String Topology

Moira Chas and Dennis Sullivan October 24, 2003

Abstract Consider two families of closed oriented curves in a manifold M d . At each point of intersection of a curve of one family with a curve of the other family, form a new closed curve by going around the rst curve and then going around the second. Typically, an i-dimensional family and a j-dimensional family will produce an i + j d + 2-dimensional family. Our purpose is to describe a mathematical structure behind such interactions.

Introduction

By the string homology of a manifold M d we mean the equivariant homology of the continuous mapping space Map(S1 , M d ) with the circle symmetry of rotating the domain. The goal of this paper is to expose the following theorem and its underpinnings: Theorem 6.1 On the string homology of a smooth or combinatorial oriented manifold M d there is a natural graded Lie algebra structure of degree (2 d). This structure called the string bracket comes from the interaction of closed oriented curves in M d . At each point of intersection of two such curves, one can form the oriented curve obtained by going around one and then around the other. Transversality will be used on the cells computing string homology to dene the string bracket of the theorem. For an i cell and a j cell of strings there will typically be a (i+1)+(j +1)d = (i+j)+(2d) dimensional set of intersection points. If we picture a bracket by two circles touching, then a triple bracket means a third circle touches the result along one arc or the other. Thus, the Jacobi identity can be viewed as in Figure 1, where the two terms on the left appear on the right with two other terms that are geometrically identical up to sign. This Lie algebra is quite non-trivial 1

Figure 1: Explanation of Jacobi identity for surfaces of genus larger than one (see reference to Goldman and Wolpert below). In the process of analyzing the above argument we found a more basic structure called the loop product on the loop homology, the ordinary homology of the free loop space Map(S1 , M d ). A version of the loop product of two cells in Map(S1 , M ) is dened for each point z on S1 by rst intersecting in M d the two cells obtained by evaluating at this point z and then composing the loops at these intersection points. There are homotopies making the loop product on loop homology into an associative graded commutative algebra. As in Gerstenhabers basic paper [1], there is a preferred homotopy for the graded commutativity (denoted ) which leads by symmetrization to a second operation {, } on loop homology. This operation, called loop bracket satises the Jacobi identity where now it is convenient to grade H by subtracting d from the usual geometric grading

so that and {, } become operations of degree zero and one respectively, Hi Hj Hi+j , Hi Hj Hi+j+1 , instead of d and (d + 1) respectively. We arrive at the following result. Theorem 4.7 The loop product with the loop bracket {, } makes the loop homology H (the ordinary homology of the free loop space) into a Gerstenhaber algebra, namely: (1) The loop product denes a graded commutative, associative algebra. (2) {, } is a Lie bracket of degree 1, which means that for each a, b, c H (i) {a, b} = (1)(|a|+1)(|b|+1) {b, a} (ii) {a, {b, c}} = {{a, b}, c} + (1)(|a|+1)(|b|+1) {b, {a, c}} (3) {a, b c} = {a, b} c + (1)|b|(|a|1) b {a, c}. Now consider the circular symmetry of the loop space Map(S1 , M ) = L(M ), and the associated degree +1 operator on homology i cycle in L(M ) i + 1 cycle in L(M ) = i cycle S1 L(M ). It turns out that is a second order operator in the sense of commutative algebra, i.e., the binary operation, the deviation of from being a derivation of , is a derivation in each variable. One also has easily that = 0. We come to Theorem 5.4 The loop product and the operator make the loop homology (the ordinary homology of Map(S1 , M )) into a Batalin Vilkovisky algebra, namely: (1) is a graded commutative associative algebra. (2) = 0. (3) (1)|a| (a b) (1)|a| a b a b is a derivation of each variable. The connection between these two results is that conditions (1), (2), and (3) of Theorem 5.4 imply formally that the binary operation dened by the deviation satises graded Jacobi so that a Batalin Vilkovisky algebra is a special type of Gerstenhaber algebra. We prove Theorem 5.4 by constructing a chain homotopy between the {x, y} of Theorem 4.7 and (1)|x| (x y) 3
action {,}

(1)|x| (x) y x (y). We note that in the discussion of these two theorems there are two independent proofs of Jacobi for the loop bracket. Finally we come again to the string bracket on string homology, the equivariant homology of the mapping space of S1 into M . The approach referred to above is to use intersection theory on chains to dene the string bracket and verify the Jacobi identity directly on the level of transversal triples of chains. The second approach to the string bracket uses the loop product discussed above. Consider the degree +1 operation lift from equivariant chains to ordinary chains corresponding to replacing an i-chain in the base of an S1 bration by the i+1 chain which is the preimage in the total space. Consider also the operation project which simply projects chains in the total space to the base. Then we dene the string bracket in terms of the loop product by the formula [x, y] = project(lift x lift y). (In our calculations, Section 6 we denote lift by M, because lift means mark a point in all possible ways on a closed curve without a mark (a string). We denote project by E because project means erase the marked point in a marked curve (loop) to get an unmarked curve (string)). The composition (lift) (project) induces the operator above on loop homology, and the Jacobi identity for [, ] follows by direct calculation from the properties of the Gerstenhaber qua Batalin Vilkovisky algebra structures above. We arrive at Theorem 6.1 stating that the string bracket denes a graded Lie algebra structure on string homology. The geometric degree is (2 d) for the usual grading of string or equivariant homology. The same argument constructing the binary operation [, ] = m2 , the string bracket, constructs ternary, etc. operators m3 , m4 , . . .. Now extend each mk to a coderivation mk to H (see Section 6). One knows the Jacobi identity for the string bracket m2 is equivalent to the relation m2 m2 = 0 for the associated coderivation m2 . The Jacobi relation for m2 generalizes to the entire collection {m2 , m3 , . . .} in the following way. Theorem 6.2 The associated coderivations {m2 , m3 , . . .} of the free commutative coalgebra H on the string homology H satisfy: (i) mk mk = 0, for k = 2, 3, 4 . . .. (ii) mk mr + mr mk = 0 for k, r = 2, 3, 4, . . ..

These coderivations mk combine in various ways to dene coderivations of square zero on H . Such a dierential is one denition of a Lie or strong homotopy Lie algebra structure on H . Thus we have Corollary 6.3 There exists an uncountable family { } of Lie structures on the string homology. Namely, for each {2, 3, . . .}, : H H dened as =

is a coderivation which satises = 0. If we examine the string bracket when d = 2 we nd a Lie bracket structure on the vector space of components of the space of closed curves in a surface. For surfaces of genus larger than zero this is the non-trivial bracket discovered in the 80s by Wolpert [12] and Goldman [3]. That discovery was strongly related to the symplectic structure on Techmuller space [12] and the symplectic structure of other spaces of at connections over a surface [3]. (For more discussion see Section 7). When the genus is zero, i.e., M is a two sphere, the loop product becomes non-trivial in the higher dimensional algebraic topology of the free loop space of S2 . The S2 -calculation is part of a general structure based on the diagram relating the intersection product on ordinary homology (H (M ), ) with the loop product on loop homology (H , ) and the Pontryagin product on the based loop space homology (H (), ), (1) (H (M ), )
constant loops

(H , )

intersection with a ber

(H (), )

Both maps are ring homomorphisms. The rst is an injection onto a direct summand showing the loop product is an extension of the classical intersection product. The image of the second map is a graded commutative subalgebra. For S2 , (H (), ) is the tensor algebra on one generator in degree one and the image is the subalgebra generated by 2 . This shows the loop product is non-trivial for S2 (Section 9). For more complete calculations we can augment the diagram ( 1) with a relation between the usual cap product operation, , and the loop product. This is described by Theorem 8.2 For each x, y homology classes and compatible pair of classes (A, a) loop product (A (x y)) = a (x y).

(see Section 8 for the denitions and the proof, and Section 9 for relevant algorithms.) The two approaches here to the string bracket, direct geometry and via the loop product, reminds one of Wittens paper [10]. There it was pointed out that closed string interactions looked at directly as in the Figure 1 are non-associative. To get around this, a marking point was introduced in [10] to facilitate the denition of an associative multiplication of (open) strings. Thus also our string bracket is non-associative but satises Jacobi and it arises from an associative product of loops (marked strings). There is also a dictionary relating our constructions with those in algebra begun by Gerstenhaber [1]. In a sequel we will discuss a rich world of general string operations in the chains of the loop space. We nd a structure like a big part of a two dimensional eld theory associated to each manifold M d . In particular we investigate the Lie structure described in Theorem 6.2 and Corollary 6.3 as well as co-versions. We also hope to follow Stashes specic suggestion to relate our structure to the work of the physicist Zwiebach [11] and [13]. Table of contents (1) Introduction (2) Denition of the loop product at the chain level and passage of to loop homology. (3) is homotopy associative and homotopy commutative where is the homotopy of commutativity. (4) Symmetrization of gives a loop bracket {, } which passes to loop homology and we have a Gerstenhaber algebra (i.e., the bracket is a biderivation of and satises Jacobi). (5) Denition of and the argument that the deviation of from being a derivation of is the loop bracket {, }. (6) Denition of M and E, string homology and the construction of a string bracket [, ] which satises Jacobi. Generalized n-brackets and generalized Jacobi. (7) The string bracket for surfaces, the work of Wolpert and Goldman, and the case of S2 . (8) Cap product and loop product. (9) Appendix 1: Sc2 and other simply connected manifolds. (10) Appendix 2: M 3 and K(, 1) manifolds.

The loop product

We think of the circle S1 as R/Z, so each point can specied by some x [0, 1). Unless otherwise stated, M is an orientable manifold of dimension d. A loop in M is a continuous map from S1 to M . Observe that a loop has a marked point: the image of 0. Map(S1 , M ), the space of all loops, will be denoted by L(M ), or by L since M is xed throughout the discussion. By an i-chain we mean a linear combination of oriented i-dimensional families of loops in M . The parameter spaces of the families are taken from any standard list of cells closed under face operators. In the algebraic topology of chains on a space there are two well known multiplications which we will combine to obtain a new structure. The rst of these multiplications is the (transversal) intersection of chains in a dmanifold: an i-chain intersected with a j-chain gives an i + j d-chain. The second is the product of an i-chain of loops with a j-chain of loops, all of whose marked points are equal to some p M . Multiplying these chains yields an i + j cartesian product chain of composed loops with marked point at p in M . Our new loop product is transversally dened at the chain level as follows (see Remark 2.2): given x, an i-chain of loops in M , and y, a j-chain of loops in M , one rst intersects the i-chain (of M ) of marked points of x with the j-chain (of M ) of marked points of y, to obtain an i + j d chain c (of M ) along which the marked point of x coincides with the marked point of y. Now we dene the chain x y, by putting at each point of c the composed loop that rst goes around the loop of x and then, around the loop of y (see Figure 2). In order to give the precise denition of the loop product we need the following remark. Remark 2.1 We will use the following orientation convention for our constructions. We will have a map of a domain cell, usually a product K1 K2 . . . Kn into M M . . . M (l factors) which is assumed to be transversal to a diagonal. We orient the product cell with the product of individual orientations. We orient the normal of the pull back by pulling back the orientation of the normal of the diagonal induced by some orientation of M . We then take the induced orientation on the pullback so that (orientation on the pullback)(orientation on the normal) = (product orientation of the cell domain)

Denote the (i d) chains of L by Li and the direct sum of all these by L . Also, if x: Kx L is a cell, we denote by Kx its underlying set. Now, for any pair of cells, x: Kx L, y: Ky L we dene the set Kxy as the transversal preimage of the diagonal of M M under the map Kx Ky M M , (kx , ky ) (x(kx )(0), y(ky )(0)) Now, dene x y: Kxy L as (x y)(kx, ky )() =
1 x(kx )(2) if [0, 2 ], 1 y(ky )(2) if [ 2 , 1).

(we keep the notation Kxy because Kxy is a manifold which can be divided into cells) . Orient Kxy by Remark 2.1. (Observe that with our new grading, if x Li and y Lj then x y Li+j . Remark 2.2 We will adopt a point of view which is used in classical intersection theory of chains in a manifold. We say that a chain operation is transversally dened if it is dened for appropriately transversal cells. We say that an identity between chain operations holds transversally if it holds on any nite subset of the chains where all constituents are appropriately transversal. Thus, the classical intersection is dened transversally at the chain level, is transversally associative and transversally graded commutative. Lemma 2.3 If x, y L is a transversal pair then (x y) = x y + (1)|x| x y. Proof. By denition, the underlying chain of x y is the oriented intersection chain in M of the marked points of x and of y respectively. Once the orientation of M is xed, the orientations of intersections behave for calculations like orientations of normal directions. Now, for each x L , we denote by x the chain of M of marked points of x. Thus, for transversal intersection one gets the familiar (x y) = (x y) + (1)|x| (x y) where |x| = dim(x) d. Since Kxy is the underlying set of x y we get Kxy = Kxy + (1)|x| Kxy . 8

The chain (xy) is the restriction of xy to Kxy . Therefore, the above formula yields the same formula for , (x y) = x y + (1)|x| x y.

Let Hi denote the i-th homology group of the loop space L with the degree shifted down by d and set H = Hd Hd+1 . . . H0 H1 . . .. We refer to H as the loop homology. Corollary 2.4 The loop product on chains passes to loop homology and denes a product Hi Hj Hi+j . Remark 2.5 The loop product is dened in all dimensions and may be nonzero way above the dimension of the manifold (e.g., M = S2 , Example 7.1, or S3 , Corollary 3.6).

Associativity of , the operator, and homotopy commutativity of

Let us rst discuss associativity. For the classical intersection product, we have that if three cycles are pairwise transversal, then the intersection product is literally associative at the chain level. The classical based loop composition is associative up to homotopy (see Stashe [7] and [8] for a complete discussion of this point). Thus both of these classical chain products yield associative multiplications in homology, and thus we will have the same associativity for , the loop product, combining intersection and based loop composition. Proposition 3.1 The loop product in loop homology Hi Hj Hi+j is associative. Proof. Assume that the three homology classes in question x, y, z are represented by cycles that are pairwise transversal. The intersection locus of (x y) z and x (y z) are literally equal with identical coorientations. The loop product is associative up to homotopy using the same considerations as in the based loop product, now parameterized by the points of the intersection x y z. 9

Let us turn to the question of commutativity. The intersection product of two transversal cycles is literally graded commutative. However, the based loop product (Pontryagin product) is often non-commutative even in homology. We will see that the product, combining these two products is homotopy commutative at the chain level. One chain homotopy is given by a new binary operation x y dened for appropriately transversal pairs x, y in the following way. Consider the chain c (of M ) where the marked point of x transversally intersects one of the images of the loops of y. Then at each point of c put the following loop: rst go around the loop of y until the intersection point with the marked point of x. Then go around x and nally, go around the rest of y (see Figure 3). The precise denition is the following: let x: Kx L, y: Ky M be two cells in L . Let Kxy be the preimage of the diagonal of M M under the map Kx [0, 1] Ky M M , (kx , s, ky ) (x(kx )(0), y(ky )(p(s))), where p: [0, 1] S1 is the usual projection. Then x y: Kxy L is dened as s if [0, 2 ). y(ky )(2) s s+1 x(kx )(2 s) if [ 2 , 2 ), (x y)(kx , s, ky )() = y(ky )(2) if [ s+1 , 1). 2

Lemma 3.2 If x, y L are appropriately transversal, then

(x y) = x y + (1)|x|+1 x y + (1)|x| (x y (1)|x||y|y x). Proof. Let x denote the chain of M of the marked points of x and let y denote the chain of M with parameter space [0, 1] Ky , given by y(s, ky ) = y(ky )(p(s)), where p: [0, 1] S1 is the usual projection. Suppose that x and y are appropriately transversal. Since the underlying sets of x y and x y coincide, (x y) is parametrized by the underlying set of (x y). Hence, for transversal intersection, one obtains (x y) = (x y) + (1)|x| (x y). Since ([0, 1] Ky ) = [0, 1] Ky [0, 1] Ky , the restriction of x y to the underlying set of (1)|x| (xy) is (1)|x| (xy (1)|x||y| y x)(1)|x| xy. One the other hand, the restriction of x y to the underlying set of x y is x y, which completes the proof. 10

In particular, if x and y are cycles, Lemma 3.2 implies that (x y) = (x y (1)|x||y|y x). Therefore, we have Theorem 3.3 (H , ) is an associative, (graded) commutative algebra. Let us compare the loop product on loop homology (H , ) with the usual homology of the manifold with intersection product (H (M ), ) and the homology of the based loop space with the based loop product or Pontryagin product, (H (), ). We have two maps, H (M ) H H () where is the inclusion of constant (or even small) loops into all loops and is the transversal intersection with one ber of the projection loop space evaluation M . If we use the usual grading on H(), our shifted grading on H , the homology of the entire loop space, and the analogous shifted grading on H (M ), then these products and the two maps have degree zero. Proposition 3.4 (H (M ), ) (H , ) (H (), ) preserve products. Proof. These follow directly from the denitions. Remark 3.5 is an injection onto a direct summand. For any Lie group manifold, is a surjection. Corollary 3.6 For M = S3 , the loop product is non-zero in innitely many degrees. Proof. By Remark 3.5, H H () is a surjection. With Q coecients, the homology of (S3 ) is the homology of CP and the Pontryagin product gives a polynomial algebra on this generator in degree 2.

The loop bracket {, } in loop homology

For easier thinking, let us now derive formulae by calculating in the chain complex of all homomorphisms L L L of degree 1, 0, 1, . . ., with the usual dened by the Leibniz rule with signs, ((a b)) = ()(a b) + (1)|| ((a b)), , which is dened by this relation, is also denoted [, ], the graded commutator of and . 11

If [, ] = 0, is usually called a chain map. As an example, if (xy) = (1)|x| x y of the previous section, [, ] evaluated on (x y) is (1)|x| (x y) + ((1)|x|+1 x y + x y) = (1)|x| ((x y) x y (1)|x|+1 x y). By Lemma 3.2 this is equal to x y (1)|x||y| y x. In other words, in the chain complex of homomorphisms L L L , the transversally dened 1-chain (1)|x| x y has boundary the transversally dened zero chain (x y (1)|x||y|y x). Symbolically, = , where x y = (1)|x| x y and (x y) = (1)|x||y| y x. Thus we consider + and calculate it is a 1-cycle [, + ] = [, ] + [, ] = [, ] + [, ] = ( ) + ( ) = 0 using [, ] = 0 and 2 = 0. Denition 4.1 The loop bracket {x, y} is dened transversally on L by the formula {x, y} = x y (1)(|x|+1)(|y|+1) y x.

Lemma 4.2 For x, y, z L the associator of is symmetric in the rst two variables, x (y z) (x y) z = (1)(|x|+1)(|y|+1) (y (x z) (y x) z). Proof. Observe that the parameter spaces of Kx(yz)(xy)z and Ky(xz)(yx)z consist in those values where the basepoint of the loops of x coincides with the one of the points of the loop of z and the basepoint of loops of y coincides with another. (see Figure 4) Symbolically, Kx(yz)(xy)z = {(kx , s, ky , t, kz ) Kx [0, 1] Ky [0, 1] Kz : x(xx )(0) = z(kz )(s), y(ky )(0) = z(kz )(t)} 12

Ky(xz)(yx)z = {(ky , t, kx , s, kz ) Kx [0, 1] Ky [0, 1] Kz : x(xx )(0) = z(kz )(s), y(ky )(0) = z(kz )(t) =}. Clearly, there is a bijection between the underlying sets of the chains in question, Kx(yz)(xy)z and Ky(xz)(yx)z , which is orientation preserving if and only if (1)(|x|+1)(|y|+1) = 1. Proposition 4.3 (L , {, }) is a graded Lie algebra (transversally) with all the degrees shifted by 1. In other words, for each x, y, z L mutually transversal, (1) {x, {y, z}} = {x, {y, z}} + (1)(|x|+1)(|y|+1) {y, {x, z}}. (2) {x, y} = (1)(|x|+1)(|y|+1) {y, x} Proof. Let us prove (1). By Lemma 4.2, {{x, y}, z} + (1)(|x|+1)(|y|+1) {y, {x, z}} {x, {y, z}} = (xy)z(1)(|x|+1)(|y|+1) (yx)z(1)(|x|+|y|)(|z|+1) z(xy)+(1)|x||z|+|y||z|+|x||y|+1)z(yx) +(1)(|x|+1)(|y|+1) y (x z) (1)(|x|+1)(|y|+|z|) y (z x) (1)(|y|+1)(|z|+1) (x z) y + (1)(|x|+|y|)(|z|+1) (z x) y x (y z) + (1)(|y|+1)(|z|+1) x (z y) + (1)(|x|+1)(|y|+|z|) (y z) x (1)|x||y|+|y||z|+|x||z|+1(z y) x = 0

Remark 4.4 An identical Proposition and proof can be found in [1] in a purely algebraic context. Corollary 4.5 The loop homology with the loop bracket,(H, {, }) is a graded Lie algebra of degree +1. Now we discuss the compatibility of loop bracket and loop product. By Lemma 3.2, the map (transversally dened) L L L L , x1 x2 y (x1 x2 ) y x1 (x2 y) (1)|x1 |(|y|+1) (x1 y) x2 is a chain map, i.e., the commutator with is zero transversally. Lemma 4.6 Let x, x1 , x2 , y, y1, y2 L be appropriately transversal, then (1) x (y1 y2 ) = (x y1 ) y2 + (1)|y1 |(|x|+1) y1 (x y2 ) 13

(2) (x1 x2 ) y x1 (x2 y) (1)|x1 |(|y|+1) (x1 y) x2 means chain homotopy.

0, where

Proof. The proof of (1) is easier because the equation holds transversally at the chain level: The set of parameters where the marked point of x coincides with one of the images of y1 y2 is the union of the set of parameters where the marked point of x coincides with one of the image of y1 union the set of parameters where the marked point of x coincides with one of the image of y2 . (see Figure 5) Let us prove (2). The idea (see Figure 6) is that the 2 dimensional chain : L L L L , x1 x2 y x1 ,x2 ,y where x1 and x2 attach to y at pairs of arbitrary points in such a way relative to the cyclic order that x1 is between the marked point and x2 , provides a chain homotopy between the two sides. More precisely, for each pair of points (s, t) T = {(s, t) [0, 1] [0, 1] : s + t 1}, we will dene a chain such that x1 is attached to y at t and x2 is attached to y at 1 s. Let K = {(kx1 , kx2 , s, ky , t) Kx1 Kx2 [0, 1] Ky [0, 1] : x1 (kx1 )(0) = y(ky )(t), x2 (kx2 )(0) = y(ky )(s), s, t [0, 1]; s + t 1} We dene a map x1 ,x2 ,y : K L. y(3) x1 (3 t) y(3) x1 ,x2 ,y (kx1 , kx2 , s, ky , t)() = x2 (3 + s 1) y(3)
t [0, 3 ] t [ 3 , t+1 ]. 3 t+1 1 [ 3 , 3] [ 1s , 2s ]. 3 3 [ 2s , 1]. 3

if if if if if

for each (kx1 , kx2 , s, ky , t) K, S1 . Since the restriction of to (i) K {(kx1 , kx2 , s, ky , t) Kx1 Kx2 [0, 1] Ky [0, 1] : t = 0} is x1 (x2 y). (ii) K {(kx1 , kx2 , s, ky , t) Kx1 Kx2 [0, 1] Ky [0, 1] : s = 1} is (1)x2 (y1) (x1 y) x2 14

(iii) K{(kx1 , kx2 , s, ky , t) Kx1 Kx2 [0, 1] Ky [0, 1] : s + t = 1} is (x1 x2 ) y then K = Ky,x1 ,x2 + (1)y Kyx1 ,x2 + (1)y+x1 Ky,x1 ,x2 + K(x1 x2 )y Kx1 (x2 y) (1)x2 K(x1 y)x2 . is the restriction of to K. Thus ( )(x1 x2 y) = (x1 x2 ) y x1 (x2 y) (1)|x1 |(|y|+1) (x1 y) x2 which completes the proof of (2). (See also [1]). By Corollary 4.5 and Lemma 4.6 we have, Theorem 4.7 The loop product with the loop bracket {, } makes the loop homology into a Gerstenhaber algebra, namely: (1) The loop product denes a graded commutative, associative algebra. (2) {, } is a Lie bracket of degree 1, which means that for each a, b, c H (i) {a, b} = (1)(|a|+1)(|b|+1) {b, a} (ii) {a, {b, c}} = {{a, b}, c} + (1)(|a|+1)(|b|+1) {b, {a, c}} (3) {a, b c} = {a, b} c + (1)|b|(|a|1) b {a, c}.

The operator
Li Li+1

Now we consider the degree +1 operation on the chains of the loop space,

given by the circle action on Map(S1 , M ). It can be dened in the following way: If x: Kx L is an i-chain then (x): S1 Kx L is the i + 1 chain such that for each (s, kx ) S1 Kx , (x)(s, kx )() = x(kx )( + s). Since commutes with the operator on chains, it passes to the loop homology, the homology of the free loop space, inducing a degree +1 operator . Moreover, if x is an i-chain and k 1 then ()k (x) has always geometric dimension i + 1. Therefore we obtain Proposition 5.1 : H H is a degree +1 operator and = 0. 15

We want to study how interacts with the above structure and {, }. In order to do it, we need to dene two auxiliary degree +1 operators on L 1 , 2 : L L . Let x: Kx L be a k-chain. Then
1 1 (x): [0, 2 ] Kx L,

2 (x): [ 1 , 1] L, 2

are the k + 1-chains dened by 1 (x)(s, kx )() = x(kx )( + s) and 2 (x)(s, kx )() = x(kx )( + s). Hence, = 1 + 2 The transversally dened map L L L , x y x y is a chain map because it is composition of (transversally dened) chain maps. On the other hand, using Lemma 3.2 one can prove that the map L L L , x y (1)|x| 2 (x y) x y is also a chain map. Moreover, these two chain maps are chain homotopic, as is shown in the next lemma. Lemma 5.2 For x, y L , (1)|x| 2 (x y) x y x y

Proof. First, the idea of the proof: consider the chain operation : L L L where the loop of x is attached to any point of the loop of y and one goes around a part of the loop of y, starting at any point between the marked points of y and x and ending where x is attached, then goes around the loop of x and nally, around the rest of y. (See Figure 7). More precisely, Let x: Kx L, y: Ky L be two cells of L . Consider Kxy , the parameter space of x y and set K = {(t, kx , s, ky ) [0, 1] Kxy : 0 t s 1} 16

Observe that x is attached to y at the image of s by y and the image of t by y is the marked point of the resultant loop. Proceeding in an analogous way as we did in the proof of Lemma 3.2, we obtain ( )(x y) = (1)|x| 2 (x y) x y x y as desired. Corollary 5.3 The loop bracket {, } on the loop homology is the deviation of from being a derivation of the loop product. In other words, for a, b H {a, b} = (1)|a| (a b) (1)|a| a b a b Theorem 5.4 The loop product and the operator make the loop homology into a Batalin Vilkovisky algebra, namely: (1) is a graded commutative associative algebra. (2) = 0. (3) (1)|a| (a b) (1)|a| a b a b is a derivation of each variable. Proof. By Theorem 4.7,Proposition 5.1 and Corollary 5.3. Remark 5.5 The alternative denition of a Batalin Vilkovisky algebra as a graded commutative algebra (A, ) with a degree +1 operator : A A such that = 0 and for each a, b, c A (a b c) = (a b) c + (1)|a| a (b c) + (1)(|a|1)b b (a c) (a) b c (1)|a| a (b) c (1)|a|+|b| a b (c) can be found in Getzler [2].

and dene a chain : K L as follows if [0, st ] , y(ky )(2 + t) 2 x(kx )(2 s + t) if [ st , st+1 ], (t, kx , s, ky )() = 2 2 y(ky )(2 + t) if [ st+1 , 1]. 2

The String Bracket

Using the circle action on the free loop space, we can dene the equivariant homology H of the entire loop space. We could describe this as the ordinary homology of the quotient space S of general smooth mappings (S1 , M d ) by the circle action of rotation in the domain circle S1 . The space S can be viewed 17

as the space of all general smooth closed curves in M . Thus we refer to the equivariant homology of the mapping space Map(S1 , M ) as string homology. The circle bration S1 (typical loops) S =string space leads to an exact sequence (geometric grading) Hi Hi Hi2 Hi1 . . . where E forgets the marked point of each member of a family of loops, M places a mark on each string in a family in all possible positions (and c is dened by cap product with the characteristic class of the circle bundle above). The operator above is the composition M E. The composition E M on homology is zero, as part of the exactness above. e Any operation Hk H given by composition of and yields an operation Hk H , = E Mk . In particular, taking = and adding a sign, gives the binary operation H H H called the string bracket, [a, b] = (1)|a| E(M(a) M(b)). where |a| = dimension a d. Theorem 6.1 String homology with the string bracket, (H , [, ]) is a graded Lie algebra of degree (2 d) for the geometric grading. Proof. By Theorem 3.3, is graded commutative. So, since M has degree +1, [a, b] = (1)|a|+(|a|+1)(|b|+1) E(M(b) M(a)) = (1)|a|.|b| [b, a]. To prove Jacobi, replace a (resp. b, c) by M(a) (resp. M(b), M(c)) in the Leibniz property (3) of Theorem 4.7 and apply E to both sides of the equation to obtain E ({M(a), M(b) M(c)} {M(a), M(b)} c (1)|a|(|b|+1) M(b) {M(a), M(c)} = 0. Now, use Theorem 5.4 and the fact that M E = to replace in the above equation each of the brackets {x, y} by the formula (1)|c| (M E)(c d)
[,] E c M

18

(1)|c| (M E)(c) d c (M E)(d). Since E M = 0 we cancel the terms where E M appear and so we obtain E (M(a) ME(M(b) M(c)) (1)|a|+1 ME (M(a) M(b)) Mc (1)|a|(|b|+1)+|a|+1 M(b) ME (M(a) M(c))) = 0 Now, replacing in the above formula each occurrence of E(M(d), M(e)) by (1)|d| [d, e] yields (1)|a|+|b| [a, [b, c]] + (1)|a|+|b| [[a, b], c] + (1)|a||b|+|a|+|b| [b, [a, c]] = 0. Hence, [a, [b, c]] = [[a, b], c] + (1)(|a||b|) [b, [a, c]].

By the above procedure, considering E Mk we dene operations mk : Hk H, by the formula Hk Hk H H. Note that if we shift the grading on string homology by (d + 1) from its geometric grading the degree of each mk becomes +1(= (k 1)d + k k(d + 1) + (d + 1)). Now extend each mk to a coderivation mk to H . Here, H is the free graded commutative coalgebra on H with the new algebraic grading and with the coalgebra structure which dualizes the usual algebra structure on H . Thus mk is the unique operation whose dual operation is the (unique) derivation on H extending the operator dual to mk . One knows the Jacobi identity for the string bracket m2 is equivalent to the relation m2 m2 = 0 for the associated coderivation m2 . The Jacobi relation for m2 generalizes to the entire collection {m2 , m3 , . . .} in the following way. Theorem 6.2 The associated coderivations {m2 , m3 , . . .} of the free commutative coalgebra H on the string homology H satisfy: (i) mk mk = 0, for k = 2, 3, 4 . . .. (ii) mk mr + mr mk = 0 for k, r = 2, 3, 4, . . ..
M k ()(k1) E

k = 2, 3, 4, . . .

19

Proof. We only have to show the equations in the case when the range of the commutator being studied lies in monomial degree one. Since the commutator of coderivations is a coderivation, this is enough to show it is identically zero. We illustrate the proof of (ii) for k = 3, r = 2. We have four families of closed curves A1 , A2 , A3 , A4 . Then (m2 m3 )(A1 A2 A3 A4 ) can be viewed a sum of twelve terms, each of them labeled with ({Ai1 , Ai2 , Ai3 }, Ai1 ), where {Ai1 , Ai2 , Ai3 } runs over all possible choices of three families from {A1 , A2 , A3 , A4 }, with a preferred element. Analogously, (m2 m3 )(A1 A2 A3 A4 ) can be viewed as a sum of twelve terms, each of them labeled with ({Al1 , Al2 }, Al1 ), where {Al1 , Al2 } runs over all possible choices of two families with a preferred element. A correspondence of the two sets of labels is given by the map ({Ai1 , Ai2 , Ai3 }, Ai1 ) ({Ai1 , An }, Ai1 ) where An is the only family not in {Ai1 , Ai2 , Ai3 }. Now, we will see that corresponding pair of terms appear in m3 m2 + m2 m3 with dierent sign. Consider, for instance, the corresponding pair of terms labeled by ({A1 , A2 , A3 }, A3 ) and ({A3 , A4 }, A3 ). For simplicity, let us denote by Ai the parameter space of the family Ai . We can assume that each Ai is a cell, and that M(Ai ) is a map M(Ai ): S1 Ai L where M(Ai )(s, a)() = Ai (a)(s + ). The parameter space of the term labeled ({A1 , A2 , A3 }, A3 ) is K , the preimage of diagM 3 diagM 2 under the map S1 S1 A1 S1 A2 S1 A3 S1 A4 M 3 M 2 , given by (t, s1 , k1 , s2 , k2 , s3 , k3 , s4 , k4 ) = ((A1 (k1 )(s1 ), A2 (k2 )(s2 ), A3 (k3 )(s3 )) , (A3 (k3 )(t), A4 (k4 )(s4 ))) 20

The parameter space of the term labeled with ({A3 , A4 }, A3 ), K is the preimage of diagM 3 diagM 2 under the map S1 S1 A1 S1 A2 S1 A3 S1 A4 M 3 M 2 , given by (s1 , t, k1 , s2 , k2 , s3 , k3 , s4 , k4 ) = ((A1 (k1 )(s1 ), A2 (k2 )(s2 ), A3 (k3 )(s3 )) , (A3 (k3 )(t), A4 (k4 )(s4 ))) . Over each point of these parameter spaces, the loops of the terms labeled with ({A1 , A2 , A3 }, A3 ) and ({A3 , A4 }, A3 ) are as in Figure 8. Observe that the only dierence between and is that s1 and t are interchanged. This produces the dierence of sign. Now, we prove (i). We illustrate for k = 3. There will be ve families, A1 , A2 , A3 , A4 , A5 . (m3 m3 )(A1 A2 A3 A4 A5 ) is the sum of thirty terms, each of them labeled with ({Ai1 , Ai2 , Ai3 }, Ai1 ) where ({Ai1 , Ai2 , Ai3 } runs over all subsets of three elements of {A1 , A2 , A3 , A4 , A5 }. We group these terms in pairs with the following correspondence ({Ai1 , Ai2 , Ai3 }, Ai1 ) ({Ai1 , Aj1 , Aj2 }, Ai1 ) where {Aj1 , Aj2 } = {A1 , A2 , A3 , A4 , A5 } \ {Ai1 , Ai2 , Ai3 }. As in the proof of case (ii), we can see that the parameter spaces corresponding to pairs cancel, so (i) holds. Corollary 6.3 There exists an uncountable family { } of Lie structures on the string homology. Namely, for each {2, 3, . . .}, : H H dened as =

is a coderivation which satises = 0. Proof. By Theorem 6.2 each is a coderivation of H of square zero and so determines (by denition) a Lie structure on H .

21

Examples

Example 7.1 Let us unwrap the string bracket for M d when d = 2. If the genus is greater than one, the equivariant homology of the loop space is concentrated in dimension zero (except for the component of the trivial loop whose higher homology we ignore at the moment). The zeroth equivariant homology group, H0 , is the vector space with basis , where denotes the set of free homotopy classes of loops in M . Let us calculate the string bracket of two elements a, b H0 . Let , : S1 M be two loops such that their free homotopy classes are a and b respectively. We consider g two elements x, y of L1 determined by two maps S1 Map(S1 , M 2 ) given by x(s)() = (s + ) y(t)() = (t + ). Hence, Ma = x and Mb = y , where x denotes the homology class of a cycle x. Assume that x and y are transversal. Thus, Kxy = {(s, t) S1 S1 : (s) = (t)}. By transversality, the set of intersection points of the curves and , is nite. Observe that it is precisely { (s) : (s, t) Kxb }. The orientation of Kxy is given by endowing each of its points with a sign. For each (s, t) Kxy , this sign is precisely, the intersection index ( , , (s)) of the loops and at (s). Then Ma Mb =
p x,y

( , , p)( )p ,

where ( )p denotes the free homotopy class which contains the loop product of and at p. Hence, [a, b] = E(Ma Mb) =
p

( , , p)E(( )p ),

It seems remarkable that this formula is well dened in the vector space of components, it is skew symmetric, and satises Jacobi. Thus the string bracket becomes for d = 2 the formula discovered by Wolpert [12] and Goldman [3] which plays a role in the symplectic structure of the Techmuller space [12](because of Thurstons earthquakes) and the symplectic structure on the at G-bundles for G a compact semisimple Lie group [3]. Indeed, trying to understand and generalize Goldmans work [3] lead us to the general theory above. 22

Example 7.2 Now take d = 3 and consider a possibly twisted circle bundle over M 3 with base a surface F of genus greater than one. It is interesting for the string bracket of M 3 to consider equivariant homology (i.e., string homology) in dimension one. Cycles are generated by maps of torii into M 3 . The projection of the torus to F is homotopic to a circle. So we only get interesting examples of torii in M 3 by taking all the bers in M 3 over some circle in F . Two of these torii A and B can be put in transversal position by putting their projections a and b in F in transversal position. We see that the string bracket of A and B in M 3 is the lift of the string bracket of a and b in F (as described in Example 7.1). So for these 3-manifolds the string bracket is just as non trivial as the string bracket on surfaces. A similar discussion applies to Seifert brations over surfaces.

Loop product and cap product

The cohomology algebra of a space (with cup product) acts on the homology of a space (called cap product) via the duality formula < a x, b >=< a b, x > where <, > is the dual pairing between homology and cohomology, is cup product, and is cap product. To relate this structure to our loop product , consider the diagram c i Map(S1 , M ) Map(gure eight, M ) Map(S1 , M ) Map(S1 , M ) where c denotes the composition of loops and i is the natural inclusion. Denition 8.1 A pair of cohomology classes (a, A) (in the appropriate spaces) is called a compatible pair if i A = c a. Theorem 8.2 For each x, y homology classes and compatible pair of classes (A, a) loop product (A (x y)) = a (x y). Proof. The loop product is the composition of intersection with image i (which is represented as a codimension d submanifold being the transverse image of the diagonal under the map Map(S1 , M ) Map(S1 , M )
marked points

M M)

with the induced transformation of c in homology. The rst process commutes with cap product with rst A then with its restriction to Map(gure eight, M ). The second process commutes with capping with a class in Map(S1 , M ) or with its pull back via c to Map(gure eight, M ) 23

Appendix 1: S2 and other simply connected manifolds

based loop maps on S2 all loops in S2 S2

Fibrations such as

have algebraic models. For example, S2 and based loops on S2 are modeled by d d (0; x; y . . .) (0; 0; x2 ) and (x; y ) (0; 0) respectively. The notation gives the generators in degrees 1; 2; 3; . . . respectively and what the dierentials are. The model is the free commutative algebra with those generators provided with a derivation d of square zero and degree 1 with the specied values. d The total space of the bration has a model (; y , x; y) (0; 2x, 0; x2 ). x x There is also a derivation of degree 1 and square zero given by (; y ; x; y) (0; 0; x; y ) and d + d = 0 is true (and in fact determines d x given and d on the base of the bration.) The models are cochain models. The obvious maps serve as cochain maps corresponding to the maps between space. The cohomology and induced transformations are derived accordingly [5]. The equivariant cohomology or string cohomology has model obtained by adding one closed generator u of degree 2 (; y, x, u; y) (0; 2x, xu, 0; x2 + y u) x x where du = 0 and dz = dz + (z)u determines the rule for the other generators z. Calculating with these models we nd the Betti numbers of the loop homology are all 1 in each dimension 1, 2, 3, . . . and the image of based loop homology in free homology is in degrees 1, 3, 5, . . .. Thus exactness for S2 of based loop homology loop homology based loop homology implies has non trivial image in the even degrees 0, 2, 4, . . .. This shows the loop product is non-trivial for S2 . Calculating further, one nds is an isomorphism on loop homology from odd to even dimensions and is zero in even dimensions. Since = M E, one nds that E is non zero in odd dimensions and M is non zero starting in odd dimensions. Since E M = 0, one nds E is zero in even dimensions and M is zero starting in even dimensions. Thus the string bracket [x, y] = E(M(x) M(y)) is zero in all dimensions for S2 . 24
inclusion d

9.1

General simply connected manifolds

This pattern works in general for simply connected manifold to describe the algorithm for calculating the rational homology of these spaces: If M has minimal model (where each dierential is quadratic + higher order terms) (0; x1 , x2 , . . . ; y1 , y2 , . . .) (dx1 , dx2 , . . . ; dy1 , dy2 , . . .) then the based loops on M has a model (x1 , x2 , . . . ; y1 , y2 , . . .) (0, 0, . . . 0; 0, . . .) and the free loop space with operations and d of degree 1 and 1 satisfying d + d = 0 has model with generators (x1 , x2 , . . . ; y1 , y2 , . . . , x1 , x2 , . . . ; y1 , y2 , . . .) is dened by xi xi , yi yi and xi , yi 0. Then d is dened so that d + d = 0 and d is given as before on the xi , yi . . . coming from M d . Thus (d + d)xi = 0 implies di = (dxi ) can be calculated since dxi and x are known. A equivariant model is obtained by adding to a free loop space model one more variable u in degree two with du = 0 and dz = dz + (z) u for the other generators z [5]. Remark 9.1 In [6] it was shown the ranks of the loop homology are unbounded for simply connected manifolds unless the minimal model has only one or two generators (like S2 or S3 ).
d d

10

Appendix 2: M 3 and K(, 1) manifolds

It is known that any closed 3-manifold is a connected sum M1 M2 M3 . . . Mn along S2 where each of the Mi s is of one of the following types (i) 1 is nite so the universal cover is homotopy equivalent to S3 . (ii) Mi is S1 S2 . (iii) 1 is innite and the universal cover is contractible [4] The technique of models plus nite group invariance can be used to treat the examples of type (i) and (ii) . If we treat examples of type (iii) it seems plausible one could develop an algorithm for the connected sum using our knowledge of S2 and free product ideas. 25

We discuss type (iii) under the hypothesis of Thurstons geometrization picture. (1) If Mi is closed hyperbolic, each centralizer of a non-zero conjugacy class is innite cyclic. Thus that component of the free loop space is a homotopy circle. By dimension reasons all loop products between these components are zero. The loop product reduces to the classical intersection product. (2) Otherwise M would be a union along torii of Seifert brations over surfaces with boundary and nite volume hyperbolic manifolds with neighborhoods of the cusps deleted. (3) If any non trivial Seifert brations are present, we have a rich structure of loop product as described in Example 7.2. (4) Finally, M could be a union along torii of hyperbolic pieces and we havent analyzed these cases.

10.1

General K(, 1) manifolds, d 3

The loop space of M is homotopy equivalent to a union over conjugacy classes in of K( , 1) homotopy types, where is the centralizer of a representative of . In particular a component of L(M ) is homotopy equivalent to a covering space of M . Thus its homological dimension is at most d. Again for hyperbolic manifolds, the string bracket is zero for dimension reasons.

References
[1] Gerstenhaber, Murray. The cohomology structure of an associative ring, Ann. Math. (2) 78, 59-73, (1963). [2] Getzler, Ezra. Batalin-Vilkovisky Algebras and Two-Dimensional Topological Field Theories Comm. Math. Phys. 159, 265-285, (1994) [3] Goldman, William M., Invariant functions on Lie groups and Hamiltonian ows of surface group representations. Invent. Math. 85 (1986), no. 2, 263-302 [4] Jaco, M and Shalen, P. Memoirs AMS 220 [5] Sullivan, Dennis. Innitesimal computations in topology, Publ. I.H.E.S. 47 (1977), 269-331. [6] Sullivan, Dennis and Vigu-Poirrier, Michelle. The homology theory of e closed geodesic problem, J. Dierential Geometry 11 (1976), 633-644. 26

[7] Stashe, James, On the homotopy associativity of H-spaces I Trans. AMS 108 (1963), 275-292. [8] Stashe, James, On the homotopy associativity of H-spaces II Trans. AMS 108 (1963), 293-312. [9] Turaev, Vladimir G., Skein quantization of Poisson algebras of loops on surfaces. Ann. Sci. Ecole Norm. Sup. (4) 24 (1991), no. 6, 635704. [10] Witten, E., Non-commutative geometry and string eld theory Nuclear Physics B 268 (1986), 253-294. [11] Witten, E and Zwiebach, B. Algebraic structures and dierential geometry in two-dimensional string theory. Nucl. Phys. B 377, (1992), 55-112. [12] Wolpert, S., On the Symplectic Geometry of Deformations of Hyperbolic Surfaces Ann. Math. Vol 117, (1983), 207-234. [13] Zwiebach, Barton. Closed string eld theory: quantum action and the B-V master equation. Nucl. Phys. B 390, (1993), 33-152.

27

Figure 2: For M 3 , the loop product of a 1-cell with a 2-cell is a 0-chain

28

Figure 3: The product

29

Figure 4: Kx(yz) and K(xy)z

30

Figure 5: Proof of Lemma 4.6(1)

31

Figure 6: Proof of Lemma 4.6(2)

32

Figure 7: Proof of Lemma 5.2

33

Figure 8: The loop of the terms labeled with ({A1 , A2 , A3 }, A3 ) and ({A3 , A4 }, A3 )

34