M = M0 ⊕M∞ (as C(q)-vector space), where K−2λ acts nilpotently on M0 and inversibly on M∞ (Fitting’s decomposition). M0 and M∞ are stable by the commutant of K−2λ in End(M), so are right ideals of M. If x ∈ F` (Uq g) is homogeneous w.r.t. the Q-graduation of F` (Uq g), x commutes (up to a non-zero scalar) with K−2λ , so M0 and M∞ are stable by left multiplication by x. Thus M0 and M∞ are also left ideals of M. 3) We now show that M0 and M∞ are Uq g-submodules of M. (a) Let {e1 , . . . , ek } be the set of central idempotents in M. The elements Kµ (µ ∈ P) of Uq g act on M (by the adjoint action) as algebra automorphisms, so permute the elements of the set {e1 , . . . , ek }. Hence for each µ, there exists an integer n ≥ 1 such that Knµ fixes each ei . Since M is, as a Uq g-module, a direct sum of modules L(ν) (without any twisting character χ), and since q is generic, we conclude that e1 , . . . , ek are fixed by the adjoint action of the elements Kµ . (b) Let e be a central idempotent in M. e is of weight zero. We consider the qn P −di n(n−1)/2 ad Ei exponential expq (ad Ei ) = n≥0 q [n]i ! (i ∈ {1, . . . , `} fixed). Then exp (ad Ei ) is a well defined operator in M. The formula 1(Ein ) = Pn n q di (n−k)k n−k k Ei Kαi ⊗ Eik enables us to see that expq (ad Ei )(e) is k=0 k i q an idempotent which we write e + x. Then 2ex + x2 = x, x(1 − 2e) = x2 , x = x(1 − 2e)2 = x3 . The weights of the Q-homogeneous components of x belong to {nαi | n ≥ 1}; so the weights of the Q-homogeneous components of x3 belong to {nαi | n ≥ 3}, and the homogeneous component of x of weight αi is null. We obtain that (ad Ei )(e) = 0. Similarly, (ad Fi )(e) = 0 for all i ∈ {1, . . . , `}. (c) M0 and M∞ are ideals in M generated by central idempotents e0 and e∞ respectively. (a) and (b) show that e0 and e∞ define the trivial Uq g-module. Hence for x ∈ M0 and u ∈ Uq g, u · x = u · (xe0 ) = (u(1) · x)(u(2) · e0 ) = (u(1) · x)ε(u(2) )e0 = (u · x)e0 ∈ M0 . The same holds for M∞ . 4) We first consider the case g = sl2 . We choose naturally λ = $ the fundamental weight, and write M0 = L0 /I and M∞ = L∞ /I . The points 2) and 3) show that L0 and L∞ are two-sided ideals and left Uq g-submodules of F` (Uq g). By definition of the Fitting decomposition, there exists an integer n ≥ 0 such that K−2n$ ∈ L∞ . Hence for all integers m ≥ n, we have K−2m$ ∈ L∞ , and thus zm$ ∈ L∞ . Let n0 ≥ 0 be the smallest integer such that for all m ≥ n0 , zm$ ∈ L∞ . Proposition 7 and the Clebsch–Gordan theorem show that if n ≥ 1, z(n+1)$ + z(n−1)$ = z$ zn$ . Thus n0 has to be equal to zero. So 1 = z0 ∈ L∞ , M∞ = M, and K−2$ acts inversibly on M.
82
P. Baumann, F. Schmitt
5) The general case is solved in the same way. We consider the decomposition of the point 2) and write M0 = L0 /I and M∞ = L∞ /I . L0 and L∞ are two-sided ideals and left Uq g-submodules of F` (Uq g), and there exists an integer n ≥ 0 such that K−2nλ ∈ L∞ . If $ ∈ P+ , then K−2(nλ+$) ∈ L∞ , and thus znλ+$ ∈ L∞ . Let ϕ be the Q-algebra morphism G ⊗Z Q → Z(Uq g), [M] 7→ I(Tr M (K2ρ )) considered at the end of Sect. 2.2. Then ϕ−1 (L∞ ) is an ideal of G ⊗Z Q, which contains all the elements [L(−w0 nλ + $)] ($ ∈ P+ ). Thus ϕ−1 (L∞ ) = G ⊗Z Q by Proposition 8, and so 1 = ϕ([L(0)]) ∈ L∞ , M∞ = M, and K−2λ acts inversibly on M. Remark. This result is a particular case of Proposition 8.4.13 in [Jo]. Accordingly, its proof is shorter than the one of Joseph’s theorem, and does not require the knowledge of the inclusions between Verma modules, nor the use of Gel 0fand–Kirillov dimensions. 2.5. Classification of some right ideals of Aq G. The notations Aq G, Uq g, V have the ∼ same meaning as in Sects. 2.1 and 2.4. The map I : (Aq G→ F` (Uq g)) was introduced in Sect. 1.3. We now specify the structure of the finite-dimensional V-modules: they are completely reducible; each Uq g-module Lχ (λ) (with λ ∈ P+ , χ : P/2Q → C× ) is (by restriction) a simple V-module; the V-modules Lχ (λ) and Lϕ (µ) are isomorphic iff λ = µ and the characters χ, ϕ restrict to the same character 2P/2Q → C× . The simple finite-dimensional V-modules will be denoted by Lχ (λ) with λ ∈ P+ and χ : 2P/2Q → C× a character. We finally remark (see [J–L1]) that a simple finitedimensional V-module is still simple as a F` (Uq g)-module. Consequently, if (Mi ) is a finite family of non-isomorphic L finite-dimensional simple V-modules, the natural ring End Mi is surjective. homomorphism F` (Uq g) → Theorem 1. 1) Let R be a finite codimensional right ideal of Aq G, which is a subcomodule of Aq G w.r.t. the right coaction δR : (Aq G → Aq G ⊗ Aq G, a 7→ a(2) ⊗ S(a(1) )a(3) ). Then there exists a finite-dimensional V-module M such that R = I−1 (annF` (Uq g) M). 2) If M is a finite-dimensional V-module, then I−1 (annF` (Uq g) M) is a finite codimensional right ideal of Aq G, stable by the right coaction δR . 3) If M and N are finite dimensional V-modules, then I−1 (annF` (Uq g) M) I−1 (annF` (Uq g) N) iff M and N have the same irreducible components.
=
4) I−1 (annF` (Uq g) M) is included in the augmentation ideal of Aq G iff M contains the trivial V-module. Proof. 1) and 2) are consequences of Propositions 4 and 9. Let M and N be two finitedimensional V-modules having the same annihilator in F` (Uq g). Then annF` (Uq g) M = annF` (Uq g) (M ⊕ N). Let M1 , . . . , Mk (respectively M1 , . . . , Mn ) be the distinct irreducible components of M (respectively M ⊕ N). Then we have: L F` (Uq g)/annF` (Uq g) (M) ' ki=1 End Mi and:
F` (Uq g)/annF` (Uq g) (M ⊕ N) '
Ln i=1
End Mi ,
and so k = n: all the irreducible components of N appear in M. 3) follows. 4) can be proved in a similar way, using the fact that the augmentation ideal of Aq G is the inverse image by I of the annihilator of the trivial V-module.
Classification of Bicovariant Differential Calculi on Quantum Groups
83
3. Differential Calculi on Quantum Groups 3.1. Woronowicz’s definition. Let A be a Hopf algebra, 0 be a bicovariant bimodule and d : A → 0 be a linear map. We say that (0, d) is a bicovariant differential calculus on A if d is a derivation, a morphism of two-sided comodules and if the image of d generates the left A-module 0. The dimension of the space 0L of left coinvariants will be supposed to be finite. When (0, d) is a differential calculus over A, we note dL the map (A → 0L , a 7→ S(a(1) ) · d(a(2) )). The subspace R = ker dL ∩ ker ε is a finite-codimensional right ideal of A, and a subcomodule for the right coadjoint coaction δR : (a 7→ a(2) ⊗ S(a(1) )a(3) ). As shown by Woronowicz, the subspace R determines (up to isomorphism) the bicovariant differential calculus (0, d): we call it the ideal associated to (0, d). Geometrically, A must be viewed as the algebra of functions over a group G, 0 is the space of 1-forms on G, 0L is the space of left-G-invariant 1-forms on G, identified with the cotangent space at the unity point of G, and dL maps a function on G to its differential at the unity point. 3.2. A construction of bicovariant differential calculi. Let A be a c.q.t. Hopf algebra over the field k, and let γ, δ be the associated maps. a basis of M, We take a finite-dimensional right A-comodule M. We note (mi )P (m∗i ) the dual basis, and Rij the elements of A such that δR (mi ) = j mj ⊗ Rji . P Then 1Rji = k Rjk ⊗ Rki and ε(Rji ) = δji (Kronecker’s symbol). Also, M is a left A∗ -module, and the Rji (viewed as linear forms on A∗ ) are the matrix coefficients θM (mi , m∗j ) of this module. SinceP (A, γ) is c.q.t., M becomes a right crossed bimodule over A for the action mi · a = j hγ(a), Rji imj (Proposition 1). M∗ is a right comodule over A too, for the P coaction δR (m∗i ) = j m∗j ⊗ S(Rij ). Using the fact that (A, δ) is a c.q.t. Hopf algebra, ∗ of a right crossed bimodule over A for the action we may endow P M with the structure ∗ ∗ hδ(a), S(R )im . Then, by making the tensor product, we obtain that mi · a = ij j j End(M) ' M ⊗ M∗ is a right crossed bimodule. We denote by 0 the bicovariant bimodule associated to this right crossed bimodule End(M). As a vector space, 0 is just the tensor product A ⊗ M ⊗ M∗ . On the basic elements, the structure maps are: b · (a ⊗ mi ⊗ m∗j ) = ba ⊗ mi ⊗ m∗j , P (a ⊗ mi ⊗ m∗j ) · b = k,l ab(1) ⊗ hγ(b(2) ), Rki imk ⊗ hδ(b(3) ), S(Rj` )im∗` ,
δL (a ⊗ mi ⊗ m∗j ) = a(1) ⊗ a(2) ⊗ mi ⊗ m∗j , P δR (a ⊗ mi ⊗ m∗j ) = k,l a(1) ⊗ mk ⊗ m∗` ⊗ a(2) Rki S(Rj` ). P ∗ It follows that the canonical element X = i 1 ⊗ mi ⊗ mi of 0 is left and right coinvariant. The linear map d : (A → 0, a 7→ X · a − a · X) is then a derivation and a morphism of two-sided comodules. Theorem 2. 1) If (A, γ) is a factorizable c.q.t. Hopf algebra and if M is a simple finitedimensional non-trivial A-comodule, then the above construction gives a bicovariant differential calculus d : (A → 0 ≡ A ⊗ End(M)). 2) Its associated ideal is R = I−1 (annA∗ (k ⊕ M)), where k is the trivial A∗ -module.
84
P. Baumann, F. Schmitt
Proof. We first compute for a ∈ A: P d(a) = k,l a(1) hI(a(2) ), Rk` i ⊗ mk ⊗ m∗` − a(1) ha(2) , δk` i ⊗ mk ⊗ m∗` P = k,l a(1) hI(a(2) ), Rk` − δk` i ⊗ mk ⊗ m∗` , and so: dL (a) = =
P P
k,l
hI(a − ε(a)), Rk` imk ⊗ m∗`
k,l
hJ(Rk` − δk` ), aimk ⊗ m∗` .
The Rji are the matrix coefficients θM (mi , m∗j ) of the A∗ -module M, which is irreducible and non-trivial. Thus, by the Jacobson density theorem, the (dim M)2 + 1 elements {1, Rji } are linearly independent in A. The (dim M)2 linear forms {J(Rk` − δk` )} are then linearly independent in A∗ , and the formula for dL (a) shows that dL maps A onto 0L = End(M). 1) is proved. The same formula shows that R is the set of elements a in the augmentation ideal of A such that I(a) is orthogonal to all the matrix coefficients Rk` of the A∗ -module M. Thus R = ker ε ∩ I−1 (annA∗ M) = I−1 (annA∗ (k ⊕ M)). We have shown 2). If we consider now a finite family (Mi ) of non-trivial non-isomorphic finitedimensional simple right A-comodules, we can do the Ldirect sum of such constructions. If (A, γ) is factorizable, then the map d : (A → (A ⊗ L End Mi )) is a bicovariant Mi )). differential calculus. The associated ideal is I−1 (annA∗ (k ⊕ 3.3. The link with the classification theorem. We are now gathering the pieces of our patchwork. According to the statements in Sect. 3.1, Theorem 1 yields a complete classification of bicovariant differential calculi on Aq G. Morally, they are all given by the construction described in Sect. 3.2. Proposition 10. Let Uq g and Aq G be the objects defined in Sect. 2.1. If the root and the weight lattices for g are equal, all the bicovariant differential calculi on Aq G can be constructed by the method described in Sect. 3.2. Proof. The results in Sect. 2.5 tell us that an ideal R associated to a bicovariant differential calculus on Aq G is a subspace I−1 (annF` (Uq g) M), where M is a V-module containing the trivial V-module. Let M1 , . . . , Mn be the distinct non-trivial irreducible components of M. The assumption on g gives us that the Mi are modules L(λi ) (without any twisting character), and so can be considered as non-trivial non-isomorphic simple right Aq G-comodules. The construction of Sect. 3.2 for this family of comodules leads to a bicovariant differential calculus whose associated L ideal is the inverse image by I Mi . It is R, and the proposition is of the annihilator of the (Aq G)∗ -module C(q) ⊕ proved. In the remainder of this section, we will discuss what happens when the root and the weight lattices differ. Up to the end of this article, we consider this case. There exist non-trivial characters χ : 2P/2Q → C× , and for any weight λ, we can look at the ideal R = I−1 (annF` (Uq g) (C(q) ⊕ Lχ (λ))), and at the associated bicovariant differential calculus. It cannot be constructed by the method of Theorem 2, since Lχ (λ) is not a right Aq G-comodule. However, one may notice that the main trick in the construction −1 of Sect. 3.2 consisted in using two different R-matrices, namely R12 and R21 . R12 was used to endow the Aq G-comodule L(λ) with the structure of a right crossed bimodule
Classification of Bicovariant Differential Calculi on Quantum Groups
85
−1 over Aq G, and R21 turned the Aq G-comodule L(λ)∗ into a right crossed bimodule over Aq G. The tensor product of these right crossed bimodules then gave the bicovariant differential calculus associated to I−1 (annF` (Uq g) (C(q) ⊕ L(λ))). When one uses the small freedom allowed in the choice of the R-matrix of Uq g (see [Ga]), one can make similar constructions for the bicovariant differential calculi associated with some of the ideals I−1 (annF` (Uq g) (C(q) ⊕ Lχ (λ))). We will not write all the details, but point out that this is the way followed by Schm¨udgen and Sch¨uler for the construction described in [S–S1], Theorem 2.2. As an example, we now describe explicitly the bicovariant differential calculus as∧ sociated with the ideal I−1 (annF` (Uq g) (C(q) ⊕ Lχ (0))). Let P/Q be the group of characters ζ : P/Q → C× . If ζ is such a character, it extends to a one-dimensional rep¯ resentation ζ¯ of Aq G by letting ζ(θ (m, m∗ )) = ζ(λ mod Q)hm∗ , mi, and this gives ∧ L(λ) an inclusion of the group P/Q into the center of (Aq G)∗ res . Since (ζ¯ ⊗ id) ◦ δR : ¯ Aq G → C(q) ⊗ Aq G is given by (x 7→ ζ(x) ⊗ 1), we can see that the kernel of ζ¯ is a one-codimensional two-sided ideal of Aq G, stable by the right coaction δR . If ζ is not trivial, the ideal R = ker ε ∩ ker ζ¯ defines a bicovariant differential calculus on Aq G. Putting χ : 2P/2Q → C× , 2λ mod 2Q 7→ ζ(λ mod Q) , we can check that R = I−1 (annF` (Uq g) (C(q) ⊕ Lχ (0))). This construction gives all the one-dimensional differential calculi on Aq G (generalizing the result of [S–S1], Remark 4 after Theorem 2.2). Finally, let X be an intermediate lattice between P and Q. The matrix coefficients of the irreducible representations of Uq g whose highest weights belong to X span a subalgebra Aq GX ⊆ Aq G. These algebras Aq GX are factorizable c.q.t. Hopf algebras. For instance, Aq GQ is the algebra of functions on the quantum adjoint group, and Aq G ≡ Aq GP is the algebra of functions on the quantum simply-connected group. Our arguments in Sect. 2.5 show that the indecomposable bicovariant differential calculi on Aq GX are classified by ideals R = Aq GX ∩ I−1 (annF` (Uq g) (C(q) ⊕ Lχ (λ))), where χ : 2X/2Q → C× is a character (extended arbitrarily to a character of the group 2P/2Q). Thus the “twisted” bicovariant differential calculi are non-local, their appearance depending of the choice of X. The bicovariant differential calculi seem localized at the central elements of GX , that is to say, at the fixed points of GX under the adjoint action.
Acknowledgement. Les deux auteurs remercient le Minist`ere Franc¸ais de l’Enseignement Sup´erieur et de la Recherche pour son soutien financier (allocations de recherche). They also warmly and sincerely thank Professor M. Rosso for explaining the subject, giving bibliographic references, and supplying us with the leading idea.
Note added in proof. P. Polo kindly communicated us the following simple proof of Proposition 8. By the formal character isomorphism, G ' Z[P]. Let Z[P]W ⊆ Z[P] be the subring of W-invariant elements. Z[P] is a module of finite type over the noetherian ring Z[P]W , hence we can choose a finite generating set (eνi )1≤i≤n from the family (eν )ν∈P . Take a weight µ such that all µ +P νi are dominant. LetPλ =∈ P+ . Then there exists some ai ∈ Z[P]W such that e−λ−µ = i ai eνi , hence 1= i ai eλ+µ+νi . Multiplying this by eρ and making the alternating sum over the Weyl group, one abtains that: P ch L(0) = i ai ch L(λ + µ + νi ). This concludes the proof. Thanks are also due to A. Joseph for some useful comments about this work.
86
P. Baumann, F. Schmitt
References [Bo] [Ca]
Bourbaki, N.: Groupes et alg`ebres de Lie, Chapitres 4, 5 et 6. Paris: Masson, 1981 Caldero, P.: El´ements ad-finis de certains groupes quantiques. C. R. Acad. Sci. Paris 316, 327–329 (1993) [Dr1] Drinfel 0d, V.G.: Quantum groups. In: Proceedings of the International Congress of Mathematicians Berkeley 1986. Providence, RI: American Mathematical Society, 1987, pp. 798–820 [Dr2] Drinfel 0d, V.G.: On almost cocommutative Hopf algebras. Leningrad Math. J. 1, 321–342 (1990) [F–P] Faddeev, L.D., Pyatov, P.N.: The differential calculus on quantum linear groups. In: Dobrushin, R.L., Minlos, R.A., Shubin, M.A., Vershik, A. M. (eds.): Contemporary Mathematical Physics (Berezin memorial volume). Amer. Math. Soc. Transl. series 2, vol. 175. Providence, RI: American Mathematical Society, 1996, pp. 35–47 [F–R–T] Faddeev, L.D., Reshetikhin, N.Yu., Takhtadzhyan, L.A.: Quantization of Lie groups and Lie algebras. Leningrad Math. J. 1, 193–225 (1990) [Ga] Gaitsgory, D.: Existence and uniqueness of the R-matrix in quantum groups. J. Algebra 176, 653– 666 (1995) [Jo] Joseph, A.: Quantum groups and their primitive ideals. Ergebnisse der Mathematik und ihrer Grenzgebiete, 3. Folge, Band 29. Berlin–Heidelberg: Springer-Verlag, 1995 [J–L1] Joseph, A., Letzter, G.: Local finiteness of the adjoint action for quantized enveloping algebras. J. Algebra 153, 289–318 (1992) [J–L2] Joseph, A., Letzter, G.: Separation of variables for quantized enveloping algebras. Amer. J. Math 116, 127–177 (1994) [Ju] Jurˇco, B.: Differential calculus on quantized simple Lie groups. Lett. Math. Phys. 22, 177–186 (1991) [L–T] Larson, R.G., Towber, J.: Two dual classes of bialgebras related to the concepts of “quantum group” and “quantum Lie algebra”. Comm. Algebra 19, 3295–3345 (1991) [Ra] Radford, D.E.: Minimal quasi-triangular Hopf algebras. J. Algebra 157, 285–315 (1993) [R–S] Reshetikhin, N.Yu., Semenov-Tian-Shansky, M.A.: Quantum R-matrices and factorization problems. J. Geom. Phys. 5, 533–550 (1988) [Ro1] Rosso, M.: Analogues de la forme de Killing et du th´eor`eme d’Harish-Chandra pour les groupes quantiques. Ann. Sci. Ecole Norm. Sup. 23, 445–467 (1990) [Ro2] Rosso, M.: Certaines formes bilin´eaires sur les groupes quantiques et une conjecture de Schechtman et Varchenko. C. R. Acad. Sci. Paris 314, 5–8 (1992) [Ro3] Rosso, M.: Alg`ebres enveloppantes quantifi´ees, groupes quantiques compacts de matrices et calcul diff´erentiel non commutatif. Duke Math. J. 61, 11–40 (1990) [S–S1] Schm¨udgen, K., Sch¨uler, A.: Classification of bicovariant differential calculi on quantum groups of type A, B, C and D. Commun. Math. Phys. 167, 635–670 (1995) [S–S2] Schm¨udgen, K., Sch¨uler, A.: Classification of bicovariant differential calculi on quantum groups. Commun. Math. Phys. 170, 315–335 (1995) [Ta] Tanisaki, T.: Killing forms, Harish-Chandra isomorphisms, and universal R-matrices for quantum algebras. In: Infinite analysis, Part B, Proceedings Kyoto 1991. Singapore: World Scientific Publishing, 1992, pp. 941–961 [Wo] Woronowicz, S.L.: Differential calculus on compact matrix pseudogroups (quantum groups). Commun. Math. Phys. 122, 125–170 (1989) [Ye] Yetter, D.N.: Quantum groups and representations of monoidal categories. Math. Proc. Camb. Phil. Soc. 108, 261–290 (1990) Communicated by A. Connes
Commun. Math. Phys. 194, 87 – 108 (1998)
Communications in
Mathematical Physics © Springer-Verlag 1998
On Orientation and Dynamics in Operator Algebras. Part I Erik M. Alfsen1 , Frederic W. Shultz2 1 2
Math. Dept. University of Oslo, P.O. Box 1053 Blindern, N-0316 Oslo, Norway. E-mail:
[email protected] Math. Dept. Wellesley College, Wellesley MA-02181, USA. E-mail:
[email protected]
Received: 16 June 1997 / Accepted: 29 September 1997
Abstract: This paper characterizes the self-adjoint part of C*-algebras and von Neumann algebras among normed Jordan algebras. It also explains how the associative product is determined by a general notion of orientation which is related to dynamics and reflects the dual role of physical variables as observables and as generators of oneparameter groups of motions of the state space (Schr¨odinger picture). This concept of orientation bridges the approach in Connes’ characterization of the natural cone of a von Neumann algebra and our own characterization of the state space of a C*-algebra. 1. Introduction An old problem in operator algebras, motivated by physics, is to determine which (partially) ordered normed linear spaces can be the self-adjoint part of a C*-algebra or a von Neumann algebra. This is implicit in several papers of Segal [28, 29] and Kadison [20, 21, 24], and it was explicitly raised for von Neumann algebras by Sakai [27] and for C*-algebras by Sherman [30]. The self-adjoint elements of such algebras are used to represent bounded observables in algebraic models of quantum mechanics. However, the self-adjoint part A of a C*algebra is not closed under the given associative product, but only under the symmetrized product (“Jordan product”) a◦b=
1 1 (ab + ba) = (a + b)2 − a2 − b2 . 2 2
(1)
This product makes A into a (real) Jordan algebra, and it has been proposed to model quantum mechanics on Jordan algebras rather than associative algebras. This approach is corroborated by the fact that many physically relevant properties of observables are adequately described by Jordan constructs. Knowing an element of A, we can express not only the expectation value of the corresponding observable, but its entire probability
88
E. M. Alfsen, F. W. Shultz
law which is given by spectral functional calculus, and in turn by the squaring operation a 7→ a2 . The Jordan algebra approach to quantum mechanics was initiated by Jordan, von Neumann and Wigner in [17] where they introduced and studied finite dimensional “formally real” Jordan algebras. The restriction to finite dimensions was removed by von Neumann [18]. Jordan operator algebras (linear spaces of self-adjoint operators on a Hilbert space closed under the Jordan product) were first studied by Topping [35] and Størmer [33]. Størmer [32] gave a characterization of those Jordan algebras of self-adjoint operators that are actually self-adjoint parts of C*-algebras. The general definitions of JB-algebras and JBW-algebras (together with a Gelfand-Naimark type representation theorem) were given by Alfsen, Shultz and Størmer [4] and by Shultz [31] respectively. These algebras are defined axiomatically as (real) Jordan algebras which are also Banach spaces, subject to suitable conditions connecting Jordan product and norm. The self-adjoint part of a C*-algebra or a von Neumann algebra is a special case of a JB-algebra or a von Neumann algebra respectively. Not all JB-algebras or JBW-algebras arise in this fashion (cf. [13, Ch. 3–4]), but nevertheless they have enough structure to effectively model quantum mechanical observables. However, it is an important feature of quantum mechanics that the physical variables play a dual role, as observables and as generators of transformation groups. The observables are random variables with a specified probability law in each state of the quantum system, while the generators determine one-parameter groups of transformations of observables (Heisenberg picture) or states (Schr¨odinger picture). Both aspects can be adequately dealt with in the C*-algebra or von Neumann algebra formulation of quantum mechanics. An element a in the self-adjoint part A of such an algebra represents an observable whose probability law is determined by spectral theory as indicated above, while an element h in A determines the one-parameter group αt : b 7→ eiht be−iht (equivalently dαt (b)/dt |t=0 = i[h, b]), which represents the time evolution of the observable b. The spectral functional calculus is a Jordan construct, but the generation of one-parameter groups cannot be expressed in terms of the symmetrized product. Instead it is determined by the anti-symmetrized product in A, which we will write as follows: i i a ? b = [a, b] = (ab − ba). (2) 2 2 Thus the decomposition of the associative product into its Jordan part and its Lie part ab = a ◦ b − i(a ? b)
(3)
separates the two aspects of a physical variable. To solve the characterization problem, we must find appropriate conditions for an ordered normed linear space A, under which it is possible to define an associative product on A + iA making this space a C*-algebra or a von Neumann algebra. By the discussion above, this problem can be divided in two parts: first to construct the Jordan part of the associative product, then the Lie part when the Jordan part is known. By a theorem of Kadison [21] the ordering and the norm of a C*-algebra determine the Jordan part of the product. However, they do not determine the product itself, since the opposite algebra has the same ordering and norm but differs in the sign of the Lie part of the product. Thus, to go from the Jordan structure to the C*-structure, we must make a choice for the Lie part of the product.
On Orientation and Dynamics in Operator Algebras. I
89
It was Connes who first realized that a concept of orientation was relevant in this context [8]. In this paper he studies ordered linear spaces associated with sigma-finite von Neumann algebras. (These are the von Neumann algebras which have a faithful normal state (cf. e.g. [34, Prop. II.3.19]), or equivalently the ones which admit a faithful representation with a separating and cyclic vector ξ.) Connes concludes with a characterization of such spaces (Theorem 5.8). Here he shows that a “complex ordered linear space with order unit” (E, E + ) is isomorphic to the pair (M, M + ) for a sigma-finite von Neumann algebra M iff there exists a self-adjoint form s on E such that the completion of the cone E + with respect to s has the following properties: (i) it is “self-polar”, (ii) it is “facially homogeneous”, (iii) it is “orientable”. (All three properties are defined in Connes’ paper.) In the development leading up to this result, it is shown that the completion of M + with respect to a self-polar form s is independent of s (Theorem 2.1), and that it can be identified with the natural cone Pξ\ of Tomita-Takesaki theory, which can be abstractly characterized by the three properties (i), (ii), (iii) above (Theorem 5.2). In [5] Bellissard and Iochum showed that the properties (i) and (ii) above characterize the natural cone associated in an analogous fashion with a (sigma-finite) JBW-algebra. (See also [14].) Thus in the context of such cones, Connes’ notion of orientation is exactly what is needed to move from (sigma-finite) JBW-algebras to (sigma-finite) von Neumann algebras. It follows from results of Kadison [20] that the self-adjoint part of a C*-algebra is isometrically isomorphic, as an ordered normed linear space, to the space A(K) of all w*-continuous affine functions on the state space K. Similarly, the self-adjoint part of a von Neumann algebra is isometrically isomorphic to the space of all bounded affine functions on the normal state space. In view of this, characterizing the self-adjoint part of a C*-algebra (von Neumann algebra) is equivalent to characterizing the state space of a C*-algebra (the normal state space of a von Neumann algebra). This was accomplished for C*-algebras by Alfsen, Hanche-Olsen and Shultz in [3] and for von Neumann algebras by Iochum and Shultz in [15]. Here too the program proceeds by way of Jordan algebras (JB-algebras and JBW-algebras respectively), but it does not involve Tomita-Takesaki theory. The result in [3] is based on two earlier papers of Alfsen and Shultz [1] and [2]. In [1] they gave conditions on the facial structure of a compact convex set guaranteeing that A(K) admits a satisfactory spectral theory and functional calculus. This gives a candidate for a Jordan product in A(K) defined in terms of squares as in Eq. (1). Then in [2, Th. 7.2] they gave necessary and sufficient conditions that this product be bilinear, in which case it makes A(K) a JB-algebra. The key condition is the “Hilbert ball axiom” which says that each face of K which is generated by two extreme points, must be affinely isomorphic to the unit ball of a Hilbert space. (These Hilbert spaces can be of arbitrary finite or infinite dimension for a general JB-algebra. But for a C*-algebra they are of dimension 3 or 1, so for the characterization of C*-algebras the relevant condition is a “3-ball axiom” rather than a general “Hilbert ball axiom”.) As in Connes’ paper, the final step is to move from Jordan structure to associative structure (here from JB-algebras to C*-algebras). Again the key role is played by a concept of “orientability”. In [3] this concept is geometric; one requires that all the “facial 3-balls” (alluded to above) can be oriented in a continuous fashion with respect to the w*-topology. This can always be done locally, so that the requirement is that these local choices can be pieced together in a continuous way to give a global orientation. (See [3, §7] for the precise definition of orientation and [3, Th. 8.4] for the main result characterizing state spaces of C*-algebras.) Note also that if K is the state space of a
90
E. M. Alfsen, F. W. Shultz
C*-algebra, then there is a 1-1 correspondence between all “global orientations” on K and all those associative products on A + iA which organize this space to a C*-algebra inducing the same Jordan product on A as the original algebra [3, Cor.8.5]. In [15], Iochum and Shultz first characterize the normal state space of a JBWalgebra by conditions closely related to Connes’ facial homogeneity axiom. Then they characterize the self-adjoint part of a von Neumann algebra among all JBW-algebras. This is more difficult than the similar problem for C*-algebras in one respect, and easier in another. Since the (generally non-compact) normal state space of a von Neumann algebra may be devoid of extreme points, one must use a modified and more complicated version of the 3-ball axiom in this case. On the other hand, no orientability condition is needed in [15] to solve the characterization problem for the normal state space of a von Neumann algebra. (Nevertheless, here too one can define a notion of “global orientation” in 1-1 correspondence with associative products in the same way as for C*-algebras, but now orientability is automatic, as we will see in Part II.) The concept of orientation introduced by Connes in [8] and that introduced by ourselves in [3] are completely different in character and relate to different contexts, but the purpose is quite similar: to pass from Jordan structure to associative structure. More specifically, Connes’ “orientability axiom” provides a passage from JBW-algebras to von Neumann algebras, while our “orientability axiom” (together with the “3-ball axiom”) provides a passage from JB-algebras to C*-algebras. One of the main purposes of the current paper is to relate these two notions. In the present Part I we will give a solution of the characterization problem which relates to dynamics and applies equally well in the C* and the von Neumann case. In the forthcoming Part II we will develop a general theory of orientation for C*-algebras and von Neumann algebras, which is of geometric nature and bridges the two original approaches to orientation for such algebras. We will now give a brief survey of the content of the paper. We begin by a short summary of definitions and results from the theory of JBalgebras and JBW-algebras (with reference to [13] for proofs). All results are direct generalization of known results for C*-algebras and von Neumann algebras, so readers mainly interested in the theory of orientation for C*-algebras and von Neumann algebras can avoid the intricacies of Jordan algebras. Then we will transfer Connes’ notion of orientation from the natural cone of a JBWalgebra A to the algebra itself, and we will call the resulting notion a “Connes orientation on A”. Like the original concept, such an orientation is a complex structure on the Lie algebra of “order derivations” modulo its center (Definition 16). Our next step will be to introduce yet another notion which is closely related to that of a Connes orientation. This notion, which makes sense both in the JB and the JBW context (and in particular in the C* and the von Neumann context) will be called a “dynamical correspondence”. It is defined to be a (suitably axiomatized) correspondence which assigns a “skew order derivation” ψa to each element a of the given algebra A (Definition 17). The skew order derivations are generators of one-parameter groups of unital order automorphisms of A, and by duality also of one-parameter groups of motions of the state space K of A. Thus a dynamical correspondance gives the elements of A a double identity, which reflects the dual role of physical variables as observables and as generators of a one-parameter group of motions of the state space. Hence the name “dynamical correspondence”. (For related notions, see [6, 7, 9, 25].) To motivate the definition of a dynamical correspondence, we explain the geometrical meaning of this notion in the case of the 2 × 2 matrix algebra (which models a 2-level quantum system, cf. e.g. [26, Ch. 15]). Here the state space is a Euclidean 3-ball, and a
On Orientation and Dynamics in Operator Algebras. I
91
self-adjoint element of the algebra acts as an affine function on the ball. This function attains its maximum and its minimum at two antipodal points, and the corresponding one-parameter group consists of rotations about the diameter between these two points (in either one of the two possible directions depending on orientation). The geometric meaning of a dynamical correspondence in the general case will be explained in Part II. Note that a Connes orientation is defined in the context of JBW-algebras, while a dynamical correspondence is defined in the general context of unital JB-algebras (which include JBW-algebras as a special case). However, for JBW-algebras, it is shown that each Connes orientation determines a unique dynamical correspondence (Proposition 21), and conversely that each dynamical correspondence on a JBW-algebra arises in this way from a unique Connes orientation on the algebra (Corollary 24). Our main result is Theorem 23 by which a unital JB-algebra A is the self-adjoint part of a C*-algebra iff there exists a dynamical correspondence on A, in which case there is a natural 1-1 map from the dynamical correspondences on A to those C*-products on A + iA which induce the given Jordan structure on A. The same conclusions hold with “JBW” in place of “JB” and “von Neumann” in place of “C*”. In Part II we will concentrate on C*-algebras and von Neumann algebras, for which we will define our general notion of orientation. Like the orientation in [3], it is defined geometrically, but here without use of extreme points. Nevertheless, the definition is “local” in that one prescribes an orientation on “small subsystems” and then requires that the choice varies continuously. This local geometric notion of orientation provides a unified framework for studying the passage from Jordan structure to associative structure in C*-algebras and von Neumann algebras. In this framework we will describe the geometry of dynamical correspondences and complete the process of relating the various notions studied in Part I.
2. Order Derivations We begin by giving the definition of JB-algebras and JBW-algebras and some of their basic properties. (A comprehensive treatment of such algebras can be found in [13]). Definition 1. A JB-algebra is a real Jordan algebra A which is also a Banach space such that the Jordan product and the norm are connected by the following conditions for a, b in A: (i) ka ◦ bk = kak ◦ kbk, (ii) ka2 k = kak2 , (iii) ka2 k ≤ ka2 + b2 k. A JB-algebra with identity element 1 is said to be unital. In this paper, we will always assume our JB-algebras are unital. Definition 2. A JBW-algebra is a JB-algebra A which is the dual of a Banach space A∗ . The space A∗ is unique [13, Th.4.4.16], and it is called the predual of A. The self-adjoint part of a C*-algebra is a JB-algebra (for the product a ◦ b = 21 (ab + ba)), and the self-adjoint part of a von Neumann algebra is a JBW-algebra. In fact, many of the basic constructs for these associative algebras can be carried over to their Jordan counterparts. Note in this connection that since a JBW-algebra is a special case of a JB-algebra, all definitions given for JB-algebras also apply to JBW-algebras.
92
E. M. Alfsen, F. W. Shultz
A unital JB-algebra A is a complete order unit space with positive cone A+ = {a2 | a ∈ A} such that for a ∈ A, −1 ≤ a ≤ 1 ⇒ 0 ≤ a2 ≤ 1,
(4)
and this actually characterizes unital JB-algebras among all real Jordan algebras with identity [13, Prop. 3.1.6]. A linear functional ρ on a JB-algebra A is said to be positive, denoted ρ ≥ 0, if ρ(a) ≥ 0 for all a ∈ A+ . The set of all positive functionals in A∗ is a w*-closed (convex) cone, denoted by (A∗ )+ . If A is a unital JB-algebra, then the set of all ρ ∈ (A∗ )+ such that ρ(1) = 1 is a w*-compact convex set called the state space of A. Elements of the state space are called states, and the extreme points of the state space are called pure states. A JB-algebra is monotone complete if every upper bounded increasing net {aα } has a least upper bound a in A. A bounded linear functional ρ on A is called normal if ρ(aα ) → ρ(a) for each net {aα } as above. The positive normal linear functionals form a separating set for A if for every non-zero a ∈ A there exists a positive normal linear functional ρ such that ρ(a) 6= 0. (This fact together with monotone completeness characterize the JBW-algebras among all JB-algebras, and this characterization is taken as the definition in [13].) The predual A∗ of a JBW-algebra A can be identified with the subspace of A∗ which consists of all normal linear functionals [13, Th. 4.4.16]. We will use the term σ-weak topology to denote the topology on A determined by the duality with A∗ (the σ(A, A∗ )-topology in Bourbaki’s terminology). Thus, the σ-weakly continuous linear functionals are precisely the normal ones. Note that every JBW-algebra A is unital [13, Lem.4.1.7]. The convex set of normal states on A is called the normal state space of A. If A is a JB-algebra, then A∗∗ can be made into a JBW-algebra in such a way that the state space of A is identified with the normal state space of A∗∗ [13, §4.4]. Furthermore, multiplication is separately σ-weakly continuous on a JBW-algebra [13, Cor. 4.1.6], so we can often make use of the σ-weak density of A in A∗∗ . We will now introduce an order theoretic concept of derivation which plays an important role in Connes’ paper [8]. It can be defined in the general context of ordered linear spaces, but we will only give the definition for JB-algebras. But first some motivating remarks. Derivations occur in many different contexts. What is common for various derivations δ, is the fact that they are linear operators generating a one-parameter group of maps etδ which preserve the algebraic structure under study. In our present context, we are focusing on the order structure, ignoring the multiplicative aspect. Therefore the Leibniz rule is not relevant here. Definition 3. A bounded linear operator δ on a JB-algebra A is called an order derivation if etδ (A+ ) ⊂ A+ for all t ∈ R, or what is equivalent, if {etδ }t∈R is a one-parameter group of order automorphisms. Lemma 4. An order derivation δ on a JBW-algebra A is σ-weakly continuous. Proof. We will first show that the map φ = etδ is σ-weakly continuous for given t ∈ R, or what is the same, that ρ ◦ φ ∈ A∗ for every ρ ∈ A∗ . Since φ is an order automorphism and ρ is a normal linear functional, then ρ ◦ φ is also a normal linear functional. Thus ρ ◦ φ ∈ A∗ as desired. The order derivation δ is the norm limit of t−1 (etδ − 1) when t → 0, so we can find a sequence {ψn } of σ-weakly continuous linear maps such that kψn − δk → 0. Let
On Orientation and Dynamics in Operator Algebras. I
93
ρ ∈ A∗ . Then ρ ◦ ψn ∈ A for every n and kρ ◦ ψn − ρ ◦ δk → 0. Since A∗ is complete, ρ ◦ δ ∈ A∗ . Since ρ ∈ A was arbitrary, δ is σ-weakly continuous. We will give a necessary and sufficient condition that a linear operator be an order derivation. The idea behind this criterion is the following: If δ is an order derivation, then the orbit of etδ through a boundary point of the cone A+ cannot proceed to the exterior of A+ , nor to the interior, so the velocity vector must lie in the tangent space. Connes turned this heuristic argument into a precise criterion for self-polar cones [8, Lemma 5.3]. Later on Hanche-Olsen and Evans generalized it to arbitrary cones with the nearest point property, i.e. to cones for which the minimum distance from an arbitrary point to a point in the cone is effectively attained [12]. In this form it can be applied also in our context, as the positive cone of a unital JB-algebra, and in fact of every order unit space, has the nearest point property. This is an elementary result, which is certainly well known. But since we have not been able to find a good reference, we include the proof. Lemma 5. The positive cone A+ of an order unit space A has the nearest point property. Proof. Let a ∈ / A+ . Observe that there exists λ ∈ R+ such that a + λ1 ∈ A+ . In fact, we can take λ = kak, since a ≥ −kak1 (by the definition of order unit [13, 1.2.1]). Set λ0 = inf {λ ∈ R | a + λ1 ∈ A+ }.
(5)
Since the positive cone A is closed, there is an element b = a + λ0 1 ∈ A . We claim that b is a nearest point for a, i.e. that kc − ak ≥ kb − ak = λ0 for every c ∈ A+ . Let c ∈ A+ and set λ = kc − ak. Then c ≤ a + λ1 (again by the definition of the order unit). Since c ∈ A+ , also a + λ1 ∈ A+ . Hence λ ≥ λ0 as claimed. +
+
Lemma 6. A bounded linear operator δ on a unital JB-algebra is an order derivation iff the following implication holds for all a ∈ A+ and ρ ∈ (A∗ )+ ρ(a) = 0 ⇒ ρ(δa) = 0.
(6)
Proof. By Theorem 1 of [12] the quoted statement holds in the context of ordered Banach spaces with the nearest point property. By Lemma 5 it can be applied in our case. We will denote the Jordan multiplier determined by an element b of a JB-algebra A by δb . Thus for all a ∈ A (7) δb (a) = b ◦ a. Lemma 7. Let A be a unital JB-algebra. Then δb is an order derivation for every b ∈ A. Proof. Suppose ρ = 0 for a ∈ A+ and ρ ∈ (A∗ )+ . By the Cauchy-Schwartz inequality for JB-algebras [13, 3.6.2], kρ(δb a)k2 = kρ(b ◦ a)k2 ≤ ρ(b2 )ρ(a2 ).
(8)
Generally a ≤ kaka for every a ∈ A . In fact, the Jordan triple product (defined in [13, 2.3.2]) determines an order preserving map a 7→ {cac} for every c ∈ A [13, 3.3.6], 1 so if we evaluate a 2 by spectral theory [13, 3.2.4], we can write 2
+
1
1
1
1
a2 = {a 2 aa 2 } ≤ {a 2 (kak1)a 2 } = kaka. Now it follows from (8) that kρ(δb a)k2 ≤ ρ(b2 )kakρ(a) = 0. By Lemma 6, δb is an order derivation.
94
E. M. Alfsen, F. W. Shultz
Definition 8. An order derivation δ on a unital JB-algebra A is self-adjoint if δ = δa for some a ∈ A, and it is skew-adjoint (or just skew) if δ(1) = 0. Our next lemma shows, among other things, that the skew order derivations are the Jordan derivations, i.e. the bounded linear operators δ which satisfy the Leibniz rule δ(a ◦ b) = (δa) ◦ b + a ◦ (δb).
(9)
Lemma 9. Let A be a unital JB-algebra with state space K and let δ be an order derivation on A. For every t ∈ R, let αt = etδ and let αt∗ be the dual map defined on A∗ by (αt∗ ρ) = ρ(αt (a)) for ρ ∈ A∗ and a ∈ A. Now the following are equivalent: (i) (ii) (iii) (iv) (v)
δ is skew, αt (1) = 1 for all t, αt is a Jordan automorphism for all t, δ is a Jordan derivation, αt∗ (K) ⊂ K for all t.
Proof. (i) ⇒ (ii) Use the exponential series for etδ . (ii) ⇒ (iii) By a known theorem of Kadison, every unital order automorphism of a C*-algebra is a Jordan automorphism [21]. The same result is in fact valid for a JB-algebra. It can most easily be obtained from Theorem 12.13 of [1], by which the Jordan square a2 = a ◦ a of an element a of A (and hence every Jordan product a ◦ b) is determined by the ordering and the order unit (via the spectral integral R completely λ2 deλ of [1, Eq. (8.28)]). Thus, if αt is a unital order automorphism, then it is also a Jordan automorphism. (iii) ⇒ (iv) Since αt is a Jordan automorphism, then α(a ◦ b) = α(a) ◦ α(b) for a, b ∈ A. By the standard argument δ(a ◦ b) = lim t−1 (αt (a) ◦ αt (b) − a ◦ b) = (δa) ◦ b + a ◦ (δb). t→0
(iv) ⇒ (i) By Leibniz’ rule δ(1) = δ(1 ◦ 1) = 2δ(1). Hence δ(1) = 0. (ii) ⇔ (v) Trivial. For our next proof we shall need two elementary results valid for elements x, y, z in a unital Banach algebra. The first is the equation lim k(1 + n−1 x)n − ex k = 0,
n→∞
(10)
which follows from the continuity of the holomorphic functional calculus. The second is the inequality ky n − z n k ≤ n · Max{kyk, kzk}n−1 ky − zk,
(11)
which follows from the decomposition y n − z n = y n−1 (y − z) + y n−2 (y − z)z + · · · + (y − z)z n−1 . We will denote the set of order derivations of a JB-algebra A by D(A). Proposition 10. The set D(A) of order derivations of a unital JB-algebra A is a real linear space closed under Lie brackets [δ1 , δ2 ] = δ1 δ2 − δ2 δ1 .
On Orientation and Dynamics in Operator Algebras. I
95
Proof. The fact that D(A) is closed under linear operations, follows directly from Lemma 6. To show that D(A) is closed under Lie brackets, it suffices to show that [δ1 , δ2 ] is an order derivation for a given pair δ1 , δ2 ∈ D(A). By looking at the first few terms of the exponential series involved, we see that etδ1 etδ2 e−tδ1 e−tδ2 = 1 + t2 [δ1 , δ2 ] + t4 φt ,
(12)
where kφt k is bounded for t in a neighbourhood of 0. 1 Set tn = n− 2 and define αn = etn δ1 etn δ2 e−tn δ1 e−tn δ2 , βn = 1 + t2n [δ1 , δ2 ] and γn = φtn for n = 1, 2, · · · . Now αn − βn = n−2 γn
(13)
for n = 1, 2, · · ·, and {kγn k} is a bounded sequence. Clearly (αn )n is an order automorphism for every n. It follows from (10) that k(βn )n − exp[δ1 , δ2 ]k → 0 when n → ∞. Thus we only have to show that k(αn )n − (βn )n k → 0 when n → ∞. We will prove this by applying (11) with αn and βn in place of y and z, and we begin by showing that {kαn kn } and {kβn kn } are bounded sequences. By (12), kαn k ≤ 1 + n−1 k[δ1 , δ2 ]k + n−2 kγn k for n = 1, 2, · · ·. We will assume [δ1 , δ2 ] 6= 0 (otherwise there is nothing to prove). Let λ > 1 be arbitrary. Then for sufficiently large n kαn k ≤ 1 + λn−1 k[δ1 , δ2 ]k ≤ exp λn−1 kδ1 , δ2 k , which gives
kαn kn ≤ exp λk[δ1 , δ2 ]k .
Since λ > 1 was arbitrary,
lim kαn kn ≤ exp k[δ1 , δ2 k .
n→∞
For every n and then also
kβn k ≤ 1 + n−1 k[δ1 , δ2 ]k, kβn kn ≤ exp k[δ1 , δ2 k .
Let M > expk[δ1 , δ2 ]k ≥ 1. By the inequalities above, kαn kn ≤ M and kβn kn ≤ M , and then also kαn kn−1 ≤ M and kβn kn−1 ≤ M , for sufficiently large n. Now it follows from (11) and (13) that k(αn )n − (βn )n k ≤ nM kαn − βn k ≤ n−1 M kγn k for large n. Thus k(αn )n − (βn )n k → 0 when n → ∞, and we are done.
Lemma 11. Every order derivation δ on a unital JB-algebra A can be decomposed uniquely as the sum of a self-adjoint and a skew derivation, namely δ = δa + δ 0 , where a = δ(1).
96
E. M. Alfsen, F. W. Shultz
Proof. Set a = δ(1) and δ 0 = δ − δa . Then δ 0 (1) = a − a ◦ 1 = 0, so δ = δa + δ 0 is a decomposition of the desired type. If δ = δb + δ 00 is another such decomposition, then evaluation at 1 gives δ(1) = b ◦ 1 = b, so b = a. Therefore the decomposition is unique. To each δ ∈ D(A) we will associate the adjoint order derivation δ ∗ defined by δ = δa − δ 0 , where δ = δa + δ 0 as above. Thus δ ∈ D(A) is self-adjoint iff δ ∗ = δ. Two elements a, b of a JB-algebra A are said to operator commute if the operators δa , δb commute, i.e. if (a ◦ x) ◦ b = a ◦ (x ◦ b) for all x ∈ A. If A is the self-adjoint part of a C*-algebra, then a, b operator commute iff ab = ba [11, Lem. 5.1]. The set of all elements in a JB-algebra A which operator commute with every other element of A is called the center of A, and it will be denoted by Z(A). Note that Z(A) is an associative subalgebra of A. We will also denote by Z(D(A)) the center of the Lie algebra D(A), i.e. the set of all δ ∈ D(A) such that [δ, δ 0 ] = 0 for every other element δ 0 of D(A). ∗
Lemma 12. If δ is a skew order derivation on a unital JB-algebra A and z ∈ Z(A), then (i) etδ z = z for all t ∈ R, (ii) δz = 0. Proof. Recall that the bidual A∗∗ is a JBW-algebra. By σ-weak continuity of multiplication in each variable separately, the bidual map (etδ )∗∗ is also a Jordan automorphism. Furthermore, again by σ-weak continuity of multiplication, the center of A will be contained in the center of A∗∗ , so it suffices to prove the lemma for the special case where A is a JBW-algebra. Then it is enough to prove the lemma for the case where z is a central idempotent, i.e. z 2 = z. Since δ is skew, then etδ is a Jordan automorphism, so etδ z is also a central idempotent for t ∈ R. Thus etδ z − z is the difference of two central projections, so ketδ z − zk is either zero or one. Since ketδ z − zk is a continuous function of t which is zero when t = 0, it must be zero for all t. This proves (i). Now also δz = limt→0 t−1 (etδ z − z) = 0, which proves (ii). Lemma 13. If A is a unital JB-algebra, then Z(D(A)) = {δz | z ∈ Z(A)}.
(14)
Proof. Assume first that δ ∈ Z(D(A)). In particular [δ, δa ] = 0 for every a ∈ A. Let z = δ(1). Then for every a ∈ A, δ(a) = δδa (1) = δa δ(1) = δa (z) = a ◦ z = z ◦ a. Hence δ = δz . Also δz δa = δa δz , so z ∈ Z(A). Assume next that z ∈ Z(A). By the definition of Z(A), δz commutes with every self-adjoint order derivation δa . Therefore we only have to show that δz commutes with every skew derivation δ. But such a derivation is a Jordan derivation, so it follows from the Leibniz rule (9) and Lemma 12 (ii) that δδz (x) = δ(z ◦ x) = (δz) ◦ x + z ◦ (δx) = δz δ(x) for every x ∈ A. Thus δδz = δz δ as desired.
On Orientation and Dynamics in Operator Algebras. I
97
If A is the self-adjoint part of a C*-algebra A, then we will assign to each d ∈ A a linear operator δd on A defined by δd (x) =
1 (dx + xd∗ ) 2
(15)
for x ∈ A. Note that δd (x) = (δd (x))∗ . Lemma 14. If A is the self-adjoint part of a C*-algebra A and δ = δd for d ∈ A, then for x ∈ A and t ∈ R, ∗ (16) exp(2tδ)(x) = etd xetd . In particular if δ = δa for a ∈ A (self-adjoint case), then exp(2tδ)(x) = eta xeta ,
(17)
and if δ = δib for b ∈ A (skew case), then exp(2tδ)(x) = eitb xe−itb .
(18)
Proof. Consider the left and right multiplication operators Ld : x 7→ dx and Rd : x 7→ xd∗ defined on A for d ∈ A. Since Ld and Rd∗ commute, then for x ∈ A, ∗
exp(2tδd )(x) = exp(tLd + tRd∗ )(x) = exp(tLd )exp(tRd∗ )(x) = etd xetd . This proves (16), from which (17) and (18) both follow.
Proposition 15. If A is the self-adjoint part of a von Neumann algebra A, then the order derivations of A are the operators δd on A defined above, and an order derivation δ is self-adjoint (skew) iff it is of the form δd for d self-adjoint (skew). Proof. An order derivation is self-adjoint iff it is of the form δa for a ∈ A (by definition). Clearly also, an order derivation is skew if it is of the form δib (i.e. of the form δd with d skew). Conversely we will show that an arbitrary skew order derivation δ is of this form. For every t ∈ R the map etδ is a Jordan automorphism of A. We extend it by (complex) linearity to all of A. By a theorem of Kadison [22] there is a central projection c such that etδ acts as a *-isomorphism from cA into A and a *-anti-isomorphism from (1 − c)A into A. By Lemma 9, etδ fixes c. Hence etδ is a *-automorphism of cA and a *-anti-automorphism of (1 − c)A. Applying etδ twice, we observe that e2tδ acts as a *-automorphism also on (1 − c)A. Since t ∈ R was arbitrary, this means that etδ is in fact a *-automorphism of A for every t ∈ R. Thus by the Kadison-Sakai theorem the generator δ of the one-parameter group {etδ }t∈R is an inner derivation on A, i.e. δ(x) = 21 (hx − xh) for some h ∈ A and all x ∈ A [19, Ex.8.7.55]. Let h = a + ib, where a, b ∈ A. Then for each x ∈ A, δ(x) = −iδia (x) + δib (x). Since δ(x) = δ(x)∗ , then δia (x) = 0 and δ(x) = δib (x). Thus δ = δib . If d ∈ A, say d = a + ib, then δd is the sum of the order derivations δa and δib , so δd is also an order derivation. Conversely we will show that every order derivation is of this form. Assume that δ is an arbitrary order derivation on A and consider the decomposition δ = δa + δ 0 established in Lemma 11. By the argument above, δ 0 = δib for some b ∈ A. Thus for every x ∈ A,
98
E. M. Alfsen, F. W. Shultz
δ(x) =
i 1 1 (ax + xa) + (bx − xb) = (a + ib)x + x(a + ib)∗ = δa+ib (x), 2 2 2
so δ is of the desired form.
The results above can easily be dualized to the predual A∗ of the von Neumann algebra A. If δ is an order derivation of A and αt = etδ , then we consider the dual operator (αt )∗ on the self-adjoint part A∗ of A∗ for t ∈ R. Generally {(αt )∗ }t∈ R is a one-parameter group of order automorphisms of A∗ , and if δ is skew then each (αt )∗ leaves the normal state space invariant (Lemma 9 (v)), so {(αt )∗ }t∈ R is a one-parameter group of affine automorphisms of the normal state space. The orbits of {(αt )∗ }t∈R can be easily visualized in the case where A is the 2 × 2 matrix algebra. Here the (normal) state space is a Euclidean 3-ball, and the pure state space is the surface of the ball, i.e. a Euclidean 2-sphere. If a ∈ A is self-adjoint and has two distinct eigenvalues λ1 < λ2 corresponding to (unit) eigenvectors ξ1 , ξ2 , then the vector states ωξ1 , ωξ2 are antipodal points on the sphere (South Pole and North Pole on Fig.1). If δ = δia (the skew case), then (αt )∗ is a rotation of the ball by an angle t(λ1 − λ2 )/2 about the diameter [ωξ1 , ωξ2 ]. Thus the one-parameter group {(αt )∗ }t∈R represents a rotational motion with rotational velocity (λ1 − λ2 )/2 about this diameter, and the orbits on the sphere are the “parallel circles" (in planes orthogonal to [ωξ1 , ωξ2 ]). If δ = δa (the self-adjoint case), then the orbits will take us out of the state space. But this can be remedied by a normalization, i.e. by considering the parametric curves t 7→ k(αt )∗ σk−1 (αt )∗ σ instead of t 7→ (αt )∗ σ. These are the “longitudinal semi-circles" on the sphere (in planes through [ωξ1 , ωξ2 ]). The proof of these facts is elementary matrix calculation and will be omitted. ωξ2
δa
δia
ωξ1 Fig. 1.
In the example above we can easily see how the one-parameter group is determined by the geometry in the self-adjoint case, and we can also see what indeterminacy there is in the skew case. The self-adjoint element a ∈ A determines a real valued affine function aˆ : ω 7→ ω(a) on the state space. This function attains its minimum λ1 at ωξ1 and its maximum λ2 at ωξ2 . In the self-adjoint case the orbits are the longitudinal semi-circles traced out in the direction from ωξ1 to ωξ2 . In the skew case the orbits are the parallel circles, but they can be traced out in two possible directions “eastbound" and “westbound". The mere knowledge of the affine function aˆ does not tell us which
On Orientation and Dynamics in Operator Algebras. I
99
direction to choose. This would require a specific orientation of the ball (right-handed −→ or left-handed around ωξ1 ωξ2 ).
3. Connes Orientations and Dynamical Correspondences We will now transfer Connes’ concept of orientation [8, Def. 4.11] to the context of JBW-algebras. The idea is to axiomatize the map δd 7→ δid of D(A) into itself where A is the self-adjoint part of a von Neumann algebra A, or rather the corresponding map which is obtained when elements of D(A) are identified modulo Z(D(A)). ˜ To simplify the notation, we will write D(A) in place of D(A)/Z(D(A)). We will ˜ Note also denote the equivalence class of an element δ of D(A)) modulo Z(D(A)) by δ. ∗ ) is well defined on D(A)), g ˜ for if δ˜1 = δ˜2 then δ1 − δ2 = δz that the involution δ˜ ∗ = (δ ∗ ∗ g g for some z ∈ Z(A) (Lemma 13). Hence δ1∗ − δ2∗ = δz∗ = δz , so (δ 1 ) = (δ2 ). Definition 16. A Connes orientation on a JBW-algebra A is a complex structure on ˜ D(A), which is compatible with Lie brackets and involution, i.e. a linear operator I on ˜ D(A) which satisfies the requirements (i) I 2 = −1 (the identity map), (ii) [I δ˜1 , δ˜2 ] = [δ˜1 , I δ˜2 ] = I[δ˜1 , δ˜2 ], ˜ ∗. (iii) I(δ˜ ∗ ) = −(I δ) If A is the self-adjoint part of a von Neumann algebra A, then it is easily verified that I : δ˜d 7→ δ˜id (with d ∈ A) is a Connes orientation of A. We call it the Connes orientation induced on A from A. (Note that the map I0 : δd 7→ δid from D(A) into itself is not well-defined, since d is not determined by δd if d is not known to be self-adjoint.) An alternative approach is to take the basic properties of the map a 7→ δia , where a is a self-adjoint element of a von Neumann algebra A as axioms for a map ψ which assigns to each element a in a general unital JB-algebra A a skew order derivation ψa on A. Geometrically ψ assigns to each real valued affine function on the state space K of A a one-parameter group of affine automorphisms of K. Such a one-parameter-group describes a motion of states, and we will call ψ a “dynamical correspondence". The precise definition is the following: Definition 17. A dynamical correspondence on a unital JB-algebra A is a linear map ψ : a 7→ ψa from A into the set of skew order derivations on A which satisfies the requirements (i) [ψa , ψb ] = −[δa , δb ] for a, b ∈ A, (ii) ψa a = 0 for all a ∈ A. A dynamical correspondence on a JB-algebra A will be called complete if it maps A onto the set of all skew order derivations on A. It is easily verified that condition (ii) above is equivalent to the statement that exp(tψa ) fixes a for all a ∈ A and all t ∈ R. This property is easily visualized in the 2 × 2 matrix example, and it will play an important role in the geometrical investigations in Part II. We will also state the definition of a dynamical correspondence in another form. In this connection we shall need the following lemma, which will also be needed later.
100
E. M. Alfsen, F. W. Shultz
Lemma 18. Let A be a unital JB-algebra and let ψ be a map from A into the set of all skew order derivations on A. Then for all pairs a, b ∈ A, [ψa , δb ] = δψa b .
(19)
Proof. Since ψa is skew, it is a Jordan derivation. Hence for all c ∈ A, ψa (b ◦ c) = (ψa b) ◦ c + b ◦ (ψa c), which can be rewritten ψa δb c − δb ψa c = δψa b c. This gives (19).
Note that linearity of ψ is not listed among the requirements in the proposition below, since it follows from the other requirements. Proposition 19. Let A be a unital JB-algebra and let ψ : a 7→ ψa be a map from A into the set of skew order derivations of A. Then ψ is a dynamical correspondence iff the following requirements are satisfied for a, b ∈ A (i) [ψa , ψb ] = −[δa , δb ], (ii) [ψa , δb ] = [δa , ψb ]. Proof. Assume first that ψ is a dynamical correspondence. Condition (i) above is trivially satisfied as it is identical with condition (i) of Definition 17. By condition (ii) of Definition 17 and Lemma 18, then for all a ∈ A, [ψa , δa ] = δψa a = δ0 = 0. By linearity of ψ, then for all a, b ∈ A, 0 = [ψa+b , δa+b ] = [ψa , δb ] + [ψb , δa ], which gives [ψa , δb ] = −[ψb , δa ] = [δa , ψb ], and proves condition (ii) above. Assume next that ψ satisfies conditions (i) and (ii) above. Condition (i) of Definition 17 is trivially satisfied, and it follows from condition (ii) above and Lemma 18 that for all a ∈ A, δψa a = [ψa , δa ] = [δa , ψa ] = −[ψa , δa ]. Hence ψa a = δψa a 1 = 0, so condition (ii) of Definition 17 is also satisfied. It remains to show that ψ is linear. By Lemma 18 and condition (ii) above, then for all a, b, δψa b (1) = [ψa , δb ](1) = [δa , ψb ](1) = −[ψb , δa ](1) = −δψb a (1). Hence ψa b = −ψb a,
(20)
from which it follows that a 7→ ψa b is a linear map from A into D(A) for each fixed b ∈ A.
On Orientation and Dynamics in Operator Algebras. I
101
In the above proof we have actually also shown that if ψ is a dynamical correspondence on A, then Eq. (20) above holds for all pairs a, b ∈ A. We shall make more use of this equation later. We will now explain the relationship between Connes orientations and dynamical correspondences, and we begin with the following: Lemma 20. If ψ is a dynamical correspondence on a unital JB-algebra A, then the kernel of ψ : A → D(A) consists of all self-adjoint order derivations δz , where z ∈ Z(A). Proof. If a ∈ ker ψ, then it follows from Definition 17 (i) that [δa , δb ] = −[ψa , ψb ] = 0 for all b ∈ A. Hence a ∈ Z(A). Conversely, if z ∈ Z(A) then δz ∈ D(Z(A)) (Lemma 13), so it follows from Lemma 18 and Proposition 19 (ii) that for all b ∈ A, δψz b = [ψz , δb ] = [δz , ψb ] = 0. Thus ψz b = 0 for all b ∈ A, so z ∈ ker ψ.
Proposition 21. A Connes orientation I on a JBW-algebra A determines a complete dynamical correspondence ψ such that ψa ∈ I(δ˜a ) for all a ∈ A, and ψ is the only dynamical correspondence with this property. Proof. Let a ∈ A. We will first show that there exists a unique δ ∈ I(δ˜a ) such that δ is skew. Choose δ1 ∈ I(δ˜a ). By Definition 16 (iii), δ˜1∗ = I(δ˜a )∗ = −I(δ˜a∗ ) = −I(δ˜a ) = −δ˜1 . ˜ so δ ∗ + δ1 = δz for some z ∈ Z(A) (Lemma 13). Now define Hence δ˜1∗ + δ˜1 = 0, 1 1 1 δ = δ1 − 2 δz = 2 (δ1 − δ1∗ ). Thus δ is skew. Also 1 δ = δ1 − δz ∈ δ˜1 = I(δ˜a ). 2
(21)
If δ1 is an arbitrary skew order derivation such that δ 0 ∈ I(δ˜a ), then δ − δ 0 is both central and skew. By Lemma 13 δ − δ 0 = 0, so δ is the unique skew order derivation in I(δ˜a ). Denote the unique skew order derivation in I(δ˜a ) by ψa for each a ∈ A. Clearly ψ : a 7→ ψa is a linear map from A into D(A). Let a, b ∈ A. By Definition 16 (i),(ii), [I δ˜a , I δ˜b ] = I[δ˜a , I δ˜b ] = I 2 [δ˜a , δ˜b ] = −[δ˜a , δ˜b ]. Thus for some z ∈ Z(A) [ψa , ψb ] = −[δa , δb ] + δz . Since ψa and ψb are skew, then [ψa , ψb ](1) = 0. Also [δa , δb ](1) = a ◦ b − b ◦ a = 0. Hence z = δz (1) = 0. Thus [ψa , ψb ] = −[δa , δb ], so ψ satisfies condition (i) of Definition 17. To show that ψ also satisfies condition (ii) of Definition 17, we first observe that for all a ∈ A, then by Definition 16 (ii), [ψ˜ a , δ˜a ] = [I δ˜a , δ˜a ] = I[δ˜a , δ˜a ] = 0.
102
E. M. Alfsen, F. W. Shultz
By Lemma 13 there exists z ∈ Z(A) such that [ψa , δa ] = δz . Now δz is a commutator of two bounded linear operators on A and it commutes with each of them, so it follows from the Kleinecke-Shirokov Theorem [10, p.128] that δz is quasi-nilpotent, i.e. 1
lim kδz n k n = 0.
n→∞
But δz n (1) = z n for all n: Hence n
kδz 2 k2
−n
n
≥ kz 2 k2
−n
= kzk,
(22)
so z = 0. (The norm-closed subalgebra generated by a single element and 1 is isometrically isomorphic to C(X) for some compact Hausdorff space X [13, 3.2.4], so n n kz 2 k = kzk2 ). Thus [ψa , δa ] = 0. Now it follows from Lemma 18 that δψa a = 0, and hence also ψa a = 0, so we have verified condition (ii) of Definition 17. Thus ψ is a dynamical correspondence. To show that ψ is complete, we consider an arbitrary skew order derivation δ and we will show that δ = ψa for some a ∈ A. More specifically, we choose an arbitrary ˜ and we will show that δ1 is self-adjoint, i.e. of the form δ1 = δa for a ∈ A, δ1 ∈ I(−δ) and that ψa = δ. By Definition 16 (iii) and the fact that δ is skew, ˜ ∗ = I(δ˜ ∗ ) = I(−δ) ˜ = δ˜1 . δ˜1∗ = −(I δ) Thus δ1∗ −δ1 = δz for some z ∈ Z(A). Applying both sides to 1, we get (δ1∗ −δ1 )(1) = z. Since δ1∗ − δ1 is skew, then z = 0. Thus δ1∗ = δ1 , so δ1 is self-adjoint, i.e. δ1 = δa for some a ∈ A. ˜ ˜ so δ˜a = −I δ. ˜ Then by Definition 16 (i), I δ˜a = −I 2 δ˜ = δ. By definition δa ∈ I(−δ), ˜ ˜ Thus δ ∈ I δa , so δ is the unique skew order derivation in I δa ; in other words δ = ψa . With this we have shown that ψ is a complete dynamical correspondence. The uniqueness is clear, since ψa is the only skew order derivation in I δ˜a . The concept of a Connes orientation is defined for JBW-algebras, while the concept of a dynamical correspondence is defined for unital JB-algebras, so the two concepts cannot be equivalent. Note however, that it will follow from our main theorem that a dynamical correspondence on a JBW-algebra is necessarily complete and is derived from a Connes orientation as in Proposition 21 (Corollary 25). Thus the two concepts are in fact equivalent in the context of JBW-algebras.
4. The Main Theorem We are now ready to prove our main theorem which relates dynamical correspondences to associative products. Definition 22. Let A be a unital JB-algebra. A C*-product (W*-product) compatible with A is an associative product (x, y) 7→ xy on the complex linear space A + iA which induces the given Jordan product on A and organizes A + iA to a C*-algebra (von Neumann algebra) with the involution (a + ib)∗ = a − ib and the norm kxk = kx∗ xk1/2 .
On Orientation and Dynamics in Operator Algebras. I
103
Note that if a JB-algebra A is the self-adjoint part of a C*-algebra A, then we can transfer the product and the norm from A to A + iA by the representation x = a + ib (where x ∈ A and a, b ∈ A). This organizes A + iA to a C*-algebra with the properties in the definition above. Thus, a JB-algebra is the self-adjoint part of a C*-algebra iff there exists a C*-product compatible with A on A + iA, similarly in the JBW-context. Theorem 23. A unital JB-algebra is (Jordan isomorphic to) the self-adjoint part of a C*-algebra iff there exists a dynamical correspondence on A. In this case each dynamical correspondence ψ on A determines a unique C*-product compatible with A such that for a, b ∈ A, i (23) ψa b = (ab − ba), 2 and each C*-product compatible with A arises in this way from a unique dynamical correspondence ψ on A. The same conclusions hold with “JBW" in place of “JB" and “W*" or “von Neumann" in place of “C*". Proof. Assume first that A admits a dynamical correspondence ψ. By Eq. (20) of Proposition 19 we can define an anti-symmetric bilinear product (a, b) 7→ a?b on A by writing a ? b = ψa b.
(24)
Next define a bilinear map (a, b) → 7 ab from A × A into A + iA (considered as a real linear space) by writing ab = a ◦ b − i(a ? b). (25) This map can be uniquely extended to a bilinear product on A + iA (considered as a complex linear space). We will show that this product is associative. By linearity, it suffices to prove the associative law a(cb) = (ac)b
(26)
for a, b, c ∈ A. Writing out (26) by means of (25), we get a ◦ (c ◦ b) − i a ? (c ◦ b) − i(a ◦ (c ? b) − a ? (c ? b) = (a ◦ c) ◦ b − i (a ◦ c) ? b − i (a ? c) ◦ b − (a ? c) ? b. Separating real and imaginary terms (and using the anti-symmetry of the ?-product), we get two equations. The first one can be written as follows: a ? (b ? c) − b ? (a ? c) = −a ◦ (b ◦ c) + b ◦ (a ◦ c),
(27)
and the second one as follows: a ? (b ◦ c) − b ◦ (a ? c) = a ◦ (b ? c) − b ? (a ◦ c).
(28)
The left-hand side of (27) is nothing but [ψa , ψb ](c), and the right-hand side of (27) is nothing but −[δa , δb ](c). Similarly the left-hand side of (28) is [ψa , δb ](c), and the right-hand side of (28) is [δa , ψb ](c). Thus these two equations follow directly from the characterization of a dynamical correspondence in Proposition 19. We must also show that the bilinear product on A + iA is compatible with the natural involution (a + ib)∗ = a − ib on A + iA, i.e. that (xy)∗ = y ∗ x∗ for x, y ∈ A + iA. By
104
E. M. Alfsen, F. W. Shultz
linearity it suffices to show that (ab)∗ = ba for a, b ∈ A. But this follows directly from the antisymmetry of the ?-product, as ∗ (ab)∗ = a ◦ b − i(a ? b) = a ◦ b + i(a ? b) = b ◦ a − i(b ? a) = ba. We have now shown that A + iA is an associative *-algebra. By the definition of the involution, the self-adjoint part of A + iA is A. Thus by (25) and the anti-symmetry of the ?-product, 1 (ab + ba) = a ◦ b 2 for all pairs a, b ∈ A. Thus the associative product in A + iA induces the given Jordan product on A. We will now show that x∗ x ∈ A+ for every x ∈ A+iA. The closed Jordan subalgebra C(x∗ x) of A generated by the self-adjoint element x∗ x and 1 is associative, hence isometrically isomorphic to the real commutative Banach algebra C(X) for a compact Hausdorff space X [13, Th.3.2.2]. Thus x∗ x = a−b, where a, b are two positive elements of C(x∗ x) such that a ◦ b = 0. By the definition of the Jordan triple product [13, 2.3.2] and the associativity of C(x∗ x), {bab} = 2b ◦ (b ◦ a) − (b ◦ b) ◦ a = b ◦ (b ◦ a) = 0. But since the Jordan product in A is equal to that induced from the associative algebra A + iA, the same is true for the Jordan triple product. Therefore also bab = 0. Calculating in the associative *-algebra A + iA, we now find that (xb)∗ (xb) = bx∗ xb = b(a − b)b = −b3 .
(29)
Since 0 ≤ b ∈ C(x∗ x), then b3 ≥ 0 and b3 = 0 iff b = 0. Thus in order to prove b = 0, it suffices to show that (xb)∗ (xb) = 0. For brevity we set y = xb, and we will show that y ∗ y = 0. By (29) y ∗ y ∈ −A+ . Write y = c + id, where c, d ∈ A and calculate yy ∗ + y ∗ y = 2(c2 + d2 ), which gives
yy ∗ = 2(c2 + d2 ) − y ∗ y ∈ A+ .
Thus we have shown that y ∗ y ∈ −A+
and yy ∗ ∈ A+ .
(30)
Following [13, 3.2.9] we define the inverse of an element a ∈ A to be its inverse in C(a) (if it exists), and we denote it by a−1 . Note that this definition is equivalent to the usual definition of inverse in Jordan algebras; namely that a0 ∈ A is the inverse of a if a0 satisfies a ◦ a0 = 1 and a2 ◦ a0 = a [4, Prop.2.4]. Furthermore it is easily shown that a ∈ A is invertible with inverse a0 in the Jordan algebra A iff a ∈ A is invertible with inverse a0 in the Jordan algebra A + iA iff a is invertible with inverse a0 in the associative *-algebra A + iA. (For the last equivalence, see [16, p.51].) By definition, a real number λ is in the spectrum of an element a of the JB-algebra A, in symbols λ ∈ sp(a), if λ1 − a is non-invertible in A. Thus, by the above, λ ∈ sp(a) iff λ1 − a is non-invertible in the associative algebra A + iA.
On Orientation and Dynamics in Operator Algebras. I
105
Calculating in the associative algebra A + iA (in the same way as in the proof of [19, Prop.3.2.8]), we find that if λ1 − y ∗ y is invertible and λ 6= 0, then (λ1 − yy ∗ ) y(λ1 − y ∗ y)−1 y ∗ + 1 = λ1, so λ1 − yy ∗ is also invertible. With this we have shown that sp(y ∗ y)\{0} = sp(yy ∗ )\{0}. By (30) sp(y ∗ y) = sp(yy ∗ ) = {0}, and by the spectral theorem for JB-algebras [13, Th.3.2.4] y ∗ y = yy ∗ = 0. Hence b = 0, and then x∗ x ∈ A+ as claimed. Define now for x ∈ A + iA, kxk = kx∗ xk 2 . 1
(31)
Extend each state ρ on A to a complex linear functional on A + iA. Note that since the states separate the points of A, their extensions will separate the points of A + iA. Define for x ∈ A + iA, (32) (x | y)ρ = ρ(y ∗ x). Since x∗ x ∈ A+ for all x ∈ A + iA, this is a semi-definite inner product. Construct now the GNS-representations (πρ , Hρ ) in the usual way, and let (π, H) be the direct sum of all such representations. If x is a non-zero element of A + iA, then there exists a state ρ such that (x | 1)ρ = ρ(x) 6= 0. Thus by the Cauchy-Schwarz inequality, ρ(x∗ x) 6= 0, and then πρ (x) 6= 0. Thus π is a *-isomorphism of A + iA into B(H). In particular, π restricts to a Jordan isomorphism of A into the self-adjoint part of B(H), so it follows from [13, Prop.3.4.3] that kπ(a)k = kak for a ∈ A. By (31) and the corresponding equation for the norm of B(H), we now have for x ∈ A + iA kπ(x)k = kπ(x∗ x)k 2 = kx∗ xk = kxk. 1
(33)
Thus (30) defines a norm on A + iA which makes π : A + iA → B(H) an isometric *-isomorphism. Clearly this norm satisfies the C*-condition kxk2 = kx∗ xk, and we will now show that A + iA is complete for this norm. Pulling back the corresponding inequalities from B(H), we have for every x = a + ib ∈ A + iA, max(kak, kbk) ≤ kxk ≤ kak + kbk. Since A is complete in the order unit norm, the space A + iA must be complete in the norm (31), and then be a C*-algebra. It follows from (24), (25) and the anti-symmetry of the ?-product that 1 i (ab − ba) = (a ? b − b ? a) = ψa (b). 2 2 Thus we have constructed a C*-product compatible with A which satisfies the requirement (23). This product is unique since the compatibility with the Jordan algebra A determines the self-adjoint part a ◦ b in (25), and the requirement (23) determines the skew part a ? b in (25). Assume next that A + iA is equipped with a C*-product (x, y) 7→ xy compatible with the JB-algebra A. Now we define a map ψ from A into D(A) by Eq. (23), and we prove by straightforward calculation that the two requirements (i),(ii) of Definition 17 are satisfied. Thus ψ is a dynamical correspondence on A. Also we can recover the given
106
E. M. Alfsen, F. W. Shultz
associative product by substituting ψa b for a ? b in (25), so this product arises from ψ by the construction in the first part of the proof. It only remains to specialize to JBW-algebras. In this connection we must prove that if A is a JBW-algebra and (x, y) 7→ xy is a C*-product on A + iA which is compatible with A, then A + iA is a von Neumann algebra. The quickest proof of this fact is based on Kadison’s theorem that a monotone complete C*-algebra with a separating family of normal states is a von Neumann algebra [23] (or [19, Ex. 7.6.38]). In fact, the conditions on monotone completeness and separation by normal states are both imposed on the self-adjoint part of the C*-algebra, and since they are satisfied for the JBW-algebra A, there is nothing more to prove. Remark. The proof above shows that if A is a JB-algebra and A + iA is equipped with an associative product such that a + ib 7→ a − ib is an involution, then A + iA can be normed (in a necessarily unique way) to become a C*-algebra. Corollary 24. A JBW-algebra A is the self-adjoint part of a von Neumann algebra iff there exists a Connes orientation on A. In this case each Connes orientation I on A determines a unique W*-product compatible with A such that for d ∈ A + iA, I(δ˜d ) = δ˜id ,
(34)
and each W*-product compatible with A arises in this way from a unique Connes orientation on A. Proof. Assume first that A admits a Connes orientation I. By Proposition 21 there exists a dynamical correspondence ψ on A such that ψa ∈ I(δ˜a ) for a ∈ A. Construct the corresponding W*-product in A + iA as in Theorem 23. Let d ∈ A, say d = a + ib, where a, b ∈ A. Then to verify (34) we observe that for all c ∈ A, i δib c = (bc − cb) = ψb c, 2 so δib = ψb . Since ψ˜ b = I(δ˜b ), then I(δ˜ib ) = I(ψ˜ b ) = I 2 (δ˜b ) = −δ˜b , which gives
I(δ˜d ) = I(δ˜a ) + I(δ˜ib ) = ψ˜ a − δ˜b = δ˜ia − δ˜b = δ˜id ,
(35)
and establishes (34). Recall that (as shown in the proof of Proposition 21) there is a unique skew-adjoint order derivation in I(δ˜id ). Thus any two W∗ -products compatible with A and satisfying (34) induce the same Lie multiplication map δid for each d ∈ A, as well as the same Jordan multiplication (inherited from A), and thus coincide. Assume next that A + iA is equipped with a C*-product compatible with the JBWalgebra A. Thus A is the self-adjoint part of a von Neumann algebra A, and by Proposition 15 each order derivation on A is of the form δd for some d ∈ A. Now we can define a ˜ map I from D(A) into itself by (34), and we prove by straightforward calculation that the requirements (i),(ii),(iii) of Definition 16 are satisfied. Thus I is a Connes orientation on A. Clearly I is the unique Connes orientation on A for which Eq. (34) is satisfied. Corollary 25. A dynamical correspondence ψ on a JBW-algebra A is necessarily complete , and there is a unique Connes orientation I on A such that ψa ∈ I(δ˜a ) for all a ∈ A.
On Orientation and Dynamics in Operator Algebras. I
107
Proof. Let ψ be a dynamical correspondence on a JBW-algebra A. By Theorem 23 there is a unique W*-product in A such that (23) holds. In other words, A is the self-adjoint part of a von Neumann algebra A which is unique (up to a *-isomorphism) under the requirement ψa = δia for all a ∈ A. Now it follows from Corollary 24 that there is a Connes orientation I on A such that I(δ˜a ) = δ˜ia = ψ˜ a . Thus ψa ∈ I(δ˜a ) for all a ∈ A. By the same argument as the one leading up to (35), I(δ˜d ) = δ˜id for all d ∈ A. Since each order derivation on A is of the form δd for some d ∈ A (Proposition 15), this proves that I is uniquely determined by the requirement ψ˜ a = I(δ˜a ) = δ˜ia
for a ∈ A.
(36)
By Proposition 21 there is a unique dynamical correspondence on A for which (36) holds, and this correspondence is complete. Thus ψ is complete, and we are done. References 1. Alfsen, E.M. and Shultz, F.W.: Non-commutative spectral theory for affine function spaces on convex sets. Memoirs A.M.S. 172, 120 (1976) 2. Alfsen, E.M.and Shultz, F.W.: On state spaces of Jordan algebras. Acta Math. 140, 155–190 (1978) 3. Alfsen, E.M., Hanche-Olsen, H.,Shultz, F.W.: On state spaces of C∗ -algebras. Acta Math. 144, 267–305 (1980) 4. Alfsen, E.M., Shultz, F.W. and Størmer, E.: A Gelfand-Neumark theorem for Jordan algebras. Adv. Math. 28, 11–56 (1978) 5. Bellissard, J. and Iochum, B.: Homogeneous self-dual cones versus Jordan algebras, the theory revisited. Ann. Inst. Fourier (Grenoble) 28, 27–67 (1978) 6. Rajarama Bhat, B.V.: On a characterization of velocity maps in the space of observables. Pac. J. Math. 152, 1–14 (1992) 7. Bunce, L.J. and Maitland-Wright, J.D.: Velocity maps in von Neumann algebras. Pac. J. Math. 170, 421–427 (1995) 8. Connes, A.: Charact´erisation des espaces vectoriels ordonn´es sous-jacent aux alg´ebres de von Neumann. Ann. Inst. Fourier (Grenoble) 24, 121–155 (1974) 9. Grgin E. and Petersen, A.: Duality of observables and generators in classical and quantum mechanics. J. Math. Phys. 15, 764–769 (1974) 10. Halmos, P.R.: Hilbert Space Problem Book. Toronto–London–Melbourne: D.Van Nostrand, 1976 11. Hanche-Olsen, H.: On the structure and tensor products of JC-algebras. Can. J. Math. 35, 1059–1074 (1983) 12. Hanche-Olsen, H. and Evans, D.E.: The generators of positive semigroups. J. Func. Anal. 32, 207–212 (1979) 13. Hanche-Olsen, H. and Størmer, E.: Jordan Operator Algebras. Boston–London–Melbourne: Pitman, 1984 14. Iochum, B.: Cˆones autopolaires et alg`ebres de Jordan. Lecture Notes in Math. 1049, Berlin–Heidelberg– New York: Springer-Verlag, 1984 p. 249 15. Iochum, B. and Shultz, F.W.: Normal state spaces of Jordan and von Neumann algebras. J. Funct. Anal. 50, 317–328 (1983) 16. Jacobson, N.: Structure and representations of Jordan algebras. Am. Math. Soc. Colloq. Publ. 39, Providence R.I.. AMS, 1968 17. Jordan, P., von Neumann, J., and Wigner, E.: On an algebraic generalization of the quantum mechanical formalism. Ann. Math. 35, 29–64 (1934) 18. von Neumann, J.: On an algebraic generalization of the quantum mechanical formalism. Mat. Sbornik, 1, 415–482 (1936) 19. Kadison, R.V. and Ringrose, J.R.: Fundamentals of the Theory of Operator Algebras, I-II. London–New York: Academic Press, 1986 20. Kadison, R.V.: A representation theory for commutative topological algebra, Memoirs A.M.S. 7, 36 (1951)
108
E. M. Alfsen, F. W. Shultz
21. Kadison, R.V.: A generalized Schwarz inequality and algebraic invariants for operators. Ann. of Math. 56, 494–503 (1952) 22. Kadison, R.V.: Isometries of operator algebras. Ann.Math. 54, 325–338 (1951) 23. Kadison, R.V.: Operator algebras with a faithful weakly-closed representation. Ann. Math. 64, 175–181 (1956) 24. Kadison, R.V.: Transformation of States in Operator Theory and Dynamics. Topology 3, Suppl. 2, 177–198 (1967) 25. Landsman, N.P.: Poisson Spaces with a transition probability. Rev. Math. Phys. 9, 29–57 (1997) 26. Mandel, L., Wolf, E.: Optical Coherence and Quantum Optics. Cambridge: Cambridge Univ. Press, 1995 27. Sakai, S.: The absolute value of W∗ -algebras of finite type. Tohoku Math. J. 8, 70–85 (1956) 28. Segal, I.E.: Irreducible representations of operator algebras. Bull. A.M.S. 53, 73–88 (1947) 29. Segal, I.E.: Postulates for general quantum mechanics. Ann. Math. 48, 930–948 (1947) 30. Sherman, S.: On Segal’s postulates for general quantum mechanics. Ann. Math. 64, 593–601 (1956) 31. Shultz, F.W.: On normed Jordan algebras which are Banach dual spaces J. Func. Anal. 31, 360–376 (1979) 32. Størmer, E.: On the Jordan structure of C∗ -algebras. Trans. Am. Math. Soc. 120, 438–447 (1965) 33. Størmer, E.: Jordan algebras of type I. Acta Math. 115, 165–184 (1966) 34. Takesaki, M.: Theory of Operator Algebras I. Berlin–Heidelberg–New York:: Springer-Verlag, 1979 35. Topping, D.: Jordan algebras of self-adjoint operators. Mem. Am. Math. Soc. 53, 48 (1965) Communicated by H. Araki
Commun. Math. Phys. 194, 109 – 134 (1998)
Communications in
Mathematical Physics © Springer-Verlag 1998
Quasi-Classical Asymptotics for the Pauli Operator Alexander V. Sobolev? Centre for Mathematical Analysis and its Applications, University of Sussex, Falmer, Brighton, BN1 9QH, UK. E-mail:
[email protected] Received: 23 July 1997 / Accepted: 30 September 1997
P Abstract: We study the behaviour of the sums Mγ = k |λk |γ , γ > 0 of the eigenvalues of the Pauli operator in L2 (Rd ), d = 2, 3, in a magnetic field µB(x) and electric field V (x) as the Planck constant ~ tends to zero and the magnetic field strength µ tends to infinity. We show that for µ~ ≤ const the sum Mγ obeys the natural Weyl type formula Z X σ 1−d |B| V−σ + 2 2kµ~|B| + V − dx, Mγ ∼ Cγ,d µ~ k>0
where σ = (d − 2)/2 + γ, with an explicit constant Cγ,d . If the field B has a constant direction, then this formula is uniform in µ ≥ 0. The method is based on Colin de Verdiere’s approach proposed in his work on “magnetic bottles” (Commun. Math Phys, 105, 327–335 (1986)).
1. Introduction The aim of the present paper is to study the negative discrete spectrum of the Pauli operator ( P = P0 + V, 2 Pd 2 (1.1) P0 = σ(−i~∇ − µa) = k=1 σk (−i~∂k − µak ) , acting in L2 () ⊕ L2 (), d = 2, 3, with the Dirichlet conditions on the boundary of the domain ⊂ Rd . Here a = (a1 , . . . , ad ) is a magnetic vector-potential, V is an electric (real-valued) potential decreasing at infinity, σ = (σ1 , σ2 , σ3 ) denotes the 3-tuple of 2×2 Pauli matrices, i.e. Hermitian matrices, such that ?
Author supported by EPSRC under grant B/94/AF/1793.
110
A. V. Sobolev
σk σl + σl σk = 2δlk I, and for any cyclic permutation (klm) of the numbers (123) one has σk σl = iσm . Up to a unitary equivalence all the realizations of these relations can be described by the standard Pauli matrices: 01 0 −i 1 0 , σ2 = , σ3 = . (1.2) σ1 = 10 i 0 0 −1 As one can see from (1.1), the operator P0 is non-negative. Moreover, it is known that under fairly broad conditions on the magnetic field B = curl a the point λ = 0 belongs to the spectrum of P0 , and in certain cases the spectrum coincides with the half-line [0, ∞) (see [3, 14] and references therein). The perturbed operator P = P0 + V may have some negative eigenvalues λk = λk (~, µB, V, ), k = 1, 2, . . . . We enumerate them counting multiplicity in the non-decreasing order. We shall study the asymptotic behaviour of the sums X |λk |γ , γ ≥ 0, (1.3) Mγ = Mγ (~, µB, V, ) = k
as ~ → 0, µ → ∞ and discuss the applicability of the natural quasi-classical asymptotic formula Mγ (~, µB, V, ) ∼ ~−d Bγ (µ~|B|, V, ), ~ → 0,
(1.4)
with the “magnetic” Weyl coefficient B(d) γ (b, V, )
Z =
βγ(d)
βγ(2)
b(x) V− (x)
d−2 2 +γ
+2
X
d−2 2 +γ
2kb(x) + V (x)
−
dx,
(1.5)
k>0 −1
= (2π)
,
βγ(3)
2 −1
= (2π )
Z
1
1
(1 − t) 2 tγ−1 dt.
γ
(1.6)
0
Notice that for b → 0 this coefficient tends to the standard Weyl coefficient Z d V− (x)γ+ 2 dx Cd,γ
(with an explicit constant Cd,γ ), so that as µ~ → 0 one recovers from (1.4) the usual Weyl asymptotics. On the contrary, if µ~ → ∞, then (1.4) reduces to Mγ (~, µB, V, ) ∼ µ~−d+1 Aγ (|B|, V, ), ~ → 0, with
Z Aγ (b, V, ) =
A(d) γ (b, V, )
=
βγ(d)
b(x)V− (x)
d−2 2 +γ
dx.
(1.7)
(1.8)
It was shown in [8, 9] that for homogeneous fields (i.e. for B = const) the asymptotics (1.4) for γ = 1 holds uniformly in the field strength µ. The first result of this kind for non-homogeneous fields was recently established in [6, 7] with the supplementary condition µ~3 → 0. In the present paper we prove that the formula (1.4) holds true if µ~ remains bounded from above (Theorem 2.1), whereas in the case of a constant direction field the asymptotics (1.4) is uniform in µ (Theorem 2.2). Notice that for d = 2 the field always has a constant direction. The strategy adopted in this paper involves the following two ingredients:
Quasi-Classical Asymptotics for the Pauli Operator
111
1. Quasi-classical asymptotics for the Pauli operator in a cube with the Dirichlet boundary conditions; 2. Appropriate a priori upper bound for the sum of the eigenvalues (usually referred to as the Lieb–Thirring inequality). First we decompose the domain into cubes. Then bit 1 yields the required asymptotics for the Pauli operator with the Dirichlet conditions on the sides of the cubes. The role of the Lieb–Thirring estimate is to ensure that the obtained asymptotic formula in the limit of small cubes turns into (1.4) for the initial problem. Our approach was inspired by Y. Colin de Verdiere’s work [2] on “magnetic bottles”, where the idea of calculating the spectral asymptotics of the magnetic Schr¨odinger operator by decomposing the configuration space into cubes with the Dirichlet conditions, was realized for the first time. Recall that the crucial point of this work was an explicit formula for the eigenvalues of the operator with a constant magnetic field on a torus, which subsequently led to appropriate spectral estimates on a cube. In fact, when implementing step 1 (see above) for µ~ ≤ const, we use these estimates directly, approximating B by a piecewise constant magnetic field (see Lemma 3.1, Proposition 3.2). In the case µ~ → ∞ (when B is assumed to have a constant direction) this approximation is too crude. Thus rather than apply the results of [2] directly, we use the strategy of this paper, modifying the “model” problem. Namely, we rely on the spectral properties of the Pauli operator on a torus with an arbitrary periodic (not necessarily constant) B having integer flux (see Sect. 4). The only piece of information we need here is that the point λ = 0 is an eigenvalue of multiplicity, which is equal to the flux of B (see [4] and Appendix A). The other eigenvalues are irrelevant, for they do not contribute into the asymptotics of Mγ as µ~ → ∞ (see Lemma 4.3). The crucial role in our approach is played by the Lieb–Thirring estimates for Mγ established in [11, 12]. In particular, those estimates are responsible for the region of values µ, ~, in which the asymptotics (1.4) is valid. Precisely, by [11, 12], Mγ ≤ C(~−3 + µ3/2 ~−3/2 ), Mγ ≤ C(~
−d
−d+1
+ µ~
),
for d = 3; for constant direction B’s, d = 2, 3
(1.9) (1.10)
(Here for the sake of discussion we ignore the dependence of these estimates on B and V and refer to (2.14), (2.15) for precise formulation). The requirement µ~ ≤ const in the case d = 3 (Theorem 2.1) is due to the fact that the estimate (1.9) shows the same behaviour in µ, ~ as the asymptotics (1.4), only for µ~ ≤ const. On the contrary, for constant direction fields (Theorem 2.2), the estimate (1.10) has the “correct” order for all µ and ~ (cf. (1.4), (1.7)). The discrepancy between (1.9) and (1.10) is connected with the fundamental property that the three-dimensional Pauli operator with a non-constant direction field may have eigenfunctions with eigenvalue zero (zero modes) (see [10]), which makes it impossible for (1.10) to hold in general. The authors of [6, 7] have developed a sophisticated method, which provides a better control over the zero modes and consequently, a Lieb–Thirring estimate having better orders of µ and ~ than (1.9), which allows for an ampler condition µ~3 → 0 of validity of (1.4). Without going into details we mention that the method of [6, 7] is concerned with the geometry of the field lines and requires more restrictive conditions on B and V than in our paper. In particular, in [6, 7] the magnetic field is supposed to be non-vanishing and bounded from above. Referring to the papers quoted above for further discussion of magnetic Lieb– Thirring inequalities and references, we mention however the most recent version of
112
A. V. Sobolev
the Lieb–Thirring inequality, obtained in [1]. It is similar to the estimate (2.14) below and reduces to (1.9) as far as the behaviour in ~ and µ is concerned. The paper is organized as follows. Section 2 contains the precise satements of main results. In Sect. 3 we prove the asymptotics (1.4) for µ~ ≤ const for the Pauli operator in a cube, assuming that V = const. Section 4 is concerned with the operator on a torus and provides necessary auxiliary estimates used in Sect. 5 for proving the asymptotics (1.4) in a cube for fields of constant direction. The proof of main results is completed in Sect. 6. Appendix A contains some known information regarding the problem on a torus (see [4]). Some technical material is collected in Appendix B. In conclusion we make some notational conventions. Sometimes we reflect the dependence of Mγ , Bγ and other quantities occurring in the paper, on the parameters of the problem (such as ~, a or V , etc). However, whenever it is convenient, we omit some or all of them from the notation. We may write, for example, Bγ (b) or Bγ (V ) instead of Bγ (b, V, ). The letters C, c with or without indices denote various positive constants, whose value is of no importance and may change from line to line. By int we denote the interior of a set ⊂ Rd . a ≈ b means that there exists a positive constant C such that C −1 ≤ a/b ≤ C. For any measurable real-valued function f we denote by f+ and f− its positive and negative parts respectively: f± = (|f | ± f )/2. This convention does not apply to operators (cf. (2.1) below). The points of Rd for both d = 2 and d = 3 as a rule are denoted by bold lower case letters: x, y, . . . . When we want to emphasize the difference between the two- and three- dimensional case we write x, y, . . . for two-dimensional variables. In particular this notation is used when representing x ∈ R3 as x = (x, x3 ), x ∈ R2 , x3 ∈ R. We shall work with two underlying Hilbert spaces: h = h() = L2 () and H = h⊕h. As a rule, elements of h and H are denoted by u, v and f, g respectively. It will be also convenient to introduce the notation d = d() = C0∞ (), D = d ⊕ d. Operators on H are usually denoted by blackboard bold letters, for instance P, H, etc. Furthermore, if W is an operator on h, then W denotes the operator W I, where I is the identity operator in C2 . This convention does not apply only to the letters N and R, C, which are reserved for the set of all positive integers and the fields of real and complex numbers respectively. For a self-adjoint operator T , R(z, T ) = (T − z)−1 denotes its resolvent. If T is semi-bounded, then T [·, ·] denotes the closed form associated with T , with the domain D[T ].
2. Conditions and Results
2.1. Basic definitions. First of all we give a precise definition of the operator (1.1) with the Dirichlet boundary conditions. We use the choice (1.2) of the Pauli matrices. Let ⊂ Rd be an open set (possibly unbounded). Let a = (a1 , . . . , ad ) ∈ L2loc () be a magnetic vector-potential with real-valued components. Define on d() the operators Q± = 51 ± i52 , 5k = −i~∂k − µak , k = 1, 2, 3.
(2.1)
Throughout the paper each statement containing the double subscript “±” must be understood separately for the sign “+” and “−”. The operators Q± , 5k are closable on
Quasi-Classical Asymptotics for the Pauli Operator
113
d(), since 5k are symmetric and Q± ⊂ Q∗∓ . We use the same letters Q+ , Q− , 5k for their closures. Consequently, the symmetric operator X 53 Q− σk 5k = (2.2) T= Q+ −53 k
acting in H() is also closed. We define the Pauli operator P0 as the operator associated with the closed quadratic form P0 [f, f ] = kTf k2 with the domain D[P0 ] = D(T). This implies that P0 = T∗ T. In the same way we define in h() the usual Schr¨odinger operator Ha with the magnetic field a: as an operator associated with the form Ha [u, v] = P 2 k h5k u, 5k vi. As a rule we assume that the vector-potential a ∈ Lloc () is such that the magnetic field defined in the distributional sense as B(x) = (B1 , B2 , B3 ) = curl a(x), Bl = ∂k aj − ∂j ak , (for any cyclic permutation l, k, j of the numbers 1, 2, 3), belongs to L∞ loc (). It is easy to check using (2.1) that A+ 0 0 B1 − iB2 − µ~ P0 = , B1 + iB2 0 0 A− ( Q∗± Q± , d = 2; (2.3) A± = 523 + Q∗± Q± , d = 3. Note the relation between A± and Ha : A± = Ha ∓ µ~B3
(2.4)
in the sense of sesqui-linear forms on d. The perturbed operator is denoted by P = P0 +V. The conditions under which this operator is well defined, will be specified later. As mentioned in the Introduction, in the case of a magnetic field having a constant direction we establish the asymptotics of Mγ in a larger region of values of the parameter µ than in the general case. We state the results for these two cases separately in Subsects. 2.2 and 2.3 below. The conditions on the potential V and the magnetic field B in both cases are chosen so as to ensure the validity of appropriate Lieb–Thirring estimates (see [11, 12]). 3 2.2. General fields, d = 3. Let `(x), b(x) be two positive functions such that b ∈ L∞ loc (R ). Define D(3) (x) = {y : |x − y| ≤ `(x)}. Assume that |`(x) − `(y)| ≤ %|x − y|, % < 1, (2.5) c ≤ b(y)b(x)−1 ≤ C, y ∈ D(3) (x), b(x)`(x)2 ≥ c.
Then we suppose that |B(x)| ≤ b(x), x ∈ . As shown in [12], if for some p > 3/2 the condition Z 3 |V (y)|p b(y) 2 + 1 dy < ∞ sup x∈
D (3) (x)∩
(2.6)
(2.7)
114
A. V. Sobolev
is fulfilled, then the operator P = P(V ) = P0 + V is well-defined as a form-sum on the domain D[P0 ]. Let the coefficient Bγ be defined in (1.5), (1.6). Theorem 2.1. Let the magnetic field B be a continuous function obeying the condition (2.6) for some function b(x) satisfying (2.5), and let (2.7) be fulfilled. Suppose that V− ∈ Lγ+3/2 (), V−γ b3/2 ∈ L1 () for some γ ≥ 1. Then the asymptotics d (2.8) lim ~ Mγ (~, µB, V, ) − Bγ (µ~|B|, V, ) = 0, ~ → 0, holds uniformly in µ~ ≤ C. 2.3. Constant direction fields. If the magnetic field has a constant direction, then under an appropriate choice of coordinates it can be written as B(x) = (0, 0, B(x)), where x = (x, x3 ), x = (x1 , x2 ). One can find a gauge such that a = (a1 , a2 , 0) with ak = ak (x), k = 1, 2, so that B = ∂1 a2 − ∂2 a1 (which is always the case for d = 2). In view of (2.3) in these coordinates the operator P0 is diagonal. The conditions on B will look the same for two- and three-dimensional situations. 2 (2) (x) = Namely, let `(x), b(x) be two positive functions such that b ∈ L∞ loc (R ). Define D {y : |x − y| ≤ `(x)}. Assume that |`(x) − `(y)| ≤ %|x − y|, % < 1, (2.9) c ≤ b(y)b(x)−1 ≤ C, y ∈ D(2) (x), b(x)`(x)2 ≥ c. Then we suppose that |B(x)| ≤ b(x), x ∈ . As shown in [11, 12], if for some p > d/2 the condition Z sup |V (y)|p b(y) + 1 dy < ∞, d = 2, x∈
or
(2.11)
D (2) (x)∩
Z sup x∈
(2.10)
D (2) (x)×[x3 ,x3 +1)∩
|V (y)|p b(y) + 1 dy < ∞, d = 3,
(2.12)
is fulfilled, then the operator P = P0 + V is well-defined as a form-sum on the domain D[P0 ]. Theorem 2.2. Let a constant direction magnetic field B be a continuous function obeying the condition (2.10) with a function b, satisfying (2.9), and let (2.11) (for d = 2) or (2.12) (for d = 3) be fulfilled. Suppose that γ+(d−2)/2
V− ∈ Lγ+d/2 (), V−
b ∈ L1 (),
for some γ ≥ 1 (for d = 2) or γ > 1/2 (for d = 3). Then the asymptotics −1 lim µ~1−d + ~−d |Mγ (~, µB, V, ) − ~−d Bγ (µ~|B|, V, )| = 0, ~ → 0, holds uniformly in µ ≥ 0.
(2.13)
Quasi-Classical Asymptotics for the Pauli Operator
115
2.4. Lieb–Thirring inequalities. The derivation of the asymptotics (2.8), (2.13) is based on the following estimates on Mγ established in [11, 12]: Proposition 2.3. 1. Let the magnetic field B obey the condition (2.6) and let (2.7) be fulfilled. Suppose that V− ∈ Lγ+3/2 (), V−γ b3/2 ∈ L1 () for some γ ≥ 1. Then Z 3 Mγ (~, µB, V, ) ≤ C~−3 V− (x)γ+ 2 dx Z 3 3 3 b(x) 2 V− (x)γ dx. (2.14) + C 0 µ 2 ~− 2
2. Let B have a constant direction, obey the condition (2.10) and let (2.11) (for d = 2) or (2.12) (for d = 3) be fulfilled. Suppose that γ+(d−2)/2
V− ∈ Lγ+d/2 (), V−
b ∈ L1 (),
for some γ ≥ 1 (for d = 2) or γ > 1/2(for d = 3). Then Z d −d Mγ (~, µB, V, ) ≤ C~ V− (x)γ+ 2 dx Z d−2 b(x)V− (x)γ+ 2 dx. + C 0 µ~−d+1
(2.15)
Initially, these estimates were proved for = Rd , d = 2, 3, which immediately implies (2.14), (2.15) for any open set in view of the variational principle by setting B = 0, V = 0 outside . Emphasize in this connection that the functions b, ` are defined in the entire space (and not only in !) to ensure the applicability of the results of [11, 12]. Note also that for B ∈ L∞ the estimates take a simpler form, which follows from (2.14), (2.15) by choosing b = ess − sup |B(x)|, `(x) = 1. x∈
Precisely, for d = 3 we have −3
Mγ (~, µB, V, ) ≤ C~
Z
3
V− (x)γ+ 2 dx + C 0 µ 2 ~− 2 b 2 3
3
3
Z
If the field has a constant direction, then Z d V− (x)γ+ 2 dx Mγ (~, µB, V, ) ≤ C~−d
+ C 0 µ~−d+1 b
V− (x)γ dx, γ ≥ 1.
Z
V− (x)γ+
d−2 2
dx.
(2.16)
(2.17)
The constants C, C 0 in (2.16) and (2.17) are the same as in (2.14) and (2.15) respectively and do not depend on . 2.5. Properties of Mγ and Bγ . Here we list some useful properties of the quantity (1.3). Below the parameters ~, µ and the field a are fixed and omitted from notation. Introduce
116
A. V. Sobolev
the discrete spectrum counting function: N (λ) = #{λk , λk < −λ}, λ ≥ 0. Note that M0 = N (0). The sum (1.3) can be represented as Z ∞ Z ∞ λγ dN (λ, V ) = γ λγ−1 N (λ, V )dλ. (2.18) Mγ (V ) = − 0
0
Furthermore, using the Weyl inequalities
N (λ, V1 + V2 ) ≤ N λ, (1 − ζ)−1 V1 + N λ, ζ −1 V2 , N (λ, V1 + V2 ) ≥ N λ, (1 + ζ)−1 V1 − N λ, −ζ −1 V2 ,
valid for any ζ ∈ (0, 1), one checks with the help of (2.18) that ( Mγ (V1 + V2 ) ≤ Mγ (1 − ζ)−1 V1 + Mγ ζ −1 V2 , Mγ (V1 + V2 ) ≥ Mγ (1 + ζ)−1 V1 − Mγ −ζ −1 V2 .
(2.19)
We conclude this section with some remarks on the coefficient Bγ . Notice, first of all that, similarly to (2.18), Z ∞ λγ−1 B0 (V + λ)dλ. (2.20) Bγ (V ) = γ 0
Now denote σ = γ + (d − 2)/2. The coefficient Bγ can be estimated with the help of the quantities R ( Md,γ = Md,γ (b, V, ) = b(x)|V (x)|σ dx, b(x) ≥ 0, (2.21) R Nd,γ = Nd,γ (V, ) = |V (x)|σ+1 dx, as follows: Bγ (b, V, ) ≤ βγ(d) Md,γ (b, V− , ) Z Z b(x) + 2βγ(d)
∞
σ
2tb(x) + V (x) 0
−
dtdx
≤ CMd,γ (b, V− , ) + C 0 Nd,γ (V− , ),
(2.22)
0
with some constants C, C depending on d, γ. Furthermore, Bγ is continuous in b and V for σ > 0. Precisely, for any functions V, V1 , V2 and non-negative functions b, b1 , b2 one has if σ ≤ 1: |Bγ (b, V1 ) − Bγ (b, V2 )| ≤ CMd,γ (b, V1 − V2 ) 1 1 σ +CNd,γ (V1 − V2 ) σ+1 Nd,γ (V1 ) σ+1 + Nd,γ (V2 ) σ+1 , and
(2.23)
|Bγ (b1 , V ) − Bγ (b2 , V )| ≤ C Md,γ (|b1 − b2 |, V ) 1 1 + Md,γ (|b1 − b2 |, V ) 2 Nd,γ (V ) 2 σ 1− σ2 + Md,γ (|b1 − b2 |, V ) 2 Nd,γ (V ) σ 1−σ + Md,γ (|b1 − b2 |, V ) Nd,γ (V ) . (2.24)
Quasi-Classical Asymptotics for the Pauli Operator
117
If σ > 1, then |Bγ (b, V1 ) − Bγ (b, V2 )| 1 σ−1 σ−1 ≤ C Md,γ (b, V1 − V2 ) σ Md,γ (b, V1 ) σ + Md,γ (b, V2 ) σ 1 σ σ + C 0 Nd,γ (V1 − V2 ) σ+1 Nd,γ (V1 ) σ+1 + Nd,γ (V2 ) σ+1 (2.25) and |Bγ (b1 , V ) − Bγ (b2 , V )| ≤ CMd,γ (|b1 − b2 |, V ) 1 1 + C 0 Md,γ (|b1 − b2 |, V ) 2 Nd,γ (V ) 2 . (2.26) The proof of the above estimates is given in Appendix B. 3. Asymptotics in a Cube. Moderate Fields Here we establish the asymptotics of Mγ for a cube = [0, R)d as ~ → 0, µ~ ≤ C for a constant potential V and arbitrary magnetic field B. The coefficient Bγ is defined in (1.5). Lemma 3.1. Let = [0, R)d be a cube such that V = const inside. Suppose that the magnetic field B is continuous in and µ~ ≤ C. Then the asymptotics (2.8) holds for all γ ≥ 1. If d = 3 and the magnetic field B has a constant direction, then (2.8) holds for all γ > 1/2. As was explained in the Introduction, the idea of the proof is borrowed from the work [2]. In particular, we use the following estimates: Proposition 3.2. Let = [0, r)d , r > 0, B be homogeneous, i.e. B = const, and V = const. Then there exists a constant C independent of B, V , such that for any δ ∈ (0, 1/2), γ ≥ 0, Mγ (~, µB, V, ) ≤ ~−d Bγ (µ~|B|, V, ); −d
Mγ (~, µB, V, ) ≥