Commun. Math. Phys. 264, 1–40 (2006) Digital Object Identifier (DOI) 10.1007/s00220-006-1538-3
Communications in
Mathematical Physics
The Virtual Class of the Moduli Stack of Stable r-Spin Curves Takuro Mochizuki1,2 1 2
Department of Mathematics, Kyoto University, Kyoto, 606-8502, Japan. E-mail:
[email protected] Max-Planck Institute for Mathematics, Vivatsgasse 7, 53111 Bonn, Germany. E-mail:
[email protected]
Received: 18 April 2004 / Accepted: 27 October 2005 Published online: 15 March 2006 – © Springer-Verlag 2006
Abstract: We recall the outline of the Seely-Singer-Witten construction of the virtual class on the moduli of stable r-spin curves. We prove that the obtained classes satisfy the axioms of Jarvis-Kimura-Vaintrob.
1. Introduction 1.1. Stable r-spin curves. Let n be a non-negative integer, r be a natural number and 1/r,m m = (m1 , . . . , mn ) be an n-tuple of integers. Let Mg,n be the moduli stack of smooth n-pointed r-spin curves of genus g, which is a tuple (C, p, F) of a smooth curve C of genus g, n-points p = {p1 , . . . , pn } contained in the curve C, and a line bundle F on C with the isomorphism F ⊗ r ωC ⊗O(−m·p), where we put m·p = mi ·pi . T. Jarvis 1/r,m constructed the moduli stack Mg,n of the stable r-spin curves, which gives the smooth 1/r,m
compactification of the stack Mg,n ([3 and 4]). Saying naively, a ‘definition’ of stable r-spin curve is a tuple (C, p, F) of an n-pointed stable curve (C, p) and a torsion-free rank 1 sheaf F with the morphism F ⊗ r −→ ωC (−m · p) with some conditions. But the moduli stacks of such naive objects are not smooth. Hence he introduced the notion of the coherent net so that the moduli stacks become smooth. Thus the moduli stack 1/r,m Mg,n over the complex number field C is topologically an orbifold. The nodal point of stable r-spin curves (C, p, F) are divided into two types by the local property of the r-spin structure F at the point: If F is locally free at the nodal point Q, then Q is called of Ramond type. Otherwise the point Q is called of Neveu-Schwarz type. By the definition, F around the point Q is described as follows: Let π : C˜ −→ C be a normalization of the curve C at the point Q. Let Qj (j = 1, 2) denote the preimage of Q. We take a coordinate neighbourhood Bi = {zi ∈ C | |zi | < 1, zi (Qi ) = 0} around points Qi .
2
T. Mochizuki
• If Q is of Ramond type, then we have the locally free sheaf F˜ generated by the sections (dzi /zi )1/r . The sheaf F is obtained from F˜ by the gluing ϕ : F˜ Q1 −→ F˜ Q2 such that ϕ((dz1 /z1 )1/r )⊗ r = −dz2 /z2 . If Q is of Ramond type, the index at Q is defined to be (−1, −1). • If Q is of Neveu-Schwarz type, then there is a pair of integers (m1 , m2 ) with the conditions 0 ≤ mi ≤ r − 2 and m1 + m2 = r − 2 such that the sheaf F is π∗ OBi (zimi dzi )1/r around the nodal point Q. The pair (m1 , m2 ) is called the index of Q. Also the number mi is called the index of Qi . 1.2. The virtual class. In the papers [6 and 7], T. Jarvis, T. Kimura and A. Vaintrob 1/r 1/r,m introduced the substack M of Mg,n for any stable decorated graph of genus g with n tails marked by the tuple m, and they formulated the axioms of the virtual class 1/r 1/r,m of M (see Definition 4.1). The virtual class of Mg,n was originally introduced by Witten ([13]), when he generalized his famous conjecture on the intersection theory of the moduli stack of stable curves, although the existence of the nice moduli stack of r-spin stable curves was established by Jarvis later. Remark 1.1. Jarvis, Kimura and Vaintrob also considered the case where one of mi is −1. In this paper, we restrict our attention to the tuples of non-negative integers. 1.3. The purpose. Our purpose of the paper is to recall the outline of the original construction of Witten, which we call the SSW-construction, and to show that the obtained virtual classes satisfy the axioms of Jarvis-Kimura-Vaintrob. We recall the SSW-construction briefly. We denote the moduli stack of smooth (resp. stable) r-spin curves by M (resp. M). From the universal curve C −→ M, we obtain the Hilbert space bundles E 0 and E 1 whose fibers on the curve (C, p, F) are L2 (C, F) and L2 (C, F ⊗ T 0,1 ) respectively. We have the family of the closed operators ∂. Then 0 1 we extend such a family ∂ : E 0 −→ E 1 to the family ∂¯ : E −→ E on the stack M due to the argument of Seely-Singer [11]. (See Sect. 3.) i Then we take finite dimensional vector subbundles E i ⊂ E satisfying the following conditions: 0 1 • ∂(E 0 ) ⊂ E 1 and the index of ∂ : E 0 −→ E 1 is same as the index of ∂ : E −→ E . • E 1 contains the orthogonal complement of Im ∂ . Such E 0 −→ E 1 is called a finite reduction. (See Definition 4.2 for a more precise state1 ment.) Let ρE 1 denote the orthogonal projection E onto E 1 . Let π denote the natural Witten gave the section φ of π ∗ E 1 over E 0 , which is given projection of E 0 on M. r−1 by φ(s) = ∂s + ρE 1 s (see Subsubsect. 4.2.3), and he shows φ −1 (0) = M (see Lemma Chern classctop (π ∗ E1 , φ), which is contained in 04.5). Hence we obtain the topd(r, ∗ H E . It can be shown that (−1) m)π∗ ctop π ∗ E 1 , φ is independent of a choice M of a finite reduction, which should be the virtual class. We will see the classes satisfy the axioms in Subsect. 4.4. One of the key points in the proof is to show ‘vanishing’. Our simple idea for vanishing is explained in the Subsubsect. 4.3.1. The author hopes that it is sufficiently clear for the readers. Remark 1.2. The algebro-geometric construction of the virtual class was given by A. Polishchuk and A. Vaintrob in [10 and 9].
The Virtual Class of the Moduli Stack of Stable r-Spin Curves
3
2. Stable r-Spin Curves We recall the definitions and the results on the stable r-spin curves due to Jarvis ([3 and 4]) to fix the notation. We refer the paper [6] as a convenient reference. We denote a tuple of points p1 , . . . , pn (resp. integers m1 , . . . , mn ) by p (resp. m). We denote the formal sum mi · pi by m · p. Definition 2.1. Let (X, p) be a nodal, n-pointed algebraic curve, and let K be rank one torsion-free sheaf on X. A d th root of K of type m = (m1 , . . . , mn ) is a pair (E, b) of rank one torsion-free sheaf E and an OX -module homomorphism b : E ⊗ d −→ K(−m · p) with the following properties: • d · deg E = deg K − mi . • b is an isomorphism on the locus of X where E is locally free. • for every point p ∈ X where E is not free, the length of the cokernel of b at p is d − 1. Jarvis used the following notion to obtain the smooth moduli space. Definition 2.2. Let K be a rank 1, torsion-free sheaf on a nodal n-pointed curve (X, p). A coherent net of r th roots of K of type m = (m1 , . . . , mn ) consists of the following data: • A rank 1 torsion-free sheaf Ed on X for every divisor d of r. ⊗ d/d
−→ Ed for every pair of divisors d
• An OX -module homomorphism cd,d : Ed
and d of r such that d divides d. These data are subject to the following restrictions: 1. E1 = K and c1,1 = id. 2. For each divisor d of r and each divisor d of d, let m = (m
1 , . . . , m
n ) be a tuple such that m
i is the unique non-negative integer less than d/d , and congruent to mi mod d . Then the homomorphism cd,d makes (Ed , cd,d ) into a d/d root of Ed of type m
. ⊗ d /d
= cd,d
holds. 3. The homomorphisms {cd,d } are compatible, i.e., cd ,d
◦ cd,d
Then Jarvis defined stable r-spin curves. Definition 2.3. An n-pointed r-spin curve of type m = (m1 , . . . , mn ) is defined to be an n-pointed nodal curve (X, p) with a coherent net of r th roots of ωX of type m, where ωX is the dualizing sheaf of X. An r-spin curve is called smooth if X is smooth, and it is called stable if X is stable. The nodal point of a stable r-spin curve (C, p, F) is divided into two types by the local property of the sheaf F at the point: If F is locally free at the nodal point Q, then Q is called of Ramond type. Otherwise Q is called of Neveu-Schwarz type. To obtain the category of the r-spin curves, the ‘morphisms’are considered as follows. ∼ Definition2.4. An isomorphism of r-spin curves X, p, {Ed , cd,d } −→ X , p ,
{Ed , cd,d
} of the same type m is defined to be a tuple (τ, β) of an isomorphism of pointed curves τ : (X, p) −→ (X , p ) and a family of isomorphisms βd : τ ∗ Ed −→ Ed with β1 being the canonical isomorphism τ ∗ ωX (−m · p) −→ ωX (−m · p), such that
the βd are compatible with all the maps cd,d and τ ∗ cd,d
.
4
T. Mochizuki
The foundational and important theorem of Jarvis is the following. Proposition 2.1 (Jarvis). The moduli functor of the stable n-pointed r-spin curves of genus g and type m is representable by a smooth proper Deligne-Mumford stack. Following Jarvis, we denote the stack of the stable n-pointed r-spin curves of genus 1/r,m 1/r,m 1/r g and type m by Mg,n . The disjoint union m,0≤mi
3. Seely-Singer Extension 3.1. Preliminary for the operators on the interval. 3.1.1. Some generality. For any real number a ∈ R, let L2 (a) denote the space of L2 functions on the open interval (0, 1) with respect to the measure R a+1 · dR. Let K(R, ρ) be a continuous function on the product of the intervals (0, 1)×(0, 1). We put as follows: 1 K(R, ρ) · (R/ρ)1+a · dρ, M1 (K) := sup R
0
1
K(R, ρ) · dR.
M2 (K) := sup ρ
0
The Virtual Class of the Moduli Stack of Stable r-Spin Curves
5
given as follows: Let us consider the integral operator K 1 K(R, ρ) · dρ. Kg(R) := 0
We recall the following lemma. naturally induces the bounded Lemma 3.1. Assume Mi (K) < ∞ for i = 1, 2. Then K 1/2 2 2 . operator L (a) −→ L (a). The operator norm is smaller than M1 (K) · M2 (K) Proof. Let g be any element of L2 (a). The following inequality holds: 1 2 K(R, ρ) · ρ −1−a · dρ × K(R, ρ) · |g|2 · ρ 1+a · dρ. Kg(R) ≤ 0
Hence we obtain the following inequality: K(R, ρ) · R/ρ 1+a · dρ Kg(R) · R 1+a · dR ≤
× K(R, ρ) · |g|2 · ρ 1+a · dρ · dR
≤ M1 (K) × K(R, ρ) · dR · |g|2 · ρ 1+a · dρ ≤ M1 (K) · M2 (K) × |g|2 · ρ 1+a · dρ. (1) Thus we are done.
3.1.2. Some functions on the product of intervals (0, 1) × (0, 1). Let χP denote the characteristic function of the subset of (R, ρ) ∈ (0, 1)2 determined bythe property P . For example, χ0
Proof. We have the following formula: 1 −k − a −1 · R k+1+a − R (k + a = 0) + 1+a K · R/ρ · dρ = δ,k 0 −R · log R (k + a = 0). The claim immediately follows from (2).
(2)
6
T. Mochizuki
Lemma 3.3. Let k be a positive number such that k ≥ −1. Let δ and δ be any positive numbers such that 0 < δ < δ < 1. Then there exists a positive constant C, which is independent of k, δ and δ such that the following holds: M2 Kδ,k ≤ C · (1 + |k|)−1 , M2 Kδ,k − Kδ ,k ≤ C · δ − δ · (1 + |k|)−1 . Proof. We have only to see the second inequality. We put ρ := min{ρ, δ}. Then we have the following inequality: 0
1
K + − K + · dR = δ,k δ ,k =
ρ
R/ρ
k
· dR
(k + 1)−1 · ρ −k · ρ k+1 − δ k+1 (k + 1 > 0) 0
(3)
ρ · log ρ − log δ .
The second inequality immediately follows from (3).
Let k be any integer, and δ be any positive number. Let us consider the following functions: k − := R/ρ · χδ<ρ
−1 − M1 Kδ,k − Kδ− ,k ≤ C · |δ − δ | · 1 + |k| .
Proof. We have only to see the second inequality. We put R := min{R, δ}. Then we have the following equality:
1
K − − K − · R/ρ 1+a · dρ δ,k δ ,k 0 −k − a −1 · R −k−a+1 · R −k−a − δ −k−a (k + a < 0) = R · log R − log δ
(k + a = 0).
The second inequality immediately follows from the formula (4).
(4)
Similarly, the following lemma can be shown elementarily. Lemma 3.5. Let k be any integer such that k ≤ 0, and let δ be any positive number. Then there exists a positive constant C, which is independent of k and δ, such that the following holds: − M2 Kδ,k ≤ C · (1 + |k|)−1 . − − Kδ− .k ≤ 2C · (1 + |k|)−1 . In particular, we have M2 Kδ,k
The Virtual Class of the Moduli Stack of Stable r-Spin Curves
7
3.1.3. Some operators on the interval (0, 1). Let a be a real number such that −2 ≤ a < 2. Let L2 (a) denote the space of L2 -functions on the interval (0, 1) with respect to + and K − induced by the measure R 1+a · dR. Let us consider the integral operators K δ,k δ,k + − Kδ,k and Kδ,k . Lemma 3.6. + gives the bounded operator on L2 (a). There exists a • In the case k ≥ −a/2, K δ,k positive constant C, which is independent of a choice of any positive number δ and any + is smaller than C · (1 + |k|)−1 . integer k as above, such that the operator norm of K δ,k − gives the bounded operator on L2 (a). There exists a • In the case k < −a/2, K δ,k positive constant C, which is independent of a choice of any positive number δ and any − is smaller than C · (1 + |k|)−1 . integer k as above, such that the operator norm of K δ,k Proof. In the case k ≥ −a/2, we have k ≥ −a/2 > −1, which implies k ≥ 0. We also have k + 1 + a ≥ −a/2 + 1 + a = 1 + a/2 ≥ 0. Thus the first claim immediately follows from Lemma 3.1, Lemma 3.2 and Lemma 3.3. In the case k < −a/2, we have k < −a/2 ≤ 1, which implies k ≤ 0. If we have k ≤ −1, then k + a < −1 + 1 = 0. If we have k = 0, then we have 0 < −a/2, which implies a < 0. Hence we have k + a < 0 + 0 = 0. Then the second claim immediately follows from Lemma 3.1, Lemma 3.4 and Lemma 3.5. Similarly, the following lemma follows from lemmas 3.1, 3.2, 3.3, 3.4 and 3.5. Lemma 3.7. • In the case k ≥ −a/2, there exists a positive constant C, which is independent of + − K + is smaller than a choice of k, δ and δ such that the operator norm of K δ,k δ ,k C · |δ − δ |1/2 · (1 + |k|)−1 . • In the case k < −a/2, there exists a positive constant C, which is independent of − − K − is smaller than a choice of k, δ and δ such that the operator norm of K δ,k δ ,k C · |δ − δ |1/2 · (1 + |k|)−1 . 3.2. Preliminary for the ∂-operator on annuli. 3.2.1. The on annuli. For any non-negative number δ, let D(δ) denote the parametrices annulus z ∈ C δ < |z| < 1 . We denote the bundle of (0, 1)-forms by 0,1 . Let us consider the line bundle F with a generator e on the disc D(0). Let us consider the ∞ of hermitian metric |e|2a = R a of F. Let us take aC ∞ -measure of D(0) and C -metric 0,1 2 2 2 on D(0). Then we have the L -spaces L D(δ), F, a and L D(δ), F ⊗ 0,1 , a for any 0 ≤ δ < 1. We denote the norm by || · ||a . The number a is called the weight. In this subsection, we fix a C ∞ -measure of the disc D(0) for simplicity. Remark 3.1. When we consider the smooth metric of F and singular measure of the form φ · R 1+a dR · dα, we obtain the same topological vector space L2 (D(δ), F, a). Here φ denotes a positive C ∞ -function on D(δ). It will be convenient to consider 1,1 -valued metrics of a vector bundle V , which is a hermitian pairing h : V ⊗ V −→ 1,1 such that h(v, v) is a positive (1, 1)-form. Then we can say that we consider the 1,1 -valued metric of F which is of the form R a · h0 ,
8
T. Mochizuki
where h0 denotes a C ∞ - 1,1 -valued metric on D(0). When we fix the coordinate, the weight, and the frame e, then a 1,1 -valued metric corresponds to a positive function. For any L2 -section f of F ⊗ 0,1 , we have the Fourier development f = fk · √ −1k·α 2 2 1+a e . Here fk are elements of L (a) = L (0, 1), R · dR . We put as follows: Qδ (f ) :=
+ (fk ) · e K δ,k
√
−1k·α
+
k≥−a/2
− (fk ) · e K δ,k
√
−1k·α
.
(5)
k<−a/2
Lemma 3.8.
• For any 0 ≤ δ < 1, Qδ give the compact operators of L2 D(δ), F ⊗ 0,1 , a to L2 D(δ), F, a . • The operator norms of Qδ are dominated by a constant, which is independent of a choice of δ. • ∂Qδ (f ) = f . Proof. The first and the second claims immediately follow from Lemma 3.6. The third claim can be checked by a formal calculation. We have the natural inclusion D(δ) ⊂ D(0). It induces the isometric embeddings: ιδ : L2 D(δ), F, a ⊕ L2 D(δ), F ⊗ 0,1 , a −→ L2 D(0), F, a ⊕ L2 D(0), F ⊗ 0,1 , a . Then the operators Qδ induce the family of the operators Qδ 0 ≤ δ < 1 of L2 D(0), F, a to L2 D(0), F ⊗ 0,1 , a . Note we have Q0 = Q0 . Lemma 3.9. The family Qδ 0 ≤ δ < 1 is continuous with respect to the variable δ and the operator norms. Proof. It immediately follows from Lemma 3.7.
Let us take a continuous family of diffeomorphisms ηδ : D(0) −→ D(δ), for ∗ example, ηδ (R, α) = (1 − δ) · R + δ, α . We take the bundle map η∞ F −→ F ∗ ∗ ∞ by η e −→ e. Then we obtain the linear map η : C D(δ), F −→ C D(0), F . Then there exists the unique positive function gδ such that gδ · η∗ preserves the hermi tian products, and hence we obtain the unitary isomorphism Pδ : L2 D(δ), F, a −→ L2 D(0), F, a . Similarly, we obtain the unitary isomorphism Pδ : L2 D(δ), F ⊗ 0,1 , a −→ L2 D(0), F ⊗ 0,1 , a . Then we obtain the family of the isometric em beddings Uδ := ιδ ◦ Pδ−1 : L2 D(0), F ⊗ 0,1 , a −→ L2 D(0), F ⊗ 0,1 , a . It is easy to see that Uδ strongly converges to Uδ0 when δ → δ0 . δ := U ∗ ◦ Qδ ◦ Uδ . Then we obtain We obtain the family of the bounded operators Q δ the following lemma by the general argument due to Seely-Singer [11]. δ 0 ≤ δ < 1 is continuous with respect to the parameter Lemma 3.10. The family Q δ and the operator norms.
The Virtual Class of the Moduli Stack of Stable r-Spin Curves
9
Proof. We see the continuity at δ = 0. The others can be seen similarly. We know the continuity of the family Qδ with respect to the parameter δ and the operator norms. ˜ 0 . Since it is compact we can approximate Q0 by an operator Note the equality Q0 = Q F of a finite rank. Let be a positive number. By the continuity of the family Qδ at δ = 0, there is a small number δ0 and the operator F of a finite rank such that the norm of the operators F − Qδ is less than for any δ < δ0 . We have the equality δ − Q 0 = U ∗ · F · Uδ − F + U ∗ · Qδ − F · Uδ + F − Q0 . Due to the strong Q δ δ continuity of the family Uδ δ , there is a positive number δ1 such that the norms of the operators Uδ∗ · F · Uδ − F are less than for any δ < δ1 . On the other hand, the multiplication of the operators U and U ∗ does not increase the norm, we obtain the desired continuity. 3.2.2. The value at the origin in the case δ = 0. Let us consider the case δ = 0.We have a straightforward generalization of Lemma 2 in [11]. We put 2 · Z := 2 · n n ∈ Z}. For any real number a ∈ R − 2 · Z, we put as follows: a˜ = min k k ∈ 2Z, k − a > 0 ,
a¯ = a˜ − a.
Lemma 3.11. Let a be an element of R − 2 · Z. Suppose that u = g · e and ∂u/∂ z¯ are ˜ g)(0). We also L2 with respect to the norm || · ||a. Then we have the value A = (za/2 − a/2 ˜ −a/2 have the estimate u(R) − z A = O(R ) when R → 0. Proof. We have only to see the case −2 < a < 0. Assume that f is contained in L2 (a). It is easy to show the existence of a positive constant C, which is independent of R, f and k, such that the following holds: k R −k R · 0 ρ f ≤ C · R −a/2 · ||f ||, (k ≤ 0), k R −k R · 1 ρ f ≤ C · R −a/2 · ||f ||, (k > 0).
(6)
Assume that u and f = ∂u/∂ z¯ are L2 , then the Fourier coefficient uk of u is of the folR R lowing form: uk (R) = R k · 1 ρ −k fk + ck in the case k > 0, uk (R) = R k · 0 ρ −k fk R in the case k < 0, and u0 (R) = 0 f0 + c0 . Here fk denote the Fourier coefficients of f . Then the claim follows from the estimate (6). Let a be a real number such that 0 < a < 2. Let f be an L2 -section of F ⊗ 0,1 2 with the weight a. Then we have the L -section Q0 (f ) of F. Since a˜ = 2, we have the value z · Q0 (f ) (0). Lemma 3.12. We have the vanishing z · Q0 (f ) (0) = 0. − (f−1 ) at R = 0 is 0, which can be Proof. We have only to see that the value of K δ,−1 checked directly.
10
T. Mochizuki
3.2.3. The r-spin bundle. Let F(l) denote the r-spin bundle on D(δ) with the generator 1/r . We take e∨ (l) := e(l)−1 ⊗ dz as a generator of the line bune(l) := zl · dz/z dual
dle L(l) := F(l)−1 ⊗ ω. We denote the ∂-operators of F(l) and L(l) by ∂ and ∂ to distinguish them. Then we can apply the generality in Subsubsect. 3.2.1 to ∂ and dual ∂ . Since we have the naturally defined pairing F(l) ⊗ L(l) ⊗ 0,1 −→ 1,1 and F(l) ⊗ 0,1 ⊗ L(l) −→ 1,1 , the 1,1 -hermitian metrics of F(l) and F(l) ⊗ 0,1 induce the 1,1 -hermitian metrics of L(l)⊗ 0,1 and L(l) respectively. When the weight for F(l) is a, we take −a as the weight for F(l)−1 ⊗ ω. Then we obtain the family of dual parametrices for both of ∂ and ∂ in the case −2 < a < 2. (Formal adjointness). We have the natural pairing F(l)⊗ 0,1 ⊗F(l)−1 ⊗ω −→ 1,1 . It induces the following perfect pairings: L2 D(δ), F(l) ⊗ 0,1 , a ⊗ L2 D(δ), F(l)−1 ⊗ ω, −a −→ C, L2 D(δ), F(l), a ⊗ L2 D(δ), F(l)−1 ⊗ ω ⊗ 0,1 , −a −→ C. We regard that they are mutually dual spaces. If we fix an appropriate metric of F(l) with weight a, then we obtain the anti-Hermitian isomorphisms: L2 D(δ), F(l) ⊗ 0,1 , a L2 D(δ), F(l)−1 ⊗ ω, −a , L2 D(δ), F(l), a L2 D(δ), F(l)−1 ⊗ ω ⊗ 0,1 , −a . Via the isomorphisms, ∂ and −∂
dual
are mutually formal adjoint.
Remark 3.2. We have the canonical isomorphism L2 D(δ), F(0)−1 ⊗ ω(O), −1 L2 D(δ), F(0)−1 ⊗ ω, 1 . Thus we can use L2 D(δ), F(0)−1 ⊗ ω(O), −1 for our argument. It may be convenient for the reader to compare it with the argument in [11]. Anyway, it does not matter which we choose. (Contribution of the origin to the integral). Let us consider the case δ = 0. Let a be any real number satisfying −2 < a < 2 and a = 0. We take a and −a as the weights of F(l) and F(l)−1 ⊗ ω respectively. Let u = f · e(l) and v = g · e∨ (l) be L2 -sections of F(l) and F(l)−1 ⊗ ω respectively, such that ∂(u) and ∂(v) are also L2 . In that case, we have ˜ ˜ the values za/2 · f (0) and z(−a)/2 · g (0) = z−a/2+1 · g (0) due to Lemma 3.11. 2 Let us consider the naturally defined pairing u, v, which is an L -section of ω such that ∂u, v is L1 . We have the value z · u, v (0), which is given by the product ˜ a/2 · g (0). The integral ∂u, v is naturally defined. It is easy to z ˜ · f (0) × z−a/2+1 see that the contribution of the origin (the singular point) to the Stokes formula is given by the value z · u, v (0). As a result, we obtain the following lemma. Lemma 3.13. origin to the Stokes Let u and v be as above. The contribution of the ˜ f (0) × z−a/2+1 ˜ formula for ∂u, v is given by the product za/2 · g (0).
The Virtual Class of the Moduli Stack of Stable r-Spin Curves
11
3.3. Preliminary for the ∂-operator around a nodal point. 3.3.1. Preliminary. Let us consider the nodal curve X0 = {(z, w) ∈ C2 | zw = 0, |z| < 1, |w| < 1}. We put X0z = X0 ∩ {w = 0}, X0w = X ∩ {z = 0}. Let O√denote the origin √0 (0, 0). We often use the real coordinate z = R · e −1α and w = S · e −1β . Let F0 be an r-spin structure on the nodal curve. Recall that we have the following two cases: 1. Neveu-Schwarz case. The sheaf F0 is a direct sum OX0z ·e(l)z ⊕OX0w ·e(r −l)w , where r r e(l)z and e(r − l)w satisfy e(l)z = zl · dz/z and e(r − l)w = wr−l · dw/w. Recall that the index at the origin O is defined to be (l − 1, r − l − 1). 2. Ramond case. The sheaf F0 is locally free, and it is the quotient of OX0z · e(0)z ⊕ OX0w · e(0)w divided by the relation (−1)1/r e(0)z = e(0)w at the origin O. Here r r e(0)z and e(0)w satisfy e(0)z = dz/z and e(0)w = dw/w. Recall that the index at the origin O is defined to be (−1, −1). Let (m1 , m2 ) denote the index of the r-spin structure F0 at the origin O. We put e := gcd(r, m1 + 1), and we put as follows: X := (z, w, t) zw = t e , |z| < 1, |w| < 1 ⊂ B 3 : = (z, w, t) |z| < 1, |w| < 1, |t| < 1 . We have the naturally defined morphism X −→ B := t ∈ C |t| < 1 . The fiber over t is given by Xt . Hence we obtain the family of the curves whose special fiber at t = 0 is the nodal curve X0 above. Recall that we have a sheaf F on X, which is a universal deformation of F0 . We denote F|Xt by Ft . We divide X into X z ∪ X w : Xz := (z, w, t) ∈ X |z| > |w| , Xw := (z, w, t) ∈ X |z| < |w| . Here · denotes the closure in X. We put Xtz := Xt ∩ X z and Xtw = Xt ∩ X w . For any z e/2 w t such that |t| < 1, we put δ(t) := |t| . Then Xt and Xt are naturally isomorphic to D δ(t) . For simplicity, we often use δ instead of δ(t). We can also take a continuous family of diffeomorphisms ηtz : X0z −→ Xtz and ηtw : X0w −→ Xtw . 3.3.2. The line bundles on curves. The sheaf F naturally induces the line bundles Ftz and Ftw on Xtz and Xtw respectively. It also induces the line bundles over Xtz and Xtw , which are also denoted by Ftz and Ftw for simplicity. We can take the trivialization 1/r of Ftz . Similarly we can take the trivialization e(m2 + e(m1 + 1)z := zm1 +1 · dz/z 1/r of Ftw . Clearly, the line bundles Ftz and Ftw with the 1)w := w m2 +1 · dw/w generators naturally induce the r-spin structure Ft on Xt in the case t = 0. −1 ⊗ ω on Xtz , and the The dual bundles are considered as follows: Lzt denotes Ftz −1 ∨ z z ⊗ dz. Note that the family of line bundles e (m1 + 1) is given by e(m1 + 1) frame z L |t| < 1 naturally gives the line bundle on X z , and e∨ (m1 + 1)z gives the frame over Xz . The dual bundle Lw in the w-side is considered similarly.
12
T. Mochizuki
3.3.3. The L2 -spaces. Here we introduce the following number for any integer l such that 0 ≤ l < r: (l) :=
2l − r . r
(7)
The following lemma can be checked easily. Lemma 3.14. • We have −1 < (l) < 1 in the case l = 0, and (0) = −1. • We have (l) + (r − l) = 0. The weights of Ftz and Lzt are defined to be (m1 + 1) and −(m1 + 1) respectively. Similarly, the weights of Ftw and Lw t are defined to be (m2 + 1) and −(m2 + 1) respectively. Let us take a positive functions φ z and φ w on B 3 . Then the restrictions of φ z to Xtz give the 1,1 -valued metrics on Xtz . (See Remark 3.1.) Similarly the positive function φ w induces the 1,1 -valued metrics on Xtw . 2 X z , F z , (m + 1) t ∈ B , Then we obtain the family of topological spaces L 1 t 2 w w L Xt , F , (m2 + 1) t ∈ B , etc. as in Subsubsect. 3.2.3. Note the following easy lemma. 2 w w z z 2 Lemma 3.15. 1 +1) ⊕L Xt ,F , For any t = 0, the topological space L Xt , F , (m (m2 + 1) is naturally isomorphic to those obtained from the C ∞ -metrics and the C ∞ measure on Xt , which we denote by L2 Xt , F . We also have the natural isomorphism for others, similarly. Remark 3.3. Due to Lemma 3.15, we can describe L2 Xtz , Ftz , (m1 +1) ⊕L2 Xtw , Ftw , (m2 +1) as L2 Xt , Ft . Therefore, we often denote L2 X0z , F0z , (m1 +1) ⊕L2 X0w , F0w , (m2 + 1) by L2 X0 , F0 for simplicity of the notation. We use the notation L2 Xt , Ft ⊗ 0,1 , L2 Xt , Lt and L2 Xt , Lt ⊗ 0,1 in a similar meaning. Once we fix the metrics and measures, the diffeomorphisms ηtz : X0z −→ Xtz induce the unitary isomorphism Ptz : L2 (Xtz , F z , (m1 + 1)) −→ L2 (X0z , F z , (m1 + 1)), as in Subsubsect. 3.2.1. Similarly we have Ptw : L2 (Xtw , F z , (m2 + 1)) −→ L2 (X0w , F z , (m2 + 1)). Thus we have the isomorphisms Ptw ⊕ Ptz of L2 (Xt , Ft ) and L2 (X0 , F0 ). Clearly, we have the isomorphisms for others. dual We obtain the family of parametrices Qzt and Qw for t for the operators ∂ and ∂ the families Ftz |t| < 1 and Ftw |t| < 1 respectively. However, we remark that Qzt ⊕ Qw parametrix for Ft on Xt , because they are not glued at the t does not give the cutting circle Xtz ∩ Xtw = (z, w, t) |z| = |w| . We modify them in the following subsubsections. 3.3.4. The matching condition at the cutting circle (Neveu-Schwarz case). Let us consider the Neveu-Schwarz case with the index (l −1, r −l −1). On each curve Xt (t = 0), we have the following relation: w r−l dw zl dz = −t el · w −r · . z w
The Virtual Class of the Moduli Stack of Stable r-Spin Curves
13
Hence we have the following relation on Xt : e(l)z = (−1)1/r · t el/r · w −1 · e(r − l)w . Therefore we have the following relation on the circle |z| = |w|, for √ √ and for z = δ · e −1α and w = δ · e −1β : e(l)z = A · e
√ −1α
te
=
δ2
·e
√
· δ (l) · e(r − l)w .
(8) −1θ
(9)
Here A denotes the constant whose norms are 1, depending on θ , l and r. Let u(z) · e(l)z and v(w) · e(r − l)w be L21 -sections on the curves Xtz and Xtw respectively, where the weights of Ftz and Ftw are (l) and (r − l) respectively. Note that u(z) · e(r)z ⊕ v(w) · e(r − l)w gives the L2,1 -section on Xt if and only if we have the relation u(z) · e(r)z = v(w) · e(l − r)w on the cutting circle |z| = |w|. The latter gives the matching condition: u(δ, α) · A · e
√
−1α
· δ (l) = v(δ, β).
Here A is the constant in (9). For the Fourier expansion u = √ −1k·β v = vk (S) · e , Eq. (10) can be reworded as follows:
uk (R) ·
√ e −1k·α
vk (δ) = Ak · δ (l) · u−k−1 (δ).
√ Here we put Ak := e− −1θ(l/r−1−k) (−1)1/r , which is a constant Similarly, we also have the relation uk (δ) = δ (r−l) · Ak · v−k−1 (δ).
(10) and (11)
whose norm is 1.
Let us consider the matching condition of the following two functions Xtz and Xtw respectively: √ ck · R k · e −1kα · e(l)z , Qt (f · e(l)z d z¯ ) + k≤−(l)/2−1
¯ + Qt (g · e(r − l)w d w)
dk · S k · e
√ −1kβ
· e(r − l)w .
k≤−(r−l)/2−1
Then we obtain the following equalities: δ ck = Ak · δ −(l)−2k−1 σ k+1 · g−k−1 · dσ,
(k ≤ −(l)/2 − 1),
1
dk = Ak · δ −(r−l)−2k−1
δ 1
σ k+1 · f−k−1 · dσ, (k ≤ −(r − l)/2 − 1).
Thus we put as follows for any t = 0: w −(l)−2k−1 Ak · δ Rt g · e(l) · d w¯ : = k≤−(l)/2−1 √ k −1kα
Rt (f · e(r − l) · d z¯ ) : = z
×R · e
×S · e
We put R0 = 0.
σ
k+1
1
· e(l)z ,
Ak δ
k≤−(r−l)/2−1 √ k −1kβ
δ
−(r−l)−2k−1
· e(r − l)w .
δ
σ 1
· g−k−1 · dσ
k+1
· f−k−1 · dσ (12)
14
T. Mochizuki
Lemma 3.16. For any t, Rt gives a compact operator. Proof. There exists a positive constant C such that the following inequality holds, for any positive number δ < 1 and any integer k < −1:
1
1/2 r 2k r 1+(l) dr
≤ C · δ k+1+(l)/2 |k|−1/2 .
δ
We also have a positive constant C such that the following inequality holds for any integer k < −1, a positive number δ < 1 and any function g: δ δ −(l)−2k−1 σ k+1 · g−k−1 · dσ ≤ C · δ k+3/2−(r−l)/r+(r−2l)/r−2k−1 1
×
1/2
|g−k−1 | · σ 2
1+(r−l)
· dσ
· |k|−1/2
= C · δ −k+1/2−l/r
1/2 2 1+(r−l) × |g−k−1 | · σ · dσ ·|k|−1/2 . (13) Since we have the equality k + 1 + (l)/2 − k + 1/2 − l/r = 1, the norm of the L2 norm of the function ck · r k (k < −1) on the interval (δ, 1) with respect to the measure r 1+(l) dr is dominated by the following:
1/2 |g−k−1 |2 · σ 1+(r−l) · dσ . C · C · δ|k|−1/2 In the case k = −1, the L2 -norm of the function c−1 · r −1 is dominated by the following, up to some constant: 1
1/2 2−2l/r −2+2l/r 2 1+(r−l) · δ σ dσ · |g0 | · σ d ·σ . δ
1/2 . Thus the first sum (12) is absolutely It is dominated by δ 2/r · |g0 |2 σ 1+(r−l) dσ convergent, which implies the compactness of the operators Rt . 3.3.5. Norm-continuous family of parametrices (in the Neveu-Schwarz case). Recall Lemma 3.15. For a pair (f, g) ∈ L2 Xtz , F z ⊗ 0,1 , (l) ⊕L2 Xtw , F w ⊗ 0,1 , (r−l) , we put as follows: t (f, g) : = Qt (f ) + Rt (g), Qt (g) + Rt (f ) Q ∈ L2 Xtz , F z , (l) ⊕ L2 Xtw , F w , (r − l) . t (f, g) = (f, g) and the matching condition at the cutting circle. Then we It satisfies ∂ Q t : L2 Xt , Ft ⊗ 0,1 −→ L2 Xt , Ft for ∂. obtain the parametrices Q Recall that we have the isomorphism Pt of L2 (Xt , F) and L2 (X0z , F0 ) etc., once we take appropriate metrics. The family of the operators Qt induces the family of the t . operators of L2 Xt , Ft to L2 X0 , F0 , which is also denoted by Q
The Virtual Class of the Moduli Stack of Stable r-Spin Curves
15
t |t| < 1 is norm continuous, and Q 0 is Lemma 3.17. The family of the operators Q the same as the operator constructed in Subsubsect. 3.2.1. Proof. We have only to check the norm continuity of the family Rt . It immediately follows from the absolute convergence of the sums in (12). See the proof of Lemma 3.16. Let us consider the parametrices for the operators ∂ relation on the circle |z| = |w| in Xt (t = 0): e∨ (l)z = −A · e
√
−1α
dual
of Lt . We have the following
· δ −(l) · e∨ (r − l)w .
Then by the same construction explained as above, we obtain the norm continuous family dual of the parametrices for ∂ : dual : L2 Xt , Lt ⊗ 0,1 −→ L2 Xt , Lt . Q t dual is same as the operator constructed in Subsubsect. 3.2.1. And Q 0 3.3.6. The Ramond case. Let us consider case. In this case, it con the Ramond will be 1−1/r 1−1/r and w−1 · e∨ (0)w = dw/w venient to use the frame z−1 · e∨ (0)z = dz/z as in [11]. Note that they are L2 -sections in the case t = 0, for the weight of Lz0 and Lw 0 are 1. Clearly they give the L2 -sections in the case t = 0, too. We have the following relation: e(0)z = (−1)1/r e(0)w ,
z−1 · e∨ (0)z = (−1)1−1/r w −1 · e∨ (0)w .
(14)
By a similar construction to that in the Neveu-Schwarz case, we obtain the continuous dual family of parametrices for the operators ∂ and ∂ : t : L2 Xt , Ft ⊗ 0,1 −→ L2 Xt , Ft , Q dual Q : L2 Xt , Lt ⊗ 0,1 −→ L2 Xt , Lt . t 0 and Q dual are same as the operators constructed in 3.2.1. See [11] for more And Q 0 detail in the Ramond case. 3.3.7. The matching condition at the origin and the vanishing. Let us consider the operdual ators ∂ and ∂ of F0 and L0 . To make them mutually adjoint, we impose the matching conditions. Recall Lemmas 3.11, 3.12 and the Subsubsect. 3.2.3. Condition 3.1. The domain of ∂ is the subspace of elements (u, v) of L2 X0z , F0z , (m1 + 1) ⊕ L2 X0w , F0w , (m2 + 1) with L2 -derivatives ∂u and ∂v such that the following holds: • In the Neveu-Schwarz case such that (m1 + 1) > 0, we impose the condition (z · u)(0) = 0. • In the Neveu-Schwarz case such that (m2 + 1) > 0, we impose the condition (w · v)(0) = 0. z w . • In the Ramond case, we impose the condition u(0) = v(0) via the gluing F0|O F0|O
16
T. Mochizuki
dual The domain of ∂ is the subspace of elements (u, v) of L2 X0z , Lz0 , −(m1 + 1) ⊕ dual dual 2 L2 X0w , Lw u and ∂ v such that the following 0 , −(m2 +1) with L -derivatives ∂ holds: • In the Neveu-Schwarz case such that −(m1 + 1) > 0, we impose the condition (z · u)(0) = 0. • In the Neveu-Schwarz case such that −(m2 + 1) > 0, we impose the condition (w · v)(0) = 0. • In the Ramond case, note that we have the isomorphism Lz0 (O)|O Lw 0 (O)|O . We also have the values (z · u)(O) ∈ Lz0 (O)|O and (w · v)(O) ∈ Lw (O) |O . Hence we 0 impose the condition z · u(0) = w · v(0) via the isomorphism. Lemma 3.18. Assume that (u1 , v1 ) ∈ L2 X0 , F0 and (u2 , v2 ) ∈ L2 X0 , L0 are condual
tained in the domains of ∂ and ∂ , and impose the matching condition as above. Then the contribution of the nodal point (0, 0) to the Stokes formula for the integral ¯ 1 · u2 ) + ∂(v ¯ 1 · v2 ) vanishes. ∂(u Proof. The claim in the Neveu-Schwarz caseimmediately follows from Lemma 3.13. ¯ 1 · u2 ) and ∂(v ¯ 1 · v2 ) vanish. In fact, both of the integrals ∂(u −1 ⊗ Let us consider the Ramond case. We consider the isomorphisms Lz0 (O) F0z w −1 z w ⊗ ωXw (O). We have the isomorphism f : F˜ 0 | O −→ ωXz (O) and L0 (O) F0 z −1 −1 w F˜ 0 | O , and thus g : F0 ⊗ ω(0)z|O −→ F0w ⊗ ω(0)w |O . The isomorphism f ⊗ g : C −→ C is −1, where we use the natural identifications ωXz (O)|O CωXw (O)|O C by the residues. Let (u1 , v1 ) and (u2 , v2 ) be elements of the L2,1 -sections of F0 and L0 . Then the matching conditions and the consideration above gives us the equality Res(u1 · u2 ) + Res(v1 · v2 ) = 0. Thus the contributions are canceled. 3.3.8. Change of norms. In the discussion above, we divide the curve Xt into Xtz and Xtw by cutting at the circle |z| = |w| for any t. We also take a tuple of positive functions φ = (φ z , φ w ). (See the Subsubsect. 3.3.3.) Let Nt (φ) denote the norm considered in this case. If we take a continuous family of the cutting circles ct and the other tuple of positive functions φ , we obtain the norm Nt (φ ). Lemma 3.19. Assume that ct is contained in the region C −1 |z| < |w| < C|z| for some constant C > 1. Then the families of the norms Nt (φ) and Nt (φ ) are mutually bounded independently of t, i.e., there is a constant C > 1, such that C −1 Nt (φ)(u) ≤ Nt (φ )(u) ≤ C Nt (φ)(u). Proof. Let f · e(l)z = g · e(r − l)w be a section of Ft . From the relation of the sections e(l)z and e(r − l)w , we obtain the relation |f | · |t|el/r · |w|−1 = |g|. Since z · w = t e , we obtain |f |2 · |z|2l/r = |g|2 · |w|2(r−l)/r . On the other hand, we have the relations dz = −w−1 · z · dw and d z¯ = −w¯ −1 · z¯ · d w¯ on Xt . Thus we have the following inequality for
The Virtual Class of the Moduli Stack of Stable r-Spin Curves
17
the integral on any subset D contained in the region (z, w, t) C −1 |z| < |w| < C|z| : |f |2 · |z|(l) · φ z · |dzd z¯ | ≤ |g|2 · |w|(r−l) · φ w · |dwd w| ¯ C −2 · D D ≤ C2 · |f |2 · |z|(l) · φ z · |dzd z¯ |. D
Hence the norms are mutually bounded.
3.3.9. Collapsing a curve. We continue to use the notation. Lemma 3.20. Let f · e(l) = g · e(r − l) be a section of Ft on Xt (t = 0). Then we have the following equality: z |f |2 · |z|(l) · dz · d z¯ = |g|2 · |w|(r−l) · · dw · d w¯ . (15) w δ 2 <|z|<δ δ<|w|<1 Let f · e(l) · d z¯ = g · e(r − l) · d w¯ be a section of Ft ⊗ 0,1 of Xt (t = 0). Then we have the following equality: 2 (l) z |f | · |z| · · dz · d z¯ = |g|2 · |w|(r−l) · dw · d w¯ . (16) 2 w δ <|z|<δ δ<|w|<1 −1 Note that w · z < 1 in the region. Proof. It can be checked by a direct calculation, by using the relation (8). L2 Xt , Ft and L2 Xt , Ft ⊗ 0,1 by the norms Nt (l), φ z We obtain the L2 -spaces and Nt (r − l), φ w on Xtz and Xtw respectively. We denote the induced norm by N1 . In
the case t = 0, we also have L2 -spaces L2 Xt , Ft and L2 Xt , Ft ⊗ 0,1 by using the isomorphism Xt −→ D(δ 2 ) and the norm N (l), φ z on D(δ 2 ). We denote the induced
norm by N0 . The spaces L2 Xt , Ft (resp. L2 Xt , Ft ⊗ 0,1 ) and L2 Xt , Ft (resp.
L2 Xt , Ft ⊗ 0,1 ) are isomorphic as topological spaces. The norm of the identity maps
Ft : L2 Xt , Ft −→ L2 Xt , Ft and Gt : L2 Xt , Ft ⊗ 0,1 −→ L2 Xt , Ft ⊗ 0,1 are less than 1 for 3.20. t by Lemma any 2 The families L Xt , Ft 0 < |t| < 1 and L2 Xt , Ft ⊗ 0,1 0 < |t| < 1 are
naturally extended to the families over the disc t |t| < 1 , by putting L2 X0 , Ft :=
L2 X0z , F0z and L2 X0 , F0 ⊗ 0,1 := L2 X0z , F0z ⊗ 0,1 . The morphism F0 :
L2 X0 , F0 −→ L2 X0 , F0 is given by the natural projection L2 X0z , F0z ⊕ L2 X0w ,
F0w −→ L2 X0z , F0z . The morphism G0 : L2 X0 , F0 ⊗ 0,1 −→ L2 X0 , F0 ⊗ 0,1 is given by the natural inclusion L2 X0z , F0z ⊗ 0,1 −→ L2 X0z , F0z ⊗ 0,1 ⊕ L2 X0w , F0w ⊗ 0,1 . We have the trivialization of the families L2 Xt , Ft |t| < 1 and L2 Xt , Ft ⊗ 0,1 |t| < 1 as in Subsubsect. 3.3.5. We also have the trivialization of the families 2 L Xt , Ft |t| < 1 and L2 Xt , Ft ⊗ 0,1 |t| < 1 by taking the diffeomor2 ) X . Then the family of the morphisms F is the family of the morphisms phisms D(δ t 0
2 Ft : L X0 , F0 −→ L2 X0 , F0 . Similarly, we obtain the family of the morphisms t : L2 X0 , F0 ⊗ 0,1 −→ L2 X0 , F0 ⊗ 0,1 . G
18
T. Mochizuki
t |t| < 0 and G t |t| < 0 Lemma 3.21. The families of the bounded operators F are strongly continuous with respect to the variable t. Proof. Let us see the strong continuity of Ft . The continuity at t = 0 is easy to see. To show the continuity at t = 0, we have only to show that the induced morphisms
w
2 Ft : L X0 , F0 −→ L2 X0 , F0 converge to 0 strongly. For diffeomorphisms ηt : 1/2 = δ · ηt∗ |w|−1 are bounded indepenD(δ) −→ Xtw , the functions ηt∗ |z| · |w|−1 dently of δ, and they converge to 0 when δ → 0. Then we obtain the strong convergence t to 0 due to (15). of F Let us see the strong continuity of Gt. We have only to see at t = 0, and we have only t : L2 X0 , F0 ⊗ 0,1 −→ L2 X w , F0 ⊗ 0,1 conto show the induced morphisms G 0 verge to 0. For diffeomorphisms ηt : Xt −→ D(δ 2 ), we put T (t) := z ηt ∈ Xtw . Let g ˜ t , we have G ˜ t (g) = G ˜ t g|T (t) . be an element of L2 X0 , F0 ⊗ 0,1 . For the morphism G The functions |z| · |w|−1 are bounded on T (t) independently of t, and the sets T (t) go ˜ (g) to 0, namely the strong convergence to {0}. Hence we obtain the convergence of G
˜ to 0 due to (16). of G Similarly we have the family of topological vector spaces L2 Xt , Lt |t| < 1 and 2
L Xt , Lt ⊗ 0,1 |t| < 1 , and we have strongly continuous families of bounded
morphisms Ft : L2 Xt , Lt −→ L2 Xt , Lt and Gt : L2 Xt , Lt ⊗ 0,1 −→ L2 Xt , Lt ⊗ 0,1 . Finally we give a remark on the morphism ψt of L2 Xt , Lt −→ L2 Xt , Ft ⊗ 0,1 induced by the norms. We consider the family of the semi-norms N s (0 ≤ s ≤ 1) on the topological vector spaces L2 (Xt , Ft ) and L2 Xt , Ft ⊗ 0,1 given by Ns = (1 − s) · N0 + s · N1 . Here we regard N0 as semi-norms on L2 X0 , F0 and L2 X0 , F0 ⊗
0,1 instead of the norms on L2 X0 , F0 and L2 X0 , F0 ⊗ 0,1 in the case t = 0. In the case s = 0, they are the norms. They induce the anti-linear continuous maps (s) ψt : L2 Xt , Lt −→ L2 (Xt , Ft ⊗ 0,1 ). It is continuous with respect to t and s. (s) Note that ψt is not isomorphic in the case (t, s) = (0, 0). In that case, the image of
(0) (0) ψ0 is L2 X0 , F0 ⊗ 0,1 = L2 X0z , F0z ⊗ 0,1 , and the map ψ0 factors through z z L2 X0 , L0 . 3.4. The global family and its graph-continuity. 3.4.1. Preliminary. Let m be an n-tuple of integers m1 , . . . , mn such that −1 ≤ mi . 1/r,m 1/r,m We denote the moduli stack Mg,n (resp. Mg,n ) simply by M (resp. M) in this subsection. Let (C, p, F) be a stable r-spin curve over the complex number field C of type m. We denote mi also by m(Pi ). We denote the set {Pi ∈ p | mi = −1} by E. Let q = Q1 , . . . , Ql denote the set of nodal points of the curve C. We denote the normalization of C by π : C˜ −→ C. Let {Qi,1 , Qi,2 } denote the preimage ofQi via π . We put p := p π −1 (q). The r-spin structure F determines the index m Qi j at the point Qi j . := m(Pi ), m(Qi j ) Pi ∈ p, Qi j ∈ π −1 (q) . Hence we obtain the tuple of integers m is naturally induced. We put E := Qi j ∈ on C The r-spin structure F of type m −1 := E E . π (q) m(Qi j ) = −1 , and E
The Virtual Class of the Moduli Stack of Stable r-Spin Curves
19
Let α = (α1 , . . . , αn ) be an element of Rn . For simplicity, we assume −2 < αi < 2 for any i. We denote αi also by α(Pi ). We assume α(P i ) = −1 for any point Pi ∈ E. For any point Qi j ∈ π −1 (q), we put α(Qi j ) := m(Qi j + 1) . Thus we obtain the tuple of the real numbers α = α(P˜i ), P˜i ∈ p˜ ∈ Rp. We put α ∨ (P˜i ) = −α(P˜i ) for any point P˜i ∈ p. Thus we obtain the tuple α ∨ ∈ Rp. Similarly we obtain the tuple α ∨ ∈ Rp. 3.4.2. The global family of the topological vector spaces over the moduli. We take an with the following conditions: 1,1 -valued hermitian metric (Remark 3.1) of F − • It is C ∞ on C p. ∞ 1,1 i ∈ • Around a point P p, it is of the form ψ · |z|α(Pi ) , where ψ denotes a C - hermitian metric, and z denotes a holomorphic coordinate such that z P˜i = 0. ⊗ 0,1 with a similar condition. Then We also take an 1,1 -hermitian metric of F we obtain the Hilbert spaces: 0 α , F, E (C,p,F ) := L2 C,
1 ⊗ 0,1 , α . F E (C,p,F ) := L2 C,
The underlying topological space is independent of a choice of 1,1 -hermitian metrics. 0 1 We also have the closable operator ∂ : E (C,p,F ) −→ E (C,p,F ). ⊗L⊗ 0,1 −→ 1,1 , the −1 ⊗ω . Since we have the natural pairing F We put L := F C 1,1 -hermitian metric of F induces the 1,1 -hermitian metrics of L ⊗ 0,1 . Similarly the 1,1 -metric of F ⊗ 0,1 induces the 1,1 -metric of L. Then we obtain the Hilbert spaces: ∨0 L, E (C,p,F ) := L2 C, α∨ ,
∨1 L ⊗ 0,1 , E (C,p,F ) := L2 C, α∨ .
The underlying topological vector spaces are independent of a choice of 1,1 -hermitian ∨0 ∨1 dual : E (C,p,F ) −→ E (C,p,F ). metrics. We also have the closable operator ∂ ∨1
0
We have the naturally defined perfect pairing of E and E . We also have the 1 ∨0 naturally defined perfect pairing of E and E . If we fix 1,1 -metrics, we obtain ∨1 0 ∨0 the anti-Hermitian isomorphisms ψ : E (C,p,F ) −→ E (C,p,F ) and ψ : E (C,p,F ) −→ 1
E (C,p,F ). They depend on choices of 1,1 -hermitian metrics. dual
are mutually formal adjoint. To 3.4.3. The adjointness. The operators ∂ and −∂ make them mutually adjoint, we have to impose the conditions to their domains. For the points P˜i ∈ π −1 (q), we impose the matching condition explained in Subsubsect. 3.3.7 (see Condition 3.1). For any points Pi ∈ p such that α(Pi ) ∈ 2 · Z, we consider the following conditions: 0 (Pi ) If u ∈ E (C,p,F ) is contained in the domain of ∂, we have the vanishing zα(Pi )/2 · u (Pi ) = 0. ∨0
(Pi , dual) If v ∈ E (C,p,F ) is contained in the domain of ∂ −α(P i )/2+1 · v (P ) = 0. z i
dual
, we have the vanishing
20
T. Mochizuki
Lemma 3.22. For each point Pi , we impose one of the conditions (Pi ) or (Pi , dual). dual Then ∂ and ∂ are mutually adjoint. Proof. Similar to Lemma 3.18.
3.4.4. The preferred norms. Let (C, p, F) be a stable r-spin curve over the complex k number field C, which can be regarded as a point of M. We take a kmulti-disc B (T ) := z := (z1 , . . . , zk ) |zi | < T , i = 1, . . . , k and the etale map B (1) −→ M such that the point (0, . . . , 0) ∈ B k (1) corresponds to the curve (C, p, F) considered. We have the universal family q : C −→ B k and the universal r-spin structure F. We denote the fiber q −1 (z) by Cz . Let σj denote the j th section of q. For any point z ∈ B k (1), we i
i
have the corresponding topological vector spaces E z and E z . We would like to take a family of their norms with some good properties. In the following argument, we shrink B k (1) to B k (T ) for an appropriate T ≤ 1 without mentioning it, if we need to. We may assume that the divisors zi = 0 (i = 1, . . . , j ≤ k) correspond to the singularities of the map q. Then the morphism q is submersive on B ∗ j × B k−j , where B ∗ denotes the punctured disc. Let Qi be the corresponding nodal point of the curve C(0,... ,0) . Let (mi,1 , mi,2 ) be the index at Qi . We can take an open subset Ui of C around Qi and an embedding Ui −→ B k+2 (T ) = (z1 , . . . , zk , xi , yi ) such that the image is the closed subset determined by xi yi = zie , where e = gcd(r, mi,1 + 1). Let fi and gi be holomorphic functions on the open subset Ui such that fi = xi f˜i , gi = yi g˜ i where f˜i (0, 0) and g˜ i (0, 0) are not 0. Namely the functions fi and gi are close to xi and yi respectively. We consider the following subsets of Ui : y Uix := |fi | > |gi | , Ui := |fi | < |gi | , Wi := |fi | = |gi | . We may assume that the index mi,1 (resp. mi,2 ) corresponds to the divisor given by xi = 0 (resp. yi = 0). We have the trivialization e(mi,1 + 1)x (resp. e(mi,2 + 1)y ) of F y on Uix (resp. Ui ) given as in Subsubsect. 3.3.2. In the following construction, we shrink Ui without mentioning it, if we need to. xi yi Remark 3.4. Let us take positive φ kand φ functions on Ui . Then we obtain the family 1,1 of -hermitian metrics of Ui ∩ Cz z ∈ B (T ) . (See Remark 3.1.) We take a family of 1,1 -hermitian metrics of the family F˜ z z ∈ B k (T ) with the following conditions: • They are smooth on C − i Ui ∪ j σj . • Around the j th section σj , it is of the form |z|αj · ψ, where ψ denotes a family of C ∞ - 1,1 -hermitian metrics, and z denotes a relative coordinate of q : C −→ B around σj such that z(σj ) = 0. • On Ui , it is of the form given as in Remark 3.4, i.e., it is induced by positive functions φ xi and φ yi . z ⊗ 0,1 z ∈ We also take a family of 1,1 -hermitian metrics of the family F Cz i B k (T ) . They induce the family of Hilbert norms on the topological vector spaces E z ∨i
and E z (i = 0, 1). Definition 3.1. Such norms are called preferred norms.
The Virtual Class of the Moduli Stack of Stable r-Spin Curves
21
Lemma 3.23. The preferred norms are mutually bounded. Proof. The change of positive functions φ xi and φ yi has no effect up to boundedness. A change of the holomorphic functions fi and gi satisfying the conditions above changes the cutting circles. Note that the circles Wi = {|fi | = |gi |} are contained in a region of the form {C −1 |xi | ≤ |yi | ≤ C|xi |} for some C > 1. Then we may apply the argument in Lemma 3.19. Thus any changes of the holomorphic functions fi and gi have no effect. Remark 3.5. A sum of preferred norms is not necessarily preferred, because the cutting circles are not same. But it is mutually bounded with any preferred norm. 3.4.5. The local trivializations and the bundle structure. We continue to use the setting in Subsubsect. 3.4.4. We take a preferred norm. Then we can take the trivializations of the families of the topological vector bundles, which we shall explain. 0, zj +1 , . . . , zk ) We denote the projection B k −→ B k−j , (z1 , . . . , zk ) −→ (0, . . . , by p. We can take a diffeomorphism η : C − i Wi p ∗ C|B k−j − i Wi|B k−j . We [δ, C] defined by χδ (a) = (C − δ) · C −1 a + δ. We use have the maps χδ : [0, C] −→ √ √ the real coordinate xi = Ri · e −1αi and yi = Si · e −1βi . Then we may assume the following: • A point (z1 , . . . , zk , xi , yi ) of Ui ∩ |xi | > |yi | is mapped to √ 0, . . . , 0, zj +1 , . . . , zk , χ −1 (Ri ) · e −1αi , 0 via η. • A point (z1 , . . . , zk , xi , yi ) of Ui ∩ |xi | < |yi | is mapped to √ 0, . . . , 0, zj +1 , . . . , zk , 0, χ −1 (Si ) · e −1βi via η. ∗ −1 We can lift the morphism P to the bundle map η : F η F such that it is of the form η∗ e(mi,1 + 1)x = e(mi,1 + 1)x on Uix , and of the form η∗ e(mi,2 + 1)y = y e(mi,2 + 1)y on Ui . Then there exist the positive functions φi (i = 0, 1) such that φi · η∗ gives the unii i i tary isomorphism z of η−1 E z = E p(z) to E z . It gives the local trivializations of the i families E z z ∈ B k . Similarly, we can take the local trivializations of the families ∨i Ez z ∈ Bk . Since the preferred metrics are mutually bounded, the change of the trivializations, raised by changes of preferred norms, are continuous. As a result, we obtain the bundle k ∨k structure of the families E and E . We can also take global families of 1,1 -hermitian k metrics of F and F ⊗ 0,1 , which induces the Hilbert norms of E over M which is mutually bounded with locally defined preferred norms. dual
3.4.6. Graph-continuity of the families of the operators ∂ and ∂ . For the trivializations, the following is shown by the argument of Seely-Singer. (See [11]). Proposition 3.1. The families of the closed operators ∂¯z and ∂¯zdual are graph-continuous. ∨0
1
∨1
0
Proof. We have the anti-Hermitian isomorphisms E E and E E induced by 1 0 the Hilbert norms. Thus we regard the operators −∂¯zdual as the closed operator E −→ E ,
22
T. Mochizuki 0
1
which is the adjoint of ∂. We put E z := E z ⊕ E z , and then we have the family of the closed operators:
id ∂¯ Dz = : E z −→ E z . −∂¯ dual id The graph continuity of the operators ∂¯ and −∂¯ dual is equivalent to the norm-continuity of the inverse Dz−1 . We constructed a family of local parametrices for the families of the operators ∂¯ and dual ∂ around the nodal points or marked points in Subsect. 3.2 and 3.3. Hence we obtain and Q dual . the continuous family of the global parametrices, which we denote by Q Then the following holds:
T Q 0 Q id −∂¯ dual = id + (17) dual 0 dual T . −Q −Q ∂¯ id In other words, we have Dz Qz = I + Tz , where Tz is a compact operator. Note that Dz · Qz is defined, since the image of Qz is contained in the domain of Dz . Since the operators Tz vanish around the nodal points and the marked points, it is easy to see that they give the continuous family. The right-hand side of (17) is Fredholm. Since the operator Dz is invertible, the operator Qz is a Fredholm operator of E z to the domain of Dz with the index 0. We take a complement Vp(z) of the image Im(Qp(z) ) in the domain Dom(Dp(z) ). In particular, we focus around the point 0 := (0, . . . , 0) ∈ B k . Let S0 be a morphism of E 0 to Dom(D0 ) which gives the isomorphism Ker(Q0 ) −→ V0 and vanishes on Ker(Q0 )⊥ . Thus D0 (Q0 + S0 ) = id + T0 + D0 S0 is invertible, and we obtain the equality D0−1 = (id + T + D0 S0 )−1 · (Q0 + S0 ). We extend S0 to the continuous family of the operators on B k−j (T ) = {(0, . . . , 0, zj +1 , . . . , zk )} ⊂ B k (T ). We have the equality Dz · (Qz + Sp(z) ) = id + Tz + Dz Sp(z) . The right hand side is norm-continuous, and hence invertible on a neighbourhood of the point 0. Thus the operator Dz has the norm-continuous inverse. 4. The Virtual Class 4.1. The axioms of Jarvis-Kimura-Vaintrob. For any n-tuple of integers m, we put d(r, m) := −2r −1 · (2g − 2 − mi ) + 2. Let Gg,n denote the set of the decorated stable graph of genus g with n-tails marked with the n-tuple m of non-negative integers. 1/r 1/r Definition 4.1 ([6 and 7]). A family of the cohomology classes c ∈ H d(r,m) M ∈ Gg,n satisfying the following axioms, are called r-spin virtual classes. If the graph 1/r 1/r,m consists of one vertex and no edge, then c is denoted by cg,n . ∗ 1/r,m 1/r r Axiom 1. If the graph is connected, the equality c = holds, e∈E() le · i cg,n 1/r
where i denotes the inclusion morphism M
1/r,m
−→ Mg,n . If the graph is the 1/r 1/r (d) disjoint union of the graphs , then it is described as the product c = d c (d) . 1/r,m
Axiom 2 (Convexity). In the case π∗ F = 0 on some connected components of Mg,n , 1/r,m then the restriction of the cohomology class cg,n to the component is the Euler 1 ∨ class of the vector bundle (R π∗ F) .
The Virtual Class of the Moduli Stack of Stable r-Spin Curves
23
Axiom 3 (Cutting edge). Let be a decorated graph and ˜ be the graph obtained by 1/r 1/r cutting all edges of . We have the morphism µ˜ : M˜ ×M M −→ M . 1/r
1/r
We also have the morphism p1 : M˜ ×M M −→ M˜ . Then the equality 1/r 1/r p1 ∗ µ˜ ∗ c = r |E()| · c ˜ holds. Axiom 4 (Descent Axiom). We have the universal i th section σi of the universal curve 1/r,m 1/r,m Cg,n −→ Mg,n . Thus we have the line bundle σi∗ F which we denote by ψ˜ i (m). Let r · δ i denote an n-tuple of integers which is r in the i-th position and 0 in all 1/r,m+rδ i 1/r,m = −ψ˜ i (m) · cg,n holds. others. Then the equality cg,n 1/r,m Axiom 4’ (Vanishing axiom). In the case mi = r − 1 for some i, the virtual class cg,n is 0. Axiom 5 (Forgetting tails). Let ˜ be a decorated stable graph with a tail ti marked by mi = 0, and let be the decorated stable graph obtained by removing the tail ti . We 1/r 1/r 1/r 1/r have the forgetting morphism π : M˜ −→ M . Then the equality c ˜ = π ∗ c holds. In the next subsection, we recall the construction due to Witten. 1/r,m
Remark 4.1. In [13], the class cg,n vis-Kimura-Vaintrob [6].
is called the top Chern class. Here we follow Jar-
Remark 4.2. Jarvis-Kimura-Vaintrob also considered the case where one of mi is −1. In this paper, we restrict our attention to the tuples of non-negative integers. Remark 4.3. In Axiom 3, if one of the nodes is of Ramond type, the right hand side is understood to be 0. Hence we have to show the vanishing of the right hand side in that case. Remark 4.4. In the case mi = r − 1, it is easy to see that ψ˜ i is 0 in the rational cohomology. Hence Axiom 4 would formally imply Axiom 4’, if we could extend the Seely-Singer-Witten construction in the case when one of mi is −1. 4.2. The construction of virtual classes. 4.2.1. A finite reduction. Let m denote a tuple of integers (m1 , . . . , mn ) such that mi ≥ −1. Let be a decorated stable graph of genus g with n-tails marked with the n-tuple m. We denote M by M for simplicity. Let α = (α1 , . . . , αn ) be an n-tuple of real numbers. We impose the following conditions for simplicity: Condition 4.1. • In the case mi = −1, the inequality −2 < αi < 0 holds. • In the case 0 ≤ mi ≤ r − 1, the inequality −2 < αi ≤ 1 holds. • In the case mi ≥ r, the inequality −2 < αi < 2 holds. 1/r,m
In the case that consists of one vertex and no edge, i.e., M = Mg,n , we have i
∨i
the bundles E and E of the topological vector spaces and the family of the closable operators over M (see Subsect. 3.4). If is connected, then the bundles are obtained by
24
T. Mochizuki 1/r,m
the pull back via the morphism M −→ Mg,n . If is a disjoint union of connected graphs (i) , the bundles are obtained as a direct sum of the pull backs via the projections M −→ M (i). By taking an appropriate family of 1,1 of the fami -hermitian metrics lies F˜ (C,p,F ) (C, p, F) ∈ M and F˜ (C,p,F ) ⊗ 0,1 (C, p, F) ∈ M , we obtain the ∨0
1
anti-Hermitian isomorphism ψ : E −→ E . When we distinguish the dependence of i i E on α and m, we use the notation E (m, α), and similarly for others. dual We take the domains of ∂ and ∂ such that they are mutually adjoint. For that purpose, we have only to impose one of the conditions (Pi ) or (Pi , dual) for each point Pi ∈ p such that α(Pi ) = 0. Condition 4.2. • We choose the condition (Pi ) (resp. (Pi , dual)) in the case αi > 0 (resp. αi < 0), if we do not mention anything. • In the case mi = −1, we do not impose either (Pi ) or (Pi , dual). i
Definition 4.2. Let us consider finite dimensional subbundles E i ⊂ E (i = 0, 1) with the following conditions: 0 ¯ • The fiber E(C, ∂. p,F ) over (C, p, F) ∈ M is contained in the domain of the operator 0 1 The image ∂(E(C,p,F )) is contained in E(C,p,F ), and the equality index(∂) = index ∂ : 0 1 E(C, p,F ) −→ E(C,p,F ) holds. 0 1 • The fiber E(C, p,F ) contains the orthogonal complement of ∂ E(C,p,F ) . 0 th r−1 gives the L2 -section of L • For any element s of E(C, (C,p,F ) p,F ), the (r −1) square s with the appropriate weight. (See Subsubsect. 3.4.2 for L.) Such E 0 −→ E 1 is called a finite reduction of ∂¯ : E −→ E . 0
1
Lemma 4.1. There is a finite reduction. Proof. First, we take a finite dimensional subbundle H of E conditions:
∨1
satisfying the following
∨1 • The naturally defined morphisms H(C,p,F ) −→ E (C,p,F )/ Im ∂¯ dual are surjective for any point (C, p, F) of M. • Any element s of H(C,p,F ) is a smooth section of L(C,p,F ). Moreover s is constantly 0 on a neighbourhood of the marked points Pi and the preimage of the nodal points. Since the family of the operators ∂¯ dual is graph-continuous, (∂¯ dual )−1 (H ) gives the ∨0
∨1
dual
). We subbundle of E because of the surjectivity of H(C,p,F ) −→ E (C,p,F )/ Im(∂ 1 dual −1 (H ) , and then it is a subbundle of E . Since ψ Ker ∂¯ dual is put E 1 := ψ ∂ the orthogonal complement of Im(∂), E 1 contains the orthogonal complement of Im(∂). We put E 0 := ∂¯ −1 (E 1 ). Since the family of the operators ∂¯ is graph-continuous, E 0 is 0 a subbundle of E . Thus we obtain the finite dimensional bundles E 0 and E 1 and the ¯ E 0 : E 0 −→ E 1 . morphism ∂| The first and second conditions hold by our construction. Let us check the third condition, i.e. the L2 -properties around the singular points. Let (C, p, F) be a stable r-spin curve. We use the notation in Subsect. 3.4. The singular points we have to consider are the following:
The Virtual Class of the Moduli Stack of Stable r-Spin Curves
25
such that α(Pi ) = 0. 1. The points Pi ∈ p−E Recall that we put −2 < α(Pi ) < 0. 2. The marked points Pi ∈ E. dual −1 Let us consider Case 1. Let s be an element of ∂ (H )(C,p,F ). Then s is holo 1/r −α(P ) i · f · zm(Pi ) · dz · d z¯ , where morphic around Pi , and thus ψ(s) is of the form |z| z denotes a holomorphic coordinate such that z(Pi ) = 0 and f denotes a C ∞ -function. 1/r 0 be an element of E (C,p,F ), which is contained in the domain Let u = g · zm(Pi ) · dz of ∂ and satisfies ∂u = ψ(s). Then it is easy to see that g is of the form z · |z|−αi · f1 + f2 , where f1 denotes a continuous function and f2 denotes a holomorphic function. In the case 0 ≤ mi ≤ r − 1, we have assumed |αi | ≤ 1, and thus g is bounded. Thus it is easy to check that ur−1 gives an L2 -section of L(C,p,F ). Let us consider the case mi ≥ r. We have only to consider the case 1 − αi < 0, and we can ignore f2 . Note that we have 1 − αi > −1. Since f1 is continuous, we have only to check the L2 -integrability of the function |z|(r−1)·(1−αi ) × |z|mi with respect to the measure |z|−α · |dz · d z¯ |, which is easy to see. dual −1 Let us consider Case 2. Let s be an element of ∂ (H )(C,p,F ). Then it is holo 1/r −α(P ) i · z−1 · f · z−1 · dz · d z¯ , morphic around Pi , and thus ψ(s) is of the form |z| −1 1/r 0 ∞ be an element of E (C,p,F ), where f denotes a C -function. Let u = g · z · dz which is contained in the domain of ∂ and satisfies ∂u = ψ(s). Then it is easy to see that g is of the form z · z−1 · |z|−α(Pi ) · f1 + f2 , where f1 denotes a continuous function and f2 denotes a holomorphic function. Hence we have only to check the L2 -integrability of −1/r (r−1)/r = z−1 · dz · dz/z , with respect to the measure |z|−α(Pi ) · |dz · d z¯ |. dz/z −2 −α(P ) i Since the function |z| · |z| is integrable with respect to the measure |dz · d z¯ |, we obtain the desired L2 -property. 4.2.2. The inner product of ∂s and ψ(s r−1 ). Let us take a finite reduction E 0 −→ E 1 . Let (C, p, F) be a point of M. Let π : C˜ −→ C denote the normalization of C. Let s be 0 r an element of E(C, of 1,1 with some singularity, let us p,F ). Since ∂s gives a section r consider the Stokes formula for the integral ∂s . We use the notation in Subsect. 3.4. Lemma 4.2. In the case Pi ∈ p − E, the contribution of the point Pi to the Stokes formula is 0. Proof. For the preimage of the nodal points, the claim is checked in Lemma 3.18. We have only to consider the case α(Pi ) = 0 for any marked point Pi . The claim immediately follows from the form of s around the points Pi , which has already been seen in the proof of Lemma 4.1. Let us consider the contribution of the points Pi ∈ E. Because of the inequality −2 < α(Pi ) < 0 and Lemma 3.11, we have the value s(Pi ). We have the isomorphism F ⊗ r ω(Pi ) around Pi , and we have the residue map ω(Pi ) −→ C at Pi . Hence we can regard s r (Pi ) as the complex number. Lemma 4.3. Let s be as above. Then we have the following formula: r−1 r · ∂s, ψ s = ∂s r = s r (Pi ). C˜
Pi ∈E
(18)
26
T. Mochizuki
Proof. The second equality follows from the standard Stokes formula and the residue formula. By definition of the anti-Hermitian isomorphism ψ, we have the following equality: ∂s, ψ s r−1 = s r−1 · ∂s = r −1 · ∂s r . C˜
Thus we obtain the first equality. E0
Let follows:
−→
E1
C˜
be a finite reduction. We define the morphism : E 0 −→ C as
(s) :=
∂s r =
s(Pi )r = r · ∂s, ψ(s r−1 ) .
(19)
Pi ∈E
4.2.3. Witten map and the 0-set. Let E 0 −→ E 1 be a finite reduction. We denote the 1 orthogonal projection of E onto the subbundle E 1 by ρE 1 . Let s be an element of E 0 . Since the (r − 1)th square s r−1 satisfies the L2 -condition, ∨0 r−1 ) ∈ E 1 . Hence we obtain the morit is an element of E . We have the element ψ(s r−1 0 1 . Then we put φ := ∂|E 0 + φ1 , phism φ1 : E −→ E given by φ1 (s) := ρE 1 ψ s which we call the Witten map. Let π denote the natural projection of the vector bundle E 0 onto M. The map φ can be regarded as the section of the vector bundle π ∗ E 1 over E 0 . We consider the restriction of the section φ to the subset −1 (a) ⊂ E 0 for a complex number a ∈ C. The following lemma is one of the key ideas for our vanishing argument (the vanishing axiom and the Ramond case of Axiom 3). Lemma 4.4. If an element s ∈ −1 (a) satisfies φ(s) = 0, then the complex number a is a non-positive real number. Proof. We have the equality 0 = φ(s) = φ1 (s) + ∂s. Taking the inner product with φ1 (s), we obtain the following equality: 2 0 = φ1 (s) + ∂s, ρE 1 ψ s r−1 . On the other hand, we have the following equalities: ∂s, ρE 1 ψ s r−1 = ρE 1 ∂s , ψ s r−1 = ∂s, ψ s r−1 = r −1 · (s) = r −1 · a. Thus a is real and non-positive.
The following lemma, due to Witten, is critical for the definition of the virtual class. Lemma 4.5 (Witten). Let s be an element of the subset −1 (0). The equality φ(s) = 0 implies the equality s = 0. Proof. Since we have the equality ∂s, ψ(s r−1 ) = 0 for any element s ∈ −1 (0), the following equality holds: ρE 1 ◦ ψ(s r−1 ), ∂s = ψ(s r−1 ), ∂s = 0. Hence the equality φ(s) = 0 implies ρE 1 ◦ ψ(s r−1 ) = ∂s = 0. In particular, s is holo dual , ψ s r−1 is contained in E 1 , by definition morphic. Since s r−1 is contained in Ker ∂ of finite reduction. Hence ρE 1 ◦ ψ(s r−1 ) = 0 implies ψ s r−1 = 0, which implies s = 0.
The Virtual Class of the Moduli Stack of Stable r-Spin Curves
27
Remark 4.5. Let x be a point of M, and let f1 , . . . , fd be an orthogonal frame of Ex1 . Then we have the formula for s ∈ Ex0 : φ1 (s) :=
d
s
r−1
· fi · fi .
i=1
The description may be useful to check the continuity of φ1 t with respect to the parameter t, when the continuous family of a vector bundles Et1 are given. 4.2.4. The construction of the cohomology class c . When the set E = i mi = −1 is empty, then we have −1 (0) = E 0 , by definition. Due to Lemma 4.5, the section φ on E 0 gives the top Chern class ctop (π ∗ E 1 , φ) of π ∗ E 1 with the support M. (See the appendix about the cohomology theory and the Chern class of orbispaces.) ∗ (E 0 ) −→ Since E 0 is smooth over M, we have the push forward morphism π∗ : Hcpt H ∗−rank E (M). Hence we obtain the cohomology class π∗ ctop (π ∗ E 1 , φ) in the cohomology group H ∗ (M). 0
Lemma 4.6 (Witten). The cohomology class is independent of a choice of finite reductions. The cohomology class is independent of a choice of Hilbert norms, when we fix the weight α. Proof. The second claim is easy, because two norms are connected by a continuous family. Let us show the first claim. We have only to compare the classes obtained from two finite reductions E 0 −→ E 1 and E˜ 0 −→ E˜ 1 in the case E i ⊂ E˜ i . The natural projections E 0 −→ M and E˜ 0 −→ M are denoted by π and π˜ respectively. We take the orthogonal complement E 1 of the subbundle E 1 in E˜ 1 . We take the complement E 0 of the subbundle E 0 in E˜ 0 such that the restriction of ∂ to E 0 gives the isomorphism of E 0 and E 1 . We have the sections φ1 : E 0 −→ π ∗ E 1 and φ˜ 1 : E˜ 0 −→ π˜ ∗ E˜ 1 , as in Subsubsect. 4.2.3. Then the sections φ = ∂ + φ1 and φ˜ = ∂ + φ˜ 1 are used for the constructions respectively. We have the section φ : E˜ 0 −→ π˜ ∗ E˜ 1 defined by φ (s1 , s2 ) = ∂(s1 + s2 ) + φ1 (s1 ) for s = (s1 , s2 ) ∈ E 0 ⊕ E 0 . It is easy to see that the cohomology classes obtained from the sections φ and φ coincide. We have the map at : E˜ 0 −→ E˜ 0 defined as at (s1 , s2 ) = (s1 , t · s2 ). We also have the map bt : E˜ 1 −→ E˜ 1 defined as bt (s1 , s2 ) = (s1 , ts2 ). Note that we have the following relation: ∂s, bt ◦ φ˜ 1 ◦ at (s) = bt ◦ ρE˜ 1 ∂s , φ˜ 1 ◦ at (s) = bt ∂s , φ˜ 1 ◦ at (s) = ∂ at (s) , φ˜ 1 at (s) = r −1 ∂(at (s)r ) = 0. (20) We put φ (t) (s) = ∂(s) + bt φ˜ 1 (at (s)) for 0 ≤ t ≤ 1. It is easy to see that φ (t) is ˜ By the same argument as the proof of a homotopy to connect the sections φ and φ. Lemma 4.5, the cohomology class is defined for any 0 ≤ t ≤ 1, and thus it is constant. Thus we obtain the well defined cohomology class c := π∗ ctop π ∗ (E 1 , φ) ∈ 1/r H ∗ M .
28
T. Mochizuki
4.2.5. Independence of the cohomology classes from a weight α. Let us see that the cohomology class c obtained in Subsubsect. 4.2.4 is independent of a choice of a weight α. (Recall Condition 4.1.) For that purpose, we have only to compare the cohomologyclasses for α and (0, . . . , 0). Hence we consider the family of the weights α(t) = αi (t) i = 1, . . . , n (0 ≤ t ≤ 1) where we put αi (t) := t · αi . i
∨i
Let E (m, α(t)) and E (m, α(t)) denote the vector bundles for the tuple m with the weight α(t). We take a continuous family of 1,1 -hermitian for weights α(t). metrics For each t, we obtain the cohomology classes c(t) ∈ H ∗ M by the construction in Subsubsect. 4.2.4. Lemma 4.7. The family of classes c(t) 0 ≤ t ≤ 1 is constant. Namely, the cohomology class is independent of a choice of a weight α. Proof. Let t0 be a real number such that 0 ≤ t0 ≤ 1. Let small us take a sufficiently positive number γ , and we put a(t0 ) := max 0, t0 − γ and b(t0 ) := min 1, t0 + γ . ∨ 1 m, α(t0 ) as in the proof of Lemma 4.1. We can regard them as Let us take H ⊂ E ∨ 1 m + r · δ i , α(t) for any t ∈ [a(t0 ), b(t0 )]. It is easy to see that the subbundle of E dual −1 the subbundles ∂ H are independent of a choice of t ∈ [a(t0 ), b(t0 )], due to dual
our choice of the domain of the operators ∂ (Condition 4.2). Since the family of the dual −1 1,1 1 -metrics is continuous for t, Et := ψ ∂ H gives a continuous family of −1 1 Et is also continthe vector bundles on M. The family of vector bundles Et0 := ∂ uous due to our choice of the domain of ∂. Then we can conclude that the family of the cohomology classes c(t) is constant on [a(t0 ), b(t0 )]. Therefore we obtain the desired constantness. 4.2.6. The definition of the virtual class. We have arrived at the following definition of the virtual class. Definition 4.3. Let m be a n-tuple of non-negative integers. Let us take a finite reduction 1/r for M with an appropriate weight α as in Condition 4.1, and we put as follows: r · c
c : = (−1)d(r,m) · le e∈E() r · π∗ (ctop (π ∗ E 1 , φ)) ∈ H d(r,m) M1/r = (−1)d(r,m) · . le e∈E()
1/r
−1 It class of the moduli stack M . Here we put d(r, m) = 2 · r · is called the virtual −2g + 2 + mi + 2.
4.2.7. Remark on the pull back. Although we discussed the construction of the coho1/r mology class from the universal family over M , the argument works for any locally trivializable orbispaces. (See the appendix for locally trivializable orbispaces.) Namely, 1/r assume we have a continuous map f : V −→ M of locally trivializable orbispac es. Then we have the family of r-spin structures, f ∗ C, p, F over V , where C, p, F
The Virtual Class of the Moduli Stack of Stable r-Spin Curves
29
1/r
denotes the universal r-spin curve over M . When we fix an appropriate weight α, i the r-spin structure f ∗ C, p, F induces bundles of topological vector spaces E V and ∨i
i
E V over V . They are isomorphic to the pull backs f ∗ E and f ∗ E
∨i
of the bundles of
1/r M . We
0
1
can take a finite reduction of ∂ : E V −→ E V , topological vector spaces over and then we obtain the cohomology class cV as in Subsubsect. 4.2.1–4.2.5, which is independent of choices of finite reductions, norms and weights. If we take a finite reduc0 1 1/r tion E 0 −→ E 1 of ∂ : E −→ E over M , then the pull back f ∗ E 0 −→ f ∗ E 1 0 1 gives a finite reduction of E V −→ E V . Lemma 4.8. In the case, we have cV = f ∗ c . 1/r
Proof. Let π : E 0 −→ M and π˜ : f ∗ E 0 −→ V denote the natural projections. We have the morphism fE : f ∗ E 0 −→ E 0 . Let φ denote the Witten map E 0 −→ 1/r M .Then fE∗ φ : E 0 −→ π˜ ∗ f ∗ E 1 denote the Witten map for V . We have π ∗ E 1 for fE∗ ctop π ∗ E 1 , φ = ctop π˜ ∗ f ∗ E 1 , fE∗ φ . Then we obtain the following: π˜ ∗ ctop π˜ ∗ f ∗ E 1 , fE∗ φ = π˜ ∗ fE∗ ctop π ∗ E 1 , φ = f ∗ π∗ ctop π ∗ E 1 , φ . It implies the claim.
4.3. The argument for vanishing. 4.3.1. Some vanishing. Let us continue to use in Subsubsect. 4.2.3, and the notation let us consider the case that the set E := i mi = −1 is not necessarily empty. We : E 0 −→ π ∗ E 1 ⊕ C, induced by the sections φ : E 0 −→ π ∗ E 1 consider the section φ 0 and : E −→ C. Here C denotes the trivial bundle on E 0 . Due to Lemma 4.5, the is M. Hence we obtain the cohomology class π∗ ctop (π ∗ E 1 ⊕ C, φ ) as in 0-set of φ Subsubsect. 4.2.4. We perturb the section to − a for a ∈ C \ {t ∈ R | t ≤ 0}, and = (φ, ) to φ −1 we perturb the section φ ∗ a 1:= (φ, a ). Then we know φa (0) = ∅ due to is trivial. Lemma 4.4. Thus the class π∗ ctop π E ⊕ C, φ 4.3.2. The vanishing axiom (the case |E| = 1). Let us consider the case |E| = 1, say E = {i}, i.e., mi = −1 and mj ≥ 0 (j = i). Let M denote the moduli stack of stable 0
r-spin curve of type m. We take a finite reduction E 0 −→ E 1 of ∂ : E (m, α) −→ 1 E (m, α) as in Subsubsect. 4.2.1 for the weights α as in Condition 4.1. Recall that we do =0 not impose either (Pi ) or (Pi , dual). We have the vanishing πE ∗ ctop π ∗ E 1 ⊕ C, φ as is shown in Subsubsect. 4.3.1. Let us impose the condition (Pi ) to the domain of ∂. To distinguish it, we denote the (1) −1 1 (1) (E ). We may assume that it is a subbundle of operator by ∂ . We put Eˆ 0 := ∂ E 0 such that E 0 /Eˆ 0 is of rank one. Let πˆ : Eˆ 0 −→ M denote the natural projection and ι : Eˆ 0 −→ E 0 denote the inclusion. Lemma 4.9. We have ctop C, = r · ι∗ 1 in H−1 (0) E 0 . Proof. We have −1 (0) = Eˆ 0 with the multiplicity r, which gives the desired equality.
30
T. Mochizuki
Let φˆ denote the restriction of φ to Eˆ 0 . Lemma 4.10. We have πˆ ∗ ctop πˆ ∗ E 1 , φˆ = 0. = ι∗ ctop πˆ ∗ E 1 , φˆ . Then Proof. From Lemma 4.9, we have r −1 · ctop π ∗ E 1 ⊕ C, φ . the claim follows from the vanishing of π∗ ctop π ∗ E 1 ⊕ C, φ Let us consider the tuple m = m+r ·δ i , where r ·δ i is given in Axiom 4 in Definition 4.1. Note m i = mi + r = r − 1. The moduli stack of stable r-spin curves of type m
is naturally isomorphic to M. We do not distinguish them. We put α = α + 2 · δ i . Let i E (m , α ) be the bundle of topological spaces for the tuple m with the weight α . We i i have the natural identification E (m, α) E (m , α ), and Eˆ 0 −→ E 1 gives a finite 0 1
reduction of E (m ∗, α 1) −→ E (m , α ) under the identification. Hence the vanishing of the class πˆ ∗ ctop πˆ E , φˆ implies the following vanishing axiom. Proposition 4.1 (The vanishing axiom). The virtual class is trivial in the case where one of mj is r − 1. 4.3.3. Preliminary for the vanishing in the Ramond case in Axiom 3. Let be a stable decorated stable graph of genus g with n-tails marked with the n-tuple m of integers such that mi ≥ −1. We put E = {i mi = −1}. Let us consider the case E is indexed as E = {(i, j ) i = 1, . . . w, j = 1, 2}. 1/r
Let (C, F) be the universal r-spin curve over M . Let σi,j denote the section of 1/r
C −→ M , corresponding to (i, j ) ∈ E. Note that we have the canonical isomorphism ⊗r ρi,j : F|σ C given by the residue. i,j 1/r
Let us take a finite reduction E 0 −→ E 1 for M . Let π : E 0 −→ M denote 0 the natural projection. We have the morphism i : E −→ C given by i (s) = r r ρi,1 (s ) + ρi,2 (s ). Note the equality (s) = i (s). Let φ : E 0 −→ E 1 denote the Witten map. 1/r Lemma 4.11. We have φ −1 (0) ∩ i −1 i (0) = M . −1 Proof. It follows from i −1 i (0) ⊂ (0). : E 0 −→ π ∗ E 1 ⊕ i C. Since the 0-set of The maps φ and i induce the section φ 1/r ∗ 1 ∗ is M1.r φ , we obtain the cohomology class π∗ ctop π E ⊕ i C, φ in H M . = 0. Lemma 4.12. We have the vanishing π∗ ctop π ∗ E 1 ⊕ i C, φ Proof. Note that we have i (s) = (s). Let a be a complex number which is not a w −1 −1 non-positive real number. Then the set φ −1 (0) ∩ −1 i=2 i (0) ⊂ φ (0) ∩ 1 (a) ∩ −1 (a) is empty, due to Lemma 4.4. Hence we are done. 4.4. The main theorem. 4.4.1. The statement 1/r,m
Theorem 4.1. The virtual classes cg,n obtained in Subsect. 4.2 satisfy the axioms of Jarvis, Kimura and Vaintrob (see Definition 4.1).
The Virtual Class of the Moduli Stack of Stable r-Spin Curves
31
Axiom 4’ (the vanishing axiom) is proved in Proposition 4.1. Let us look at Axiom 2. Under the assumption in Axiom 2, the operators ∂ are injective for any point (C, p, F) of the component, and R 1 π∗ F is a locally free sheaf on the component. Hence we can take 0 −→ R 1 π∗ F as a finite reduction. Therefore the claim in Axiom 2 holds. We will see that the other axioms are satisfied in the following subsubsections. 4.4.2. Axiom 1. Axiom 1 follows from the following two lemmas. Lemma 4.13. If is a connected decorated stable graph of genus g with n-tails marked ∗ 1/r,m with the n-tuple m, the virtual class c is same as e∈E() r/ le · i cg,n , where i 1/r,m
1/r
denotes the natural morphism M −→ Mg,n . Proof. It immediately follows from Lemma 4.8.
(i) Lemma 4.14. In the case when is a disjoint sum of connected graphs , we have c = i c (i) . 1/r
Proof. We take finite reductions E0 (i) −→ E1 (i) for M (i) . Let E 0 −→ E 1 denote the 1/r
direct sum of the pull backs of E0 (i) −→ E1 (i) via the projections (i) : M
−→
1/r 1/r M (i) . Since the bundle of topological vector spaces on M is a direct sum of the pull 1/r 1/r backs of the bundles over M (i) , E 0 −→ E 1 is a finite reduction for M . 1/r Let φ (i) : E0 (i) −→ E1 (i) denote the Witten map on M (i) , then the direct sum φ of 1/r the pull backs of φ (i) is the Witten map on M . 1/r 1/r Let π : E 0 −→ M and π (i) : E0 (i) −→ M (i) denote the natural projections. We (i) have the naturally defined morphism E : E 0 −→ E0 (i) . Then we have the following:
ctop π ∗ E 1 , φ = ctop π ∗ (i) ∗ E1 (i) , (i) ∗ φ (i) i
=
(i) ∗ ctop π (i) ∗ E1 (i) , φ (i) . E ctop π (i) ∗ E1 (i) , φ (i) =
i
The first equality follows from (π ∗ E 1 , φ) i π ∗ (i) ∗ E0 (i) , (i) ∗ φ (i) and the prod uct formula of the top Chern classes, and the third equality follows from E 0 E0 (i) and the Kunneth formula. Then we obtain the following: (i) π∗ ctop π (i) ∗ E1 (i) , φ (i) . π∗ ctop π ∗ E 1 , φ = i
Then the claim immediately follows.
4.4.3. Axiom 4. We have only to consider the case where the graph consists of one vertex and no edge. We compare the classes for the tuples m = (m1 , . . . , mn ) and m+r ·δ i . 1/r,m+rδ i 1/r,m Recall that we have the natural isomorphism between Mg,n and Mg,n . We do not distinguish them. Let α(t) be the family of the weights given as follows: αj (t) = 0
(j = i),
αi (t) = t.
32
T. Mochizuki ∨i
i
Let E (m + r · δ i , α(t)) and E (m + r · δ i , α(t)) denote the vector bundles for the tuple m +r · δ i with the weight α(t). For each t, we obtain the cohomology classes c(t) ∈ H ∗ M by the construction in Subsubsect. 4.2.4. In the case t = 0, the coho1/r,m+r·δ
i mology . Due to Lemma 4.7, the family of classes (t) class is the virtual class cg,n c 0 ≤ t < 2 is constant. Let β(t) (0 < t ≤ 2) denote the weight such that βi (t) = t − 2 and βj (t) = 0 i ∨ i for i = j . Let E m, β(t) and E m, β(t) for the tuple m with the weight β(t). We obtain the family of the cohomology classes c (t) 0 < t ≤ 2 of H ∗ M . In the case 1/r,m t = 2, the class is the same as the virtual class cg,n . The family c (t) 0 < t ≤ 2 is constant due to Lemma 4.7. Hence we have only to compare c(t) and c (t) for some t ∈ (0, 2). Let us fix an element t ∈ (0, 2). It is easy to see the following equalities:
i i E m + r · δ i , α(t) = E m, β(t) , dual
E
∨ i
∨ i m + r · δ i , α(t) = E m, β(t) . (1)
(1) dual
We denote ∂ and ∂ for the tuple m+r ·δ i by ∂ and ∂ respectively. We denote dual (2) (2) dual (1) ∂ and ∂ for the tuple m by ∂ and ∂ respectively. The difference of ∂ and (2) (1) dual (2) dual ∂ and the difference of ∂ and ∂ are the choices of the domains. Recall (1) that we impose the condition (Pi ) for the domains of ∂ . If we impose the condition (2) (2) dual . We use the identification in (Pi , dual), then we obtain the operators ∂ and ∂ the following argument without mention. ∨ 1 Let H be a subbundle of E m + r · δ i , α(t) as in the proof of Lemma 4.1. We (i) dual −1 (H ) (i = 1, 2). We have H2 ⊂ H1 . We may assume that H2 put Hi := ∂ is a subbundle of H1 such that the rank of H1 /H2 is 1, by replacing H with a bigger subbundle if we need it. For any point (C, p, F) ∈ M, we have the naturally defined morphism: ! 0 1 (1) (21) ψ(H2 )(C,p,F ) −→ Cok ∂ : E (C,p,F ) −→ E (C,p,F ) . Note that the cokernel of the map (21) is at most one dimensional, because ψ(H2 )(C,m,F ) −→ Cok ∂
(2)
is surjective.
Lemma 4.15. We may assume that the map (21) is surjective for each (C, p, F), by replacing H with a bigger subbundle if we need it. Proof. Let s be any section of F ⊗ 0,1 . It gives an element of the cokernel of (21). We remark that we can take a section s of F ⊗ 0,1 with the following conditions: • The induced element in the cokernel of (21) is not 0. • s is C ∞ , and it vanishes around the singular points and the marked points of (C, p). If we add an element ∂ψ −1 (s) to H(C,p,F ), then we obtain the desired property at x. It is also elementary to globalize the procedure. (i) −1 1 We put Et1i := ψ Hi (i = 1, 2), and Et0i := ∂ Et i . Note that c(t) and c (t) 0 1 0 are obtained from finite reductions Et 1 −→ Et 1 and Et 2 −→ Et12 respectively.
The Virtual Class of the Moduli Stack of Stable r-Spin Curves
33
(1) −1 1 (1) We put Eˆ t0 := ∂ (Et 2 ). Due to Lemma 4.15 and the graph-continuity of ∂ , 1 0 , and we have the Eˆ t0 is a subbundle of E m + r · δ i . It is also a subbundle of Et,2 following exact sequence: 0 −→ σi∗ F(m) −→ 0. 0 −→ Eˆ t0 −→ Et,2
(22)
Here σi denotes the i th section, and F(m) denotes the universal r-spin structure for the tuple m. The morphism E 0 −→ σi∗ F(m) gives a section ρ : E 0 −→ π ∗ σi∗ F(m). 1 . Let cˆ denote The construction in Subsubsect. 4.2.4 can be applied to Eˆ t0 −→ Et,2 the obtained cohomology class. By the same argument as the proof of Lemma 4.6, we can show that cˆ is the same as c (t) . 0 −→ E 1 be the Witten map. Let φˆ denote the restriction of φ to E ˆt . Let φ : Et,2 t,2 0 We denote the projections Et,2 and Eˆ t onto M by π and πˆ respectively. Let ι denote 0 . Then we have the following relation from the exact sequence the injection Eˆ t −→ Et,2 (22): 1 1 ), φˆ = ctop π ∗ σi∗ F, ρ · ctop π ∗ (Et,2 ), φ . ι∗ ctop πˆ ∗ (Et,2 Hence we obtain the following: 1 1 , φˆ = c1 σi∗ F · π∗ ctop π ∗ (Et,2 , φ) . πˆ ∗ ctop πˆ ∗ Et,2 It implies cˆ = ψ˜ i · c(t) . By noting the signature in Definition 4.3, we obtain the formula 1/r,m+r·δ i 1/r,m cg,n = −ψ˜ i (m) · cg,n . Thus we obtain Axiom 4. 4.4.4. Axiom 5. We consider the case that the nth point Pn is forgotten. Since we already know Axiom 4, we can assume that mi is less than r − 2 for each marked point. Let m denote a tuple such that mn = 0, and let m denote a tuple obtained from m by forgetting mn . Let M(m) and M(m ) denote the corresponding moduli stacks. i ∨i ¯ ∂¯ dual on M(m). We also We have the bundles E (m), E (m) and the operators ∂, i ∨i dual
on M(m ). Let πn : have the bundles E (m ), E (m ) and the operators ∂, ∂
th M(m) −→ M(m ) denote the morphism forgetting the n marked point. Let (C, p, F) be a point of M(m). The image of the point via πn is denoted by (C , p , F ). We have C = C or C = C ∪ P1 . We have the continuous map F(C,p,F ) : 0
0
E (C,p,F ) −→ E (C ,p ,F ), by forgetting the section on P1 . We have the continuous map 1
1
G(C,p,F ) : E (C ,p ,F ) −→ E (C,p,F ), by putting 0 on P1 . Then the families of the maps 0
1
F : E (m) −→ πn∗ E(m ) and G : πn∗ E (m ) −→ E 1 (m) are strongly continuous, due to Lemma 3.21. Lemma 4.16. We have the natural isomorphism of the L2 -cohomology groups of F˜ (resp. L) on C˜ and F˜ (resp. L ) on C˜ with the given weights (see Subsect. 3.4). Proof. Let us consider the case when the forgotten point Pn is a nodal point. It means that Pn is on the rational curve collapsed, which contains no other marked points and exactly two nodal points Q1 , Q2 . The indices m(Qi ) at the nodal points are less than r − 2, and thus one of the following holds: (i) m(Q1 ) + m(Q2 ) = r − 2, or (ii) m(Q1 ) = m(Q2 ) = −1. In both cases, the claims can be checked easily.
34
T. Mochizuki
Let us consider the case that the point Pn is on the marked point. It means that Pn is on a rational curve P1 which contains exactly one other marked point Pi and exactly one other nodal point Q. As in Case 1, it is easy to check the claims. We take families of 1,1 -hermitian metrics F and F ⊗ 0,1 on M(m). It induces i the family of the Hilbert norms N1 for E (m). We also take a family of 1,1 -hermitian i metrics for M(m ), which induces the family of the Hilbert norms for E (m ). It also i induces the family of the semi-norms N0 on E (m), as in Subsubsect. 3.3.9. We put ∨0 1 Ns := (1 − s) · N0 + s · N1 . It induces the anti-linear maps ψ (s) : E −→ E . ∨1 We take a finite dimensional vector subbundle H of the bundle E (m ) as in the proof ∨ 1 ∨1 of Lemma 4.1. Since we have the strongly continuous map πn∗ E (m ) −→ E (m), we can regard πn∗ H as the subbundle of E
∨1
(m). dual −1 (m), which is given by ∂ (H ) on M(m). On dual −1 ∨0
(H ) the other hand, we have a subbundle H2 of E (m ), which is given by ∂
on M(m ). We take a subbundle H2 of E
∨0
Lemma 4.17. The restriction of the strongly continuous map E to H2 gives an isomorphism of H2 and πn∗ (H2 ). Proof. It follows from Lemma 4.16.
∨0
∨0 (m) −→ πn∗ E (m )
1 0 −1 1 Es ⊂ E (m) for 0 ≤ s ≤ 1. We put Es1 := ψ (s) H2 ⊂ E (m) and Es0 := ∂ 0
1
On the other hand, we have a finite reduction E 0 −→ E 1 for E (m ) −→ E (m ) over M(m ), which is obtained from H . Lemma 4.18. We can naturally identify ∂ : E00 −→ E01 and ∂ : πn∗ E 0 −→ πn∗ E 1 , via 0
0
1
1
the morphisms E (m) −→ πn∗ E (m ) and πn∗ E (m ) −→ E (m). Proof. It follows from Lemma 4.16 and Lemma 4.17.
We have the Witten maps φ (s) : Es0 −→ Es1 (0 < s ≤ 1) given by φ (s) (u) := (s) ∂u + ρEs1 ψ (ur−1 ) as in Subsubsect. 4.2.3. By our construction, φ (s) is continuous with respect to the parameter s. Let φ denote the Witten map for the finite reduction E 0 −→ E 1 . Under the identifications πn∗ E 0 E00 and πn∗ E 1 E01 , πn∗ φ gives the map E00 −→ E01 , which we denote by φ (0) . Lemma 4.19. The family φ (s) 0 ≤ s ≤ 1 is continuous. Proof. We have only to check at s = 0. It can be easily seen by using the description given in Remark 4.5. The maps φ (s) (0 ≤ s ≤ 1) induce the virtual class of M(m). The map φ (0) = πn∗ φ induces the pull back of the virtual class of M(m ) via πn . Thus we can conclude that the pull back of the virtual class of M(m ) is the same as the virtual class of M(m), due to Lemma 4.19. Thus we obtain Axiom 5.
The Virtual Class of the Moduli Stack of Stable r-Spin Curves
35
4.4.5. Axiom 3. Let be a decorated stable graph of genus g with n-tails marked with the n-tuple m. We may assume that is connected. We consider the morphisms: 1/r
1/r
1/r
µ : M ×M M˜ −→ M ,
1/r
p : M ×M M˜ −→ M˜ .
We denote the universal curve over M and M˜ by C and C˜ respectively. We have the canonical map λ : p∗ C˜ −→ µ∗ C . Let F and F˜ denote the universal r-spin structures for C and C˜ . Then we have the canonical map : µ∗ F −→ λ∗ p ∗ F˜ . It is isomorphic outside the nodal points of Ramond type. On the nodal point of Ramond type, we have the matching condition. More precisely, let e denote the edge of which is of Ramond type. We have the cor1/r responding divisor De of µ∗ C , which is a nodal point of µ∗ C −→ M ×M M˜ . Then we have the line bundle Ie on De , and we can take a morphism λ∗ p ∗ F˜ −→ Ie such that the kernel of the morphism λ∗ p ∗ F˜ −→ e Ie is isomorphic to µ∗ F . 1/r
Let i : M
1/r,m
−→ Mg,n
i
be the natural morphism. Let E (m) and E
∨i
(m) be the
1/r,m bundle over Mg,n obtained from the universal r-spin structure as in Subsubsect. 3.4.2. 1/r,m The pull back of cg,n via the morphism i ◦ µ is obtained from a finite reduction for ∗ ∗ i ∗ ∗ ∨i
µ i E (m) and µ i E (m). (See Subsubsect. 4.2.7.) i ∨i 1/r,m ˜ and E ˜ (α) ˜ over M˜ . Here, the On the other hand, we have the bundles E ˜ (α) ˜ weight α˜ is given as follows: Let t be a tail of , which is obtained from a tail of , then ˜ which is obtained from an edge of , then we put we put α(t) ˜ = 0. Let t be a tail of , α(t) ˜ := (mt + 1), where mt denotes the index of the tail. We take the domains of ∂ 1/r,mv dual and ∂ as in Condition 4.2. We remark that E ˜ is a product of the stacks Mgv ,nv for ∨i ˜ and that E i˜ (α) ˜ and E ˜ (α) ˜ are direct sums of the pull backs of the the vertices v of , 1/r,mv
∨i
i
bundles of the topological vector spaces E v (α˜ v ) and E v (α˜ v ) over Mgv ,nv . 0
1
Let Ev0 −→ Ev1 be finite reductions of E v (α˜ v ) −→ E v (α˜ v ). Let E 0 −→ E 1 be a i ˜ and direct sum of the pull backs of Ev0 −→ Ev1 , which is a finite reduction for E ˜ (α) ∨i
˜ E ˜ (α). Lemma 4.20. Assume that all the edge of is of Neveu-Schwarz type. Then the induced 1/r cohomology class from E 0 −→ E 1 is the virtual class of M˜ . ˜ Then the claim follows from Proof. We remark that |(mi + 1)| < 1 for any tails of . Lemma 4.7. Let us consider the pull back p ∗ E 0 −→ p∗ E 1 . Then we have the canonical map −→ e Ie , which may be assumed to be surjective. Let Eˆ 0 denote the kernel.
p∗ E 0
0 1 Lemma 4.21. Eˆ 0 −→ p ∗ E 1 is a finite reduction of µ∗ i ∗ E −→ µ∗ i ∗ E . i
0
Proof. It is clear from our construction of E and the choice of the domains ∂ : E −→ 1 E .
36
T. Mochizuki
Let t be a tail of ˜ obtained from an edge of of Ramond type. Then we have the map Rest : E 0 −→ C, by taking the residue of s r at the marked points corresponding to the tail t. Let te,i (i = 1, 2) denote the tails of ˜ corresponding to the edge e of of Ramond type. Then we obtain the map e (s) := Reste,1 (s r ) + Reste,2 (s r ). Then we have the top Chern class ctop C, e ∈ H ∗ −1 E 0 for each edge e of Ramond type. e (0)
On the other hand, we have the map pE : Eˆ 0 −→ E 0 of the orbifolds. Then we have the cohomology classes pE ∗ (1) ∈ H ∗ ˆ 0 (E 0 ), where 1 denotes the unit element of pE (E )
the cohomology ring of Eˆ 0 . The following lemma is obtained from our construction by a geometric consideration.
Lemma 4.22. We have pE Eˆ 0 = e −1 e (0), and the equality e ctop C, e = pE ∗ (1) in H ∗ ˆ 0 E 0 . pE (E )
Proof. For simplicity of the explanation, we give the argument in the case where consists of one edge of Ramond type. The general case can be discussed similarly. Let dv denote the rank of Ev0 , and let U (dv ) denote the dvth unitary group. We put V := v P (Ev0 ), where P (Ev0 ) denote the U (dv )-principal bundles associated to Ev0 . Note their isotopy groups are trivial, i.e. P (Ev0 ) are the usual manifolds. Then we have 1/r the morphism V −→ M . Let Eˇ 0 denote the pull back of E 0 . We have two subsets pE (Eˆ 0 ) ×E 0 Eˇ 0 and −1 (0) ×E 0 Eˇ 0 . We have only to show the coincidence of them. Let x be any point of V . We have only to compare the fibers of pE (Eˆ 0 ) ×E 0 Eˇ 0 and −1 (0) ×E 0 Eˇ 0 over x. The point x ∈ V naturally gives a stable r-spin curve (C, p, F) ˜ Let Pe,i denote the points corresponding to the tails te,i . Then we have the of type . ⊗r canonical isomorphism F|P C given by the residue. They induce the isomorphism e,i ρ : F|Pe,1 F|Pe,2 . 1/r 1/r M ˜ over M ˜ . Let pˇ denote Let Vˇ denote the fiber product of V and M × M
−1 ˇ to the the natural morphism V −→ V⊗. rThen the fiber pˇ (x) is naturally 0isomorphic set I := κ : F|Pe,1 F|Pe,2 κ = −ρ . Hence the fiber of pE (Eˆ ) ×E 0 Eˇ 0 over x is as follows: " 0 0 −1 s ∈ E(C, p,F ) κ(s|P1 ) = s|P2 = s ∈ E(C,p,F ) s ∈ e (0) . κ∈I
ˇ0 Hence we obtain −1 (0)×E 0 Eˇ 0 = pE (Eˆ 0 )× E 0 E . We obtain the equality of the cohomology classes, by comparing the induced v U (dv ) equivariant cohomology classes. (See the Appendix.) Let φ : E 0 −→ E 1 denote the Witten map as in Subsubsect. 4.2.3. Let π : E 0 −→ denote the projection. The maps φ and e induce the section φ˜ of π ∗ E 1 ⊕ ! ∗ 1 ˜ e C. Then we obtain the top Chern class ctop π E ⊕ e C, φ = e ctop (C, e ) · ∗ 1 ctop π E , φ . On the other hand, let φˆ denote the restriction of p∗ φ to Eˆ 0 . Let πEˆ : Eˆ 0 −→ 1/r,m M × M ˜ be the projection. Then we obtain the top Chern class ctop π ∗ p ∗ E 1 , φˆ . 1/r M˜
M
Eˆ
The Virtual Class of the Moduli Stack of Stable r-Spin Curves
37
Lemma 4.23. We have the following equality: ! # ! pE ∗ ctop πE∗ˆ p ∗ E 1 , φˆ = ctop π ∗ E 1 ⊕ C, φ˜ . e
Proof. It follows from Lemma 4.22 and the following equality: ∗ ∗ 1 pE ∗ ctop πE∗ˆ p ∗ E 1 , φˆ = pE ∗ ctop pE π E , φˆ = pE ∗ (1) · ctop π ∗ E 1 , φ . Note we have the equality p ◦ πEˆ = π ◦ pE .
1/r,m
Corollary 4.1. We have p∗ µ∗ c = 0, if one of the edge of is of Ramond type. Namely, Axiom 3 holds in the Ramond case. Proof. Due to Lemma 4.23, we have only to show the vanishing of π∗ ctop π ∗ E 1 ⊕ e C, φ , which follows from Lemma 4.12. Let us consider the case that all of the edge of is of Neveu-Schwarz type. Then 0 1 p∗ E 0 −→ p ∗ E 1 is a finite reduction of µ∗ i ∗ E (m) −→ µ∗ i ∗ E (m). Hence the induced 1/r, m 1/r,m cohomology class is the same as µ∗ i ∗ cg,n . Then we obtain the equality µ∗ c = ∗ 1/r,m · p r/ l c , due to Lemma 4.20. Then Axiom 3 in the Neveu-Schwarz e e∈E() ˜ case immediately follows. Therefore the proof of Theorem 4.1 is accomplished. 5. Appendix 5.1. Locally trivial orbispaces. In the following, orbispaces are assumed to be paracompact. For an orbispace V , [V ] denote the associated topological space. Let V be an orbispace and E be a vector bundle over V . Once we take a hermitian metric of E, we obtain the U (rE )-principal bundle P (E), where rE denotes the rank of E. We only consider the rational cohomology groups. Remark 5.1. We do not intend to develop the general theory.
Definition 5.1. Let V be an orbispace. It is called locally trivial, if the isotopy groups of any points are finite abelian group and independent of a choice of a point up to isomorphisms. Lemma 5.1. If V is a locally trivial orbispace, then there exist a topological space V , a finite abelian group A and a homomorphism ρ : π1 (V ) −→ Aut (A) such that V U /Tρ . Here Aut (A) denotes the automorphism group of A, and Tρ denotes the torsor associated to ρ. Let E be a complex vector bundle over a locally trivial orbispace V , then we have the vector bundle E over V and the Tρ -action on E . We do not distinguish (V , E ) and (V , E) in the sequel, when V is a locally trivial orbispace. Let V be a locally trivial orbispace, and E be a vector bundle over V . In that case, the following lemma can be checked easily. Lemma 5.2. P (E) is locally trivial.
38
T. Mochizuki
5.2. Locally trivializable orbispaces. Definition 5.2. • An orbispace V is called locally trivializable, if there is a pair (L, P ) of a compact Lie group L and a principal L-bundle P over V such that P is locally trivial. • Let E be a vector bundle on V . If the associated unitary principal bundle P (E) is locally trivial as orbispaces, then V is called s-locally trivializable. If V is a locally trivializable orbispace, it is convenient to consider the Borel construc tion B(P , L) := P ×L EL and the equivariant cohomology HL∗ (P ) := H ∗ B(P , L) . Here (L, P ) denotes a pair of compact Lie group L and a principal L-bundle over V in Definition 5.2, and EL denotes the universal bundle over a classifying space BL of L. over For example, a vector bundle E over V induces the L-equivariant vector bundle E P . Hence we have the equivariant Chern class in HL∗ (P ). Since we consider the rational cohomology groups, we have the isomorphism HL∗ (P ) H ∗ ([V ]). If (Li , Pi ) are pairs of compact Lie groups Li and Li -principal bundles Pi over V as in Definition 5.2, we have the L1 × L2 -principal bundle P1 × V P2 which is a locally trivial orbispace. We have the natural isomorphisms: HL∗1 P1 HL∗1 ×L2 (P1 ×V P2 ) HL∗2 (P2 ). In this sense, the group HL∗ (P ) is independent of a choice of (L, P ), and hence we often denote it by H ∗ (V ), and the Chern class is given in this group. Similarly, we can consider the relative cohomology groups H ∗ (V , S) (S ⊂ V ) and the cohomology groups with support HS∗ (V ). 1/r
Lemma 5.3. For any decorated graph , the orbispace M is locally trivializable. 1/r,m
Proof. The claim can be reduced to the case Mg,n . Since it is smooth, we have the 1/r,m
principal bundle associated with the tangent bundle T Mg,n . Since the isotopy group of a general point is finite abelian, the claim immediately follows.
5.3. Pull back. Let f : V1 −→ V2 be a morphism of s-locally trivializable orbispaces. Then we have vector bundles Ei over Vi such that P (Ei ) are locally trivial. We have the following morphisms: P (E1 ) ←−−−− P (E1 ) ×V1 f ∗ P (E2 ) −−−−→ P (E2 ). Note that P (E1 ) ×V1 f ∗ P (E2 ) is locally trivial. Hence we have the naturally defined morphism:
HU∗ (dE
P (E2 ) −→ HU∗ (dE
2)
1 )×U (dE2
P (E1 ) ↓ ∗ ) P (E1 ) ×V1 f P (E2 )
HU∗ (dE
1)
In other words, we have the naturally defined morphism H ∗ (V2 ) −→ H ∗ (V1 ).
The Virtual Class of the Moduli Stack of Stable r-Spin Curves
39
5.4. Thom isomorphism. Let V be an s-locally trivializable orbispace, and P (E1 ) is locally trivial for a vector bundle E1 over V . Let E be a vector bundle over V . Then P (E1 ) ×V P (E) and Eˆ := P (E1 ) ×V P (E) ×V E also locally trivial, and the natural projection Eˆ −→ P (E1 ) ×V P (E) is a vector bundle. Then we have the Thom isomor ∗+2rE ˆ L) H ∗ B(P (E1 ) ×V P (E), L) , where L denotes the group phism Hcpt B(E, ∗+2rE U (dE1 ) × U (dE ). In other words, we have the Thom isomorphism π∗ : Hcpt (E) ∗ H (V ). Similarly, we obtain the Gysin morphism by using the Borel construction. The functorial properties can be easily reduced to the ordinary case. 5.5. The top Chern class. Let U (r) be the r th unitary group, and p : EU (r) −→ BU (r) denote the universal U (r)-principal bundle. We have the associated universal vector bundle E over BU (r). Then we have the universal section s : EU (r) −→ p ∗ E, whose 0-set is BU (r). Let E be a vector bundle over a locally trivializable orbispace V with section φ. Let (L, P ) be a pair of compact Lie group and P be a L-principal bundle over V as over B(P , L) with the in Definition 5.2. Then we have the induced vector bundle E induced section φ . We can take a continuous map : B(P , L) −→ E such that E ∗ ∗ ∗ ∗ p E and φ = s under the identification. Hence we obtain the map HBU (r) (E) −→ ∗ Hφ∗−1 (0) B(P , L) Hφ∗−1 (0) V . The image of the Thom class in HBU (r) E is called the top Chern class, and it isdenoted by ctop (E, φ). Standard properties ctop E1 ⊕ E1 , φ1 ⊕ φ2 = ctop E1 , φ1 · ctop E2 , φ2 and f ∗ ctop E, φ = ctop f ∗ E, f ∗ φ can be shown by reducing to the ordinary case.
Acknowledgement. The author is heartily grateful to Professor T. Jarvis for his interest, comments and encouragement. He also acknowledges Professors T. Kimura and Y. P. Lee for their kindness. He thanks the referees for their readings. The author wrote the original version of this paper at Osaka City University, the revision at Kyoto University, and the final version at Max-Planck Institute for Mathematics. The author expresses his gratitude to the institutions.
References 1. 2. 3. 4. 5. 6. 7. 8. 9. 10.
Chen, W.: A Homotopy theory of orbispaces, http://arxiv.org/list/math.AT/0102020, 2001 Fujita, H., Kuroda, S.T.: Functional Analysis I (in Japanese). Tokyo: Iwanami Shoten, 1983 Jarvis, T.: Geometry of the moduli of higher spin curves. Internat. J. Math. 11, 637–663 (2000) Jarvis, T.: Torsion-free sheaves and moduli of generalized spin curves. Composito Math. 110: 291– 333 (1998) Jarvis, T.: The Picard group of the moduli of higher spin curves. New York J. Math. 7, 23–47 (2001) Jarvis, T., Kimura, T., Vaintrob, A.: Moduli spaces of higher spin curves and integrable hierarchies. Composito Math. 126, 157–212 (2001) Jarvis, T., Kimura, T., Vaintrob, A.: Gravitational descendents and the moduli space of higher spin curves. In: Advances in algebraic geometry motivated by physics (Lowell, MA, 2000). Contemp. Math. 276, pp. 167–177, 2001 Jarvis, T., Kimura, T., Vaintrob, A.: Spin Gromov-Witten Invariants. Commun. Math. Phys. 259, no. 3, 511–543 (2005) Polishchuk, A.: Witten’s top Chern class on the moduli space of higher spin curves. In : Frobenius manifolds, Aspects Math. Wiesbaden: Vieweg, E36, pp. 253–264, 2004 Polishchuk, A., Vaintrob, A.: Algebraic construction of Witten’s top Chern class. In: Advances in algebraic geometry motivated by physics (Lowell, MA, 2000). Contemp. Math. 276, pp. 229–249, 2001
40
T. Mochizuki
11. Seely, R., Singer, I.M.: Extending ∂¯ to Singular Riemann Surfaces. J. Geom. Phys. 5, 121–136 (1989) 12. Witten, E., Two dimensional gravity and intersection theory on the moduli space. Surveys in Diff. Geom. 1, 243–310 (1991) 13. Witten, E.: Algebraic geometry associated with matrix models of two dimensional gravity. In: Topological methods in modern mathematics (Stony Brook, NY, 1991), Houston, TX: Publish or Perish, pp. 235–269, 1993 Communicated by N.A. Nekrasov
Commun. Math. Phys. 264, 41–69 (2006) Digital Object Identifier (DOI) 10.1007/s00220-005-1501-8
Communications in
Mathematical Physics
Nonassociative Tori and Applications to T-Duality Peter Bouwknegt1 , Keith Hannabuss2 , Varghese Mathai3 1
Department of Theoretical Physics, Research School of Physical Sciences and Engineering, and Department of Mathematics, Mathematical Sciences Institute, The Australian National University, Canberra, ACT 0200, Australia. E-mail:
[email protected] 2 Balliol College, Oxford OX1 3BJ, England. E-mail:
[email protected] 3 Department of Pure Mathematics, University of Adelaide, Adelaide SA 5005, Australia. E-mail:
[email protected] Received: 17 December 2004 / Accepted: 19 September 2005 Published online: 9 March 2006 – © Springer-Verlag 2006
Abstract: In this paper, we initiate the study of C ∗ -algebras A endowed with a twisted action of a locally compact abelian Lie group G, and we construct a twisted crossed product A G, which is in general a nonassociative, noncommutative, algebra. The duality properties of this twisted crossed product algebra are studied in detail, and are applied to T-duality in Type II string theory to obtain the T-dual of a general principal torus bundle with general H-flux, which we will argue to be a bundle of noncommutative, nonassociative tori. Nonassociativity is interpreted in the context of monoidal categories of modules. We also show that this construction of the T-dual includes the other special cases already analysed in a series of papers. 1. Introduction Recent work has revealed the strong connections between T-duality in string theory and Takai duality for C ∗ -algebras (as for instance discussed in the introduction to [3, 28]), but for general H-fluxes C ∗ -algebras are no longer adequate. In this paper we present a generalisation which permits a very precise description of the general T-dual. Let T be a compact connected Abelian Lie group of rank with Lie algebra t and let t be the dual of t. Let E → M be a principal T-bundle with connection. By the Chern-Weil construction the space (E)T of T-invariant forms on E is isomorphic to the space of forms on M with values in ∧ t, i.e. k (E)T ∼ =
k
p (M, ∧k−p t) ,
(1.1)
p=0
and by a classical result of Chevalley and Koszul, the de-Rham complex (• (E), d) is chain homotopy equivalent to the complex (• (M, ∧• t), D) with a modified de-Rham differential D, and hence the associated cohomologies are isomorphic. Furthermore, we
42
P. Bouwknegt, K. Hannabuss, V. Mathai
know that any class in H k (E) can be represented by a form in k (E)T . Thus, under this isomorphism, an H -flux in H 3 (E) can be considered as a 4-tuple H = (H3 , H2 , H1 , H0 ) with Hp ∈ p (M, ∧3−p t) for p = 0, 1, 2, 3, closed under the action of D (cf. [18, 5] for more details).1 T-duality for principal circle bundles was treated geometrically in [2, 3], and dimensional considerations force the H0 and H1 components of the H-flux to vanish in this case. The T-dual turns out to be another principal circle bundle with T-dual H-flux. The arguments were extended in [4] to principal T -bundles with H-flux satisfying the condition that the H0 and H1 components vanish. Then the T-dual turns out to be another principal T -bundle with T-dual H-flux having vanishing H0 and H1 components. The analysis in [28, 29] shows that if one considers principal T -bundles with H-flux satisfying the condition that just the H0 component vanishes, then one arrives at the surprising conclusion that the T-dual bundle has to have noncommutative tori as fibres, provided the H1 component is non-zero. The weaker condition in [28, 29] permits non-vanishing H1 , but still excludes non-zero H0 . In this paper we shall remove the last of these constraints, to allow a non-vanishing H0 component. In this case, we arrive at the astonishing conclusion that the T-dual bundle has to have nonassociative tori as fibres, taking it even beyond the normal range of noncommutative geometry. The key step is provided by a new explicit construction of a continuous trace algebra B having a given Dixmier-Douady invariant, whose spectrum is the total space E of the principal T-bundle, together with automorphisms βg for g ∈ G = t, which transform the spectrum in a way compatible with the G-action on E. The new features arise because in general g → βg is not a homomorphism but satisfies βx βy = ad(v(x, y))βxy , where ad(v) denotes conjugation by a unitary element of the multiplier algebra MB. One expects the algebra associated with the T-dual to be the crossed product B β G, but the twisting forces us to take a suitably twisted crossed product B β,v G. Such (Leptin-)Busby-Smith twistings have long been known but, for non-trivial H0 , v(x, y) is not a cocycle and that means that associativity fails in the twisted crossed product. In Sect. 2 we give the relationship between the differential forms and the multicharacters on G which will be used in our later constructions. Section 3 reviews the generalised Busby-Smith twisted crossed products. An example of a twisting is given in Sect. 4, together with a proof that it is the only type up to stability. This is followed by an example of a nonassociative generalisation of the compact operators. The theory of twisted induced representations is developed in Sect. 6 and then used to construct examples of algebras with given spectrum and Dixmier–Douady class in Sect. 7. In Sect. 8 it is shown that the twisted crossed product of a twisted induced algebra is isomorphic to a generalisation of the twisted compact operators. In Sect. 9, the double dual is shown to be the tensor product of the original algebra with the twisted compact operators, that is, Morita equivalent in this category to the original algebra. In Sect. 10, the mathematical results of the previous sections are used to justify the assertion that the T-dual to a general principal torus bundle with H-flux, is a bundle of nonassociative tori. The final section outlines how associativity can be restored by working in a different category, an idea which will be explored in more detail in a subsequent paper. 1 The conclusions in this paper are valid for integral classes H ∈ H 3 (E, Z) as well, since this paper deals exclusively with the introduction of the additional ‘degree of freedom’ H0 , which does not carry torsion, on top of established results which hold in the case of torsion H . Note that H0 can be identified with the restriction of H to a fibre. For simplicity we have chosen to formulate some of the results in terms of differential forms.
Nonassociative Tori and Applications to T-Duality
43
2. Differential Forms and Multicharacters Let T be a compact connected Abelian Lie group of rank , and E → M a principal T-bundle. In essence the action of T on E provides a map from the Lie algebra t to vector fields on E, which we write X → ξX , and then a p-form f ∈ p (E)T in the fibre directions defines the antisymmetric multilinear form valued function on M, (ξ ∗ f )(X1 , X2 , . . . , Xp ) = f (ξX1 , ξX2 , . . . , ξXp ).
(2.1)
For abelian Lie groups the form exponentiates to a multicharacter on G = t, φ exp(X1 ), . . . , exp(Xp ) = exp(−2πi(ξ ∗ f )(X1 , . . . , Xp )).
(2.2)
(A multicharacter is a character in each variable, and this property follows from the additivity of ξ ∗ f , and φ is also antisymmetric in the sense that even permutations of its variables leave it unchanged whilst odd permutations invert it.) The multicharacter property ensures that this is always a (Moore) cocycle in Z p (t, T), since, for example when p = 3, φ(y, z, w)φ(x, yz, w)φ(x, y, z) = φ(y, z, w)φ(x, y, w)φ(x, z, w)φ(x, y, z) = φ(xy, z, w)φ(x, y, zw). (It is known that every cohomology class in H 2 (Rn , T) can be represented by an antisymmetric bicharacter of the form φ [23, 19], and for p = 3 it is certainly true that smooth cocycles are cohomologous to antisymmetric tricharacters.) Example. Consider the torus bundle T3 over a point, with H0 the class defined by k times the volume form dx 1 ∧dx 2 ∧dx 3 . The associated antisymmetric form on a, b, c ∈ t = R3 is then given by f (a, b, c) = k[a, b, c] ≡ ka.(b × c),
(2.3)
whence φ(a, b, c) = exp(−2kπ i[a, b, c]). Although we are mainly interested in the abelian groups T = Tn , G = t = Rn , and N∼ = Zn the kernel of the exponential map t → T, the constructions which we present in Chapters 3 to 8 are valid for general unimodular separable locally compact groups with a tricharacter φ. For that reason we shall write the group composition multiplicatively, and be careful not to commute terms. However, this generalisation is less sweeping than may appear because the tricharacter φ on G defines a homomorphism of each variable into the abelian group T, and so must be lifted from a tricharacter on the abelianisation G/[G, G]. Finally we note that, by definition, the Dixmier-Douady class has components which are integral 3-forms, and that means, in particular, that the tricharacter φ constructed from H0 is identically 1 on N × N × N, where N is the kernel of the exponential map. We shall assume this to be true in the general case.
44
P. Bouwknegt, K. Hannabuss, V. Mathai
3. Generalised Busby-Smith Twisted Crossed Products As noted above we now work with a general unimodular separable locally compact group G and a closed subgroup N on which the tricharacter φ has trivial restriction. For any group G one can interpret H 2 (G, A) as classifying central extensions of G by A, whilst, for p > 3, H p (G, A) is usually interpreted in terms of crossed modules [15, 16, 6, 20, 21, 34, 26], with elements of H 3 (G, A) known as MacLane–Whitehead obstructions [35, 27]. However, we shall see that these classes also arise in a C ∗ -algebraic context. The H -field is usually linked to the equivariant Brauer group of a continuous trace C∗ -algebra A with spectrum E, on which a group G acts as automorphisms so that the dual action E agrees with the bundle structure. However, the equivariant Brauer group can be described entirely in terms of cohomology classes in H p (M, H 3−p (G, T)) for p = 1, 2, 3, [11], leaving no room for H -fields with a component in H 0 (M, H 3 (G, T)), (for example, any non-trivial H -field on G considered as a principal G-bundle over a point). Since the representatives of H 0 are locally constant functions we shall concentrate our attention on H 3 (G, T). This suggests that we must consider a wider class of algebras or actions. In fact, inner automorphisms automatically act trivially on the spectrum, so that only homomorphisms of G to the outer automorphisms Out(A) = Aut(A)/Inn(A) are interesting. However, to work with these one needs a lifting α : G → Aut(A). The problem then is that αx αy and αxy can differ by an inner automorphism ad(u(x, y)) : a → u(x, y)au(x, y)−1 , that is αx αy = ad(u(x, y))αxy . We can take u(x, y) = 1 whenever x or y is the identity. This is almost precisely the data needed to define a Busby–Smith (or Leptin) twisted crossed product A α,u G of A and G, [24, 7, 32]. Assuming that u is a measurable function on G×G, we can define a twisted convolution product and adjoint on C0 (G, A) by (f ∗ g)(x) =
f (y)αy [g(y −1 x)]u(y, y −1 x) dy,
G
f ∗ (x) = u(x, x −1 )−1 αx [f (x −1 )]∗ , and complete this to get a new algebra. The link with the algebraists’ picture of H 3 (G, A) arises because u is no longer a cocycle since the condition linking α and u tells us only that the adjoint actions of u(x, y)u(xy, z) and αx [u(y, z)]u(x, yz) coincide, so that one has a modified cocycle condition: φ(x, y, z)u(x, y)u(xy, z) = αx [u(y, z)]u(x, yz)
(3.1)
for some central unitary element φ(x, y, z) ∈ U Z(A). It is easy to check that φ is a cocycle defining an element of H 3 (G, U Z(A)). (Essentially the same argument is used in [8] to explain the origin of the Gauss anomaly and Jackiw’s nonassociative anomaly in quantum field theory.) When φ is a tricharacter one can still form a twisted crossed product. Proposition 3.1. When φ defined as above is an antisymmetric tricharacter the twisted crossed product A α,u G satisfies the ∗-algebra identity (f ∗ g)∗ = g ∗ ∗ f ∗ , and is
Nonassociative Tori and Applications to T-Duality
45
associative if and only if φ ≡ 1. In fact, we have ((f ∗ g) ∗ h)(x) = f (z)αz [g(z−1 y)]αy [h(y −1 x)]u(z, z−1 y)u(y, y −1 x) dydz, G×G f (z)αz [g(z−1 y)]αy [h(y −1 x)]φ(z, z−1 y, y −1 x) (f ∗ (g ∗ h))(x) = G×G
×u(z, z−1 y)u(y, y −1 x) dydz. Proof. Using the modified cocycle identity, one calculates that (f ∗ g)∗ (x) = φ(x, x −1 y, y −1 )∗ u(y, y −1 )∗ αy [g(y −1 )]∗ G
×u(x, x −1 y)∗ αx [f (x −1 y)]∗ dy, whilst ∗
∗
(3.2)
φ(y, y −1 x, x −1 y)∗ u(y, y −1 )∗ αy [g(y −1 )]∗
(g ∗ f )(x) = G
×u(x, x −1 y)∗ αx [f (x −1 y)]∗ dy,
(3.3)
and for antisymmetric tricharacters both factors involving φ are 1. The twisted crossed product algebra A α,u G has ((f ∗ g) ∗ h)(x) = (f ∗ g)(y)αy [h(y −1 x)]u(y, y −1 x) dy, G = f (z)αz [g(z−1 y)]u(z, z−1 y)αy [h(y −1 x)]u(y, y −1 x) dydz G×G f (z)αz [g(z−1 y)]αz αz−1 y [h(y −1 x)]u(z, z−1 y) = G×G
×u(y, y −1 x) dydz
= G×G
f (z)αz [g(z−1 y)]αz−1 y [h(y −1 x)]u(z, z−1 y)
×u(y, y −1 x) dydz, and, using the modified cocycle identity, (f ∗ (g ∗ h))(x) = f (z)αz [(g ∗ h)(z−1 x)]u(z, z−1 x) dz G = f (z)αz [g(z−1 y)]αz−1 y [h(y −1 x)]u(z−1 y, y −1 x) G×G
×u(z, z−1 y) dydz,
= G×G
f (z)αz [g(z−1 y)]αz−1 y [h(y −1 x)]αz [u(z−1 y, y −1 x)]
×u(z, z−1 y) dydz
= G×G
f (z)αz [g(z−1 y)]αz−1 y [h(y −1 x)]φ(z, z−1 y, y −1 x)
×u(z, z−1 y)u(y, y −1 x) dydz,
46
P. Bouwknegt, K. Hannabuss, V. Mathai
so that the Busby-Smith twisted crossed product is nonassociative except in the case φ ≡ 1. Henceforth we shall always take φ to be an antisymmetric tricharacter. Usually Busby-Smith products are only defined when φ = 1, but we shall see that much of the theory goes through without that assumption, so that this provides a means of constructing nonassociative from associative algebras. The nonassociativity becomes even more transparent when one considers a covariant representation (U, π ) of (G, A) satisfying the conditions U (x)π(a)U (x)−1 = π(αx (a)),
U (x)U (y) = π(u(x, y))U (xy).
(3.4)
These give U (x)[U (y)U (z)] = U (x)[π(u(y, z))]U (yz) = π(u(x, yz))π(αx u(y, z))U (x(yz)),
[U (x)U (y)]U (z) = π(u(x, y))U (xy)U (z) = π(u(x, y))π(u(xy, z))U ((xy)z),
so that φ(x, y, z)U (x)[U (y)U (z)] = [U (x)U (y)]U (z).
(3.5)
In fact, H 3 (G) was already interpreted as defining a nonassociative structure in [16], and this has resurfaced in the physics literature [22, 13]. 4. Generalised Packer–Raeburn Stabilisation For a given antisymmetric tricharacter φ on G there is a simple example of an algebra with twisting on which G acts as automorphisms. It is derived from the imprimitivity algebra generated by multiplication and translation operators on L2 (G). The right regular representation ρ acts on ψ ∈ L2 (G) by (ρ(x)ψ)(v) = ψ(vx), and we define (uρ (y, z)ψ)(v) = φ(v, y, z)ψ(v).
(4.1)
(We shall often be interested in the case when G is a closed subgroup of a group H, with φ defined on H × H × H and trivial on G × G × G. Then uρ (y, z) can be defined as a multiplication operator on L2 (G) for general y, z ∈ H, and the restriction of uρ to G × G is identically 1.) Now φ(x, y, z)(uρ (x, y)uρ (xy, z)uρ (x, yz)−1 ψ)(v) = φ(x, y, z)φ(v, x, y)φ(v, xy, z)φ(v, x, yz)−1 ψ(v) = φ(vx, y, z)ψ(v)
Nonassociative Tori and Applications to T-Duality
47
and (ρ(x)uρ (y, z)ρ(x)−1 ψ)(v) = (uρ (y, z)ρ(x)−1 ψ)(vx) = φ(vx, y, z)(ρ(x)−1 ψ)(vx) = φ(vx, y, z)ψ(v),
(4.2)
so that, setting αx = ad(ρ(x)), we get φ(x, y, z)uρ (x, y)uρ (xy, z)uρ (x, yz)−1 = αx [uρ (y, z)]. This gives an explicit realisation of an algebra C = C0 (G) with an action of G by automorphisms and with the appropriate Busby-Smith obstruction. (When G is non-compact uρ (x, y) is in the multiplier algebra rather than the algebra itself.) The same idea can be extended to the right regular σ -representation on L2 (G) given by (ρ(x)ψ)(v) = σ (v, x)ψ(vx) for any borel multiplier σ , and this makes no difference to the cocycle identity. Naturally, one can also work with the left regular representation (λ(x)ψ)(v) = ψ(x −1 v), and (uλ (y, z)ψ)(v) = φ(v, y, z)−1 ψ(v).
(4.3)
This is useful because it links directly to the formulation used by [30] to show that twisted crossed products defined by cocycles are stably equivalent to normal crossed products. In our case u is not a cocycle, but there is nonetheless a nice generalisation of the Packer–Raeburn Theorem. Theorem 4.1. Let A, G, α, u be as above. There exists a strongly continuous action β of G on A ⊗ K(L2 (G)) and a twisting uλ such that (β, uλ ) are exterior equivalent to (α ⊗ id, u ⊗ 1), that is, there exists vs = (1 ⊗ λs )(id ⊗ M(u(s, ·))∗ ) such that βs = ad(vs )(αs ⊗ id),
∗ id ⊗ uλ (s, t) = vs αs (vt )(u(s, t) ⊗ 1)vst .
(4.4)
Proof. The proof of Theorem 3.4 in [30] is still valid as far as the last line on p. 301. At that point the original argument uses the fact that u is a cocycle to show that it has been untwisted by the exterior equivalence. In our case u satisfies a modified cocycle identity, so that that last line (in the original notation) gives φ(s, t, r)−1 ξ(r) = (uλ (s, t)ξ )(r). One similarly obtains: Corollary 4.2. In the situation of Theorem 4.1 one has (A α,u G) ⊗ K(L2(G)) ∼ = (A ⊗ K) β,uλ G .
(4.5)
The Packer-Raeburn stabilisation trick is used to show that up to stabilisation by tensoring with compact operators A α,u G and A β,uλ G are isomorphic. The original idea derived from Quigg’s generalisation of Takai duality for twisted crossed products, and their β is just equivalent to Quigg’s double dual α . One may therefore obtain a duality theorem by the same procedure. We postpone discussion of this until Sect. 9 where we shall give a much more detailed account of duality for abelian groups.
48
P. Bouwknegt, K. Hannabuss, V. Mathai
5. Twisted Compact Operators Before developing the theory further it is useful to give a very simple example of a nonassociative algebra, obtained by twisting the algebra of compact operators on L2 (G) using the factor φ(x, y, z). We start with the Hilbert-Schmidt operators, realised as kernels K(x, y) for x, y ∈ G, with the involution K ∗ (x, y) = K(y, x), and norm 2 |K(x, y)|2 dxdy, (5.1) KH S = G×G
and define the new multiplication
(K1 ∗ K2 )(x, z) =
φ(x, y, z)K1 (x, y)K2 (y, z) dy.
(5.2)
G
This is consistent with the involution because (K1 ∗ K2 )∗ (x, z) = (K1 ∗ K2 )(z, x) = φ(z, y, x)−1 K1 (z, y) K2 (y, x) dy G = φ(x, y, z)K2∗ (x, y)K1∗ (y, z) dy G
= (K2∗ ∗ K1∗ )(x, z). The fact that φ(x, y, x) = 1 also means that we still have K2H S = (K ∗ ∗ K)(x, x) dx.
(5.3)
G
As usual, one can define a C ∗ -norm using the left regular representation K1 = sup(K1 ∗ K2 H S /K2 H S ). By using the Cauchy-Schwarz inequality and by considering rank one projections one sees that this is equivalent to the ordinary operator norm. The twisted compact operators Kφ (L2 (G)) are the completion of the Hilbert-Schmidt operators with respect to that norm. Unfortunately the new multiplication is not associative unless φ(x, y, z)φ(x, z, w) = φ(x, y, w)φ(y, z, w),
(5.4)
for all x, y, z, w ∈ G. In that case φ(x, y, z) = φ(x, y, w)φ(y, z, w)/φ(x, z, w), but conversely, whenever φ(x, y, z) has the form ψ(x, y)ψ(y, z)/ψ(x, z), for some function ψ, the algebra is associative. The twisted multiplication of kernels was used in [9] to study the hyperbolic quantum Hall effect, but in that two-dimensional situation the multiplication is automatically associative. Once one gets to three dimensions the analogous algebra is nonassociative. Proposition 5.1. The group G acts on the twisted algebra Kφ (L2 (G)) with multiplication φ(x, y, z)K1 (x, y)K2 (y, z) dy (5.5) (K1 ∗ K2 )(x, z) = G
by natural ∗-automorphisms θx [K](z, w) = φ(x, z, w)K(zx, wx),
(5.6)
and θx θy = ad(σ (x, y))θxy , where ad(σ (x, y))[K](z, w) = φ(x, y, z)φ(x, y, w)−1 K(z, w) comes from the multiplier σ (x, y)(v) = φ(x, y, v).
Nonassociative Tori and Applications to T-Duality
49
Proof. Using the tricharacter property of φ, we see that (θx [K1 ]∗θx [K2 ])(z, w) = φ(z, v, w)φ(x, z, v)φ(x, v, w)K1 (zx, vx)K2 (vx, wx)dv G = φ(zx, vx, wx)φ(x, z, w)K1 (zx, vx)K2 (vx, wx) dv G
= φ(x, z, w)(K1 ∗ K2 )(zx, wx), = θx [K1 ∗ K2 ](z, w). We also have θx [K]∗ (z, w) = θx [K](w, z) = φ(x, z, w)K(w, z) = φ(x, w, z)K ∗ (w, z) = θx [K ∗ ](z, w).
(5.7)
Moreover, θx θy [K](z, w) = φ(x, z, w)(θy [K])(zx, wx) = φ(x, z, w)φ(y, zx, wx)K(zxy, wxy) = φ(x, y, z)φ(x, y, w)−1 θxy [K](z, w) = ad(σ (x, y))[θxy [K]](z, w), with σ (x, y)(z, w) = φ(x, y, z)δ(z − w).
Note. There is also a left-handed version of this which uses the multiplication φ(x, y, z)−1 K1 (x, y)K2 (y, z) dy (K1 ∗ K2 )(x, z) =
(5.8)
G
and automorphisms τx [K](z, w) = φ(x, z, w)K(x −1 z, x −1 w),
(5.9)
and it is this version which will appear later, in Sect. 9. Some further automorphisms of the original algebra Kφ (L2 (G)) given by γx [K](z, w) = φ(x, z, w)−1 K(x −1 z, x −1 w)
(5.10)
will be useful in Sect. 8. When G is a contractible group the twisted compact operators are just a deformation of the usual ones. Proposition 5.2. When G is a contractible group Kφ (L2 (G)) is a continuous deformation of K(L2 (G)). Proof. Let {t : t ∈ [0, 1]} give a contraction of G onto the identity, that is t : G → G is continuous and satisfies 0 (x) = x and 1 (x) is the identity for all x ∈ G. We then define (for the right-handed version) φ(x, t (y), z)K1 (x, y)K2 (y, z) dy, (5.11) (K1 ∗t K2 )(x, z) = G
so that at t = 0 we have the twisted and at t = 1 the untwisted product.
50
P. Bouwknegt, K. Hannabuss, V. Mathai
6. Twisted Induced Algebras The introduction of u and φ has far-reaching consequences, because almost all the standard procedures have to be deformed, and we shall now investigate these in more detail. Suppose that A is a C∗ -algebra on which C(M) acts as double centralisers and the subgroup N of G acts by automorphisms αr , (r ∈ N), and that for each s, t ∈ N there are unitaries u(s, t) in the multiplier algebra M(A) such that αs αt = ad(u(s, t))αst , and also the modified cocycle condition (for a specified continuous tricharacter φ) αr [u(s, t)]u(r, st) = φ(r, s, t)u(r, s)u(rs, t).
(6.1)
Such algebras always exist, since we can take the algebra freely generated by a collection of symbols {u(s, t) : s, t ∈ N} and define the automorphism αr by the formula αr [u(s, t)] = φ(r, s, t)u(r, s)u(rs, t)u(r, st)−1 .
(6.2)
We shall suppose that u and φ extend to continuous functions on G satisfying the same relations: αr [u(x, y)]u(r, xy) = φ(r, x, y)u(r, x)u(rx, y),
(6.3)
for r ∈ N and x, y ∈ G. (We are mainly interested in the case when N is a maximal rank lattice in a vector group G, and then φ automatically extends, and in the interesting examples u does too.) Normally one would induce an algebra admitting a G-action from that containing an N-action, but that will no longer work since the induced algebra is trivial. Instead we consider the u-induced algebra B = u-indG N A described in the next result. Proposition 6.1. The space B = u-indG N A of functions f ∈ C0 (G, A) which satisfy −1 f (rx) = ad(u(r, x)) αr [f (x)] for all x ∈ G and r ∈ N is not trivial, and is closed under pointwise multiplication of functions (f1 f2 )(x) = f1 (x)f2 (x) and the involution f ∗ (x) = f (x)∗ . The norm f = sup f (x) is a C ∗ -norm. Proof. We first note that (using the cocycle condition and remembering that the adjoint action is unaffected by the central factor φ) functions in the space satisfy f (rsx) = ad(u(r, sx))−1 αr [f (sx)] = ad(u(r, sx))−1 αr ([ad(u(s, x))]−1 αs [f (x)]) = ad(u(r, sx))−1 αr [ad(u(s, x))]−1 αr αs [f (x)] = ad(u(rs, x))−1 ad(u(r, s))−1 αr αs [f (x)] = ad(u(rs, x))−1 αrs [f (x)], showing consistency of the condition. Without u this consistency check would fail and the induced algebra would be trivial, showing why one cannot use the normal induced algebra. We shall exhibit some useful explicit functions in the induced algebra in Proposition 6.4, but there is also a general construction which is useful. For f a function C0 (G) and a an element in A, we define an A-valued function (f ♦a) on G by (f ♦a)(x) = f (nx)αn−1 [ad(u(n, x))[a]] dn. (6.4) N
Nonassociative Tori and Applications to T-Duality
51
Using the cocycle identity for ad(u) (the obstruction φ is central and so disappears in the adjoint action) we then check that (f ♦a)(rx) = f (nrx)αn−1 [ad(u(n, rx))[a]] dn N = f (nrx)αn−1 [ad(αn [u(r, x)])−1 ad(u(n, r))ad(u(nr, x))[a]] dn N −1 = ad(u(r, x)) f (nrx)αn−1 [ad(u(n, r))ad(u(nr, x))[a]] dn N −1 = ad(u(r, x))−1 αr f (nrx)αnr [ad(u(nr, x))[a]] dn N
= ad(u(r, x))−1 αr [(f ♦a)(x)], showing that (f ♦a) defines an element of B. Using the fact that αr and ad(u(x, r)) are automorphisms, we see that (f1 f2 )(xr) = f1 (xr)f2 (xr) = ad(u(x, r))αr [f1 (x)]ad(u(x, r))αr [f2 (x)] = ad(u(x, r))αr [f1 (x)f2 (x)] = ad(u(x, r))αr [(f1 f2 )(x)], so that the space of u-induced functions is closed under the product. Finally, exploiting the unitarity of u(r, x), we have f ∗ (rx) = f (rx)∗ = [u(r, x)−1 αr [f (x)]u(r, x)]∗ = u(r, x)−1 αr [f (x)]∗ u(r, x) = u(r, x)−1 αr [f ∗ (x)]u(r, x), so that the involution respects the constraint. Finally f ∗ f = sup (f ∗ f )(x) = sup f (x)∗ f (x) = sup f (x)2
(6.5)
gives a C ∗ -norm. This shows that the induced space B is actually a ∗-algebra (moreover, an associative algebra, since A was associative). A C ∗ -norm can be defined much as in the usual case. Exploiting the ideas of the ordinary induced representations we can improve on this. then B is a continTheorem 6.2. If A is a continuous trace algebra with spectrum A = N\(G × A), where the action of r ∈ N on uous trace algebra with spectrum B is defined by r(x, π) = (rx, π ◦ αr−1 ad(u(r, x))). (x, π ) ∈ G × A Proof. We should start by checking that the above formula does indeed define an action, but that is essentially the same calculation just done to show that f ♦a satisfies the equivariance condition for B. The rest of the proof follows the ideas in [33, 6.16 to 6.21], but with the actions on the other sides, and using our definition of f ♦a, and noting that in the proof of 6.18 one needs to restrict to neighbourhoods of both s and x. This enables
52
P. Bouwknegt, K. Hannabuss, V. Mathai
(x, π ) : F → π(F (x) of us to show that the representation defined by (x, π ) ∈ G × A, B is irreducible, because, for any (x, a) ∈ G × A we can use appropriate f ♦a to find F ∈ B such that F (x) = a, and then use irreducibility of π . The rest of the proof in [33] is independent of the twisting. We next introduce automorphisms of the twisted induced algebra. Proposition 6.3. For y ∈ G and a function f : G → A define βy [f ](x) = ad(u(x, y)) [f (xy)]. Then βy preserves the subalgebra B and defines a ∗-automorphism of it. Proof. βy [f ](rx) = ad(u(rx, y))[f (rxy)] = ad(u(rx, y))ad(u(r, xy))−1 αr [f (xy)] = ad(u(r, x))−1 ad(αr [(u(x, y)])αr [f (xy)] = ad(u(r, x))−1 αr ([ad(u(x, y))][f (xy)]) = ad(u(r, x))−1 αr [βy [f ](x)], showing that βy satisfies the equivariance condition. To see that these are automorphisms we need only note that (βy [f1 ]βy [f2 ])(x) = (βy [f1 ](x)βy [f2 ])(x) = ad(u(x, y))[f1 (xy)]ad(u(x, y))[f2 (xy)] = ad(u(x, y))[f1 (xy)f2 (xy)] = ad(u(x, y))[(f1 f2 )(xy)] = βy [f1 f2 ](x), as required, and compatibility with the involution is also easily checked.
Naturally the map y → βy is not a homomorphism. Proposition 6.4. The functions v(y, z) : x → φ(x, y, z)u(x, y)u(xy, z)u(x, yz)−1 , lie in the multiplier algebra of B and satisfy βx [v(y, z)]v(x, yz) = φ(x, y, z)v(x, y)v(xy, z).
(6.6)
The automorphisms defined by β satisfy the relations βy βz = ad(v(y, z))βyz .
(6.7)
Proof. When x ∈ N we can write v(y, z)(x) = αx [u(y, z)], but otherwise αx is undefined, and we cannot reduce v(y, z)(x) = φ(x, y, z)−1 ad(u(x, y))ad(u(xy, z)) ad(u(x, yz))−1 . To check that v(y, z) satisfies the equivariance condition for membership of B, we calculate u(rx, y)u(rxy, z)u(rx, yz)−1 = φ(r, x, y)−1 u(r, x)−1 αr [u(x, y)]u(r, xy)u(rxy, z)u(rx, yz)−1 = φ(r, x, y)−1 φ(r, xy, z)−1 u(r, x)−1 αr [u(x, y)]αr [u(xy, z)]u(r, xyz)u(rx, yz)−1 = φ(r, x, y)−1 φ(r, xy, z)−1 φ(r, x, yz)u(r, x)−1 αr [u(x, y)u(xy, z)]αr [u(x, yz)]−1 ×u(r, x) = φ(r, x, y)−1 φ(r, xy, z)−1 φ(r, x, yz)ad(u(r, x))−1 αr [u(x, y)u(xy, z)u(x, yz)−1 ] = φ(r, y, z)−1 ad(u(r, x))−1 αr [u(x, y)u(xy, z)u(x, yz)−1 ].
Nonassociative Tori and Applications to T-Duality
53
From this we see that [v(y, z)](rx) = φ(rx, y, z)φ(r, y, z)−1 ad(u(r, x))−1 αr [φ(x, y, z)−1 v(y, z)(x)] = ad(u(r, x))−1 αr [v(y, z)(x)]. We cannot conclude that v(y, z) lies in B since it does not satisfy the analytic condition of vanishing outside compact sets, but it is certainly in the multiplier algebra. By making some cancellations, we obtain φ(s, x, y)φ(s, xy, z) v(x, y)v(xy, z)v(x, yz)−1 (s) = u(s, x)u(sx, y) φ(s, x, yz) ×u(sxy, z)u(s, xyz)−1 u(s, x)−1 = φ(s, y, z)ad(u(s, x))[u(sx, y)u(sxy, z)u(sx, yz)−1] = φ(x, y, z)−1 ad(u(s, x))[v(y, z)(sx)] = φ(x, y, z)−1 βx [v(y, z)](s). Finally we see that (βy βz [f ])(x) = ad(u(x, y))[(βz [f ](xy)] = ad(u(x, y))ad(u(xy, z)[f (xyz)] = ad(u(x, y))ad(u(xy, z)ad(u(x, yz)−1 )[(βyz f (x)] = ad(v(y, z))(x)[(βyz f (x)]. defined by β has orbit space B/G = A/N, Corollary 6.5. The action of G on B so that β defines the principal G/N-bundle (G × A)/N → A/N. to ◦ βy . In the earlier notation we have Proof. The action of y ∈ G sends ∈ B (x, π )(βy [f ]) = π(βy [f ](x)) = π(ad(u(x, y))[f (xy)]),
(6.8)
and, since inner automorphisms don’t affect the class of a representation, this is equivalent to (xy, π ). For r ∈ N this reduces to (x, π )(βr [f ]) = π(ad(u(x, r)u(r, x)−1 )ad(u(r, x)αr [f (x)]),
(6.9)
which is equivalent to r(x, π), showing that the subgroup N stabilises the irreducible representations of B, so that we have a G/N bundle, and that the orbit space is A/N, as claimed. 7. Algebras with Prescribed Spectrum and Dixmier-Douady Class Before tackling the general case it is useful to consider what happens for the principal bundle G/N over a point. In that case only the component H0 can be non-trivial, and we assume that it defines the tricharacter φ as before. In fact [10] gives a universal construction for a principal projective unitary bundle over a group, but we shall give an alternative description of the algebra as a twisted induced algebra. We shall induce from N to G, and in order that the spectrum should be just G/N we induce the algebra K(L2 (N)) of compact operators. That carries the
54
P. Bouwknegt, K. Hannabuss, V. Mathai
right regular representation ρ of N and twisting uρ described in Sect. 4. Although the restriction of uρ to N × N is 1, the extension to G × G gives the induced algebra a twist. 2 At this point it is instructive to consider why uρ -indG N (K(L (N))) has a non-vanishing Dixmier–Douady obstruction when the action of N on K(L2 (N)) is given by 2 αρ = adρ. One might expect that the induced algebra uρ -indG N (K(L (N)) simply acts on the induced Hilbert space of square-integrable functions ψ : G → L2 (N), which satisfy the equivariance condition ψ(rx) = uρ (r, x)−1 ρ(r)ψ(x),
(7.1)
so that there is no obstruction. However, this is incorrect because the suggested equivariance condition on ψ is inconsistent, when uρ is not a cocycle: ψ(rsx) = uρ (r, sx)−1 ρ(r)ψ(sx) = uρ (r, sx)−1 ρ(r)uρ (s, x)−1 ρ(s)ψ(x) = uρ (r, sx)−1 αr [uρ (s, x)]−1 ρ(r)ρ(s)ψ(x) = φ(r, s, x)−1 uρ (rs, x)−1 uρ (r, s)−1 ρ(rs)ψ(x). Now, as we noted in the last paragraph, uρ (r, s) = 1, but the presence of the argument x∈ / N means that φ(r, s, x) = 1, and we end up with constraints uρ (rs, x)−1 ρ(rs)ψ(x) = ψ(rsx) = φ(r, s, x)−1 uρ (rs, x)−1 ρ(rs)ψ(x)
(7.2)
which can be satisfied only by ψ = 0. Proposition 7.1. Let φ be the tricharacter of G constructed from H0 , uρ , ρ defined on 2 L2 (N) as in Sect. 4, and αρ = adρ. The algebra uρ -indG N (K(L (N))) has spectrum G/N, the G action on the spectrum is transitive with stabiliser N, and the Dixmier-Douady class is described by the 3-form H0 . Proof. We assume that φ is obtained from the class f = H0 by the procedure described in Sect. 2. Choose an open set F ⊆ G on which the projection π to G/N is one-one, and translates Fi = F xi whose projections give a cover of G/N. These translates share the property that the projection to G/N is injective, and so we may choose sections γi : π(Fi ) → G. The differences γij (v) = γi (v)γj (v)−1 lie in N. The restriction of the induced algebra to algebra-valued functions on π(Fi ) ⊆ G/N is Morita equivalent to C(π(Fi )) via the bimodule Xi = L2 (π(Fi ), L2 (N)), (restriction to the subsets enables us to sidestep the earlier problem with L2 (G, L2 (N))). The actions are the obvious pointwise multiplicative actions, (f ψ)(v) = f (γi (v))ψ(v)
(7.3)
and this is a imprimitivity bimodule in the sense of [33]. Over the intersection π(Fi ) ∩ π(Fj ) there is an equivalence of the two bimodules Xi and Xj , given by the map (gij ψ)(v) = uρ (γij (v), γj (v))−1 ρ(γij (v))ψ(v)
(7.4)
Nonassociative Tori and Applications to T-Duality
55
from Xj to Xi . To see why this works we note that (gij f ψ)(v) = uρ (γij (v), γj (v))−1 ρ(γij (v))f (γj (v))ψ(v) = ad(uρ (γij (v), γj (v))−1 ad(ρ(γij (v))[f (γj (v))] uρ (γij (v), γj (v))−1 ρ(γij (v))ψ(v). Since γij (v) ∈ N this simplifies to f (γij (v)γj (v))uρ (γij (v), γj (v))−1 ρ(γij (v))ψ(v) = (f gij ψ)(v).
(7.5)
To compute the obstruction we must compare gij gj k and gik on π(Fi ) ∩ π(Fj ) ∩ π(Fk ). Now we have (gij gj k φ)(v) = uρ (γij (v), γj (v))−1 ρ(γij (v))(gj k ψ)(v) = uρ (γij (v), γj (v))−1 ρ(γij (v)) uρ (γj k (v), γk (v))−1 ρ(γj k (v))ψ(v) = uρ (γij (v), γj (v))−1 ad(ρ(γij (v)))[uρ (γj k (v), γk (v))−1 ] ρ(γij (v))ρ(γj k (v))ψ(v). Applying the modified cocycle identity we have (gij gj k φ)(v) = φ(γij (v), γj k (v), γk (v))−1 uρ (γik (v), γk (v))−1 uρ (γij (v), γj k (v))−1 ρ(γik (v))ψ(v), and finally, since γij (v), γj k (v) ∈ N we deduce that uρ (γij (v), γj k (v)) = 1, giving (gij gj k φ)(v) = φ(γij (v), γj k (v), γk (v))−1 (gik ψ)(v).
(7.6)
ˇ This shows that the Dixmier-Douady class can be described by the Cech cocycle φij k = φ(γij (v), γj k (v), γk (v))−1 = exp[2πif (γij (v), γj k (v), γk (v))] = exp[2πif (γi (v), γj (v), γk (v))], where the antisymmetry of f has been used in the final line. To find a form describing the de Rham cocycle we note that, since locally γk (v) and v are the same and the differences γij are in N, df (γij (v), γj k (v), γk (v)) = f (γij (v), γj k (v), dγk (v)) = f (γij (v), γj k (v), dv) = f (γij (v), γj (v), dv) − f (γij (v), γk (v), dv), giving an explicit expression as the difference of one-forms. Repeating this process twice we arrive at the de Rham form f (dv, dv, dv) giving the class H0 . (The antisymmetry of f compensates for the antisymmetry of the exterior product to give a non-vanishing answer.) The same ideas can now be used to deal with the general case. However, since we want to use results for the untwisted case, we shall at this point restrict ourselves to the case of an abelian group G.
56
P. Bouwknegt, K. Hannabuss, V. Mathai
Theorem 7.2. Let G be abelian and A a continuous trace algebra with an action of N ⊂ G by locally projectively unitary automorphisms. (This includes the assumption that the restriction of u to N × N takes the constant value 1). Both u-indG N (A) and (A) are continuous trace algebras, and the difference of their Dixmier-Douady indG N G 3 invariants δ(u-indG N (A)) − δ(indN (A)) is the class defined by the form f ∈ (G, Z) associated to the tricharacter φ: G δ(u-indG N (A)) − δ(indN (A)) = [f ].
(7.7)
Proof. We first note that by Theorem 4.1 u is exterior equivalent to uλ which by (4.3) has trivial restriction to N × N. Since A has continuous trace, it is locally Morita equivalent to an algebra of compact operators, and all our calculations will be local. In fact we by open sets {Uλ } and find Hilbert spaces Hλ such that the restriction AUλ may cover A of A to Uλ is Morita equivalent to C(Uλ , K(Hλ )) via some bimodule Yλ . Combining these with the bimodules Xl used in the previous proof we take X(l,λ) = Xl ⊗ Yλ for the restriction of the algebra to π(Fl ) × Uλ . We may assume that the cover is fine enough that αn (a) is equivalent to ad(ρnλ )(a) (m,µ) for a ∈ AUλ , with ρ λ a θ λ -representation. We now define an equivalence G(l,λ) on the overlap of the sets π(Fl ) × Uλ and π(Fm ) × Uµ by setting (G(l,λ) ψ)(v) = u(γlm (v), γm (v))−1 hλµ ρ µ (γlm (v))ψ(v), (m,µ)
(7.8)
where hλµ describes the equivalences of Yλ and Yµ , (which are assumed to satisfy the ˇ relationship hλµ hµν = λµν hλν , where is a Cech cocycle describing the DixmierDouady class of A.) On overlaps the projective representations ρ λ are equivalent in the sense that ρ µ (n)hµν = hµν κµν (n)ρ ν (n), for some character κµν ∈ N. The adjoint actions of ρ µ and ρ ν are both equivalent to α. We now calculate that (G(l,λ) G(m,µ) ψ)(v) = u(γlm (v), γm (v))−1 hλµ ρ µ (γlm (v))u(γmn (v), (m,µ)
(n,ν)
γn (v))−1 hµν ρ ν (γmn (v))ψ(v) = u(γlm (v), γm (v))−1 αγlm )[u(γmn (v), γn (v))−1 ]hλµ hµν ×κµν (γlm )ρ ν (γlm (v))ρ ν (γmn (v))ψ(v). The first two terms combine as before to give u(γln (v), γn (v))−1 u(γlm (v), γmn (v))−1 = φ(γlm , γmn , γn )−1 u(γln (v), γn (v))−1 , (7.9) whilst the projective representions give ρ ν (γlm (v))ρ ν (γmn (v)) = θ ν (γlm (v), γmn (v))ρ ν (γln (v)), giving (m,µ)
(n,ν)
(G(l,λ) G(m,µ) ψ)(v) = φ(γlm , γmn , γn )−1 κµν (γlm )θ ν (γlm (v), γmn (v))λµν u(γln (v), γn (v))−1 hλν ρ ν (γln (v))ψ(v) = φ(γlm , γmn , γn )−1 κµν (γlm )θ ν (γlm (v), γmn (v))λµν G(l,λ) ψ(v), (n,ν)
(7.10)
Nonassociative Tori and Applications to T-Duality
57
from which we deduce the obstruction. All the factors except the first would be present for indG N (A), so that the difference between the Dixmier-Douady obstructions for G −1 u-indN (A) and indG N (A) is just given by φ , or as forms by f . Our formula shows that the obstruction has four contributions: the MacLane-Whitehead obstruction φ −1 , the Mackey obstruction θ , the Phillips-Raeburn obstruction κ, and the Dixmier-Douady obstruction for A, corresponding to H0 , H1 , H2 and H3 . Corollary 7.3. Let G be abelian and E be a principal G/N-bundle over M with prescribed Dixmier-Douady invariant associated with H . Then there is a u-induced algebra with spectrum E, and the correct action of G, having the Dixmier-Douady invariant H . This u-induced algebra is not necessarily unique.2 Proof. We know from [2, 4, 28] that there is an algebra indG N (A) associated with the principal G/N-bundle E and having Dixmier-Douady form H − H0 . This algebra is not necessarily unique, cf. [28]. With f as in the theorem we see that u-indG N (A) has Dixmier-Douady class described by the form H . An alternative approach would be to note that a continuous trace algebra with Dixmier–Douady class given by H , can be constructed as the tensor product of two continuous trace algebras with classes H − H0 and with H0 , and then to apply [28] for the former and our earlier result for the class H0 . Although we have shown how to construct an algebra with a given Dixmier-Douady class, it is natural to wonder whether one could also find another algebra with twisting given by an ordinary cocycle. The next result shows that this is not possible. Theorem 7.4. Every system with the Dixmier-Douady invariant H is described by automorphisms whose twisting gives the same tricharacter φ. Proof. By the general theory we know that any other system must be exterior equivalent to the u-induced system above, that is it is described by automorphisms λx = ad(W (x))βx , and w(x, y) = W (x)βx [W (y)]v(x, y)W (xy)−1 for some MB-valued function W on G. Since the twistings are cohomologous they must define the same class φ, but more explicitly we calculate that λx [w(y, z)]w(x, yz) = ad(W (x))βx [W (y)βy [W (z)]v(y, z)W (yz)−1 ] W (x)βx [W (yz)]v(x, yz)W (xyz)−1 = W (x)βx [W (y)]βx βy [W (z)]βx [v(y, z)]v(x, yz)W (xyz)−1 = φ(x, y, z)W (x)βx [W (y)]ad(v(x, y)) ×βxy [W (z)]v(x, y)v(xy, z)W (xyz)−1 = φ(x, y, z)W (x)βx [W (y)]v(x, y)W (xy)−1 W (xy) ×βxy [W (z)]v(xy, z)W (xyz)−1 = φ(x, y, z)w(x, y)w(xy, z), showing that the same cocycle φ arises. 2
The non-uniqueness reflects the fact that there may exist several liftings of the action of N on A to G. So, strictly speaking, u-ind does not define a functor, but u-indG N (A) just indicates the induced algebra with a prescribed action of G.
58
P. Bouwknegt, K. Hannabuss, V. Mathai
8. The Twisted Crossed Product Algebra In this section we can return to the case of a general group G. We have argued that the dual of a bundle described by B with a twisted group action should be given by the twisted crossed product. For the induced algebra u-indG N (A) its twisted crossed product with G can be calculated explicitly, and the following result shows that it is a generalisation of the twisted compact operators in Sect. 5. Theorem 8.1. The twisted crossed product u-indG N (A) β,v G is isomorphic to the ∗-algebra of A-valued kernels on G × G satisfying K1 (rz, rw) = φ(r, z, w)−1 u(r, z)−1 αr [K1 (z, w)]u(r, w), with K ∗ (z, w) = K(w, z)∗ and product K1 (z, v)K2 (v, w)φ(z, v, w) dv. (K1 K2 )(z, w) =
(8.1)
(8.2)
G
Proof. The twisted crossed product consists of functions Fj : G → B with twisted convolution F1 (y)βy [F2 (y −1 x)]v(y, y −1 x) dy. (8.3) (F1 ∗ F2 )(x) = G
Identifying B with functions from G to A, we may give the elements of the twisted crossed product a second argument in G and, using the explicit form of the induced action and of v, get F1 (z, y)ad(u(z, y))[F2 (zy, y −1 x)] (F1 ∗ F2 )(z, x) = G
×φ(z, y, y −1 x)u(z, y)u(zy, y −1 x)u(z, x)−1 dy,
(8.4)
which can be rearranged as
φ(z, y, y −1 x)F1 (z, y)u(z, y)F2 (zy, y −1 x)
(F1 ∗ F2 )(z, x)u(z, x) = G
×u(zy, y −1 x) dy.
(8.5)
We now define K1 (z, zy) = F1 (z, y)u(z, y), K2 (z, zy) = F2 (z, y)u(z, y) and (K1 K2 )(z, zx) = (F1 ∗ F2 )(z, x)u(z, x) to obtain K1 (z, zy)K2 (zy, zx)φ(z, y, y −1 x) dy. (8.6) (K1 K2 )(z, zx) = G
Setting w = zx, v = zy and exploiting the antisymmetry of φ the result follows. We readily check that F ∗ (z, x)u(z, x) = v(x, x −1 )∗ ad(u(z, x))[F (zx, x −1 )]∗ u(z, x) = u(zx, x −1 )∗ F (zx, x −1 )∗ , from which it follows that K ∗ (z, w) = K(w, z)∗ .
(8.7)
Nonassociative Tori and Applications to T-Duality
59
The kernels inherit an equivariance condition from the inducing process, K1 (rz, rzy) = F1 (rz, y)u(rz, y) = ad(u(r, z))−1 αr [F1 (z, y)]u(rz, y) = u(r, z)−1 αr [K1 (z, zy)u(z, y)−1 ]u(r, z)u(rz, y) = φ(r, z, y)−1 u(r, z)−1 αr [K1 (z, zy)]u(r, zy). Exploiting the antisymmetry of φ this gives K1 (rz, rw) = φ(r, z, w)−1 u(r, z)−1 αr [K1 (z, w)]u(r, w).
(8.8)
We note that since the original product F1 ∗ F2 respected this equivariance condition, so does the product on kernels. The norm on the twisted crossed product can be defined from the left regular representation and so agrees with that on kernels. The description of the crossed product algebra in Theorem 8.1 can be made more precise for a bundle over a point described by A = K(L2 (N)) as in Proposition 7.1. However, this time it is more useful to take the left handed version of the automorphisms, that is αr [k](s, t) = k(r −1 s, r −1 t) ,
(8.9)
and set (u(s, t)ψ)(r) = φ(r, s, t)−1 ψ(r). Theorem 8.2. The nonassociative torus describing the dual of a bundle over a point is isomorphic to the algebra Aφ = Kφ (L2 (G)) γ ,u N, where γ is defined by Eq. (5.10). Proof. We abbreviate notation for the K(L2 (N))-valued kernels by setting K(z, w) : (s, t) → K(z, w; s, t),
(8.10)
so that the equivariance condition can be given explicitly as K(rz, rw; s, t) = φ(r, z, w)−1 φ(r, z, s)K(z, w; r −1 s, r −1 t)]φ(r, w, t)−1 , (8.11) or, replacing the first two arguments, as K(z, w; s, t) = φ(r, z, w)−1 φ(r, z, s)K(r −1 z, r −1 w; r −1 s, r −1 t)]φ(r, w, t)−1 . (8.12) Taking r = s in this formula we get K(z, w; s, t) = φ(s, z, w)−1 K(s −1 z, s −1 w; 1, s −1 t) ×φ(w, s, t).
(8.13)
Writing u(x, y) for the multiplier associated with the automorphism of twisted kernels γx defined by Eq. (5.10), we have K(z, w; s, t) = γs [K](z, w; 1, s −1 t)u(s, t) = γs [K](z, w; 1, s −1 t)u(s, s −1 t) . (8.14)
60
P. Bouwknegt, K. Hannabuss, V. Mathai
This shows that the kernels can be reconstructed from their values when the third argument is 1, and in this case the product formula can be simplified. The product (K1 K2 )(z,w; r, t) =
K1 (z, v; r, s)K2 (v, w; s, t)φ(z, v, w) dsdv , (8.15) N×G
reduces to (K1 K2 )(z, w; 1, t) =
K1 (z, v; 1, s)K2 (v, w; s, t)φ(z, v, w) dsdv . N×G
(8.16) The second kernel can be rewritten using the equivariance condition to give
(K1 K2 )(z, w; 1, t) =
K1 (z, v; 1, s)γs [K2 ](v, w; 1, s −1 t)u(s, s −1 t) ds}
{ G
N
×φ(z, v, w) dv .
(8.17)
Identifying the kernel Kj with the Kφ (L2 (G))-valued function s → {(z, w) → Kj (z, w; 1, s)} on N, this is just the twisted crossed product of the two functions. In other words we can identify the algebra with Kφ (L2 (G)) γ ,u N. In the general case of a bundle over M one needs to take A = C(M, K(L2 (N)), but since the products of functions on M are all taken pointwise, there is no essential change in the calculations. The only point for caution is that, in those cases where φ depends on M (and consequently so also do γ and u), one is really looking at an algebra of continuous sections of a bundle over M rather than just functions. Observe that by Theorem 2 8.1, the nonassociative torus Aφ is canonically isomorphic to u-indG N (K(L (N)) β,v G. Theorem 8.3. The nonassociative torus describing the dual of a bundle over M is isomorphic to the algebra C(M, Kφ (L2 (G)) γ ,u N). When N is trivial this gives the twisted algebra of kernels introduced in Sect. 5, though with φ replaced by its inverse. More generally it provides an extension of Green’s Generalised Imprimitivity Theorem [17] to the case of these twisted induced algebras. In contrast to the twisted crossed product algebra for the dual, the correspondence algebra is associative. Provided that there is no Mackey obstruction the correspondence space is associated with the algebra u-indG N (A) β,v N. Theorem 8.4. The algebra u-indG N (A) β,v N is associative. Proof. By Propositions 3.1 and 6.4 u-indG N (A) β,v N is associative if and only if the restriction of φ to N × N × N is 1, and we have seen that this is a consequence of the integrality of the form H0 . For non-trivial Mackey obstruction one replaces N by the projection onto G of the centre of the central extension defined by the multiplier [19, 14], and that algebra is similarly associative.
Nonassociative Tori and Applications to T-Duality
61
9. The Dual Action In this section we shall assume that G is abelian. Following the development in Sect. 5, we may define automorphisms of the kernels by τx [K](z, w) = φ(x, z, w)K(x −1 z, x −1 w),
(9.1)
u(x, y))−1 τxy , where and these satisfy τx τy = ad( ad( u(x, y)[K])(z, w) = φ(x, y, z)φ(x, y, w)−1 K(z, w).
(9.2)
However, we are primarily interested in the case of abelian groups, and there is also a on the dual algebra. Its definition is motivated by more useful action of the dual group G ξ on F ∈ u-indG (A) β,v G given by β ξ [F ](z, y) = ξ(y)F (z, y). Rewritten the action β N in terms of kernels this leads to the following idea. define β ξ by Proposition 9.1. For ξ ∈ G ξ [K](z, w) = ξ(z−1 w)K(z, w). β
(9.3)
ξ is an automorphism of u-indG (A) β,v G, and β η = β ξ β ξ η . Then β N ξ is well-defined, that is that it preserves the subspace Proof. We must first show that β of kernels K satisfying the equivariance condition of Theorem 8.1. In fact we see that ξ [K](rz, rw) = ξ(z−1 w)K(rz, rw) β = ξ(z−1 w)φ(r, z, w)−1 u(r, z)−1 βr [K(z, w)]u(r, w) ξ [K](z, w)]u(r, w), = φ(r, z, w)−1 u(r, z)−1 βr [β which gives the required equivariance condition. It is also an automorphism since ξ(z−1 v)K1 (z, v)ξ(v −1 w)K2 (v, w)φ(z, v, w)−1 dv (βξ [K1 ] βξ [K2 ])(z, w) = G = ξ(z−1 w)K1 (z, v))K2 (v, w)φ(z, v, w)−1 dv G
= ξ(z−1 w)(K1 K2 )(v, w). η = β ξ η . ξ β There is no twisting involved since it is easy to check that β
The We now form the crossed product of the twisted crossed product algebra with G. to the algebra of kernels described in the last section, and elements are functions from G with multiplication so may be regarded as A-valued functions on G × G × G, η [ ( k1 k2 )(z, w, ξ ) = k2 ](v, w, η−1 ξ )φ(z, v, w)−1 dηdv k1 (z, v, η)β G×G k1 (z, v, η)η(v −1 w) k2 (v, w, η−1 ξ )φ(z, v, w)−1 dηdv. = G×G
We shall denote the group-theoretic Fourier transform of k with respect to its third argument by k (which is now a function on G × G × G): k(z, w, ξ )ξ(x) dξ. (9.4) k(z, w, x) = G
62
P. Bouwknegt, K. Hannabuss, V. Mathai
The multiplication obtained by Fourier transform is (assuming appropriate normalisation of the measures) ξ(x) k1 (z, v, η)η(v −1 w) k2 (v, w, η−1 ξ ) (k1 k2 )(z, w, x) = G G×G×
×φ(z, v, w)−1 dξ dηdv
=
G G×G×
k2 (v, w, η−1 ξ )(η−1 ξ )(x) k1 (z, v, η)η(v −1 wx)
×φ(z, v, w)−1 dξ dηdv
k1 (z, v, v −1 wx)k2 (v, w, x)φ(z, v, w)−1 dv.
= G
We now introduce another transformation by k(x; z, w) = φ(x, z, w)−1 u(x, z)k(xz, xw, w −1 )u(x, w)−1 .
(9.5)
Theorem 9.2. There is an isomorphism 2 ∼ (u-indG = u-indG G N (A) β,v G) β N (A) ⊗ Kφ (L (G)).
(9.6)
Proof. One important property which follows from the equivariance condition for the A-valued kernels is that k(rx; z, w) = φ(rx, z, w)−1 u(rx, z)k(rxz, rxw, w −1 )u(rx, w)−1 = [φ(rx, z, w)φ(r, xz, xw)]−1 u(rx, z)u(r, xz)−1 αr [k(xz, xw, w −1 )] ×u(r, xw)u(rx, w)−1 = [φ(rx, z, w)φ(r, xz, xw)]−1 φ(r, x, z)−1 φ(r, x, w) u(r, x)−1 αr [u(x, z)]αr [k(xz, xw, w −1 )]αr [u(x, w)]−1 u(r, x) = [φ(rx, z, w)φ(r, xz, xw)]−1 φ(r, z, x)φ(r, x, w) ad(u(r, x))−1 αr [u(x, z)k(xz, xw, w −1 )u(x, w)−1 ] = ad(u(r, x))−1 αr [ k(x; z, w)], so that the kernels k satisfy the induced algebra condition with respect to x, and so can be considered elements of the algebra induced from N to G by the A-valued kernels. Next we consider the product on these functions k2 )(x; z, w) = φ(x, z, w)−1 u(x, z)(k1 k2 )(xz, xw, w−1 )u(x, w)−1 ( k1 = φ(x, z, w)−1 u(x, z)k1 (xz, xv, v −1 )k2 (xv, xw, w−1 ) G
−1
×u(x, w)
=
φ(xz, xv, xw)−1 dv
k2 (x; v, w)φ(z, v, w)−1 dv. k1 (x; z, v)
G
Thus we have the pointwise product with respect to x, with the twisted multiplication of Kφ (L2 (G)) in the fibres.
Nonassociative Tori and Applications to T-Duality
63
We shall show in the final section that the twisted compact operators are in a certain sense Morita equivalent to the ordinary compact operators, so that this result provides a very precise analogue of the normal duality theorem. is To complete the argument we really need to know that the double dual action β equivalent to the original β to within the action on the twisted compact operators. The double dual action is defined on (u-indG G by the same procedure used N (A) β,v G) β variable and the to obtain the dual action, that is multiplication by the pairing of the G group element: ( β g [ k](z, w, ξ ) = ξ(g) k(z, w, ξ ).
(9.7)
Theorem 9.3. The double dual action of G can be written as ( β g [ k])(x; z, w) = τg [(v(g, g −1 z)−1 βg [ k]v(g, g −1 w)(x; z, w)].
(9.8)
Proof. Fourier transforming the definition of the action we get g [k](z, w, x) = ( β
G
ξ(g)ξ(x) k(z, w, ξ ) = k(z, w, xg).
(9.9)
Using the same notation for the equivalent action on k, g [k])(xz, xw, w −1 )u(x, w)−1 k])(x; z, w) = φ(x, z, w)−1 u(x, z)( β ( β g [ = φ(x, z, w)−1 u(x, z)k(xz, xw, w −1 g)u(x, w)−1 = φ(xg, g −1 z, g −1 w)φ(x, z, w)−1 u(x, z)u(xg, g −1 z)−1 k(xg; g −1 z, g −1 w)u(xg, g −1 w)u(x, w)−1 . In terms of the induced twisting we now we have u(x, z)u(xg, g −1 z)−1 = φ(x, g, g −1 z)−1 v(g, g −1 z)(x)−1 u(x, g),
(9.10)
and substituting this (and the analogous expression in w), we arrive at g [ k])(x; z, w) = φ(xg, g −1 z, g −1 w)φ(x, z, w)−1 φ(x, g, g −1 z)−1 φ(x, g, g −1 w) ( β ×v(g, g −1 z)(x)−1 u(x, g) k(xg; g −1 z, g −1 w)u(x, g)−1 v(g, g −1 w)(x) k(xg; g −1 z, g −1 w)] = φ(g, z, w)v(g, g −1 z)(x)−1 ad(u(x, g))[ −1 ×v(g, g w)(x) = φ(g, z, w)v(g, g −1 z)(x)−1 (βg [ k](x; g −1 z, g −1 w)]v(g, g −1 w)(x). This can be rewritten in terms of the twisted action τg on kernels and the adjoint action of v in the form
64
P. Bouwknegt, K. Hannabuss, V. Mathai
g [ ( β k])(x; z, w) = τg [(v(g, g −1 z)−1 βg [ k]v(g, g −1 w))(x; z, w)],
(9.11)
showing that to within an inner automorphism of the kernels one has βg ⊗ τg , and up to an action on the twisted kernels one recovers βg . In the application to principal Tn -bundles one takes G = Rn . As this is contractible, Proposition 5.2 shows that Kφ (L2 (G)) is a deformation of K(L2 (G)), so that this is a close substitute for the usual duality theorem. 10. Applications to T-Duality In this section, we apply the mathematical results of the earlier sections to determine the T-dual of principal torus bundles with general H-flux, thus generalizing earlier results in [2–4, 28, 29]. Let G = R , and N = Z in the setup of Corollary 7.3. Let E → M be a principal T -bundle, and H ∈ H 3 (E) be an integral H-flux on E.3 Then we can identify H = (H3 , H2 , H1 , H0 ), where Hp ∈ p (M, ∧3−p t), under the isomorphism (1.1), and closed under D. Then by the results in [28, 29], there is a continuous trace C ∗ -algebra indG N (A) with spectrum equal to E and with Dixmier-Douady invariant equal to (H3 , H2 , H1 , 0), which has an action of G that covers the given action of G on E. This action is not necessarily unique. Then by Corollary 7.3, we know that there is another continuous trace C ∗ -algebra u-indG N (A) with spectrum equal to E and with Dixmier-Douady invariant equal to H = (H3 , H2 , H1 , H0 ), which has a twisted action of G that covers the given action of G on E. Our main definition in this section is: Definition 10.1. The twisted crossed product u-indG N (A) β,v G is defined to be the T-dual to the principal T -bundle E with H-flux H . We justify this definition as follows. Firstly, the T-dual of u-indG N (A) β,v G is the which by the twisted Takai duality Theorem crossed product (u-indG G, (A) G) β,v β N G 2 9.2 is isomorphic to u-indG N (A) ⊗ Kφ (L (G)). That is, the T-dual of u-indN (A) β,v G G is Morita equivalent to the continuous trace algebra u-indN (A), so that T-duality applied twice returns us to where we started, up to Morita equivalence. In the special case when H = H0 , the fibre of this bundle over the point z ∈ M is equal to the nonassociative torus Aφ of rank with tricharacter φ corresponding to H0 (see Theorem 8.2). In the general case, but when H0 is zero, the fibre is a stabilized noncommutative torus with invariant H1 , [28, 29], and when H0 = 0 and H1 = 0, then the fibre is the stabilized algebra of continuous functions on a torus, [2–4]. Thus we have the following theorem. Theorem 10.2 (T-duality for principal torus bundles). Let E → M be a principal T -bundle over M, and H ∈ H 3 (E) be an integral H-flux on E. Then H = (H3 , H2 , H1 , H0 ), where Hp ∈ p (M, ∧3−p t). Let c1 (E) ∈ H 2 (M, t) denote the first Chern class of E, which determines E up to isomorphism. Then: which is a principal (1) If H0 = 0 and H1 = 0, then there is a canonical T-dual E = H2 ∈ H 2 (M, has a T-dual HT -bundle over M first Chern class c1 (E) t). E = (H 3 , H 2 , 0, 0) given by H 3 = H3 and H 2 = c1 (E). T-duality is neatly flux H encapsulated in the commutative diagram, 3 The conclusions in this section are valid for integral classes H ∈ H 3 (E, Z) as well since the component H0 does not carry torsion, and the remainder of the arguments is based on the results of [28, 29], which also hold for torsion H . For simplicity we state the results for differential forms only.
Nonassociative Tori and Applications to T-Duality
65
E ×M ?E ?? ?? ?? ? p pˆ ??? ?? E? E ?? ?? ?? ? π ?? ?? πˆ ? M
(10.1)
E A
f(z)
p
M z
Fig. 10.1. In the diagram, the fiber over z ∈ M is the noncommutative torus Af (z) , which is represented by a foliated torus, with foliation angle equal to f (z)
(2) If H0 = 0 and H1 = 0, then the T-dual is a continuous field of (stabilized) noncommutative tori Af over M, where the fiber over the point z ∈ M is equal to the rank k noncommutative torus Af (z) (see Fig. 10.1 above). Here f : M → T(2) is t). This map is not a continuous map representing H1 ∈ [M, T(2) ] ⊂ H 1 (M, ∧2 , unique, but the nonuniqueness does not affect its K-theory. (3) If H0 = 0 and if H = H0 , then the T-dual is a bundle of nonassociative tori Aφ (cf. Theorem 8.2) over M, where φ is the tricharacter associated to H0 . For general H , the T-dual is a continuous field of algebras that contains both the noncommutative torus and the nonassociative torus, and moreover, the T-dual is not unique, but the nonuniqueness occurs exactly as in part (2) above.
Part (1) was proved in [2, 3] when = 1 and in [4] for general . Part (2) was proved in [28] when = 2 and in [29] for general . Part (3) is what has been proved in this paper. A particular, but important case of Theorem 10.2 above is the following. 1. The T-dual of the torus T3 with no background flux is the dual torus T3 . This remains true if the background flux is topologically trivial. 2. (T3 , k dx ∧ dy ∧ dz) considered as a trivial circle bundle over T2 . The T-dual of (T3 , k dx ∧ dy ∧ dz) is the nilmanifold (HR /HZ , 0), where HR is the 3 dimensional
66
P. Bouwknegt, K. Hannabuss, V. Mathai
Heisenberg group and HZ the lattice in it defined by 1 x k1 z HZ = 0 1 y : x, y, z ∈ Z . 00 1
(10.2)
3. (T3 , k dx ∧ dy ∧ dz) considered as a trivial T2 -bundle over T. The T-dual of (T3 , k dx ∧ dy ∧ dz) is a continuous field of stabilized noncommutative tori, C ∗ (HZ ) ⊗ K, since k dx ∧ dy ∧ dz = 0. T2
4.
(T3 , k (T3 , k
dx ∧ dy ∧ dz) considered as a trivial T3 -bundle over a point. The T-dual of dx ∧ dy ∧ dz) is a nonassociative torus, Aφ (cf. Theorem 8.2), where φ is the
tricharacter associated to k dx ∧ dy ∧ dz, since
T3
k dx ∧ dy ∧ dz = 0.
We end with some speculations and open problems related to the results of the paper. In Sect. 11, we propose a natural definition of K-theory for the special nonassociative algebras that are considered in this paper. These are of the form A β,v G, where A is a C ∗ -algebra admitting a twisted action of the Abelian group G = R . We expect an analogue of Connes–Thom isomorphism theorem in K-theory to hold, showing that the K-theories of A and A β,v G are naturally isomorphic. This would then give further evidence that our definition of the T-dual of a principal torus bundle with H-flux is indeed correct. Finally, the K-theory of our special nonassociative algebras should be Morita invariant in our context, namely invariant under tensor product with twisted compact operators. Then the twisted Takai duality Theorem 9.2 would prove that T-duality applied twice returns us to the torus bundle with H-flux that we started out with. It remains to also determine the topological invariants of continuous fields of noncommutative tori and bundles of nonassociative tori as in the paper. This would then enable one to give a more symmetric characterization to the T-dual, similar to part (1) of the theorem above. We have an explicit conjecture for this, the explanation for which is in [5], namely, for the continuous field of noncommutative tori Af , there should be a “Chern class” invariant c1 (Af ) = (H2 , H1 , 0) satisfying dH2 + c1 (E) ∧ H1 = 0. In this case, we can add the following to part (2) of the theorem above. The T-dual Af is classified by its Chern class invariant c1 (Af ) = (H2 , H1 , 0) satis = (H 3 , H 2 , H 1 , 0), fying dH2 + c1 (E) ∧ H1 = 0 and dH1 = 0 and has T-dual H-flux H 3 = H3 , H 2 = c1 (E) and H 1 = 0. given by H Similarly, for the bundle of nonassociative tori Aφ with tricharacter φ associated to H0 , there should also be a “Chern class” invariant c1 (Aφ ) = (H2 , H1 , H0 ) satisfying dH2 + c1 (E) ∧ H1 = 0 and dH1 + c1 (E) ∧ H0 = 0. In this case, we can add the following to part (3) of the theorem above. The T-dual Aφ is classified by its Chern class invariant c1 (Aφ ) = (H2 , H1 , H0 ) satisfying dH2 + c1 (E) ∧ H1 = 0, dH1 + c1 (E) ∧ H0 = 0 and dH0 = 0. It has T-dual = (H 3 , H 2 , H 1 , H 0 ) given by H 3 = H3 , H 2 = c1 (E), H 1 = 0 and H 0 = 0. H-flux H What also remains to be done is T-duality for nonabelian principal bundles, where some of the ideas of this paper and [5] apply.
Nonassociative Tori and Applications to T-Duality
67
11. Nonassociative Algebras and Monoidal Categories – An Outlook Although the nonassociativity of the crossed product algebra appears to present a serious amendment to the notion of duality, that is not really the case. The fact that the same obstruction φ appears throughout is a signal that one should rather work in the monoidal category of C0 (G)-modules in which the isomorphism : (U ⊗ V ) ⊗ W → U ⊗ (V ⊗ W ) is given by the action of φ ∈ C(G × G × G), the multiplier algebra of C0 (G) ⊗ C0 (G) ⊗ C0 (G), [25, 12]. The cocycle identity for φ is equivalent to commutativity of the fundamental pentagonal diagram which ensures that all higher associators are consistent. By Fourier transforming we could identify the category as G-modules rather than C(G)-modules, which fits more directly into the framework of the duality theorem. The identity object is the trivial G-module 1 on C, which certainly has the property that, for any G-module U , U ⊗ 1 and 1 ⊗ U are naturally isomorphic to U . This is equivalent on C(G) to evaluating a function at the identity. Because φ vanishes when an argument is set equal to the identity, the two obvious maps from U ⊗ (1 ⊗ V ) = [(U ⊗ 1) ⊗ V ] to U ⊗ V are consistent. An algebra A is a monoid in this category, and the identification automatically takes care of the associativity. We can also define a left A-module M if one has a morphism A ⊗ M → M. A left A-module M is said to be projective if given any surjective morphism of left A-modules a : E → N and any morphism of left A-modules b : M → N , there is a morphism of left A-modules c : M → E such that a ◦ c = b. If A has a unit, then one can define the monoid V (A) consisting of isomorphism classes of finitely generated projective left A-modules under the direct sum operation. Then K0 (A) is defined as the Grothendieck group of V (A). If A does not have a unit, and A+ denotes A with a unit adjoined to it, then K0 (A) is defined as the kernel of the canonical morphism K0 (A+ ) → K0 (C) ∼ = Z. This will be studied in detail in a subsequent paper. modules is the algebra K (L2 (G)) of An example of a monoid in the category of G φ action twisted compact operators with the G (ξ · K)(x, y) = ξ(xy −1 )K(x, y) (ξ · ψ)(x) = ξ(x)ψ(x), is a module with K ⊗ ψ → and L2 (G), which has the G-action K ∗ ψ, where K(x, z)ψ(z) dz. (K ∗ ψ)(x) = G
actions are compatible since The G ξ(xz−1 )K(x, z)ξ(z)ψ(z) dz = (ξ(x)(K ∗ ψ))(x). ((ξ · K) ∗ (ξ · ψ))(x) = G
Then
(K1 ∗ (K2 ∗ ψ))(x) =
K1 (x, y)K2 (y, z)ψ(z) dydz, G
and the alternate bracketing (K1 ∗ K2 ) ∗ ψ must be computed as the image of (K1 ⊗ (K2 ⊗ ψ)), giving φ(x, y, z)−1 K1 (x, y)K2 (y, z)ψ(z) dydz, ((K1 ∗ K2 ) ∗ ψ)(x) = G
consistent with the multiplication law on the twisted kernels.
68
P. Bouwknegt, K. Hannabuss, V. Mathai
but then the action on One can alternately work with the C0 (G) action rather than G kernels requires a use of the coproduct (f )(x, y) = f (xy), so that (f · K)(x, y) = f (xy −1 )K(x, y). One can similarly define right A-modules, and also bimodules for two algebras A1 and A2 . It is also possible to look at A1 -A2 -bimodules X which have an action of a prod1 × G 2 with maps 1 and 2 defining the associativity properties, and are uct group G 1 -modules as left A1 -modules and in the category of G 2 -modules as in the category of G right A2 -modules. Such a bimodule can be used to set up a Morita equivalence between left A2 -modules and left A1 -modules, by mapping a left A2 -module M to the quotient of X ⊗ V by the equivalence relation (x.b) ⊗ ψ ∼ 2 (x ⊗ (b.ψ)), for x ∈ X, b ∈ A2 and ψ ∈ M. This allows us to define Morita equivalence between algebras with different kinds of associativity. In particular, if we take A1 = Kφ (L2 (G)), A2 = K(L2 (G)), with X = K(L2 (G)), equipped with the usual right multiplication action of A2 and the left multiplication action of A1 defined above, then we have Morita equivalence between the twisted and untwisted algebras. Clearly this is only an outline of some of the ideas arising out of this new perspective on nonassociativity, and we shall explore these in more detail in the sequel to this paper. Since posting this paper on the arXives [1] has come to our attention, which also investigates some nonassociative algebras albeit in a rather different context. Acknowledgements. We would like to thank Edwin Beggs, Alan Carey and Ulrike Tillmann for useful comments. KCH would like to thank the University of Adelaide for hospitality during the initial stages of the project. PB and VM were financially supported by the Australian Research Council.
References 1. Akrami, S.E., Majid, S.: Braided cyclic cocycles and nonassociative geometry. J. Math. Phys. 45, 3883–3911 (2004) 2. Bouwknegt, P., Evslin, J., Mathai, V.: T-duality: Topology change from H-flux. Commun. Math. Phys. 249, 383–415 (2004) 3. Bouwknegt, P., Evslin, J., Mathai, V.: On the topology and H-flux of T-dual manifolds. Phys. Rev. Lett. 92, 181601 (2004) 4. Bouwknegt, P., Hannabuss, K.C., Mathai, V.: T-duality for principal torus bundles. J. High Energy Phys. 03, 018 (2004) 5. Bouwknegt, P., Hannabuss, K.C., Mathai, V.: T-duality for principal torus bundles and dimensionally reduced Gysin sequences. Adv. Theor. Math. Phys. 9(5), 749–773 (2005) [hep-th/0412268] 6. Brown, K.S.: Cohomology of groups. New York–Berlin: Springer Verlag, 1982 7. Busby, R.C., Smith, H.A.: Representations of twisted group algebras. Trans. Amer. Math. Soc. 149, 503–537 (1970) 8. Carey, A.L.: The origin of three-cocycles in quantum field theory. Phys. Lett. B194, 267–270 (1987) 9. Carey, A.L., Hannabuss, K.C., Mathai, V., McCann, P.: The quantum Hall effect in hyperbolic space. Commun. Math. Phys. 190, 629–673 (1998) 10. Carey, A.L., Mickelsson, J.: The universal gerbe, Dixmier–Douady class, and gauge theory. Lett. Math. Phys. 59, 47–60 (2002) 11. Crocker, D., Kumjian, A., Raeburn, I., Williams, D.: An equivariant Brauer group and actions of groups on C∗ algebras. J. Funct. Anal. 146, 151–184 (1997) 12. Chari, V., Pressley, A.: A guide to quantum groups. Cambridge: Cambridge University Press, 1994 13. Cornalba, L., Schiappa, R.: Nonassociative star product deformations for D-brane worldvolumes in curved backgrounds. Commun. Math. Phys. 225, 33–66 (2002) 14. Echterhoff, S., Rosenberg, J.: Fine structure of the Mackey machine for actions of abelian groups with constant Mackey obstruction. Pacific J. Math. 170, 17–52 (1995) 15. Eilenberg, S., MacLane, S.: Cohomology theory in abstract groups II. Ann. Math. 48, 326–341 (1947) 16. Eilenberg, S., MacLane, S.: Algebraic cohomology and loops. Duke Math. J. 14, 435–463 (1947) 17. Green, P.: The structure of imprimitivity algebras. J. Funct. Anal. 36, 88–104 (1980)
Nonassociative Tori and Applications to T-Duality
69
18. Greub, W., Halperin S., Vanstone, R.: Connections, curvature, and cohomology. Vols II and III, New York: Academic Press, 1973 19. Hannabuss, K.C.: Representations of nilpotent locally compact groups. J. Funct. Anal. 34, 146–165 (1979) 20. Holt, D.F.: Cohomology groups H n (G, M). J. Alg. 60, 307–320 (1979) 21. Huebschmann, J.: Crossed n-fold extensions of groups and cohomology. Commun. Math. Helv. 55, 302–313 (1980) 22. Jackiw, R.: Three-cocycles in mathematics and physics. Phys. Rev. Lett. 54, 159–162 (1985) 23. Kleppner, A.A.: Multipliers on abelian groups. Math. Ann. 158, 11–34 (1965) 24. Leptin, H.: Verallgemeinerte L1 -Algebren. Math. Ann. 159, 51–76 (1965) 25. MacLane, S.: Categories for the working mathematician. New York: Springer Verlag, 1971 26. MacLane, S.: Historical Note. J. Alg. 60, 319–320 (1979) 27. MacLane, S., Whitehead, J.H.C.: On the 3-type of a complex. Proc. Nat. Acad. Sci. U.S.A. 30, 41–48 (1950) 28. Mathai, V., Rosenberg, J.: T-duality for torus bundles with H-fluxes via noncommutative topology. Commun. Math. Phys. 253, 705–721 (2005) 29. Mathai, V., Rosenberg, J.: On mysteriously missing T-duals, H-flux and the T-duality group. http://arxiv.org/list/hep-th/0409073, 2004, to appear in “Proceedings of the XXXIII International Conference of Differential Geometric Methods in Mathematical Physics” (August 2005), editors Mo-Lin Ge and Weiping Zhang, World Scientific 2006 [hep-th/0409073] 30. Packer, J.A., Raeburn, I.: Twisted crossed products of C∗ -algebras. Math. Proc. Camb. Phil. Soc. 106, 293–311 (1989) 31. Quigg, J.C.: Duality for reduced twisted crossed products of C ∗ -algebras. Indiana Univ. Math. J. 35, 549–571 (1986) 32. Raeburn, I., Sims, A., Williams, D.: Twisted actions and obstructions in group cohomology. In: Cuntz, J., Echterhoff S. (ed.,) C ∗ -algebras. Berlin: Springer Verlag, 2000 33. Raeburn, I., Williams, D.: Morita equivalence and continuous trace C ∗ algebras. Mathematical Surveys and Monographs of the A.M.S. 60, Providence, RI: Amer. Math. Soc., 1998 34. Ratcliffe, J.G.: Crossed extensions. Trans. Amer. Math. Soc. 257, 73–89 (1980) 35. Whitehead, J.H.C.: Combinatorial homotopy II. Bull. Amer. Math. Soc. 55, 453–496 (1949) Communicated by M.R. Douglas
Commun. Math. Phys. 264, 71–85 (2006) Digital Object Identifier (DOI) 10.1007/s00220-005-1509-0
Communications in
Mathematical Physics
Homological Mirror Symmetry for Toric del Pezzo Surfaces Kazushi Ueda Research Institute for Mathematical Sciences, Kyoto University, Kyoto 606-8502, Japan E-mail:
[email protected] Received: 21 January 2005 / Accepted: 10 August 2005 Published online: 15 Febraury 2006 – © Springer-Verlag 2006
Abstract: We prove the homological mirror conjecture for toric del Pezzo surfaces. In this case, the mirror object is a regular function on an algebraic torus (C× )2 . We show that the derived Fukaya category of this mirror coincides with the derived category of coherent sheaves on the original manifold.
1. Introduction Mirror symmetry started as a mysterious relationship between complex geometry of a Calabi-Yau 3-fold and symplectic geometry of another Calabi-Yau 3-fold called the mirror manifold. In 1994, Kontsevich [11] proposed a program to understand various mirror phenomena as a consequence of the following homological mirror conjecture: Calabi-Yau manifolds always come in pairs in such a way that the derived category of coherent sheaves on one manifold is equivalent as a triangulated category to the derived Fukaya category of the other. Although mirror symmetry was first discovered for Calabi-Yau manifolds, there are also variants of these phenomena for other classes of manifolds. One such example is Givental’s theorem [8] giving integral representations of J -functions of toric Fano manifolds. See also [3, 19, 6, 10]. We can also formulate the homological mirror conjecture for toric Fano manifolds. Let X be a toric Fano manifold of dimension n. Then the mirror partner is a regular function W on an algebraic torus (C× )n of the same dimension equipped with a symplectic structure (along with additional data called a grading). Here, the function W is a Newton polynomial for the convex hull of the generators of a 1-dimensional cone of the fan of X. Coefficients of this polynomial do not matter as long as they are chosen
Supported by JSPS Fellowships for Young Scientists No.15-5561.
72
K. Ueda
general enough. By Kouchnirenko [12], W has exactly dim H ∗ (X, C) critical points. Take a regular value t of W and a distinguished basis of vanishing cycles in W −1 (t). We also have to choose a grading and a spin structure on each of these vanishing cycles. The directed Fukaya category Fuk→ W of W (along with the choice of gradings and spin structures) is an A∞ -category whose objects are vanishing cycles and whose morphisms are Floer complexes. Roughly speaking, the Floer complex between two vanishing cycles are the vector space spanned by intersection points between them, and the compositions of morphisms are given by “counting polygons.” By Seidel [16], the derived category D b Fuk→ W of Fuk→ W is independent of the choice of a distinguished basis of vanishing cycles. The following is the main result of this paper: Theorem 1. The derived category of coherent sheaves on a toric del Pezzo surface X is equivalent as a triangulated category to the derived category of the directed Fukaya category of the mirror W of X; D b coh(X) ∼ = D b Fuk→ W. Our proof is based on an explicit computation. The above Theorem 1 extends the works of Seidel [17] and Auroux, Katzarkov and Orlov [1], where the cases of P2 , P1 × P1 , and P2 blown-up at one point are treated. See also the paper by Hori, Iqbal and Vafa [9], where homological mirror symmetry for Fano manifolds are discussed from a physics point of view. Note added. After the submission of this paper, a preprint [2] appeared which contains the proof of homological mirror symmetry for del Pezzo surfaces which are not necessarily toric.
2. Derived Category of Coherent Sheaves We describe the structure of the bounded derived category D b coh(Y ) of coherent sheaves on Y in this section, where Y is the projective plane blown-up at three points p1 , p2 and p3 in general position. Let φ : Y → P2 be this blow-up and E1 , E2 and E3 be the exceptional divisors corresponding to p1 , p2 and p3 respectively. It is a toric surface and the generators of one-dimensional cones of its fan is drawn in Fig. 1.
v2
v4
v3
v6
v1
v1 v2 v3 v4 v5 v6
= = = = = =
v5 Fig. 1. The toric data of Y
(1, 0), (0, 1), (−1, −1), (−1, 0), (0, −1), (1, 1).
Homological Mirror Symmetry for Toric del Pezzo Surfaces
73
Definition 1. 1. An object E in a triangulated category is exceptional if C if i = 0, i Ext (E, E) = 0 otherwise. 2. An ordered set of objects (Ei )N i=1 in a triangulated category is an exceptional collection if each Ei is exceptional and Ext k (Ei , Ej ) = 0 for any i > j and for any k. By combining theorems of Beilinson [4] and Orlov [13], we have the following generators of D b coh(Y ): Theorem 2. Let C = (E1 , E2 , E3 , E4 , E5 , E6 ), where E1 = OE1 (−1)[−1], E2 = OE2 (−1)[−1], E3 = OE3 (−1)[−1], E4 = φ ∗ OP2 (−1), E5 = φ ∗ P2 (1), E6 = OY .
(1)
Then C is an exceptional collection generating D b coh(Y ). Here, OEi (−1) is the sheaf supported on Ei and is isomorphic to the tautological sheaf OP1 (−1) on Ei ∼ = P1 , [•] is the shift operator in the derived category, φ ∗ is the derived pull-back, OP2 (−1) is the tautological sheaf on P2 , and P2 (1) is the cotangent sheaf of P2 tensored with the hyperplane sheaf OP2 (1). Proposition 1. All the non-zero Ext-groups within the exceptional collection C are Hom(Ei , E4 ) Hom(Ei , E5 ) Hom(Ei , E6 ) Hom(E4 , E5 ) Hom(E4 , E6 ) Hom(E5 , E6 )
= = = = = =
Cpi , Ker(V ∨ → (Cpi )∨ ), C, 2 V ∨ , V ∨, V,
where i = 1, 2, 3, V ∼ = C3 is the three-dimensional vector space such that P2 = P(V ), the check denotes the dual vector space, Cpi ⊂ V is the one-dimensional subspace corresponding to pi ∈ P(V ), and the map V ∨ → (Cpi )∨ is the dual of the inclusion Cpi → V . Compositions of morphisms are given by
(v, ω)
∈
∈
Hom(Ei , E4 ) × Hom(E4 , E5 ) −→ Hom(Ei , E5 ) −→
−ιv ω
,
(v, ω)
∈
∈
Hom(Ei , E4 ) × Hom(E4 , E6 ) −→ Hom(Ei , E6 ) −→
ω(v)
,
74
K. Ueda
∈
∈
Hom(Ei , E5 ) × Hom(E5 , E6 ) −→ Hom(Ei , E6 ) −→
(ω, v)
,
ω(v)
∈
∈
Hom(E4 , E5 ) × Hom(E5 , E6 ) −→ Hom(E4 , E6 ) −→
(ω, v)
ιv ω,
where ιv : ∧2 V ∨ → V ∨ for v ∈ V is the interior product. Proof. All the computations reduce to those on P2 in the following way: Let E and F be vector bundles on P2 . Then RHom(φ ∗ E, φ ∗ F) = RHom(E, Rφ∗ φ ∗ F) = RHom(E, F ⊗ Rφ∗ OY ) = RHom(E, F), RHom(φ ∗ E, OEi (−1)) = RHom(E, Rφ∗ OEi (−1)) = 0, and RHom(OEi (−1)[−1], φ ∗ E) = RHom({OY → OY (Ei )}, φ ∗ E) = R({OY (−Ei ) → OY } ⊗ φ ∗ E) = R({Rφ∗ OY (−Ei ) → Rφ∗ OY } ⊗ E) = R({Ipi → OP2 } ⊗ E) = R(Opi ⊗ E) = E|pi , where Ipi is the ideal sheaf of pi and the last line is the fiber at pi . Here, we have used the exact sequence 0 → OY → OY (Ei ) → OEi (−1) → 0. We can further use the exact sequence 0 → P2 (1) → V ∨ ⊗ OP2 → OP2 (1) → 0 to reduce the computations involving P2 (1) to those involving OP2 (i), i ∈ Z: RHom(OEi (−1)[−1], φ ∗ P2 (1)) = R(Opi ⊗ {V ∨ ⊗ OP2 → OP2 (1)}) = {V ∨ → (Cpi )∨ }, RHom(φ ∗ OP2 (−1), φ ∗ P2 (1)) = RHom(OP2 (−1), {V ∨ ⊗ OP2 → OP2 (1)}) = R(OP2 (1) ⊗ {V ∨ ⊗ OP2 → OP2 (1)}) = R({V ∨ ⊗ OP2 (1) → OP2 (2)}) = {V ∨ ⊗ V ∨ → Sym2 V ∨ } = ∧2 V ∨ ,
Homological Mirror Symmetry for Toric del Pezzo Surfaces
75
RHom(φ ∗ P2 (1), OY ) = RHom({V ∨ ⊗ OP2 → OP2 (1)}, OP2 ) = R({OP2 (−1) → V ⊗ OP2 }) = V. Compositions of morphisms can be easily read off from the above computations.
3. Fukaya Category First we recall the definition of an A∞ -category. For a Z-graded vector space V and i ∈ Z, V [i] denotes the shift of V by i; V [i]j = V i+j . Definition 2. An A∞ -category A consists of 1. the set of objects Ob(A), 2. for any c1 , c2 ∈ Ob(A), a Z-graded C-vector space homA (c1 , c2 ) called the set of morphisms, 3. for any positive integer k and for any set of objects {ci }ki=0 , the composition mk : homA (c0 , c1 )[1] ⊗ · · · ⊗ homA (ck−1 , ck )[1] −→ homA (c0 , ck )[1] which is a linear map of degree 1 such that for any positive integer k, any set of objects {ci }ki=0 and any set of morphisms {ai }ki=1 , ai ∈ homA (ci−1 , ci ), the following A∞ -relations hold: k k−1
(−1)deg a1 +···+deg ai mi−j +k+1 (a1 ⊗ · · · ⊗ ai
i=0 j =i+1
⊗mj −i (ai+1 ⊗ · · · ⊗ aj ) ⊗ aj +1 ⊗ · · · ⊗ ak ) = 0. Here degrees are counted after shifts, i.e., if a ∈ V i , deg a = i − 1 in V [1]. Since m21 = 0, we define HomA (c1 , c2 ) = H 0 (homA (c1 , c2 ), m1 ). The Fukaya category of Lagrangian submanifolds in a symplectic manifold is defined in [7]. We use the following adaptation for exact Morse fibrations by Seidel [16]. Let W : Z → C be a regular function on an affine algebraic manifold Z of complex dimension n with a K¨ahler structure. Assume the following conditions: – The K¨ahler metric is complete. – The symplectic form ω of Z is exact, i.e., there exists a one form θ on Z such that ω = dθ. – At any critical points of W , the Hessian of W is non-degenerate. – All the critical values are distinct. Such W gives rise to an exact Morse fibration in the terminology of [16]. Using the K¨ahler structure, we can define the lift cp : [0, 1] → Z of a path c : [0, 1] → C starting from a point p ∈ Z such that W (p) = c(0) by using the horizontal distribution defined as the orthogonal complement of the tangent space along the fiber of W . Assume that the origin is a regular value of W and fix an order (pi )N i=1 on the set of of vanishing paths is a set of smooth critical points of W . A distinguished set (ci )N i=1 paths ci : [0, 1] → C satisfying
76
1. 2. 3. 4. 5.
K. Ueda
ci (0) = 0, ci (1) = W (pi ), ci has no self-intersection, images of ci and cj intersect only at the origin, ci (0) = 0, and
(0) < arg c (0), i = 1, . . . N − 1, for some choice of the branch of arg(•). arg ci+1 i
Given a distinguished set (ci )N i=1 of vanishing paths, the corresponding vanishing cycles are defined by (Ci )N i=1 cp (t) = pi }. Ci = {p ∈ W −1 (0) | lim t→1
They are Lagrangian submanifolds of W −1 (0). The directed Fukaya category Fuk→ W for W is roughly an A∞ -category whose objects are vanishing cycles and whose morphisms are Lagrangian intersection Floer complexes. To define a Z-grading on the Floer complex, we need the concept of grading on Lagrangian submanifolds introduced by Kontsevich [11], which we now recall. See also Seidel [15]. Let (M, ω) be a symplectic manifold of dimension 2n. An almost complex structure on M is a section J ∈ (M, End(T M)) such that J 2 = − id. J is called ω-compatible if g(V1 , V2 ) = ω(V1 , J V2 ) defines a Hermitian metric on the tangent bundle. Fix an ω-compatible almost complex structure J of M and let S be the principal U (1)-bundle associated to the complex line bundle (n (T ∗ M, J ))⊗2 . The fiber of S at p ∈ M is (n (Tp∗ M, J ))⊗2 /R>0 . We assume that the first Chern class of this complex line bundle vanishes, so that S has a section. A grading of M is a choice of a section : M → S. Fix a grading on M. Let LagM → M be the Lagrangian Grassmannian bundle on M, whose fiber at p ∈ M is the Grassmannian of Lagrangian subspaces in the symplectic vector space Tp M. Define det 2 : LagM → U (1) as follows: For a Lagrangian subspace L ⊂ Tp M, pick any basis {ei }ni=1 of L and take the square of their exterior product; (e1 ∧ e2 ∧ · · · ∧ en )⊗2 ∈ (n (Tp M, J ))⊗2 . det 2 (L) is the image of this element by (p). A Lagrangian submanifold L ⊂ M gives a canonical section sL : L p → Tp L ∈ LagM |p . Denote the composition of sL and det 2 by φL . A grading of a Lagrangian submanifold is a lift φ L : L → R of φL to the universal cover R of U (1) ∼ = R/Z. Now we define the Maslov index. A smooth path : [0, 1] → Lag(V , ω) in the Lagrangian Grassmannian of a fixed symplectic vector space (V , ω) of dimension 2n is called crossingless if (0) ∩ (t) = (0) ∩ (1) for all t ∈ (0, 1]. For a crossingless path t , its differential (0) at t = 0 gives an element of the tangent space T(0) Lag(V , ω) ⊂ T(0) Gr(n, V ) ∼ = Hom((0), V /(0)), where Gr(n, V ) is the Grassmannian of n-dimensional subspaces in V . The composition of (0) and ω defines a quadratic form (0) v → ω(v, (0)v) on (0), which descends to a quadratic form on (0)/(0) ∩ (1) since it vanishes on (0) ∩ (1). The resulting form on (0)/(0) ∩ (1) is called the crossing form [14]. For an intersection p ∈ L1 ∪ L2 of two graded Lagrangian submanifolds (L1 , φ L1 ) and (L2 , φ L2 ), its Maslov index is defined as follows: Choose a crossingless path : [0, 1] → LagM |p from Tp L0 to Tp L1 such that the corresponding crossing form at t = 0 is negative definite. There is a unique lift α : [0, 1] → R of the composition α of : [0, 1] → LagM |p and α (0) = φ det2 (p) : LagM |p → U (1) such that L0 (p), and the Maslov index I (p) of the intersection p of two graded Lagrangian submanifolds is defined by α (1). I (p) = φ L1 (p) −
Homological Mirror Symmetry for Toric del Pezzo Surfaces
77
Now we come back to our exact Morse fibration W . A relative Maslov map is a sec⊗2 of the top exterior product of the relative tion of the second tensor power (n−1 Z/C ) cotangent bundle away from the critical points. Since 0 is a regular value, the restriction of to W −1 (0) gives a grading of W −1 (0). Assume that all the vanishing cycles can be graded, and fix a grading on each vanishing cycle. We also need to choose a spin structure on each vanishing cycle in order to orient the moduli spaces of pseudoholomorphic maps [7]. Since each vanishing cycle is homeomorphic to a sphere, the choice of a spin structure is unique except for dimC W −1 (0) = 1, where one has as many as H 1 (S 1 , Z/2Z) = Z/2Z choices. We take the non-trivial spin structure in such a case. Let Ci denote the vanishing cycle Ci endowed with the above grading and spin structure. We assume that vanishing cycles intersect each other transversally. This condition can always be met by moving vanishing cycles within their Hamiltonian isotopy classes if necessary. Definition 3. Given a function W on an affine K¨ahler manifold Z together with a relative Maslov map and a choice of a distinguished basis of vanishing cycles (Ci )N i=1 with → gradings and spin structures, its directed Fukaya category Fuk W is an A∞ -category such that – the set of objects is the distinguished basis of vanishing cycles; Ob(Fuk→ W ) = (C1 , . . . , CN ),
– the set of morphisms between Ci and Cj is the Z-graded vector space 0 · idC homFuk→ W (Ci , Cj ) = C i
p∈Ci ∩Cj
i > j, i = j, spanC {p} i < j,
where deg p = I (p) (the Maslov index), and – for a positive integer k, a set of objects (Ci0 , . . . , Cik ) and morphisms pl ∈ Cil−1 ∩ Cil for l = 1, . . . , k, the composition mk is given by mk (p1 , . . . , pk ) =
#Mk+1 (Ci0 , . . . , Cik ; p0 , . . . , pk )p0 .
p0 ∈Ci0 ∩Cik
Here, Mk+1 (Ci0 , . . . , Cik ; p0 , . . . , pk ) is the stable compactification of Mk+1 (Ci0 , . . . , Cik ; p0 , . . . , pk ) defined as follows: A disk with k + 1 marked points on the boundary is a pair (D 2 , (z0 , . . . , zk )) of a closed unit disk D 2 = {z ∈ C | |z| ≤ 1} and an ordered set (z0 , . . . , zk ) of k + 1 points on the boundary respecting the cyclic order. Let ∂l D 2 ∈ ∂D 2 be the interval between zl and zl+1 , where we set zk+1 = z0 . Fix an almost complex structure J on W −1 (0). A smooth map ϕ : D 2 → M is called pseudoholomorphic if dϕ ◦ JD 2 = J ◦ dϕ, where JD 2 is the canonical complex structure on D 2 . Mk+1 (Ci0 , . . . , Cik ; p0 , . . . , pk ) is the moduli space of pairs ((D 2 , (z0 , . . . , zk )), ϕ) such that
78
1. 2. 3. 4.
K. Ueda
(D 2 , (z0 , . . . , zk )) is a disk with k + 1 marked points on the boundary, ϕ : D 2 → W −1 (0) is a pseudoholomorphic map, ϕ(∂l D 2 ) ⊂ Cil for l = 0, . . . , k, and ϕ(zl ) = pl for l = 0, . . . , k.
Although Mk+1 (Ci0 , . . . , Cik ; p0 , . . . , pk ) is not an honest manifold in general due to subtleties in the definition of the moduli space of pseudoholomorphic maps, it has a Kuranishi structure with corners and it makes sense to “count” numbers #Mk+1 (Ci0 , . . . , Cik ; p0 , . . . , pk ) of points. See [7] for details. These numbers are counted with signs determined by the orientation of the moduli space, which in turn is determined by the spin structures on vanishing cycles. #Mk+1 (Ci0 , . . . , Cik ; p0 , . . . , pk ) is zero if dim Mk+1 (Ci0 , . . . , Cik ; p0 , . . . , pk ) = 0. These moduli spaces must be compact so that it makes sense to count the numbers of points. To prove compactness of the moduli space, one needs the boundedness of the energy. Usually, this is achieved by introducing the Novikov ring as the coefficient ring of the Floer complex in order to control energies of pseudoholomorphic maps. This is not necessary in our case since vanishing cycles are exact Lagrangian submanifolds, i.e., [θ ] = 0 ∈ H 1 (Ci , R) for the primitive θ of ω; ω = dθ . To see this, let C be the vanishing cycle
corresponding to a vanishing path c : [0, 1] → C. Then C bounds the disk D = p∈C cp ([0, 1]); ∂D = C. Since dθ|D = ω|D = 0, [θ] = 0 ∈ H 1 (∂D, R). Therefore, there exists a function Ki on each vanishing cycles such that θ|Ci = dKi . Then for any ϕ ∈ Mk+1 (Ci0 , . . . , Cik ; p0 , . . . , pk ), the energy of ϕ is E(ϕ) = =
D2 k
ϕ∗ω =
∂D 2
ϕ∗θ =
k l=0
ϕ∗θ
∂i D 2
(Kil (pl+1 ) − Kil (pl ))
l=0
and does not depend on the homotopy class of ϕ. This also implies the vanishing of m0 . Non-compactness of W −1 (0) does not cause any problem either. Since W −1 (0) is Stein, the maximum principle prevents pseudoholomorphic disks from running away to infinity, assuring the compactness of the moduli space. Seidel proved that although the directed Fukaya category Fuk→ W depends on the choice of the distinguished basis of vanishing cycles, different choices are related by mutations, hence its derived category is invariant. Now we explain the mirror construction of toric Fano manifolds after Givental [8]. Given a fan of an n-dimensional toric Fano manifold, let {vi }ri=1 , vi = (vi1 , . . . , vin ) ∈ Zn , be the set of generators of its one-dimensional cones. Then the mirror object for this toric Fano manifold is the regular function W (x1 , . . . , xn ) =
r
qi x1vi1 · · · xnvin
(2)
i=1
on the algebraic torus SpecC[xi±1 ]ni=1 . Here, qi ’s are parameters corresponding to the deformation of symplectic structures on the toric Fano manifold. Therefore, the mirror of our Y is the regular function W (x, y) = q1 x + q2 y +
q4 q5 q3 + + + q6 xy xy x y
(3)
Homological Mirror Symmetry for Toric del Pezzo Surfaces
79
on the algebraic torus (C× )2 = SpecC[x, x −1 , y, y −1 ] (see Fig. 1). The Fukaya category does not depend on a general choice of qi ’s. We have used (q1 , q2 , q3 , q4 , q5 , q6 ) = (1, 1, 1, 0.215, 0.25, 0.3) to draw the figures appearing below. Equip (C× )2 with the d|y| symplectic form d|x| |x| ∧ d(arg x) + |y| ∧ d(arg y) and the relative Maslov map
= (Res
dx dx ∧ dy ⊗2 ) =( )⊗2 . xyW (x, y) ∂y (xyW (x, y))
W −1 (0) is an affine elliptic curve, which can be compactified by adding six points. We depict the critical values of W and our choice of a distinguished set of vanishing paths in Fig. 2. Figure 3 shows the corresponding vanishing cycles. The opposite sides of the square in Fig. 3 are identified to form a two-dimensional torus. The open circles denote the points which are missing due to the non-compactness of (C× )2 . We have drawn Fig. 3 in the following way: First regard W −1 (t) as a branched double cover of C× by the projection πt : W −1 (t) (x, y) → y ∈ C× . For any i = 1, . . . , 6, the branch points of πci (t) moves in C× as we vary t, until two of them finally collides at t = 1. The vanishing cycle Ci is the circle in W −1 (0) over this trajectory of collision. This determines the vanishing cycle Ci up to isotopy. These vanishing cycles can be straightened within their Hamiltonian isotopy classes. Here, being straight refers to the flat metric on the torus W −1 (0), where • denotes the completion. Note that this flat metric on W −1 (0) has nothing to do with the K¨ahler metric on W −1 (0) induced from that of (C× )2 . This determines Ci up to translation. This translational ambiguity can be fixed by imposing exactness, which requires the knowledge of the primitive θ of the symplectic form ω on W −1 (0). We do not try to carry this out since translations of straight Ci ’s do not alter the combinatorial structure of intersections, which is all we need in our computation of the Fukaya category below. The grading on W −1 (0) given
c4
c5
c3
c2 c1 c6
Fig. 2. Distinguished basis of vanishing paths
80
K. Ueda
by is the grading coming from the restriction of the second tensor power of the holo morphic 1-form on W −1 (0). Then one can choose a grading φ Ci on each vanishing cycle Ci so that all the Maslov indices of Ci ∩ Cj are zero for i < j . To give a spin structure on Ci is the same as to give a two-fold covering of Ci . We take the non-trivial cover such that two branches interchange on the black dots in Fig. 3. Different choices for spin structures lead to different categories, and it turns out that the above choice gives a category derived equivalent to the category of coherent sheaves on the del Pezzo surface Y . Let Ci denote the vanishing cycle Ci with the above grading and spin structure. Since the Maslov indices of all the intersection points are zero, the sign for the counting of a triangle is −1 if the boundary of the triangle hits odd numbers of these dots, and +1 otherwise (see Seidel [18]). Note that although the choice of spin structures is important, the choice of the positions of the black dots is irrelevant. The change of the signs caused by a change of the positions of the black dots can be absorbed by a redefinition of the signs of the basis for the Floer cohomologies. mk is non-zero only for k = 2 (polygons with more than three edges in Fig. 3 do not contribute since homFuk→ W (Ci , Cj ) = 0 for i > j ). Theorem 3. There exists an isomorphism
φij : HomD b coh(Y ) (Ei , Ej ) → HomFuk→ W (Ci , Cj ) −x1∨
x3
e3 x2∨ x1∨
x3∨ e1
C3 C5
x2∨
x1 x1∨ ∧ x2∨
x3
C1 C4
x2∨ x3∨
x2
x1∨
e2
C6 C2
x2∨
∧ x3∨
x1
x3∨ x2 x3∨ ∧ x1∨ Fig. 3. Vanishing cycles in the fiber at the origin
Homological Mirror Symmetry for Toric del Pezzo Surfaces
81
as a C-vector space for i, j = 1, . . . , 6 such that the diagrams HomD b coh(Y ) (Ei , Ej ) × HomD b coh(Y ) (Ej , Ek ) −−−−→ HomD b coh(Y ) (Ei , Ek ) φ ×φ φ
ij j k
ik
(4)
HomFuk→ W (Ci , Cj ) × HomFuk→ W (Cj , Ck ) −−−−→ HomFuk→ W (Ci , Ck ) commute. Proof. We omit the lower suffix of Hom(•, •) since there is no danger of confusion. We construct the above φij ’s explicitly. Since m1 = 0, Hom(Ci , Cj ) is isomorphic
to hom(Ci , Cj ) and spanned by intersection points of Ci and Cj , to which we assign the basis of Hom(Ei , Ej ) as in Fig. 3. Notations for the basis of Hom(Ei , Ej ) is given in Appendix. This defines the linear isomorphisms φij ’s. The commutativity of (4) is verified by counting triangles. Let us illustrate this with a few examples. Take x1 ∈ Hom(E1 , E4 ) and x1∨ ∈ Hom(E4 , E6 ), whose composition is e1 ∈ Hom(E1 , E6 ). On the Fukaya side, we count the number of triangles whose edges are contained in C1 ∪C4 ∪C6 and two of whose vertices are x1 and x1∨ . Such a triangle exists uniquely, whose remaining vertex is e1 (Fig. 4). Next we try to compose x3 ∈ Hom(C3 , C4 ) and x1∨ ∈ Hom(C4 , C6 ) in the Fukaya category. The only possibility would be the triangle with e3 as the remaining vertex. However, this is not allowed because of the missing point (Fig. 5). This shows that the composition of x3 and x1∨ in the Fukaya category is zero, which matches the table in Appendix calculated in D b coh(Y ). Finally, as an example of counting with signs, we compute Hom(C1 , C4 ) × Hom(C4 , C5 ) (x1 , x1∨ ∧ x2∨ ) → −x2∨ ∈ Hom(C1 , C5 ).
C6 x1∨ e1
x1 C1 C4
Fig. 4. Counting triangle
82
K. Ueda
x3
e3 C3 x1∨
C6
C4
Fig. 5. No triangles
C5
x2∨ x1 x1∨ ∧ x2∨
C1 C4
Fig. 6. Counting triangles with signs
Indeed, the triangle whose vertices are x1 , x1∨ ∧ x2∨ , and x2∨ contains one black dot on its edge (Fig. 6). We can do similar analysis for the rest of the triangles. The result perfectly agrees with the computations in D b coh(Y ). Although the derived category of an A∞ -category is usually defined using twisted complexes, we adopt the following definition in this paper since our Fuk→ W satisfies homkFuk→ W (Ci , Cj ) = 0 for k = 0 and mk = 0 for k = 2.
Homological Mirror Symmetry for Toric del Pezzo Surfaces
83
Definition 4. Let A = ⊕i,j HomFuk→ W (Ci , Cj ) be the total morphism algebra. The derived Fukaya category D b Fuk→ W is the bounded derived category D b (mod-A) of the category of right finite-dimensional modules over the algebra A. Now we can state our main theorem: Theorem 4. There exists an equivalence of triangulated categories D b coh(Y ) ∼ = D b Fuk→ W. Proof. From Theorem 3, we have A∼ = ⊕i,j HomD b coh(Y ) (Ei , Ej ). Theorem 4 follows immediately from the theorem of Bondal [5] that D b coh(Y ) ∼ = D b (mod-(⊕i,j Hom(Ei , Ej ))). We can perform a similar analysis for all the other toric del Pezzo surfaces to obtain Theorem 1. Acknowledgement. We thank O. Fujino, K. Fukaya, H. Iritani, T. Kawai, K. Saito and A. Takahashi for valuable discussions and comments.
4. Appendix Here we give the table of compositions of morphisms in the derived category of coherent sheaves on the projective plane P(V ) blown-up at p1 = [1 : 0 : 0], p2 = [0 : 1 : 0], p3 = [0 : 0 : 1]. We use the following notations: Ei ’s are defined in (1), {xi }3i=1 is the basis of V , {xi∨ }3i=1 is the dual basis of V ∨ , and ei is the generator of Hom(Ei , E6 ) = C for i = 1, 2, 3: Hom(E1 , E4 ) × Hom(E4 , E5 ) → Hom(E1 , E5 ) x1
x1∨ ∧ x2∨ x2∨ ∧ x3∨ x3∨ ∧ x1∨ , −x2∨ 0 x3∨
Hom(E1 , E4 ) × Hom(E4 , E6 ) → Hom(E1 , E6 ) x1∨ x2∨ x3∨ , x1 e1 0 0 Hom(E1 , E5 ) × Hom(E5 , E6 ) → Hom(E1 , E6 ) x1 x2 x3 x2∨ 0 e1 0 , x3∨ 0 0 e1 Hom(E2 , E4 ) × Hom(E4 , E5 ) → Hom(E2 , E5 ) x1∨ ∧ x2∨ x2∨ ∧ x3∨ x3∨ ∧ x1∨ , x2 x1∨ −x3∨ 0
84
K. Ueda
Hom(E2 , E4 ) × Hom(E4 , E6 ) → Hom(E2 , E6 ) x1∨ x2∨ x3∨ , x2 0 e2 0 Hom(E2 , E5 ) × Hom(E5 , E6 ) → Hom(E2 , E6 ) x1 x2 x3 x1∨ e2 0 0 , x3∨ 0 0 e2 Hom(E3 , E4 ) × Hom(E4 , E5 ) → Hom(E3 , E5 ) x1∨ ∧ x2∨ x2∨ ∧ x3∨ x3∨ ∧ x1∨ , x3 0 x2∨ −x1∨ Hom(E3 , E4 ) × Hom(E4 , E6 ) → Hom(E3 , E6 ) x1∨ x2∨ x3∨ , x3 0 0 e3 Hom(E3 , E5 ) × Hom(E5 , E6 ) → Hom(E3 , E6 ) x1 x2 x3 x1∨ e2 0 0 , x2∨ 0 e3 0 Hom(E4 , E5 ) × Hom(E5 , E6 ) → Hom(E4 , E6 ) x1 x2 x3 x1∨ ∧ x2∨ x2∨ −x1∨ 0 . x2∨ ∧ x3∨ 0 x3∨ −x2∨ ∨ ∨ ∨ ∨ x3 ∧ x1 −x3 0 x1 References 1. Auroux, D., Katzarkov, L., Orlov, D.: Mirror symmetry for weighted projective planes and their noncommutative deformations. http://arxiv.org/list/math.AG/0404281, 2004. 2. Auroux, D., Katzarkov, L., Orlov, D.: Mirror symmetry for Del Pezzo surfaces: Vanishing cycles and coherent sheaves. http://arxiv.org/list/math.AG/0506166, 2005 3. Batyrev, V.: Quantum cohomology rings of toric manifolds. Ast´erisque 218, 9, 34 (1993) 4. Beilinson, A.: Coherent sheaves on Pn and problems in linear algebra. Funkts. Anal. i. Prilozhen. 12, no. 3, 68–69 (1978) 5. Bondal, A.: Representation of associative algebras and coherent sheaves. Izv. Ross. Akad. Nauk Ser. Mat. 53, no. 1, 25–44 (1989) 6. Eguchi, T., Hori, K., Xiong, C.-S.: Gravitational quantum cohomology. Int. J. Mod. Phys. A 12, no. 9, 1743–1782 (1997) 7. Fukaya, K., Oh, Y.-G., Ohta, H., Ono, K.: Lagrangian intersection Floer theory. Preprint, available at http://www.math.kyoto-u.ac.jp/˜fukaya, 2000 8. Givental, A.: Homological geometry and mirror symmetry. In: Proceedings of the International Congress of Mathematicians. Proceedings, Zurich 1994. Basel: Birkh¨auser, 1994, pp. 472–480 9. Hori, K., Iqbal, A., Vafa, C.: D-Branes and mirror symmetry. http://arxiv.org/list/hep-th/0005247, 2000 10. Hori, K., Vafa, C.: Mirror symmetry. http:// arxiv.org/list/hep-th/0002222, 2000 11. Kontsevich, M.: Homological algebra of mirror symmetry. In: Proceedings of the International Congress of Mathematicians. Proceedings, Zurich 1994. Basel: Birkh¨auser, 1994, pp. 120–139 12. Kouchnirenko, A.: Poly´edres de Newton et nombres de Milnor. Invent. Math. 32, no. 1, 1–31 (1976)
Homological Mirror Symmetry for Toric del Pezzo Surfaces
85
13. Orlov, D.: Projective bundles, monoidal transformations and derived categories of coherent sheaves. Izv. Ross. Akad. Nauk Ser. Math. 56, no. 4, 852–862 (1992) 14. Robbin, J., Salamon D.: Maslov index for path. Topology 32, no. 4, 827–844 (1993) 15. Seidel, P.: Graded Lagrangian submanifolds. Bull. Soc. Math. France 128, no. 1, 103–149 (2000) 16. Seidel, P.: Vanishing cycles and mutations. In: European Congress of Mathematics, Vol. II. Proceedings, Barcelona 2000. Basel: Birkh¨auser, 2001, pp. 65–85 17. Seidel, P.: More about vanishing cycles and mutations. In: Symplectic geometry and mirror symmetry. Proceedings, Seoul 2000. Singapore:World Sci. Publishing, 2001, pp. 429–465 18. Seidel, P.: Homological mirror symmetry for the quartic surface. http://arxiv.org/list/math.AG/ 0310414 (2003) 19. Witten, E.: Phases of N = 2 theories in two dimensions. Nucl. Phys. B 403, 159–222 (1993) Communicated by N.A. Nekrasov
Commun. Math. Phys. 264, 87–114 (2006) Digital Object Identifier (DOI) 10.1007/s00220-005-1513-4
Communications in
Mathematical Physics
Drinfeld Twists and Algebraic Bethe Ansatz of the Supersymmetric Model Associated with Uq (gl (m|n)) Wen-Li Yang1,2 , Yao-Zhong Zhang2 , Shao-You Zhao2,3 1 2 3
Institute of Modern Physics, Northwest University, Xian 710069, P.R. China Department of Mathematics, University of Queensland, Brisbane, QLD 4072, Australia Department of Physics, Beijing Institute of Technology, Beijing 100081, China
Received: 1 March 2005 / Accepted: 22 August 2005 Published online: 31 January 2006 – © Springer-Verlag 2006
Abstract: We construct the Drinfeld twists (or factorizing F -matrices) of the supersymmetric model associated with quantum superalgebra Uq (gl(m|n)), and obtain the completely symmetric representations of the creation operators of the model in the F -basis provided by the F -matrix. As an application of our general results, we present the explicit expressions of the Bethe vectors in the F -basis for the Uq (gl(2|1))-model (the quantum t-J model). 1. Introduction It was realized in [1] that for the XXX or XXZ spin chain systems, there exists a nondegenerate lower-triangular F -matrix (the Drinfeld twists) [2] in terms of which the R-matrix of the system is factorized: −1 R12 (u1 , u2 ) = F21 (u2 , u1 )F12 (u1 , u2 ),
(1.1)
where the R-matrix acts on the tensor space V ⊗ V with V being a 2-dimensional Uq (gl(2))-module. In the basis provided by the N-site F -matrix, i.e. the so-called F -basis, the entries of the monodromy matrices of the models appear in completely symmetric forms. As a result the Bethe vectors of the models are dramatically simplified and can be written down explicitly. These results enabled the authors in [3, 4] to compute form factors, correlation functions [5] and spontaneous magnetizations of the systems analytically and explicitly. The results of [1] were generalized to other systems. In [6], the Drinfeld twists associated with any finite-dimensional irreducible representations of Yangian Y [gl(2)] were investigated. In [7], Albert et al constructed the F -matrix of the rational gl(m) Heisenberg model, obtained a polarization free representation of its creation operators and resolved the hierarchy of its nested Bethe vectors. In [8, 9], the Drinfeld twists of the elliptic XYZ and Belavin models were constructed. Recently we have successfully constructed the Drinfeld twists for the rational gl(m|n) supersymmetric model and resolved
88
W.-L. Yang, Y.-Z. Zhang, S.-Y. Zhao
the hierarchy of its nested Bethe vectors in the F -basis [10, 11]. Quantum integrable models associated with Lie superalgebras [12–14] are physically important because they give strongly correlated fermion models of superconductivity (e.g. [15, 16]). In this paper, we extend our results in [10, 11] to the quantum (or q-deformed) supersymmetric model associated the quantum superalgebra Uq (gl(m|n)) (including quantum supersymmetric t-J model as a special case). Such a generalization is non-trivial due to the following fact. It is well-known that the gl(m|n) rational model has gl(m|n) symmetry which enables one to express the creation operators Ci (u) in terms of the element Tm+n,m+n (u) of the monodromy matrix T (u) and the generators of gl(m|n) by (anti)commutation relations [7, 11]. However, the corresponding quantum model is not Uq (gl(m|n)) invariant (unless appropriate boundary conditions are imposed). One of the consequences is that the creation operators Ci (u) of the quantum model cannot be expressed in terms of Tm+n,m+n (u) and the generators of Uq (gl(m|n)) by simple q-(anti)commutation relations. Indeed, it is found in this paper that extra quantum correction terms are needed, due to the non-trivial coproduct structure of the quantum superalgebra. Having found such a new recursive relation (3.37) and constructed the factorizing F -matrices of the quantum model, we obtain the symmetric representations of the creation operators of the monodromy matrix in the F -basis. These results make possible a complete resolution of the hierarchy of the nested Bethe vectors of the Uq (gl(m|n)) model. As an example, we give the explicit expressions of the Bethe vectors of the quantum t − J model associated with Uq (gl(2|1)). The present paper is organized as follows. In Sect. 2, we introduce some basic notation on the quantum superalgebra Uq (gl(m|n)). In Sect. 3, we derive the recursive relation between the elements of the monodromy matrix and the generators of Uq (gl(m|n)). In Sect. 4, we construct the F -matrix and its inverse of the Uq (gl(m|n)) model. In Sect. 5, we obtain the symmetric representations of the creation operators in the F -basis. As an application of our general results, the hierarchy of the nested Bethe vectors of the Uq (gl(2|1)) model is resolved in Sect. 6. We conclude the paper by offering some discussions in Sect. 7. Some detailed technical derivations are given in Appendices A-B. 2. Quantum Superalgebra Uq (gl(m|n)) Let us fix two non-negative integers n, m such that n + m ≥ 2 and a positive integer N(N ≥ 2), and a generic complex number η such that the q-deformation parameter, which is defined by q = eη , is not a root of unity. Let V be a Z2 -graded (n + m)-dimensional vector space with the orthonormal basis {|i, i = 1, . . . , n + m}. The Z2 -grading is chosen as: [1] = · · · = [m] = 1, [m + 1] = · · · = [m + n] = 0. Definition 1. The quantum superalgebra Uq (gl(m|n)) is a Z2 -graded unital associative superalgebra generated by the generators E i,i , (i = 1, . . . , n + m) and E j,j +1 , E j +1,j (j = 1, . . . , n + m − 1) with the Z2 -grading [E i,i ] = 0, [E j +1,j ] = [E j,j +1 ] = [j ] + [j + 1] by the relations:
[E i,i , E i ,i ] = 0, [E i,i , E j,j +1 ] = (δi,j −δi,j +1 )E i,j+1 , i = 1, . . . , n+m, (2.1) [E i,i , E j+1,j ] = (δi,j+1 −δi,j )E j +1,j ,
[E j,j +1 , E j +1,j ] = (−1)[j ] δj,j
hj
(2.2) −hj
q −q , j = 1, . . . , n+m−1, q − q −1
(2.3)
Drinfeld Twists and Algebraic Bethe Ansatz
89
and the Serre relations: (E m,m+1 )2 = (E m+1,m )2 = 0,
[E j,j +1 , E j ,j +1 ] = [E j +1,j , E j +1,j ] = 0, |j − j | ≥ 2, (E j,j +1 )2 E j ±1,j ±1+1 − (q + q −1 )E j,j +1 E j ±1,j ±1+1 E j,j +1 +E j ±1,j ±1+1 (E j,j +1 )2 = 0, j = m, (E j +1,j )2 E j ±1+1,j ±1 − (q + q −1 )E j +1,j E j ±1+1,j ±1 E j +1,j +E j ±1+1,j ±1 (E j +1,j )2 = 0, j = m,
(2.4)
where hj = (−1)[j ] E j,j − (−1)[j +1] E j +1,j +1 . In addition to the above Serre relations, there exist also extra Serre relations [17] which we omit. Here and throughout, we adopt the convention: [x, y] = xy − (−1)[x][y] yx, x, y ∈ Uq (gl(m|n)). One can easily see that the Z2 -graded vector space V supplies the fundamental Uq (gl(m|n))-module and the generators of Uq (gl(m|n)) are represented in this space by π(E i,i ) = ei,i , π(E j,j +1 ) = ej,j +1 , π(E j +1,j ) = ej +1,j ,
(2.5)
where ei,j ∈ End(V ) is the elementary matrix with elements (ei,j )lk = δj k δil . Uq (gl(m|n)) is a Z2 -graded triangular Hopf superalgebra endowed with Z2 -graded algebra homomorphisms that are coproduct : Uq (gl(m|n)) −→ Uq (gl(m|n)) ⊗ Uq (gl(m|n)) defined by (E i,i ) = 1 ⊗ E i,i + E i,i ⊗ 1, i = 1, . . . , n + m,
(2.6)
(E j,j +1 ) = 1 ⊗ E j,j +1 + E j,j +1 ⊗ q ,
(2.7)
hj
(E
j +1,j
)=q
−hj
⊗E
j +1,j
+E
j +1,j
⊗ 1,
(2.8)
and counit : Uq (gl(m|n)) −→ C defined by (E j,j +1 ) = (E j +1,j ) = (E i,i ) = 0, (1) = 1, and a Z2 -graded algebra antiautomorphism (antipode) S: Uq (gl(m|n)) −→ Uq (gl(m|n)) given by S(E j,j +1 ) = −E j,j +1 q −h , S(E j +1,j ) = −q h E j +1,j , S(E i,i ) = −E i,i . j
j
Multiplications of tensor products are Z2 graded:
(x ⊗ y)(x ⊗ y ) = (−1)[y][x ] xx ⊗ yy , for homogeneous elements x, y, x , y ∈ Uq (gl(m|n)) and where [x] ∈ Z2 denotes the grading of x. It should be pointed out that the antipode satisfies the following equation, for homogeneous elements x, y ∈ Uq (gl(m|n)), S(xy) = (−1)[x][y] S(y)S(x),
90
W.-L. Yang, Y.-Z. Zhang, S.-Y. Zhao
and generalizes to inhomogeneous elements through linearity. The coproduct, counit and antipode satisfy the following relations, ∀x ∈ Uq (gl(m|n)): ( ⊗ id)(x) = (id ⊗ )(x), ( ⊗ id)(x) = x = (id ⊗ )(x), m(S ⊗ id)(x) = m(id ⊗ S)(x) = (x),
(2.9)
where m denote the product of any two elements of Uq (gl(m|n)), i.e., m(x ⊗ y) = xy for x, y ∈ Uq (gl(m|n)). The generators {E j,j +1 } ( {E j +1,j }) are the simple raising (lowering) generators of Uq (gl(m|n)) associated with the simple roots. Thanks to the Serre relations, the other generators associated with the non-simple roots (called the non-simple generators) can be uniquely constructed through the simple ones by the following relations: E α,γ = E α,β E β,γ − q −(−1) E β,γ E α,β , 1 ≤ α < β < γ ≤ n + m, [β]
E
γ ,α
=E
γ ,β
E
β,α
−q
(−1)[β]
E
β,α
E
γ ,β
, 1 ≤ α < β < γ ≤ n + m.
(2.10) (2.11)
The coproduct, counit and antipode of the non-simple generators can be obtained through those of the simple ones. Here, we give the coproduct of non-simple generators which will be used later. Lemma 1. The coproduct of the non-simple generators is (E γ ,γ −l ) = q −
l
γ −k k=1 h
⊗ E γ ,γ −l + E γ ,γ −l ⊗ 1
l−1 l−i γ−k [γ−l+i] + (1−q 2(−1) )q− k=1 h E γ −l+i,γ−l ⊗E γ ,γ−l+i , i=1
γ −l ≥ 1 and l≥2, (E
γ ,γ +l
) = 1⊗E l−1
+
γ ,γ +l
+E
γ ,γ +l
(1−q−2(−1)
[γ+i]
⊗q
(2.12)
l−1
γ +k k=0 h
)E γ +i,γ +l ⊗E γ ,γ+i q
l−1 k=i
hγ +k
,
i=1
γ +l≤n+m and l ≥ 2.
(2.13)
Proof. This lemma can be proved by induction using the coproducts of the simple generators (2.6)-(2.8), the definitions of the non-simple generators (2.10)-(2.11) and the fact that the coproduct is an algebra homomorphism, as well as (2.1)-(2.3) and the Serre relation (2.4).
3. Recursive Relation between Monodromy Matrix Elements and Uq (gl(m|n)) Generators Let R ∈ End(V ⊗ V ) be the R-matrix associated with the fundamental Uq (gl(m|n))module V . The R-matrix depends on the difference of two spectral parameters u1 and u2 associated with two copies of V , and is, in the present grading, given by [13, 14, 18]
Drinfeld Twists and Algebraic Bethe Ansatz
91
R12 (u1 , u2 ) = R12 (u1 − u2 ) m+n m m+n = c12 ei,i ⊗ ei,i + ei,i ⊗ ei,i + a12 ei,i ⊗ ej,j i=1 − + b12
i=j =1
i=m+1 m+n
+ (−1)[j ] ei,j ⊗ ej,i + b12
i>j =1
m+n
(−1)[j ] ei,j ⊗ ej,i , (3.1)
j >i=1
where sinh(u1 −u2 ) e±(u1−u2 ) sinh η ± = b± (u1 , u2 ) ≡ , b12 , sinh(u1 −u2 +η) sinh(u1 −u2 +η) sinh(u1 − u2 − η) = c(u1 , u2 ) ≡ , sinh(u1 − u2 + η)
a12 = a(u1 , u2 ) ≡
(3.2)
c12
(3.3)
and η is the so-called crossing parameter. One can easily check that the R-matrix satisfies the unitary relation R21 R12 = 1.
(3.4) N+1
Let us introduce the (N + 1)-fold tensor product space V ⊗ V · · · ⊗ V , whose components are labelled by 0, 1 . . . , N from the left to the right. As usual, the 0th space, denoted by V0 ( Vi for the i th space), corresponds to the auxiliary space and the other N
N spaces constitute the quantum space H =V ⊗ V · · · ⊗ V . Moreover, for each factor space Vi , i = 0, . . . , N, we associate a complex parameter zi . The parameter associated with the 0th space is usually called the spectral parameter which is set to z0 = u in this paper, and the other parameters are called the inhomogeneous parameters. In this paper we always assume that all the complex parameters u and {zi |i = 1, . . . , N} are generic ones. Hereafter we adopt the standard notation: for any matrix A ∈ End(V ), Aj (or A(j ) ) is an embedding operator in the tensor product space, which acts as A on the j th space and as an identity on the other factor spaces; Rij = Rij (zi , zj ) is an embedding operator of R-matrix in the tensor product space, which acts as an identity on the factor spaces except for the i th and j th ones. The R-matrix satisfies the graded Yang-Baxter equation (GYBE) R12 R13 R23 = R23 R13 R12 .
(3.5)
In terms of the matrix elements defined by ij R(u)(|i ⊗ |j ) = R(u)ij (|i ⊗ |j ), i,j
the GYBE reads ij j k R(u1 − u2 )ij R(u1 − u3 )ii kk R(u2 − u3 )j k (−1)[j ]([i ]+[i ]) i ,j ,k
=
i ,j ,k
j k
i j
R(u2 − u3 )j k R(u1 − u3 )iikk R(u1 − u2 )i j (−1)[j ]([i]+[i ]) .
92
W.-L. Yang, Y.-Z. Zhang, S.-Y. Zhao
Besides the GYBE, the R-matrix satisfies the following relation: −1 R12 , R12 (x) = P12 (x)P12
(3.6)
where P12 is the superpermutation operator, i.e., P12 (|i ⊗ |j = (−1)[i][j ] |j ⊗ |i. Using the coproduct structure of Uq (gl(m|n)), one can define the action of Uq (gl(m|n)) on the (N + 1)-fold tensor product space. For any x ∈ Uq (gl(m|n)), let us denote the action of x on the (N + 1)-fold tensor product space by (x)0...N : (x)0...N = (N) (x) = (id ⊗ (N−1) )(x).
(3.7)
By a straightforward calculation, one has Lemma 2. (E )0...N = i,i
N
i,i E(k) , i = 1, . . . , n + m,
(3.8)
k=0
(E
j,j +1
)0...N =
N
j,j +1
E(k)
N
j i=k+1 h(i)
q
, j = 1, . . . , n + m − 1,
(3.9)
, j = 1, . . . , n + m − 1,
(3.10)
k=0
(E j +1,j )0...N =
N
q−
k−1 i=0
j
h(i)
j +1,j
E(k)
k=0 i,j
where E(k) is the embedding of ei,j in the tensor product space, which acts as ei,j on the k th space and as identity on the other factor spaces. The actions of non-simple generators can be obtained from those of simple ones through (2.10) and (2.11). Let SN+1 denote the permutation group of the N + 1 space labels (0, . . . , N ). The GYBE (3.5) and unitary relation (3.4) of R-matrix allow one to introduce the following mapping. Definition 2. One can define a mapping from SN+1 to End(V0 ⊗ H) which associate in σ a unique way an element R0...N ∈ End(V0 ⊗ H) to any element σ of the permutation group SN+1 . The mapping has the following composition law:
σσ σ σ σ R0...N = P σ R0...N (P σ )−1 R0...N = Rσσ (0...N) R0...N , ∀σ, σ ∈ SN+1 ,
(3.11)
where P σ is the Z2 -graded permutation operator on the tensor product space, i.e., P σ |i0 (0) . . . |iN (N) = |i0 (σ (0)) . . . |iN (σ (N)) . For any elementary permutation σj with σj (0, . . . , j, j + 1, . . . , N ) = (0, . . . , j + 1, j, . . . , N ), j = 0, . . . , N, the σj corresponding R0...N is σ
j = Rj j +1 . R0...N
(3.12)
σ can be constructed through (3.11) For any element σ ∈ SN+1 , the corresponding R0...N and (3.12) as follows. Let σ be decomposed in a minimal way in terms of elementary permutation as σ = σβ1 . . . σβp , where the positive integer p is the length of σ . The comσ position law enables one to obtain the expression of the associated R0...N . The GYBE σ (3.5) and (3.4) guarantee the uniqueness of R0...N . For the special element σc of SN+1 ,
σc = σ0 σ1 . . . σN−1 , namely, σc (0, 1, . . . , N ) = (1, 2, . . . , N, 0),
(3.13)
Drinfeld Twists and Algebraic Bethe Ansatz
93
σc the associated R0...N is given by σc T (u) ≡ T0 (u) = T0,1...N (u) = R0...N = R0 N R0 N−1 . . . R01 .
(3.14)
σc Thus R0...N is the quantum monodromy matrix T (u) of the Uq (gl(m|n)) spin chain on an N-site lattice. By the GYBE, one may prove that the monodromy matrix satisfies the GYBE
R00 (u − v)T0 (u)T0 (v) = T0 (v)T0 (u)R00 (u − v).
(3.15)
Define the transfer matrix t (u) t (u) = str0 T (u),
(3.16)
where str0 denotes the supertrace over the auxiliary space. Then the Hamiltonian of our model is given by d ln t (u) (3.17) |u=0 . du This model is integrable thanks to the commutativity of the transfer matrix for different parameters, H =
[t (u), t (v)] = 0,
(3.18)
which can be verified by using the GYBE. The fundamental relation (3.6) and the co-associativity (2.9) of the coproduct of Uq (gl(m|n)) enable one to prove the following result, using the procedure similar to that in [1] for the non-super case, Proposition 1. The mapping defined in Definition 2 satisfies the following relation: σ σ (x)0...N = P σ (x)0...N (P σ )−1 R0...N , ∀x ∈ Uq (gl(m|n)), σ ∈ SN+1 . R0...N
(3.19)
One may decompose the monodromy matrix T (u) in terms of the basis of End(V0 ) as n+m
T (u) =
i,j
Ti,j (u)E(0) ≡
i,j =1
n+m
Ti,j (u)ei,j ,
(3.20)
i,j =1
where the matrix elements Ti,j (u) are operators acting on the quantum space H and have the Z2 -grading: [Ti,j (u)] = [ei,j ] = [i] + [j ]. Similarly, for the quantities defined in Lemma 2, we have the decomposition: i,i + (E i,i )0...N = E(0)
N
i,i E(k) = ei,i + (E i,i )1...N , i = 1, . . . , n + m,
(3.21)
k=1 j,j +1
(E j,j +1 )0...N = E(0)
N
q
j k=1 h(k)
+
N
j,j +1
E(k)
N
q
j i=k+1 h(i)
k=1
= ej,j +1 q j +1,j
(E j +1,j )0...N = E(0)
(hj )1...N j
+ (E j,j +1 )1...N , j = 1, . . . , n + m − 1,
+ q −h(0)
N
q−
k−1 i=1
j
h(i)
(3.22)
j +1,j
E(k)
k=1
= ej +1,j + q −hj (E j +1,j )1...N , j = 1, . . . , n + m − 1,
(3.23)
94
W.-L. Yang, Y.-Z. Zhang, S.-Y. Zhao
where hj = (−1)[j ] ej,j − (−1)[j +1] ej +1,j +1 . Without confusion, hereafter we adopt the following convention:
i,i , i, i = 1, . . . , n + m, ei,i = E(0)
hj =
(3.24)
j h(0) = (−1)[j ] ej,j −(−1)[j +1] ej +1,j +1 ,
Ei,i = (E i,i )1...N =
N
j = 1, . . . , n+m−1,
i,i E(k) , i = 1, . . . , n + m,
(3.25) (3.26)
k=1
Hj = (−1)[j ] Ej,j − (−1)[j +1] Ej +1,j +1 , j = 1, . . . , n + m − 1, N
Ej,j +1 = (E j,j +1 )1...N =
j,j +1
E(k)
N
j i=k+1 h(i)
q
(3.27)
, j = 1, . . . , n + m − 1, (3.28)
k=1
Ej +1,j = (E
j +1,j
)1...N =
N
q−
k−1
j i=1 h(i)
j +1,j
E(k)
, j = 1, . . . , n + m − 1, (3.29)
k=1
and a similar convention for the non-simple generators. Then the operators {Ei,j } are the operators which act non-trivially on the quantum space H and trivially (i.e. as an identity) on the auxiliary space V0 . From (3.13), we have P σc (hj )0...N (P σc )−1 = hj + Hj , j = 1, . . . , n + m − 1, P
σc
P
σc
(E
j,j +1
(E
j +1,j
σc −1
)0...N (P )
σc −1
)0...N (P )
=q
hj
Ej,j+1 +ej,j+1 , j = 1, . . . , n+m−1,
= ej,j+1 q
−Hj
+ Ej +1,j , j = 1, . . . , n+m−1.
(3.30) (3.31) (3.32)
Using (3.21)-(3.23), (3.30)-(3.32) and Lemma 1, we have Proposition 2. (E γ ,γ −l )0...N = q − +
l−1
l
k=1 hγ−k
[γ −l+i]
(1 − q 2(−1)
Eγ ,γ −l +eγ ,γ −l
)q −
l−i
k=1 hγ −k
eγ −l+i,γ −l Eγ ,γ −l+i , γ − l ≥ 1 and l ≥2,
i=1
P (E σc
γ ,γ −l
σc −1
)0...N (P )
= eγ ,γ −l q
−
l
(3.33)
k=1 Hγ −k
+ Eγ ,γ −l
l−1 l−i [γ−l+i] ) eγ ,γ −l+i q − k=1 Hγ −k Eγ −l+i,γ −l , γ − l ≥ 1 and l ≥ 2, + (1−q 2(−1) i=1
(E
γ ,γ+l
+
)0...N = Eγ ,γ +l +eγ ,γ +l q
l−1
[γ +i]
(1−q −2(−1)
(3.34)
l−1
k=0 Hγ +k
)eγ +i,γ +l Eγ ,γ +i q
l−1 k=i
Hγ +k
, γ +l ≤ n+m and l ≥ 2,
i=1
(3.35)
Drinfeld Twists and Algebraic Bethe Ansatz
95
P σc (E γ ,γ +l )0...N (P σc )−1 = eγ ,γ+l + q +
l−1
[γ +i]
(1 − q −2(−1)
)eγ ,γ +i q
l−1
l−1 k=i
k=0
hγ +k
hγ +k
Eγ ,γ +l
Eγ +i,γ +l , γ +l ≤ n+m and l ≥ 2.
i=1
(3.36) Substituting (3.35) and (3.36) into Proposition 1, we obtain our main result in this section: Theorem 1. The matrix elements Tn+m,n+m−l (u) (l = 1, . . . , n + m − 1) of the monodromy matrix can be expressed in terms of Tn+m,n+m (u) and the generators of Uq (gl(m|n)) by the following recursive relation: Tn+m,n+m−l (u) = l [n+m] = q −(−1) En+m−l,n+m Tn+m,n+m (u)−Tn+m,n+m (u)En+m−l,n+m q − k=1 Hn+m−k −
l−1
(1 − q −2(−1)
[n+m−α]
)Tn+m,n+m−α (u)En+m−l,n+m−α q −
l
k=α+1 Hn+m−k
.
α=1
(3.37) The proof of this theorem is relegated to Appendix A. We call the second term in the R.H.S. of (3.37) a quantum correction term, which vanishes in the rational limit (q → 1). Moreover, such a nontrivial correction term only occurs in the higher rank models (i.e., when n + m ≥ 3). In the rational limit: q → 1, (3.37) reduces to the (anti)commutation relations used in [7, 11]. For some special values of m and n, the associated recursive relations in the present grading become: • For the Uq (gl(1|1)) case:
T2,1 (u) = q −1 E1,2 T2,2 (u) − T2,2 (u)E1,2 q −H1 .
• For the Uq (gl(2|1)) case which corresponds to the quantum t − J model:
T3,2 (u) = q −1 E2,3 T3,3 (u) − T3,3 (u)E2,3 q −H2 ,
T3,1 (u) = q −1 E1,3 T3,3 (u) − T3,3 (u)E1,3 q −H2 −H1 −(1 − q 2 )T3,2 (u)E1,2 q −H1 .
(3.38)
(3.39)
(3.40)
• For the Uq (gl(2|2)) case which corresponds to the quantum EKS model [15]:
T4,3 (u) = q −1 E3,4 T4,4 (u) − T4,4 (u)E3,4 q −H3 , (3.41)
T4,2 (u) = q −1 E2,4 T4,4 (u) − T4,4 (u)E2,4 q −H3 −H2 −(1 − q −2 )T4,3 (u)E2,3 q −H2 ,
T4,1 (u) = q −1 E1,4 T4,4 (u) − T4,4 (u)E1,4 q −H3 −H2 −H1
(3.42)
−(1 − q −2 )T4,3 (u)E1,3 q −H2 −H1 −(1 − q 2 )T4,2 (u)E1,2 q −H1 .
(3.43)
96
W.-L. Yang, Y.-Z. Zhang, S.-Y. Zhao
4. Factorizing F-Matrices and Their Inverses In this section, we construct the Drinfeld twists [2] (factorizing F-matrices) on the N -fold tensor product space (i.e. the quantum space H) associated with the quantum superalgebra Uq (gl(m|n)). 4.1. Factorizing F-matrix. Let SN be the permutation group associated with the indiσ σ ces (1, . . . , N ) and R1...N the N-site R-matrix associated with σ ∈ SN . R1...N acts non-trivially on the quantum space H and trivially (i.e as an identity) on the auxiliary space. Definition 3. The F-matrix F1...N (z1 , . . . , zN ) is an operator in End(H) and satisfies the following three properties: I. lower-triangularity; II. non-degeneracy; III. factorization, namely, σ Fσ (1)...σ (N) (zσ (1) , . . . , zσ (N) ) R1...N = F1...N (z1 , . . . , zN ), ∀σ ∈ SN . (4.1)
Define the N-site F -matrix: F1...N ≡ F1...N (z1 , . . . , zN ) =
∗ n+m
N
α
σ (j ) σ Pσ (j ) S(σ, ασ )R1...N ,
(4.2)
σ ∈SN ασ (1) ...ασ (N ) =1 j =1
embedding of the project operator P α in the i th space with (P α )kl = where Piα is the δkl δkα , the sum ∗ in (4.2) is over all non-decreasing sequences of the labels ασ (i) : ασ (i+1) ≥ ασ (i) ασ (i+1) > ασ (i)
if σ (i + 1) > σ (i), if σ (i + 1) < σ (i),
(4.3)
and S(σ, ασ ) is a c-number function of σ, ασ and the element cij of the R-matrix, defined by
N 1 [ασ (k) ] 1 − (−1) δασ (k) ,ασ (l) ln(1 + cσ (k)σ (l) ) . (4.4) S(σ, ασ ) ≡ exp 2 l>k=1
Proposition 3. The F-matrix F1...N given by (4.2)-(4.4) satisfies Properties I, II, III. Proof. The definition of F1...N (4.2) and the summation condition (4.3) imply that F1...N is a lower-triangular matrix. Moreover, one can easily check that the F -matrix is nondegenerate because all diagonal elements are non-zero. We now prove that the F -matrix (4.2) satisfies Property III. Any given permutation σ ∈ SN can be decomposed into elementary ones of the group SN as σ = σi1 . . . σik . By (3.11), we have, if Property III holds for any elementary permutation σi , σ Fσ (1...N) R1...N = σi
= Fσi1 ...σik (1...N) Rσi k ...σi 1
k−1
σi
σik−1 σi1 (1...N) Rσi1 ...σik−2 (1...N) . . . R1...N
σi1 . . . R1...N (1...N) k−2
= Fσi1 ...σik−1 (1...N) Rσi k−1 ...σi 1
σi
1 = . . . = Fσi1 (1...N) R1...N = F1...N .
(4.5)
Drinfeld Twists and Algebraic Bethe Ansatz
97
For the elementary permutation σi , we have σi Fσi (1...N) R1...N =
=
∗
N
σ ∈SN ασi σ (1) ...ασi σ (N ) j =1
=
∗
N
σ ∈SN ασi σ (1) ...ασi σ (N ) j =1
=
N ∗(i)
ασ
σ (j )
ασ
σ (j )
σi Pσi σi (j ) S(σi σ, ασi σ )Rσσi (1...N) R1...N
σi σ Pσi σi (j ) S(σi σ, ασi σ )R1...N
α
σ˜ ∈SN ασ˜ (1) ...ασ˜ (N ) j =1
σ˜ (j ) σ˜ Pσ˜ (j ˜ , ασ˜ )R1...N , ) S(σ
where σ˜ = σi σ , and the summation sequences of ασ˜ in ασ˜ (j +1) ≥ ασ˜ (j ) ασ˜ (j +1) > ασ˜ (j )
(4.6)
∗ (i)
now have the form
if σi σ˜ (j + 1) > σi σ˜ (j ), if σi σ˜ (j + 1) < σi σ˜ (j ).
(4.7)
Comparing (4.7) with (4.3), we find that the only difference between them is the transposition σi factor in the “if” conditions. For a given σ˜ ∈ SN with σ˜ (j ) = i and σ˜ (k) = i + 1, we now examine how the elementary transposition σi will affect the inequalities (4.7). If |j − k| > 1, then σi does not affect the sequence of ασ˜ at all, that is, the sign of inequality “>” or “≥” between two neighboring root indexes is unchanged with the action of σi . If |j − k| = 1, then in the summation sequences of ασ˜ , when σ˜ (j + 1) = i + 1 and σ˜ (j ) = i, sign “≥” changes to “>”, while when σ˜ (j + 1) = i and σ˜ (j ) = i + 1, “>” changes to “≥”. Thus (4.3) and (4.6) differ only when equal labels ασ˜ appear. With the help of the relation c21 c12 = 1, one may prove that in this case the σi product Fσi (1...N) R1...N still equals F1...N (see [10] for a more detailed proof). Thus, we obtain σ R1...N (z1 , . . . , zN ) = Fσ−1 (1...N) (zσ (1) , . . . , zσ (N) )F1...N (z1 , . . . , zN ),
(4.8)
and the factorizing F -matrix F1...N of Uq (gl(m|n)) is proved to satisfy all three properties.
From the expression of the F -matrix, one knows that it has an even grading, i.e., [F1...N ] = 0.
(4.9)
4.2. Inverse of the F-matrix. The non-degenerate property of the F -matrix implies that −1 we can find the inverse matrix F1...N . To do so, we first define ∗ F1...N
=
∗∗ n+m
−1
S(σ, ασ )Rσσ (1...N)
σ ∈SN ασ (1) ...ασ (N ) =1
N
α
σ (j ) Pσ (j ) ,
(4.10)
j =1
where the sum ∗∗ is taken over all possible αi which satisfies the following nonincreasing constraints: ασ (i+1) ≤ ασ (i) ασ (i+1) < ασ (i)
if σ (i + 1) < σ (i), if σ (i + 1) > σ (i).
(4.11)
98
W.-L. Yang, Y.-Z. Zhang, S.-Y. Zhao
Proposition 4. The inverse of the F -matrix is given by −1 ∗ F1...N = F1...N −1 ij ,
(4.12)
i<j
where
ββ
[ij ]αii αjj =δαi βi δαj βj
sinh(zi −zj ) sinh(zi −zj +η) sinh(zj −zi ) sinh(zj −zi +η) 1 2 (z −z ) cosh2 η −4 sinh i j sinh(zi −zj+η) sinh(zi −zj −η)
if αi > αj if αi < αj , if αi =αj =m+1, . . . , n+m, if αi =αj =1, . . . , m. (4.13)
∗ . Substituting (4.2) and (4.10) into Proof. We compute the product of F1...N and F1...N the product, we have ∗ = F1...N F1...N
∗
∗∗
S(σ, ασ )S(σ , βσ )
σ ∈SN σ ∈SN ασ1 ...ασN βσ ...βσ 1
×
N
N
N
−1
α
σ (i) σ Pσ (i) R1...N Rσσ (1...N)
i=1
=
×
N
i=1
∗
σ ∈SN
β
Pσ σ(i)(i)
σ ∈S α
N
∗∗
S(σ, ασ )S(σ , βσ )
ασ1 ...ασN βσ ...βσ 1
−1
σ (i) σ σ Pσ (i) Rσ (1...N)
i=1
N
N
β
Pσ σ(i)(i) .
(4.14)
i=1
To evaluate the R.H.S., we examine the matrix element of the R-matrix −1 ασ (N ) ...ασ (1) σ Rσσ (1...N) . βσ (N ) ...βσ (1)
(4.15)
Note that the sequence {ασ } is non-decreasing and {βσ } is non-increasing. Thus the non-vanishing condition of the matrix element (4.15) requires that ασ and βσ satisfy βσ (N) = ασ (1) , . . . , βσ (1) = ασ (N) .
(4.16)
One can verify [7] that (4.16) is fulfilled only if σ (N ) = σ (1), . . . , σ (1) = σ (N ).
(4.17)
Let σ¯ be the maximal element of the SN which reverses the site labels σ¯ (1, . . . , N ) = (N, . . . , 1).
(4.18)
σ = σ σ¯ .
(4.19)
Then from (4.17), we have
Drinfeld Twists and Algebraic Bethe Ansatz
99
Substituting (4.16) and (4.19) into (4.14), we have ∗ F1...N F1...N
=
∗
σ ∈SN ασ1 ...ασN
N
N ασ (i) σ¯ ασ (i) S(σ, ασ )S(σ, ασ ) Pσ (i) Rσ (N...1) Pσ (i) . i=1 i=1
(4.20) The decomposition of R σ¯ in terms of elementary R-matrices is unique modulo the GYBE. One reduces from (4.20) that F F ∗ is a diagonal matrix: ∗ F1...N F1...N =
(4.21)
ij .
i<j
Then (4.12) is a simple consequence of the above equation.
5. Monodromy Matrix in the F -Basis In the previous section, we have constructed the F -matrix and its inverse which act on the quantum space H. The non-degeneracy of the F -matrix means that its column vectors also form a complete basis of H, which is called the F -basis. In this section, we study the generators of Uq (gl(m|n)) and the elements of the monodromy matrix in the F -basis. 5.1. Uq (gl(m|n)) generators in the F -basis. The Cartan generators {E i,i } (or {H j }) and the simple generators {E j,j +1 }, {E j +1,j } of Uq (gl(m|n)) are realized on H by {Ei,i } (or {Hj }), {Ej,j +1 } and {Ej +1,j }, respectively, as (3.26)-(3.29). The other non-simple generators {E i,j } can be obtained from the simple ones by (2.10) and (2.11), and denote their realizations on H by {Ei,j }. Introduce the generators in the F -basis: −1 E˜ i,j = F1...N Ei,j F1...N , i, j = 1, . . . , n + m.
(5.1)
Theorem 2. In the F -basis the Cartan and the simple generators of Uq (gl(m|n)) are given by E˜ i,i = Ei,i =
N
i,i E(k) , i = 1, . . . , n + m,
(5.2)
k=1
E˜ j,j +1 =
N
j,j +1
⊗γ =k G(γ ) (k, γ ), j = 1, . . . , n + m − 1,
j +1,j
⊗γ =k G(γ )
E(k)
j,j +1
(5.3)
k=1
E˜ j +1,j =
N
E(k)
j +1,j
(k, γ ), j = 1, . . . , n + m − 1.
k=1 γ ,γ ±1
Here the diagonal matrices G(j )
(i, j ) are:
(5.4)
100
W.-L. Yang, Y.-Z. Zhang, S.-Y. Zhao
• For 1 < γ + 1 ≤ m, k = γ, 2e−η cosh η, = δkl (2aij cosh η)−1 eη , k = γ + 1, 1, otherwise, −η k = γ + 1, 2e cosh η, γ +1,γ (G(j ) (i, j ))kl = δkl (2aj i cosh η)−1 eη , k = γ , 1, otherwise, γ ,γ +1 (G(j ) (i, j ))kl
(5.5)
(5.6)
• For γ = m, 2e−η cosh η, k = γ , k = γ + 1, = δkl e−η , 1, otherwise, (2aj i cosh η)−1 eη , k = γ , γ +1,γ (G(j ) (i, j ))kl = δkl (aj i )−1 eη , k = γ + 1, 1, otherwise, γ ,γ +1 (G(j ) (i, j ))kl
(5.7)
(5.8)
• For 1 + m ≤ γ < n + m, (aij )−1 eη , k = γ , = δkl e−η , k = γ + 1, 1, otherwise, (aj i )−1 e−η , k = γ + 1, γ +1,γ (G(j ) (i, j ))kl = δkl eη , k = γ, 1, otherwise. γ ,γ +1 (G(j ) (i, j ))kl
(5.9)
(5.10)
Proof. Using (3.26)-(3.29), Proposition 1 and the composition law of R-matrices (3.11), one can prove the theorem. Here, without losing generality, we give the proof for the generator E˜ 1,2 as an example. From the expressions of F1...N and its inverse, we have E˜ 1,2 =
∗
∗∗
S(σ, ασ )S(σ , βσ )
σ,σ ∈SN ασ (1) ...ασ (N ) βσ (1) ...βσ (N )
×
N
−1
α
σ (i) σ Pσ (i) R1...N (E 1,2 )1...N Rσσ (1...N)
i=1
=
N
β
Pσ σ(i)(i)
i=1
∗
∗∗
−1 ij
i<j
S(σ, ασ )S(σ , βσ )
σ,σ ∈SN ασ (1) ...ασ (N ) βσ (1) ...βσ (N )
×
N i=1
α
−1
−1
σ (i) σ σ σ Pσ (i) [(P1...N (E 1,2 )1,...,N P1...N )]Rσσ (1...N)
N i=1
β
Pσ σ(i)(i)
i<j
−1 (5.11) ij
Drinfeld Twists and Algebraic Bethe Ansatz
=
N σ,σ ∈SN k=1 α
σ (1) ×Pσ (1)
=1
1,2 E(σ (l)) q
101
∗
N
1 i=l+1 h(σ (i))
∗∗
S(σ, ασ )S(σ , βσ )
ασ (1) ...ασ (N ) βσ (1) ...βσ (N )
N α =1→2 β −1 ασ (N ) σ −1 σ σ (l) . . . Pσ (l)=k Rσ (1...N) Pσ σ(i)(i) ij , . . . Pσ (N) i=1
(5.12)
i<j
where in (5.11) we have used (3.19) and l stands for indices between 1 and N such that σ (l) = k. Here and below, 1 → 2 means that 1 is changed to 2; thus ασ (l) = 1 → 2 in (5.12) means that ασ (l) = 1 is replaced by ασ (l) = 2 on the site σ (l) = k. The ele −1
α
σ (1) σ between Pσ (1) ment of Rσσ (1...N) is denoted as
=1
α
=1→2
α
σ (l) . . . Pσ (l)=k
ασ (N ) ... 1→2 ... 1 −1 σ Rσσ (1...N) βσ (N ) ...βσ (1)
σ (N )
σ (l)=k
β
β
σ (N ) (N ) and Pσ σ(N) . . . Pσ σ(1)(1) . . . Pσ (N)
σ (1)
(5.13)
.
We call the sequence {ασ (l) } normal if it is arranged according to the rules in (4.3), otherwise, we call it abnormal. It is now convenient for us to discuss the non-vanishing condition of the R-matrix element (5.13). Comparing (5.13) with (4.15), we find that the difference between them lies in the k th site. Because the group label in the k th space has been changed, the sequence {ασ } is now an abnormal sequence. However, it can be permuted to the normal sequence by some permutation σˆ k . Namely, α1→2 in the abnormal sequence can be moved to a suitable position by using the permutation σˆ k according to rules in (4.3). (It is easy to verify that σˆ k is unique by using (4.3).) Thus, by procedure similar to that in the previous section, we find that when σ = σˆ k σ σ¯
and
βσ (N) = ασ (1) , . . . , βσ (1) = ασ (N) ,
(5.14)
the R-matrix element (5.13) is non-vanishing. ij Because the non-zero condition of the elementary R-matrix element Rij is i + j = i + j , the following R-matrix elements: ασ (N ) ... 1 ...1→2... 1 −1 σ Rσσ (1...N) βσ (N ) ...βσ (1)
σ (N )
σ (l)=k
σ (p)
σ (1)
,
with 1 ≤ p ≤ l are also non-vanishing. Therefore, (5.12) becomes E˜ 1,2 =
N
∗
S(σ, ασ )S(σˆ k σ, ασˆ k σ ) σ ∈SN k=1 ασ1 ...ασN N 1 α =1 ασ (l) =1→2 ασ (N ) 1,2 i=l+1 h(σ (i)) P σ (1) . . . Pσ (l)=k . . . Pσ (N) × E(σ σ (1) (l)) q + ...
(5.15)
102
W.-L. Yang, Y.-Z. Zhang, S.-Y. Zhao 1,2 + E(σ (p)) q + ... 1,2 +E(σ (1)) q σ¯ σ −1 σˆ k−1 σ k σ (N...1)
×Rσˆ
N
1 i=n+1 h(σ (i))
N
1 i=2 h(i)
N i=1
=
N
α
σ (1) Pσ (1)
ασˆ k σ (i) k σ (i)
Pσˆ
α
σ (1) Pσ (1)
=1
=1→2
α
σ (p) . . . Pσ (p)
α
=1→2
=1
α
=1
α
σ (l) σ (N ) . . . Pσ (l)=k . . . Pσ (N)
α
σ (l) σ (N ) . . . Pσ (l)=k . . . Pσ (N)
−1 ij
(5.16)
i<j
1,2 E(k) ⊗j =k G1,2 (j ) (k, j ),
(5.17)
k=1
where index p runs between 1 and l and σˆ k is the element of SN which permutes the first abnormal sequence in the square bracket of (5.16) to normal sequence. Using the similar procedure, one can prove the theorem for other generators.
The non-simple generators E˜ γ ,γ ±l (for l ≥ 2) can be obtained through the simple ones by (2.10) and (2.11). Some remarks are in order. Equation (5.2) implies that H˜ j = Hj , j = 1, . . . , N,
(5.18)
which will be used frequently. In the rational limit: η → 0 (or q → 1), our results reduce to those in [11] and those in [7] for the special case m = 0.
5.2. Creation operators in the F -basis. Among the matrix elements of the mondromy matrix Ti,j (u), the operators Tn+m,n+m−l (u) (l = 1, . . . , n + m − 1) are called creation operators [5] and are usually denoted by Cn+m−l (u) = Tn+m,n+m−l (u), l = 1, . . . , n + m − 1.
(5.19)
In the F -basis, they become −1 C˜ n+m−l (u) = F1...N Cn+m−l (u)F1...N , l = 1, . . . , n + m − 1.
(5.20)
Let us denote Tn+m,n+m (u) by D(u) and the corresponding operator in the F -basis by −1 ˜ . D(u) = F1...N D(u)F1...N
(5.21)
˜ Proposition 5. D(u) is a diagonal matrix given by ˜ D(u) = ⊗N i=1 diag (a0i , . . . , a0i , 1)(i) .
(5.22)
Proof. From (3.20), we derive that D(u)P0n+m = Tn+m,n+m (u)en+m,n+m = P0n+m T0,1...N (u)P0n+m .
(5.23)
Drinfeld Twists and Algebraic Bethe Ansatz
103
Acting the F -matrix from the left on the both sides of the above equation, we have ∗
F1...N D(u)P0n+m =
N
S(σ, ασ )
σ ∈SN ασ (1) ...ασ (N )
∗
=
σ σ Pσα(i) R1...N P0m+n T0,1...N (u)P0m+n
i=1 N
S(σ, ασ )
σ ∈SN ασ (1) ...ασ (N )
σ σ Pσα(i) P0m+n T0,σ (1...N) (u)P0m+n R1...N .
i=1
(5.24) Following [7], we can split the sum index m + n,
according to the number of occurrences of the
∗
N
F1...N D(u)P0n+m =
∗
σ ∈SN k=0 ασ (1) ...ασ (N )
×
N−k
N
S(σ, ασ )
α
σ (j ) δασ (j ) ,m+n Pσ (j )
j =N−k+1
α
σ (j ) m+n σ Pσ (j T0,σ (1...N) (u)P0m+n R1...N . ) P0
(5.25)
j =1 σ . We have Consider the prefactor of R1...N N−k j =1
=
N−k
=
m+n Pσm+n T0,σ (1...N) (u)P0m+n (j ) P0
j =N −k+1 N
α
σ (j ) Pσ (j )
j =1 N−k
N
α
σ (j ) Pσ (j )
R0 σ (j )
m+n m+n
j =N −k+1
P m+n T0,σ (1...N−k) (u)P0m+n m+n m+n 0
α
σ (j ) m+n Pσ (j T0,σ (1...N −k) (u)P0m+n ) P0
j =1
=
N−k
=
i=1
Pσm+n (j )
j =N−k+1
Pσm+n (j )
j =N−k+1
R0 σ (i)
m+nασ (i) m+nασ (i)
i=1 N−k
N
N
a0 σ (i)
N−k j =1
α
σ (j ) Pσ (j )
N−k
α
σ (j ) Pσ (j )
j =1 N
N
n+m Pσm+n (j ) P0
j =N−k+1 n+m Pσm+n , (j ) P0
(5.26)
j =N−k+1
where a0i = a(u, zi ). Substituting (5.26) into (5.25), we have F1...N D(u) = ⊗N i=1 diag (a0i , . . . , a0i , 1)(i) F1...N . This completes the proof of the proposition.
(5.27)
By means of the expressions of the generators of Uq (gl(m|n)) in the previous subsection, combining with Theorem 1 and Proposition 5, we have
104
W.-L. Yang, Y.-Z. Zhang, S.-Y. Zhao
Theorem 3. In the F -basis the creation operators Cn+m−l (u)(l = 1, . . . , n + m − 1) are given by l [n+m] ˜ ˜ E˜ n+m−l,n+m q − k=1 Hn+m−k C˜ n+m−l (u) = q −(−1) E˜ n+m−l,n+m D(u) − D(u) −
l−1
(1−q −2(−1)
[n+m−α]
)C˜ n+m−α (u)E˜ n+m−l,n+m−α q −
l
k=α+1 Hn+m−k
.
α=1
(5.28) For some special values of m and n, we have: • For m = 0 and n = 2 (namely, the Uq (gl(2) case), our general results reduce to those in [1]. • For the Uq (gl(2|1)) case, ˜ ˜ E˜ 2,3 q −H2 − D(u) C˜ 2 (u) = q −1 E˜ 2,3 D(u) N e−(u−zi ) sinh η 2,3 E sinh(u − zi + η) (i) i=1 sinh(u − zj ) 2 sinh(u − zj ) cosh η ⊗j =i diag , (5.29) , ,1 sinh(u − zj + η) sinh(u − zj + η) (j ) ˜ ˜ E˜ 1,3 q −H1 −H2 − (1 − q 2 )C˜ 2 (u)E˜ 1,2 q −H1 C˜ 1 (u) = q −1 E˜ 1,3 D(u) − D(u)
=
=
N 2 sinh(u − zj ) cosh η e−(u−zi ) sinh η 1,3 E(i) ⊗j =i diag , sinh(u − zi + η) sinh(u − zj + η) i=1 sinh(u − zj ) sinh(zi − zj + η) ,1 sinh(zi − zj ) sinh(u − zj + η) (j ) N ezj −u sinh(u − zi ) sinh2 η[ezj −zi + 2 sinh(zi − zj )] 1,2 2,3 E(i) ⊗ E(j ) sinh(u − zi + η) sinh(u − zj + η) sinh(zi − zj ) i=j =1 2 sinh(u − zk ) cosh η sinh(u − zk ) sinh(zi − zk + η) , ,1 ⊗k=i,j . sinh(u − zk + η) sinh(zi − zk ) sinh(u − zk + η) (k) (5.30)
+
6. Bethe Vectors in the F -Basis The explicit expressions of the creation operators of the Uq (gl(m|n)) monodromy matrix in the F -basis, given in the previous section, enable us to resolve the hierarchy of the nested Bethe vectors of the model associated with Uq (gl(m|n)). Here, we take the quantum t − J model (i.e. the Uq (gl(2|1))-model) as an example to demonstrate the procedure. The generalization to the general Uq (gl(m|n)) case is straightforward. In the framework of the standard algebraic Bethe ansatz method [19], the Bethe vector of the quantum supersymmetric t-J model is given by d1 ...dα
N = ( (1) Cd1 (v1 ) . . . Cdα (vα )|vac, (6.1) α ) d1 ...dα
Drinfeld Twists and Algebraic Bethe Ansatz
105
where α is a positive integer and di = 1, 2, |vac is the pseudo-vacuum state 0 0 , |vac = ⊗N i=1 1 (i)
(6.2)
(1)
and ( α )d1 ...dα are functions of the spectral parameters vj , and are the vector components of the nested Bethe vector (1) . The nested Bethe vector (1) is given by (1)
(1)
(1)
(1) (1) (1) (1)
(1) α = C (v1 )C (v2 ) · · · C (vβ )|vac ,
(6.3)
where β is a positive integer and |vac(1) is the nested pseudo-vacuum state |vac
(1)
=
⊗αi=1
0 . 1 (i)
(6.4)
The creation operator of the nested Uq (gl(2)) system C (1) (u) is the lower-triangular entry of the nested monodromy matrix T (1) (v (1) ), T (1) (v (1) ) = r0α (v (1) − vα )r0 α−1 (v (1) − vα−1 ) . . . r01 (v (1) − v1 ) (1) (1) A (v ) B (1) (v (1) ) ≡ , C (1) (v (1) ) D (1) (v (1) )
(6.5)
and the nested R-matrix is
c12 0 0 0 a12 −b+ 12 r12 (u1 , u2 ) ≡ r12 (u1 − u2 ) = 0 −b− a12 12 0 0 0
0 0 . 0 c12
(6.6)
Equation (6.5) implies that the nested system is a Uq (gl(2)) spin chain on α-site lattice and the corresponding inhomogeneous parameters are {vi |i = 1, . . . , α}. So, in the following, we shall adopt the same convention for the nested system as that in previous sections but the inhomogeneous parameters will be replaced by {vi }. Acting the associated F -matrix on the pseudo-vacuum state (6.2), one finds that the pseudo-vacuum state is invariant. It is due to the fact that only terms with roots equal to 3 will produce non-zero results. Therefore, the Uq (gl(2|1)) Bethe vector (6.1) in the F -basis can be written as ˜ N (v1 , . . . , vα ) ≡ F1...N N (v1 , . . . , vα )
d1 ...dα ˜ = ( (1) Cd1 (v1 ) . . . C˜ dα (vα )|vac. α )
(6.7)
d1 ...dα (1)
The c-number coefficient ( α )d1 ...dα has to be evaluated in the original basis, not in the F -basis.
106
W.-L. Yang, Y.-Z. Zhang, S.-Y. Zhao
6.1. Nested Bethe vectors in the F -basis. Let us first compute the nested Bethe vectors in the F -basis. For the nested R-matrix (6.6), we define the α-site F and F ∗ -matrices by (1) F1...α
=
2
∗
α
σ (j ) ¯ (1) σ Pσ (j ) S (c, σ, ασ )r1...α ,
α
σ ∈Sα ασ (1) ...ασ (α) =1 j =1 2 ∗∗
∗(1)
F1...α =
−1 S¯ (1) (c, σ, ασ )rσσ(1...α)
σ ∈Sα ασ (1) ...ασ (α) =1
α
(6.8)
α
σ (j ) Pσ (j ) ,
(6.9)
j =1
respectively, where the c-number function S¯ (1) is given by S¯ (1) (c, σ, ασ ) ≡ exp{
α
δασ (k) ,ασ (l) ln(1 + cσ (k)σ (l) )}.
(6.10)
l>k=1
Therefore the inverse of the F -matrix can be represented in terms of the F ∗ -matrix as (1) (1) ∗(1) ¯ )−1 , (F1...α )−1 = F1...α ( (6.11) ij i<j
with
¯ (1) = diag ij
sin(vj − vi ) 4 sinh2 (vi − vj ) cosh2 η , , sinh(vi − vj + η) sinh(vi − vj − η) sinh(vj − vi + η)
sinh(vi − vj ) 4 sinh2 (vi − vj ) cosh2 η , . sinh(vi − vj + η) sinh(vi − vj + η) sinh(vi − vj − η)
With the help of the F -matrix (6.8) and its inverse, one may compute the α-site Uq (gl(2)) generator E˜ 1,2 in the F -basis, α eη sinh(vi − vj + η) 1,2 −η ˜ E(i) ⊗j =i 2e cosh η, . (6.12) E1,2 = 2 sinh(vi − vj ) cosh η (j ) i=1
(1) (1) Define D˜ (1) (u) = F1...α D (1) (u)(F1...α )−1 . From Proposition 5, we obtain sinh(u − vi ) sinh(u − vi − η) D˜ (1) (u) = ⊗αi=1 , . sinh(u − vi + η) sinh(u − vi + η) (i)
(6.13)
The creation operator in the F -basis is then obtained with the help of the nested R-matrix (6.6) and Theorem 3 in the case of m = 2 and n = 0, and is given by ˜ ˜ E˜ 1,2 )q −H1 C˜ (1) (u) = (q E˜ 1,2 D(u) − D(u) α e−(u−vi ) sinh η 1,2 2 sinh(u − vj ) cosh η =− E ⊗j =i , sinh(u − vi + η) (i) sinh(u − vj + η) i=1 sinh(vi − vj + η) sinh(u − vj − η) . 2 sinh(vi − vj ) sinh(u − vj + η) cosh η (j )
(6.14)
Drinfeld Twists and Algebraic Bethe Ansatz
107
(1)
(1)
Applying F1...α to the nested Bethe vector α (6.3), we obtain (1) (1) (1) (1) (1) (1) ˜ (1)
α (v1 , . . . , vβ ) ≡ F1...α (v1 , . . . , vβ )
= s(c)C˜ (1) (v1 )C˜ (1) (v2 ) . . . C˜ (1) (vβ )|vac(1) , (6.15) (1)
(1)
(1)
(1) where we have used F1...α |vac(1) = i<j (1 + cij )|vac(1) ≡ s(c)|vac(1) . Substituting C˜ (1) (v) into (6.15), we obtain (1) (1) (1) ˜ (1) (1) ˜ (1) (1) ˜ (1)
α (v1 , . . . , vβ ) = s(c)C (v1 ) . . . C (vβ ) |vac (1) (1) (1) = s(c) Bβ (v1 , . . . , vβ |vi1 , . . . , viβ )E(i1,2 . . . E(i1,2 |vac(1) , (6.16) β) 1) i1 <...
where (1)
(1)
(1)
Bβ (v1 , . . . , vβ |v1 , . . . , vβ ) (1) β e−(vk −vσ (k) ) sinh η = − (1) sinh(vk − vσ (k) + η) σ ∈Sβ k=1 ×
×
α
sinh(vσ (k) − vj + η) sinh(vk − vj − η)
j =σ (k),... ,σ (β)
2 sinh(vσ (k) − vj ) sinh(vk − vj + η) cosh η
(1)
(1)
β (1) 2 cosh η sinh(vk − vσ (l) ) (1)
l=k+1
sinh(vk − vσ (l) + η)
(6.17)
.
6.2. Bethe vectors of the quantum supersymmetric t-J model in the F -basis. Now back to the Bethe vector (6.7) of the quantum supersymmetric t-J model. As is shown in Appendix B, the Bethe vector is invariant (modulo overall factor) under the exchange of arbitrary spectral parameters: ˜ N (vσ (1) , . . . , vσ (α) ) =
1
˜
N (v1 , . . . σ c1...α
, vα ), σ ∈ Sα ,
(6.18)
σ has the decomposition law where c1...α
σ σ σ c1...α = cσσ (1...α) c1...α ,
(6.19)
σi and c1...α = ci i+1 ≡ c(vi , vi+1 ) for an elementary permutation σi . This result is a generalization of that in [20, 21]. This invariance enables us to concentrate on a particularly simple term in the sum (6.7) of the following form with p1 number of di = 1 and α − p1 number of dj = 2,
C˜ 1 (v1 ) . . . C˜ 1 (vp1 )C˜ 2 (vp1 +1 ) . . . C˜ 2 (vα ).
(6.20)
In the F -basis, the commutation relation between Ci (v) and Cj (u), i.e. (B.4), becomes C˜ i (v)C˜ j (u) = −
1 b(u, v) ˜ C˜ j (u)C˜ i (v) + Cj (v)C˜ i (u). a(u, v) a(u, v)
(6.21)
108
W.-L. Yang, Y.-Z. Zhang, S.-Y. Zhao
Then using (6.21), all C˜ 1 ’s in (6.20) can be moved to the right of all C˜ 2 ’s, yielding C˜ 1 (v1 ) . . . C˜ 1 (vp1 )C˜ 2 (vp1 +1 ) . . . C˜ 2 (vα ) = = g(v1 , . . . , vα )C˜ 2 (vp1 +1 ) . . . C˜ 2 (vα )C˜ 1 (v1 ) . . . C˜ 1 (vp1 ) + . . . ,
(6.22)
p1 α where g(v1 , . . . , vα ) = k=1 l=p1 +1 (−1/a(vl , vk )) is the contribution from the first term of (6.21) and “. . . ” stands for other terms contributed by the second term of (6.21). It is easy to see that these other terms have the form C˜ 2 (vσ (p1 +1) ) . . . C˜ 2 (vσ (α) )C˜ 1 (vσ (1) ) . . . C˜ 1 (vσ (p1 ) ),
(6.23)
with σ ∈ Sα . Substituting (6.22) into the Bethe vector (6.7), we obtain 11...12...2 ˜ p1 (v1 , . . . , vα ) = ( (1)
α ) N
p1
α
k=1 l=p1 +1
1 − a(vl , vk )
× C˜ 2 (vp1 +1 ) . . . C˜ 2 (vα )C˜ 1 (v1 ) . . . C˜ 1 (vp1 )|vac + . . . , (6.24) where and below, we use the up-index p1 to denote the Bethe vector corresponding to the quantum number p1 . All other terms in (6.24) (denoted as “ . . . ”) are to be obtained from the first term by the permutation (exchange) symmetry. Then we have, ˜ p1 (v1 , . . . , vα ) =
N =
p1 1 σ 11...12...2 c1...α ( (1),σ ) α p1 !(α − p1 )! σ ∈Sα
α
−
k=1 l=p1 +1
1 a(vσ (l) , vσ (k) )
× C˜ 2 (vσ (p1 +1) ) . . . C˜ 2 (vσ (α) )C˜ 1 (vσ (1) ) . . . C˜ 1 (vσ (p1 ) ) |vac ,
(6.25)
(1),σ (1) where ( α )11...12...2 ≡ (fˆσ α )11...12...2 with fˆσ defined by (B.1) in Appendix B. (1) 1...12...2 in (6.25), which has to be evaluated in the original baWe now show that ( α ) sis, is invariant (modulo an overall factor) under the action of the Uq (gl(2)) F -matrix, i.e. 11...12...2 11...12...2 ˜ (1) ( (1) = (t (c))−1 (
, α ) α )
(6.26)
where the scalar factor t (c) is
t (c) =
p1
(1 + c¯ij )
j >i=1
α
(1 + c¯ij ), c¯ij = c(vi , vj ),
j >i=p1 +1
so that it can be expressed in the form of (6.16).
(6.27)
Drinfeld Twists and Algebraic Bethe Ansatz
109
Write the nested pseudo-vacuum vector in (6.4) as |vac(1) ≡ |2 · · · 2(1) ,
(6.28)
where the number of 2 is α. Then the nested Bethe vector (6.15) can be rewritten as
(1)
(1) (1)
(1) α (v1 . . . vp1 ) ≡ | α =
d1 ...dα ( (1) |d1 . . . dα (1) . α )
(6.29)
d1 ...dα (1)
Acting the Uq (gl(2)) F -matrix F1...α from the left on the above equation, we have (1) (1) (1) (1) ˜ (1) ˜ (1)
α (v1 . . . vp1 ) ≡ | α = F1...α | α =
d1 ...dα ˜ (1) (
|d1 . . . dα (1) . α )
(6.30)
d1 ...dα
It follows that (1) 1...12...2 (1) ˜ (1) ˜ (1) = 1 . . . 12 . . . 2|
(
α ) α = 1 . . . 12 . . . 2|F1...α | α
= 1 . . . 12 . . . 2|
∗
α
σ (j ) ¯ (1) σ (1) Pσ (j ) S (c, σ, ασ )R1...α | α
α
σ ∈Sα ασ (1) ...ασ (α) j =1
" α " ∗ ασ (j ) "" Pσ (j ) " = 1 . . . 12 . . . 2| " α ...α σ (1) σ (α) j =1
(6.31) S¯ (1) (c, σ, ασ )| (1) α
σ =id
=
t (c)1 . . . 12 . . . 2| (1) α
=
1...12...2 t (c)( (1) , α )
(6.32) (6.33)
where the scalar factor t (c) is given by (6.27). Summarizing, we propose the following form of the Uq (gl(2|1)) Bethe vector ˜ p1 (v1 , . . . , vα ) =
N =
p1 α 1 σ c1...α p1 !(α − p1 )! σ ∈Sα
i=1 j =p1 +1
2 sinh(vσ (i) − vσ (j ) ) cosh η sinh(vσ (i) − vσ (j ) + η)
(1) ×Bp(1) (v1 , . . . , vp(1) |vσ (1) , . . . 1 1 p α 1
×
k=1 l=p1 +1
, vσ (p1 ) ) 1 C˜ 2 (vσ (p1 +1) ) . . . C˜ 2 (vσ (α) ) − a(vσ (l) , vσ (k) )
× C˜ 1 (vσ (1) ) . . . C˜ 1 (vσ (p1 ) ) |vac . Substituting (5.29) and (5.30) into the above relation, we finally have
(6.34)
110
W.-L. Yang, Y.-Z. Zhang, S.-Y. Zhao
Proposition 6. The nested Bethe vector of the quantum t − J model is given by ˜ p1 (v1 , . . . , vα )
N 1 = p1 !(α − p1 )!
(1)
Bα,p1 (v1 , . . . , vα ; v1 , . . . , vp(1) |zi1 , . . . , ziα ) 1
i1 <...
α
×
E(i2,3 j)
j =p1 +1
p1 j =1
E(i1,3 |vac, j)
(6.35)
where {i1 , i2 , . . . , ip1 } ∩ {ip1 +1 , ip1 +2 , . . . , iα } = ∅ and (1)
Bα,p1 (v1 , . . . , vα ; v1 , . . . , vp(1) |zi1 , . . . , ziα ) = 1 =
σ c1...α
σ ∈Sα p1
i=1 α
×
k=1
p1
α j =p1 +1
l=p1 +1
2 sinh(vσ (i) − vσ (j ) ) cosh η sinh(vσ (i) − vσ (j ) + η)
sinh(vσ (l) − zik ) sinh(vσ (l) − vσ (k) + η) − sinh(vσ (l) − vσ (k) ) sinh(vσ (l) − zik + η)
∗ ×Bα−p (vσ (p1 +1) , . . . , vσ (α) |zip1 +1 , . . . , ziα ) 1
(v1 , . . . , vp(1) |vσ (1) , . . . , vσ (p1 ) )Bp∗1 (vσ (1) , . . . , vσ (p1 ) |zi1 , . . . , zip1 ). ×Bp(1) 1 1 (6.36) (1)
Here the function Bp∗ (v1 , . . . , vp |z1 , . . . , zp ) is given Bp∗ (v1 , . . . , vp |z1 , . . . , zp ) = =
σ ∈Sp
sign(σ )
p p e−(vk −zσ (k) ) sinh η 2 sinh(vk − zσ (l) ) cosh η . sinh(vk − zσ (k) + η) sinh(vk − zσ (l) + η)
k=1
l=k+1
(6.37)
7. Discussions We have constructed the factorizing F -matrices for the supersymmetric model associated with quantum superalgebra Uq (gl(m|n)) with generic m and n, which includes the quantum supersymmetric t-J model as a special case. We have obtained the completely symmetric representations for the creation operators of the model in the F -basis. Our results make possible a complete resolution of the hierarchy of its nested Bethe vectors. As an example, we have given the explicit expressions of the Bethe vectors of the quantum t − J model in F -basis. Our results are new even for the special m = 0 case, which give the results of the general model associated with Uq (gl(n)). (For m = 0 and n = 2, our results reduce to those in [1].) Authors in [22] solved the quantum inverse problem of the supersymmetric t-J model i,j in the original basis. Namely, they reconstructed the local operators (E(k) ) in terms of
Drinfeld Twists and Algebraic Bethe Ansatz
111
operators figuring in the gl(2|1) monodromy matrix. Their results should be generalizable to the Uq (gl(2|1)) case. Then together with the results of the present paper in the F -basis one should be able to get the exact representations of form factors and correlation functions of the quantum supersymmetric t-J model. These are under investigation and results will be reported elsewhere. Acknowledgements. This work was financially supported by the Australia Research Council. S. -Y. Zhao has also been supported by the UQ Postdoctoral Research Fellowship.
Appendix A: Proof of Theorem 1 Proposition 2 allows one to derive the following equations: (E n+m−l,n+m )0...N = En+m−l,n+m + en+m−l,n+m q +
l−1
(1−q −2(−1)
[n+m−α]
l
k=1 Hn+m−k
)en+m−α,n+m En+m−l,n+m−α q
α
k=1 Hn+m−k
,
α=1
l = 1, . . . , n + m − 1, P σc (E n+m−l,n+m )0...N (P σc )−1 = en+m−l,n+m + q +
l−1
(1−q −2(−1)
[n+m−α]
)en+m−l,n+m−α q
(A.1)
l
k=1 hn+m−k
α
k=1 hn+m−k
En+m−l,n+m
En+m−α,n+m ,
i=1
l = 1, . . . , n + m − 1.
(A.2)
Taking x = E n+m−l,n+m and σ = σc and using (3.14), then (3.19) becomes T (u) (E n+m−l,n+m )0...N = P σc (E n+m−l,n+m )0...N (P σc )−1 T (u).
(A.3)
Substituting (3.20), (A.1) and (A.2) into the above equation, we have, for the L.H.S. of the resulting relation,
L.H.S. =
n+m
(−1)([i]+[n+m−l])([n+m−l]+[n+m]+1) ei,n+m Ti,n+m−l (u)q
l
k=1 Hn+m−k
i=1
+
n+m
ei,j (−1)[i]+[j ] Ti,j (u)En+m−l,n+m
i,j =1
+
n+m l−1
(1 − q −2(−1)
[n+m−α]
)(−1)([i]+[n+m−α])([n+m−α]+[n+m]+1)
i=1 α=1
×ei,n+m Ti,n+m−α (u)En+m−l,n+m−α q
α
k=1 Hn+m−k
.
(A.4)
112
W.-L. Yang, Y.-Z. Zhang, S.-Y. Zhao
Similarly for the R.H.S. of the resulting relation, we obtain R.H.S. =
n+m
(−1)([n+m−l]+[n+m]+1)([i]+[j ]) q
l
k=1 hn+m−k
ei,j En+m−l,n+m Ti,j (u)
i,j =1
+
n+m
(−1)[n+m]+[i] en+m−l,i Tn+m,i (u)
i=1
+
n+m l−1
(1 − q −2(−1)
[n+m−α]
i=1 α=1
×en+m−l,n+m−α q
α
)(−1)([n+m−α]+[i])([n+m−α]+[n+m]+1)
k=1 hn+m−k
en+m−α,i En+m−α,n+m Tn+m−α,i (u). (A.5)
Comparing the coefficients of the en+m,n+m term on both sides, we obtain q −(−1)
[n+m]
En+m−l,n+m Tn+m,n+m (u) =
= Tn+m,n+m (u)En+m−l,n+m + Tn+m,n+m−l (u)q +
l−1
(1 − q −2(−1)
[n+m−α]
l
k=1 Hn+m−k
)Tn+m,n+m−α (u)En+m−l,n+m−α q
α
k=1 Hn+m−k
,
α=1
(A.6) which leads to the recursive relation (3.37). Appendix B: The Exchange Symmetry of the Bethe Vector For the Bethe vector N (v1 , . . . , vα ) of the quantum supersymmetric t-J model, we define the exchange operator fˆσ = fˆσil . . . fˆσik by fˆσ N (v1 , v2 , . . . , vα ) = N (vσ (1) , vσ (2) , . . . , vσ (α) ),
(B.1)
where σ ∈ Sα , and {σi } are the elementary permutations of Sα . We first study the exchange symmetry for the elementary exchange operator fˆσi which exchanges the parameter vi and vi+1 . Acting fˆσi on the Bethe vector of Uq (gl(2|1)) (6.7), we have fˆσi N (v1 , v2 , . . . , vα ) = N (v1 , . . . , vi+1 , vi , . . . , vα ) i )d1 ...dα C (v ) . . . C (v = ( (1),σ d1 1 di i+1 )Cdi+1 (vi ) . . . Cdα (vα )|vac, (B.2) α d1 ,... ,dα (1),σi
(1),σ
where {( α i )d1 ...dα } are the vector components of the nested Bethe vector α constructed by the nested monodromy matrix (1)
(1)
(1)
T (1),σi (u) = L(1) α (u, vα ) . . . Li+1 (u, vi )Li (u, vi+1 ) . . . L1 (u, v1 ), (1)
where the local L-operator is defined by Li (u, v) = r0i (u, v).
(B.3)
Drinfeld Twists and Algebraic Bethe Ansatz
113
From the GYBE (3.15), one can derive the commutation relation between Ci (u) and Cj (v), which is given by Ci (u)Cj (v) =
rˇ (u, v)kl ij Ck (v)Cl (u).
(B.4)
k,l
Here the braided r-matrix rˇ (u, v) ≡ Pr(u, v), P permutes the tensor product spaces of the 2-dimensional Uq (gl(2))-module. Then, by (B.4), (B.2) becomes fˆσi N (v1 , v2 , . . . , vα ) =
i )d1 ...dα C (v ) . . . ( (1),σ d1 1 α
d1 ,... ,dα
×(ˇr (vi+1 , vi ))kdi dli+1 Ck (vi )Cl (vi+1 ) . . . Cdα (vα )|vac. (B.5) We now compute the action of (ˇr (vi+1 , vi ))kdi dli+1 on ( (1),σi )d1 ...dα . One checks that the rˇ -matrix satisfies the YBE, (1)
(1)
rˇi i+1 (vi+1 , vi )Li+1 (u, vi )Li (u, vi+1 ) (1)
(1)
= Li+1 (u, vi+1 )Li (u, vi )ˇri i+1 (vi+1 , vi ) .
(B.6)
Therefore, acting rˇ on T (1),σi (u), we have rˇi i+1 (vi+1 , vi )T (1),σi (u) = T (1) (u)ˇri i+1 (vi+1 , vi ).
(B.7)
Thus, because rˇi i+1 (vi+1 , vi ) v2 ⊗ v2 = ci i+1 (vi+1 , vi )v2 ⊗ v2 =
1 v2 ⊗ v 2 , ci i+1
we obtain
i )d1 ...di di+1 ...dα = (ˇr (vi+1 , vi ))kdi dl i+1 ( (1),σ α
di di+1
1 ci i+1
d1 ...kl...dα ( (1) . α )
(B.8)
Changing the indices k, l to di , di+1 , respectively, and substituting the above relation into (B.5), we obtain the exchange symmetric relation of the Bethe vector of Uq (gl(2|1)), fˆσi N (v1 , v2 , . . . , vα ) =
1 ci i+1
N (v1 , v2 , . . . , vα ),
(B.9)
for the elementary permutation operator σi . It follows that under the action of the exchange operator fσ , fˆσ N (v1 , v2 , . . . , vα ) = σ where c1...α is defined in (6.19).
1
N (v1 , v2 , . . . σ c1...α
, vα ),
(B.10)
114
W.-L. Yang, Y.-Z. Zhang, S.-Y. Zhao
References 1. Maillet, J.M., Sanchez de Santos, I.: Drinfel’d twists and algebraic Bethe ansatz. http://arxiv.org/qalg/9612012, 1996 2. Drinfeld, V.G.: Constant quasi-classical solutions of the Yang-Baxter quantum equation. Sov. Math. Dokl. 28, 667 (1983) 3. Kitanine, N., Maillet, J.M., Terras, V.: Form factors of the XXZ Heisenberg spin–1/2 finite chain. Nucl. Phys. B 554, 647 (1999) 4. Izergin, A.G., Kitanine, N., Maillet, J.M., Terras, V.: Spontaneous magnetization of the XXZ Heisenberg spin-1/2 chain. Nucl. Phys. B 554, 679 (1999) 5. Korepin, V.E., Bogoliubov, N.M., Izergin, A.G.: Quantum Inverse Scattering Method and correlation Function. Cambridge: Cambridge Univ. Press, 1993 6. Terras, V.: Drinfel’d twists and functional Bethe ansatz. Lett. Math. Phys. 48, 263 (1999) 7. Albert, T.-D., Boos, H., Flume, R., Ruhlig, K.: Resolution of the nested hierarchy for the rational sl(n) models. J. Phys. A 33, 4963 (2000) 8. Albert, T.-D., Boos, H., Flume, R., Poghossian, R.H., Ruhlig, K.: An F -twisted XYZ model. Lett. Math. Phys. 53, 201 (2000) 9. Albert, T.-D., Ruhlig, K.: Polarization free generators for the Belavin model. J. Phys. A 34, 1569 (2001) 10. Yang, W.-L., Zhang, Y.-Z., Zhao, S.-Y.: Drinfeld twists and algebraic Bethe Ansatz of the supersymmetric t-J model. JHEP 12, 038 (2004) 11. Zhao, S.-Y., Yang, W.-L., Zhang, Y.-Z.: Drinfeld twists and symmetric Bethe vectors of supersymmetric Fermion model. JSTAT, P04005 (2005) 12. Perk, J.H.H., Schultz, C.L.: New families of commuting transfer matrices in q-state vertex models, Phys. Lett. A84, 407 (1981) 13. Kulish, P.P., Sklyanin, E.K.: On solutions of the Yang-Baxter equation. J. Soviet Math. 19, 1596 (1982) 14. Kulish, P.P.: Integrable graded magnets. J. Soviet Math. 35, 2648 (1986) 15. Essler, F.H.L., Korepin, V.E., Schoutens, K.: New exactly solvable model of the strongly correlated electrons motivated by high-Tc superconductivity. Phys. Rev. Lett. 68, 2960 (1992) 16. Bracken, A.G., Gould, M.D., Links, J.R., Zhang, Y.-Z.: A new supersymmetric and exactly solvable model of correlated electrons. Phys. Rev. Lett. 74, 2768 (1995) 17. Yamane, H.: On defining relations of the affine Lie superalgebras and their quantized universal enveloping superalgebras. http://arxiv.org/abs/q-alg/9603015, 1996 18. Bazhanov, V.V., Shadrikov, A.G.: Trigonometric sotions of the triangle equations and simple Lie superalgebras. Theor. Math. Phys. 73, 1302 (1988) 19. Faddeev, L.D., Sklyanin, E.K., Takhtajan, L.A.: The quantum inverse scattering method. Theor. Math. Phys. 40, 688 (1979) 20. Takhtajan, L.A.: Quantum inverse scattering method and algebraized matrix Bethe-Ansata. J. Sov. Math. 23, 2470 (1983) 21. de Vega, H.J.: Yang-Baxter algebras, integrable theories and quantum groups. Int. J. Mod. Phys. A 44, 2371 (1989) 22. G¨ohmann, F., Korepin, V.E.: Solution of the quantum inverse problem. J. Phys. A 33, 1199 (2000) Communicated by L.Takhtajan
Commun. Math. Phys. 264, 115–144 (2006) Digital Object Identifier (DOI) 10.1007/s00220-006-1541-8
Communications in
Mathematical Physics
2-Matrix versus Complex Matrix Model, Integrals over the Unitary Group as Triangular Integrals B. Eynard1 , A. Prats Ferrer2 1 2
Service de Physique Th´eorique de Saclay, CEA/DSM/SPhT - CNRS/SPM/URA 2306, 91191 Gif-sur-Yvette Cedex, France. E-mail:
[email protected] Universitad Barcelona, Departament d’Estructura i constituents de la Mat`eria, Av. Diagonal 647, 08028 Barcelona, Spain. E-mail:
[email protected]
Received: 2 March 2005 / Accepted: 20 September 2005 Published online: 9 March 2006 – © Springer-Verlag 2006
Abstract: We prove that the 2-hermitian matrix model and the complex-matrix model obey the same loop equations, and as a byproduct, we find a formula for Itzykzon-Zuber type integrals over the unitary group. Integrals over U (n) are rewritten as gaussian integrals over triangular matrices and then computed explicitly. That formula is an efficient alternative to the former Shatashvili’s formula. 1. Introduction It has been noticed for a long time now, that the so-called “Two-Hermitian-MatrixModel” (introduced in particular for quantum gravity [19, 9]) and the so-called “Complex-Matrix-Model” (used in particular for its applications to Laplacian growth models [27, 28], and string theory) share lots of similarities: They have the same leading large N expansion properties, and, both are associated to some ensembles of biorthogonal polynomials which have formally the same properties. Here, we add a new piece to make this correspondence more precise, we prove that both models have the same loop equations. Both models are not defined for the same weights, in fact, the set of weights for which one model is well defined has no intersection with the set of weights for which the other model is well defined. However, each model can be analytically continued to a larger set of weights, and in that sense, the two models coincide. When written in terms of eigenvalues, this identification of the 2-hermitian-matrixmodel and complex-matrix-model has some interesting corollary: it gives a formula for computing integrals (of the Itzykzon-Zuber type) over the unitary group, as gaussian integrals over triangular matrices. Therefore, we obtain a very explicit formula for all correlators of Shatashvili type [24]. In [24] S. Shatashvili found a formula for all U (n) correlation functions, but his formula still contains integrals, is not explicitly symmetric in all variables, and is very difficult to use for practical purposes, such as [5]. In the particular case of the 2-point correlation function, Morozov has found a much simpler
116
B. Eynard, A.P. Ferrer
formula [23]. In [23] A. Morozov computed it for U (n) with n ≤ 3 and conjectured it for n > 3. Morozov’s formula was later proven for all n in [5], and written in an even simpler form [12]. Here, we find a natural generalization of Morozov’s formula. The formula we find here, contains no integration, it gives the U (n) correlation functions as the sum of a finite number of terms, and is very efficient for effective computations. It also provides an alternative new proof of the Itzykzon-Zuber formula. The derivations proposed in this article are elementary, and it would be interesting to put them in the more general framework of group representation theory [20, 16]. The main results presented in this paper are: • Theorem 3.3 and in particular Remark 3.3, which states the equivalence between the Hermitian-2-matrix model and the complex-matrix model: Hn ×Hn
dM1 dM2 F (M1 , M2 ) e−[
≡
GLn (C)
dZ F (Z, Z † ) e−[
α1 2 α2 2 2 M1 + 2 M2 +γ
α1 2 α2 † 2 2 Z + 2 Z +γ
Tr M1 M2 ]
Tr ZZ † ]
(1.1)
.
The definitions of each term and the meaning of that equality are explained in Sect. 3.3. • Theorem 4.1, which allows to compute U (n) integrals as triangular integrals:
dU F (X, U Y U † ) e− Tr XU Y U
U (n)
∝
σ,τ ∈n2 (−1)
†
σ (−1)τ e− Tr Xσ Yτ
T (n) F (Xσ
+ T , Yτ + T † ) e− Tr T T
†
(X)(Y )
(1.2)
for any polynomial invariant function F . • Theorem 5.2, which gives a formula for computing triangular matrix gaussian integrals. We parametrize polynomial invariant functions by pairs of permutations (of some size R), and a basis is written Fπ,π . Theorem 5.2 gives the result of integration over triangular matrices:
T (n) dT
e− Tr T T Fπ,π ( x , y, X + T , Y + T † ) − Tr T T † T (n) dT e †
= M(R) ( x , y, Xn , Yn ) M(R) ( x , y, Xn−1 , Yn−1 ) . . . M(R) ( x , y, X1 , Y1 )
π,π
, (1.3)
where M(R) ( x , y, Xn , Yn ) is the matrix of size R!, indexed by pairs of permutations: M(R) x , y, Xn , Yn ) π,ρ (
=
R i=1
1 . δπ(i),ρ(i) + (xi − Xn )(yπ(i) − Yn )
(1.4)
Theorem 5.3 shows that the matrices in Eq.1.3 commute together, and can be simultaneously diagonalized.
2-Matrix vs. Complex Matrix Model
117
• Theorem 6.1, which gives a formula for computing correlation functions in terms of biorthogonal polynomials: x, y , M1 , M2 ) e− Tr (V1 (M1 )+V2 (M2 )+M1 M2 ) Hn ×Hn dM1 dM2 Fπ,π ( − Tr (V1 (M1 )+V2 (M2 )+γ M1 M2 ) Hn ×Hn dM1 dM2 e = Mdet M(R) ( x , y, Q, P t ) , (1.5) π,π
• • • • •
where notations are explained in Sect. 6.2. Outline: In Part 2 we give definitions of groups and measures. In Part 3, we prove the equivalence between the Hermitian-2-matrix model and the complex-matrix model, by showing that they have the same loop equations. In Part 4, we prove the identity between U (n) integrals and triangular integrals, and give some examples. In particular we rederive Itzykson-Zuber’s formula and Morozov’s formula. In Part 5, we compute the triangular integrals, by parametrizing polynomial invariant functions with pairs of permutations. In particular we compute explicitly all four point functions. In Part 6, we integrate over eigenvalues using biorthogonal polynomial techniques, and get expressions for correlation functions.
2. Definitions 2.1. Ensembles. Let • U (n) := group of n × n unitary matrices, with the normalized Haar measure. • Hn := group of n × n hermitian matrices, with the Lebesgue measure: dM := dMii dReMij dImMij . (2.1) i
i<j
• GLn (C) := group of n × n complex matrices, with the Lebesgue measure: dZ := dReZij dImZij .
(2.2)
i,j
• Tn := group of n × n strictly upper triangular complex matrices, with the Lebesgue measure: dT := dReTij dImTij . (2.3) i<j
• Dn (R) := group of n × n real diagonal matrices, with the Lebesgue measure: dX := dXii . (2.4) i
• Dn (C) := group of n × n complex diagonal matrices, with the Lebesgue measure: dX := dReXii dImXii . (2.5) i
• (n) := group of permutations of n elements.
118
B. Eynard, A.P. Ferrer
2.2. Vandermonde determinant. For any diagonal matrix X = diag(X1 , . . . , Xn ) ∈ Dn (C), one writes: (X) := (Xi − Xj ) (2.6) i<j
and, for any permutation σ ∈ (n), we define the diagonal matrix: Xσ := diag(Xσ (1) , . . . , Xσ (n) ).
(2.7)
(Xσ ) := (−1)σ (X).
(2.8)
Notice that
2.3. Invariant functions. Definition 2.1. F (A, B) defined on GLn (C) × GLn (C) → C is an analytical invariant function if: • F is analytical in each variable, • ∀U ∈ GL∗n (C), F (U AU −1 , U BU −1 ) = F (A, B). Examples. F (A, B) =
p
Tr
Rt
(xt,rt − A)(yt,rt − B) ,
(2.9)
rt =1
t=1
F (A, B) = e− Tr V1 (A) e− Tr V2 (B) . Definition 2.2. Monomial invariant functions are functions of the form: p Rt k l F (A, B) = Tr (A t,rt B t,rt ) , t=1
(2.10)
(2.11)
rt =1
where the kt,rt ’s and lt,rt ’s are integers such that kt,rt + lt,rt > 0. The total degree is deg F :=
p Rt
kt,rt + lt,rt .
(2.12)
t=1 rt =1
Definition 2.3. Polynomial invariant functions are finite complex linear combinations of monomial invariant functions. Examples of polynomial invariant functions. F (A, B) = Tr Ak1 B l1 Ak2 B l2 , 1 + Tr Ak2 B l2 , F (A, B) = 1 + Tr Ak1 B l1 F (A, B) =
p t=1
det(xt − A)kt
q
det(yu − B)lu ,
(2.13)
(2.14)
u=1
F (A, B) = det(A ⊗ 1 − 1 ⊗ B).
(2.15)
2-Matrix vs. Complex Matrix Model
119
2.4. Decompositions. 2.4.1. Diagonalization. It is a standard result in algebra (see [22, 16, 20] for instance), that any hermitian matrix M ∈ Hn can be written: M = U XU † ,
(2.16)
dM = J˜n 2 (X) dU dX,
(2.17)
where U ∈ U (n) and X ∈ Dn (R). The measure is then:
where the Jacobian is n(n−1)
π 2 J˜n = n−1 k=0
k!
.
(2.18)
This decomposition is not unique. It is unique up to a permutation of eigenvalues, and up to multiplication of U by a diagonal matrix whose elements are on the unit circle. In other words, M = U XU † provides a mapping between Hn and U (n) × Dn (R)/(U (1)n × (n)). 2.4.2. Jordanization. A less standard result (see [22, 26, 20, 16] for instance), is that any complex matrix Z ∈ GLn (C) can be written: Z = U (X + T )U † ,
(2.19)
where U ∈ U (n), T ∈ Tn and X ∈ Dn (C). The measure is then: dZ = Jn |(X)|2 dU dT dX,
(2.20)
where the Jacobian is π n(n−1) 2
2 Jn = n−1 k=0
k!
.
(2.21)
This decomposition is not unique. It is unique up to a permutation of eigenvalues, and up to multiplication of U by a diagonal matrix whose elements are on the unit circle. In other words, Z = U (X + T )U † provides a mapping between Gln (C) and U (n) × Tn × Dn (C)/(U (1)n × (n)). 3. Gaussian Matrix Integrals In all that follows, we consider 3 complex numbers α1 , α2 and γ , and we define δ := α1 α2 − γ 2 and assume that δ = 0.
(3.1)
120
B. Eynard, A.P. Ferrer
3.1. Gaussian Hermitian Model. Consider the measure on Hn × Hn : e− Tr
α1
M12 +
2
α2 2 2 M2 +γ M1 M2
(3.2)
dM1 dM2 .
Definition 3.1. The partition function is: ZH (n, γ , α1 , α2 ) :=
Hn ×Hn
α1 2 α2 2 2 M1 + 2 M2 +γ M1 M2 )
dM1 dM2 e− Tr (
.
(3.3)
Notice that the integral ZH is absolutely convergent only if Re(α1 eφ + α2 e−φ ± 2γ ) > 0
∀φ ∈ R
(3.4)
which implies that Reα1 > 0, Reα2 > 0, (Reγ )2 < Reα1 Reα2 . An easy gaussian integral computation gives: ZH = 2
n
π √ δ
n2 (3.5)
.
Definition 3.2. The expectation value of an invariant function F (A, B) is:
F H :=
Hn ×Hn
dM1 dM2 F (M1 , M2 ) e− Tr (
Hn ×Hn
dM1 dM2 e− Tr (
α1 2 α2 2 2 M1 + 2 M2 +γ M1 M2 )
α1 2 α2 2 2 M1 + 2 M2 +γ M1 M2 )
.
(3.6)
Remark 3.1. It is clear, from Wick’s theorem, that if F is a monomial invariant function, then < F >H is a polynomial in αδ1 , αδ2 and γδ , and can be analytically continued to every complex α1 , α2 , γ , provided that δ = 0. 3.1.1. Gaussian Hermitian loop equations. Consider a monomial matrix valued function, of the form: f (A, B) = f0 (A, B)
p
Tr ft (A, B) ,
∀t = 0, . . . , p,
t=1
ft (A, B) =
Rt
Akt,rt B lt,rt .
(3.7)
rt =1
Define G0 (A, B) :=
Tr fu (A, B) ,
and if t ≥ 1 ,
u=0
Gt (A, B) :=
u=0,t
Tr fu (A, B).
(3.8)
2-Matrix vs. Complex Matrix Model
121
Theorem 3.1. One has the “loop equations”: α1 G0 (M1 , M2 ) Tr M1 f0 (M1 , M2 )H + γ G0 (M1 , M2 ) Tr M2 f0 (M1 , M2 ))H r−1 −1 R0 k0,r
k0,u j l0,u G0 (M1 , M2 ) Tr M1 = M1 M2 r=1 j =0
p Rt k t,r −1
R0
k −j −1 M1 0,r M2 l0,r
× Tr
+
u=1
k M1 0,u M2 l0,u
u=r+1
r−1
Gt (M1 , M2 ) Tr
t=1 r=1 j =0
H
k
j
M1 t,u M2 lt,u M1 f0 (M1 , M2 )
u=1
×
k −j −1 M1 t,r M2 lt,r
Rt
k M1 t,u M2 lt,u
u=r+1
(3.9) H
and α2 G0 (M1 , M2 ) Tr M2 f0 (M1 , M2 )H + γ G0 (M1 , M2 ) Tr M1 f0 (M1 , M2 ))H r−1 −1 R0 l0,r
k0,u k0,r l0,u j G0 (M1 , M2 ) Tr M 1 M2 = M1 M2 r=1 j =0
× Tr
+
u=1
−1 p Rt lt,r
M2
l0,r −j −1
R0
k M1 0,u M2 l0,u
u=r+1
Gt (M1 , M2 ) Tr
r−1
t=1 r=1 j =0
k
M1 t,u M2 lt,u M1 kt,r M2 j
u=1
× f0 (M1 , M2 )M2
H
lt,r −j −1
Rt u=r+1
k M1 t,u M2 lt,u
.
(3.10)
H
Notice that the RHS is a linear combination of invariant polynomial functions of degree strictly lower than the LHS. Loop equations are a standard method for finding recursion relations among expectation values [9], they were first studied by [25] for the 2-matrix model, and solved more explicitly by [14, 11, 13]. Proof. Write that the integral of a total derivative is zero:
0=
i
dM1 dM2
α1 ∂ 2 α2 2 fi,i (M1 , M2 ) e− Tr ( 2 M1 + 2 M2 +γ M1 M2 ) , (3.11) ∂M1ii
122
B. Eynard, A.P. Ferrer
i.e.
i
=
dM1 dM2
∂ fi,i (M1 , M2 ) ∂M1ii
e− Tr (
α1 2 α2 2 2 M1 + 2 M2 +γ M1 M2 )
dM1 dM2 fi,i (M1 , M2 ) (α1 M1ii + γ M2ii )
i
×e− Tr (
α1 2 α2 2 2 M1 + 2 M2 +γ M1 M2 )
(3.12)
.
Similarly:
i<j
=
dM1 dM2
∂ fi,j (M1 , M2 ) ∂ReM1ij
e− Tr (
α1 2 α2 2 2 M1 + 2 M2 +γ M1 M2 )
dM1 dM2 fi,j (M1 , M2 ) α1 (M1j i + M1ij ) + γ (M2j i + M2ij )
i<j
×e− Tr (
α1 2 α2 2 2 M1 + 2 M2 +γ M1 M2 )
(3.13)
and
dM1 dM2
i<j
=i
∂ fi,j (M1 , M2 ) ∂ImM1ij
e− Tr (
α1 2 α2 2 2 M1 + 2 M2 +γ M1 M2 )
dM1 dM2 fi,j (M1 , M2 ) α1 (M1j i − M1ij ) + γ (M2j i − M2ij )
i<j
×e− Tr (
α1 2 α2 2 2 M1 + 2 M2 +γ M1 M2 )
(3.14)
.
Taking 3.12 plus 3.13 minus i 3.14 , we get:
α1 ∂ 2 α2 2 fi,i (M1 , M2 ) e− Tr ( 2 M1 + 2 M2 +γ M1 M2 ) ∂M1ii ∂ 1 ∂ fi,j (M1 , M2 ) dM1 dM2 + −i 2 ∂ReM1ij ∂ImM1ij
dM1 dM2
i
i<j
α1
α2
×e− Tr ( 2 M1 + 2 M2 +γ M1 M2 )
1 ∂ ∂ dM1 dM2 + fj,i (M1 , M2 ) +i 2 ∂ReM1ij ∂ImM1ij 2
2
i<j
=
×e− Tr (
α1 2 α2 2 2 M1 + 2 M2 +γ M1 M2 )
dM1 dM2 ( Tr f (M1 , M2 )(α1 M1 + γ M2 )) e− Tr (
α1 2 α2 2 2 M1 + 2 M2 +γ M1 M2 )
,
(3.15)
2-Matrix vs. Complex Matrix Model
123
i.e. one can proceed as if all the M1ij were n2 real independent variables, i.e., by abuse of notation we write:
α1 ∂ 2 α2 2 dM1 dM2 fi,j (M1 , M2 ) e− Tr ( 2 M1 + 2 M2 +γ M1 M2 ) ∂M1ij i,j = dM1 dM2 ( Tr f (M1 , M2 )(α1 M1 + γ M2 )) ×e− Tr (
α1 2 α2 2 2 M1 + 2 M2 +γ M1 M2 )
(3.16)
.
Now, one can use the following rules: • Split rule: if f (M1 , M2 ) = AM1k B (where A and B are matrices), one has:
∂f (M1 , M2 )ij ∂M1 ij
i,j
=
Tr AM1k−1−l Tr M1l B .
k−1
(3.17)
l=0
• Merge rule: if f (M1 , M2 ) = A Tr (M1k B) (where A and B are matrices), one has:
∂f (M1 , M2 )ij i,j
∂M1 ij
=
k−1
Tr AM1k−1−l BM1l .
(3.18)
l=0
Then, if A and B depend on M1 , one has to use the chain rule. When one considers f given by 3.7, one gets Eq. 3.9. Equation 3.10 is obtained by doing the same for M2 . We find again that F H is a polynomial in
α1 α2 δ , δ
and γδ .
3.2. Gaussian complex model. Consider the measure on GLn (C): e
− Tr
α1 2 α2 † 2 † 2 Z + 2 Z +γ ZZ
dZ.
Definition 3.3. The partition function is: α1 2 α2 † 2 † dZ e− Tr ( 2 Z + 2 Z +γ ZZ ) . ZC (n, γ , α1 , α2 ) := Gln (C)
(3.19)
(3.20)
Notice that the integral ZC is absolutely convergent only if ∀θ ∈ R
Re(α1 eiθ + α2 e−iθ + 2γ ) > 0.
(3.21)
One can see that with θ = π, this condition can never be compatible with 3.4 (with φ = 0). Therefore, if ZH is an absolutely convergent integral then ZC is not, and vice–versa. An easy gaussian integration gives (where δ = α1 α2 − γ 2 ): ZC =
π √ −δ
n2
which can be analytically continued to every α1 , α2 , γ , provided that δ = 0.
(3.22)
124
B. Eynard, A.P. Ferrer
Definition 3.4. The expectation value of an invariant function F (A, B) is:
F C :=
Gln (C) dZ F (Z, Z
Gln (C) dZ e
† ) e− Tr (
− Tr (
α1 2 α2 † 2 † 2 Z + 2 Z +γ ZZ )
(3.23)
.
α1 2 α2 † 2 † 2 Z + 2 Z +γ ZZ )
Remark 3.2. It is clear, from Wick’s theorem, that if F is a monomial invariant function, then < F >C is a polynomial in αδ1 , αδ2 and γδ , and can be analytically continued to every complex α1 , α2 , γ , provided that δ = 0. 3.2.1. Gaussian complex loop equations. Consider a monomial matrix valued function, of the form: p
f (Z, Z † ) = f0 (Z, Z † )
Tr ft (Z, Z † ) ,
t=1 Rt
ft (Z, Z † ) =
Z kt,rt Z †
lt,rt
(3.24)
,
rt =1
define: G0 (Z, Z † ) :=
and if t ≥ 1 ,
Tr fu (Z, Z † ) ,
u=0
Gt (Z, Z † ) :=
Tr fu (Z, Z † ).
(3.25)
u=0,t
Theorem 3.2. One has the same loop equations as Theorem 3.1, replacing the subscript H by C, α1 G0 (Z, Z † ) Tr Zf0 (Z, Z † ) + γ G0 (Z, Z † ) Tr Z † f0 (Z, Z † )) C C r−1 k0,r −1 R 0
l0,u = G0 (Z, Z † ) Tr Zj Z k0,u Z † r=1 j =0
× Tr
+
u=1
p Rt k t,r −1
Z
k0,r −j −1
Z
R0
† l0,r
Z
k0,u
Z
u=r+1
Gt (Z, Z † ) Tr
r−1
t=1 r=1 j =0
† l0,u C
Z kt,u Z †
lt,u
Z j f0 (Z, Z † )
u=1
× Z
kt,r −j −1
Z
† lt,r
Rt u=r+1
Z
kt,u
Z
† lt,u
(3.26) C
2-Matrix vs. Complex Matrix Model
and
125
α2 G0 (Z, Z † ) Tr Z † f0 (Z, Z † ) + γ G0 (Z, Z † ) Tr Zf0 (Z, Z † )) C C r−1 −1 R0 l0,r
l0,u j G0 (Z, Z † ) Tr Z k0,r Z † = Z k0,u Z † r=1 j =0
× Tr
+
u=1
−1 p Rt lt,r
Z
† l0,r −j −1
R0
Z
k0,u
Z
† l0,u
u=r+1
Gt (Z, Z † ) Tr
r−1
t=1 r=1 j =0
Z kt,u Z †
u=1
× f0 (Z, Z )Z †
C
lt,u
Rt
† lt,r −j −1
Z kt,r Z †
j
Z
kt,u
Z
† lt,u
.
u=r+1
(3.27)
C
Notice that the RHS is a linear combination of invariant polynomial functions of degree strictly lower than the LHS. Proof. The proof is very similar to that of Theorem 3.1. Write that the integral of a total derivative is zero: 2 α α ∂ † − Tr ( 21 Z 2 + 22 Z † +γ ZZ † ) fr,s (Z, Z ) e (3.28) 0 = dZ ∂ReZij and
0 = −i
dZ
∂ ∂ImZij
fr,s (Z, Z † ) e− Tr (
α1 2 α2 † 2 † 2 Z + 2 Z +γ ZZ )
.
(3.29)
Taking the sum of both lines, one can proceed as if all the Zij and Z † ij were real independent variables, and from there, follow the proof of Theorem 3.1. Remark 3.3. We see that the loop equations of both models are identical. It is clear from the above derivation that this is general, even for non gaussian measures. When the measure is gaussian, the loop equations determine completely every expectation value, while for non-gaussian measures, the loop equations give recursion relations for expectation values, but don’t give the initial conditions. Let us consider in particular the “semi-classical case” [3, 6], i.e. with a measure of the type ∂µ(M1 , M2 ) = e− Tr [V1 (M1 )+V2 (M2 )+M1 M2 ] ,
(3.30)
where V1 and V2 are rational functions. In that case, the initial conditions which allow to determine all polynomial expectation values recursively, are in one–to–one correspondence with homology classes of integration paths for pairs of eigenvalues [6], therefore, there exists a choice of integration path such that one can write: dM1 dM2 e− Tr [V1 (M1 )+V2 (M2 )+M1 M2 ] (Hn ×Hn )( ) † † ≡ dZ e− Tr [V1 (Z)+V2 (Z )+ZZ ] (3.31) GLn (C)
126
B. Eynard, A.P. Ferrer
and one can consider that this equality defines the RHS. Somehow, the complex matrix model is nothing but the analytical continuation of the 2-matrix model defined on some classes of contours.
3.3. Relation between the two models. Theorem 3.3. For any polynomial invariant function F (A, B), one has:
F H = F C .
(3.32)
Notice that F H and F C have been defined for different range of values of α1 , α2 , and γ , but, as we have explained above, both are polynomials of αδ1 , αδ2 and γδ (and can be analytically continued to any α1 , α2 , and γ ). Theorem 3.3 is thus an equality between polynomials. Proof. It is sufficient to prove it for monomial invariant functions. The proof is clearly obtained from the loop equations, by recursion on deg F . It is obviously true for deg F = 0, i.e. F = 1. And the loop equations of both models are identical. Definition 3.5. For any two given complex diagonal matrices X and Y , and any polynomial invariant function F , define: W˜ F (X, Y ) := 2 (X) 2 (Y )
dU F (X, U Y U † ) e−γ Tr XU Y U ,
(3.33)
F (X + T , Y + T † ) e−γ Tr T T , −γ Tr T T † T (n) dT e
(3.34)
†
U (n)
ωF (X, Y ) := (X) (Y )
T (n) dT
†
which is a polynomial in all its variables Xi , Yj , and a polynomial in 1/γ , and: WF (X, Y ) :=
1
(Xσ )(Yτ ) e−γ Tr Xσ Yτ n!2 σ τ
×
dT F (Xσ + T , Yτ + T † ) e−γ Tr T T . †
(3.35)
T (n)
Theorem 3.4. For any polynomial invariant function F (A, B), one has: α1 α2 J˜n2 2 2 dXdY e− 2 Tr X e− 2 Tr Y W˜ F (X, Y ) 2 n! ZH (n, γ , α1 , α2 ) Dn (R)×Dn (R) 2 α1 α2 Jn 2 = dX e− 2 Tr X e− 2 Tr X WF (X, X). (3.36) n! ZC (n, γ , α1 , α2 ) Dn (C)
2-Matrix vs. Complex Matrix Model
127
Proof. Start from Theorem 3.3, diagonalize M1 and M2 on the hermitian side, and jordanize Z on the complex side, α1 α2 J˜n2 2 2
F H = 2 dXdY e− 2 Tr X e− 2 Tr Y W˜ F (X, Y ) n! ZH (n, γ , α1 , α2 ) Dn (R)×Dn (R) 2 α1 α2 Jn 2 = F C = dX e− 2 Tr X e− 2 Tr X n! ZC (n, γ , α1 , α2 ) Dn (C) ×e−γ Tr XX ωF (X, X) 2 α1 α2 Jn 2 = dX e− 2 Tr X e− 2 Tr X e−γ Tr Xσ Xτ ωF (Xσ , X τ ) n! ZC (n, γ , α1 , α2 ) Dn (C) 2 α1 α2 Jn 2 = dX e− 2 Tr X e− 2 Tr X WF (X, X). (3.37) n! ZC (n, γ , α1 , α2 ) Dn (C) The equality in the first line is obtained by diagonalizing M1 and M2 (with Jacobian given in Eq. 2.17), the equality in the second line is obtained by Jordanizing Z (with Jacobian given in Eq. 2.19), the equality between the second and third line holds for any pair of permutations σ and τ (it can be proven with Lemma A.1 given in the Appendix), and the equality of the last line comes from the definition of WF . 4. Unitary Group Integrals Here is one of the most important theorems of this paper: 4.1. Unitary integrals and triangular integrals. Theorem 4.1. For any invariant function F (A, B) one has: † dU F (X, U Y U † ) e−γ Tr XU Y U U (n)
cn = n!
σ
τ (−1)
σ (−1)τ e−γ Tr Xσ Yτ
T (n) F (Xσ
(4.1)
+ T , Yτ + T † ) e−γ Tr T T dT †
(X)(Y )
where
n−1 cn =
k=0
(−2π)
k!
n(n−1) 2
,
(4.2)
,
i.e. W˜ F (X, Y ) = n! cn WF (X, Y ).
(4.3)
Proof. Using Lemma A.1 given in the Appendix, and using Theorem 3.4, we have: α1 α2 2 2 dX dY e− 2 Tr X e− 2 Tr Y W˜ F (X, Y ) Dn (R)×Dn (R) n!2 ZH Jn = J˜n2 n! ZC
Dn (C)
dX e−
α1 2
Tr X2 −
e
α2 2
Tr X
2
WF (X, X)
128
B. Eynard, A.P. Ferrer
2 α α 1 n!2 ZH Jn − 21 Tr X2 − 22 Tr X −γ Tr Xσ X τ = dX e e e ωF (Xσ , X τ ) J˜n2 n! ZC n!2 σ,τ Dn (C) n √π 2 J n α1 α2 1 n! ZH 2 2 −δ dX dY e− 2 Tr X e− 2 Tr Y = n 2 2π n! Z n! C J˜2 √ σ,τ Dn (R)×Dn (R) n
δ
×e−γ Tr Xσ Yτ ωF (Xσ , Yτ ) n √π 2 J n α1 α2 n! ZH 2 2 −δ = dX dY e− 2 Tr X e− 2 Tr Y WF (X, Y ) n 2π n! Z 2 C Dn (R)×Dn (R) J˜n √ δ α1 α2 2 2 = n!cn dX dY e− 2 Tr X e− 2 Tr Y WF (X, Y ). (4.4) Dn (R)×Dn (R)
Notice that if f (A) and g(B) are polynomial invariant functions, i.e. f (U AU −1 ) = f (A) for all A and U (resp. g(U BU −1 ) = g(B) for all B and U ), one has: Wf (X)g(Y )F (X,Y ) (X, Y ) = f (X)g(Y )WF (X, Y ) , W˜ f (X)g(Y )F (X,Y ) (X, Y ) = f (X)g(Y )W˜ F (X, Y ).
(4.5)
Thus, for any symmetric polynomials f (x1 , . . . , xn ) and g(y1 , . . . , yn ), one has: α1 α2 2 2 0= dX dY e− 2 Tr X e− 2 Tr Y f (X)g(Y ) Dn (R)×Dn (R)
×(n! cn WF (X, Y ) − W˜ F (X, Y )). α − 21
α − 22
(4.6)
Tr Tr Notice that e e (n! cn WF (X, Y ) − W˜ F (X, Y )) is a continuous function fastly decreasing at ∞ in all variables, and it is symmetric in the x’s and in the y’s. Using the Stone–Weierstrass theorem (polynomials are dense in the space of continuous functions), one gets that n! cn WF (X, Y ) = W˜ F (X, Y )). X2
Y2
4.2. Examples. Let us illustrate Theorem 4.1 on some simple examples and recover some classical results. 4.2.1. Harish-Chandra–Itzykson–Zuber’s formula. We can use Theorem 4.1, to find a new proof of the famous Harish-Chandra–Itzykson–Zuber formula [18, 17, 20]. Indeed, consider F (A, B) = 1, Theorem 4.1 gives: −γ Xσ Yτ σ τ i i cn † σ,τ (−1) (−1) ie −γ Tr XU Y U † e = dT e−γ Tr T T n! (X)(Y ) U (n) Tn n(n−1) 2 det E π = cn , (4.7) γ (X)(Y ) which is the famous Harish Chandra-Itzykzon-Zuber integral. Here E is the matrix Eij := e−γ Xi Yj .
(4.8)
2-Matrix vs. Complex Matrix Model
129
4.2.2. Morozov’s formula. Consider TrAk B l for any integers k and l. It is in fact simpler to introduce a generating function: