September2,2004 9:14 WSPC/TrimSize: 9.75inx6.5inforProceedings witten SL(2,Z) ACTION ON THREE-DIMENSIONAL CONFORMAL FIELD THEORIES WITH ABELIAN SYMMETRY EDWARD WITTEN Institute For Advanced Study, Princeton NJ 08540 USA On the space of three-dimensional conformal field theories with U(1) symmetry and a chosen coupling to a background gauge field, there is a natural action of the group SL(2,Z). ThegeneratorSofSL(2,Z)actsbylettingthebackgroundgaugefieldbecome dynamical, an operation considered recently by Kapustin and Strassler in explaining three-dimensionalmirrorsymmetry. TheothergeneratorT actsbyshiftingtheChern– Simons coupling of the background field. This SL(2,Z) action in three dimensions is related by the AdS/CFT correspondence to SL(2,Z) duality of low energy U(1) gauge fieldsinfourdimensions. Table of Contents 1 Introduction 1174 2 AbelianChern–SimonsInteractionsInThreeDimensions 1177 2.1 Generalities . . . . . . . . . . . . . . . . . . . . . . . . . . . 1177 2.2 A Trivial Theory . . . . . . . . . . . . . . . . . . . . . . . . 1178 2.3 An Almost Trivial Theory . . . . . . . . . . . . . . . . . . . 1182 3 Action Of SL(2,Z) On Conformal Field Theories 1184 4 Modular Action On Current Two-Point Function 1190 5 Interpretation In AdS/CFT Correspondence 1192 References 1199 1173 September2,2004 9:14 WSPC/TrimSize: 9.75inx6.5inforProceedings witten 1174 Edward Witten 1. Introduction In this paper, we will consider objects of the following kind: conformal field theoriesa inthreedimensionsthathaveaU(1)symmetry, withanassociated conserved current J. The goal of the discussion is to show that there is a natural action of SL(2,Z) mapping such theories to other theories of the same type. This is not a duality group in the usual sense; an SL(2,Z) transformation does not in general map a theory to an equivalent one or even to one that is on the same component of the moduli space of conformal theories with U(1) symmetry. It simply maps a conformal field theory with U(1) symmetry to another, generally inequivalent conformal field theory with U(1) symmetry. Roughlyspeaking,wewillgivethreeapproachestounderstandingtherole ofSL(2,Z). InSection3,afterexplaininghowthegeneratorsofSL(2,Z)act, we perform a short formal computation using properties of Chern–Simons theories to show that the relevant operations do in fact generate an action of SL(2,Z). In Section 4, for theories in which the current two-point function is nearly Gaussian, we show directly the SL(2,Z) action on this two-point function. Finally, in Section 5, we explain the origin of the SL(2,Z) action from the point of view of the AdS/CFT correspondence. After submitting the original hep-th version of this paper, I learned that inthecontextoffractionalquantumHallsystems,essentiallythesamedefini- tionsweremadesometimeago,andacomputationsimilartothatinSection 4 was performed, by Burgess and Dolan [1]. Their motivation came from indications [2-8] of a duality group underlying the fractional quantum Hall effect. More generally, Chern–Simons gauge fields and operations adding and removing them are extensively used in understanding the quantum Hall effect. There is an extensive literature on this; an introduction for particle physicists is [9]. The starting point for the present paper was work by Kapustin and Strassler [10] on three-dimensional mirror symmetry [11]. Kapustin and Strassler considered an operation that we will call S, since it turns out to „ « correspond to the generator 0 1 of SL(2,Z) that usually goes by that −10 name. TheS operationisdefinedasfollows. OnegaugestheU(1)symmetry, introducing a gauge field A that is coupled to J. Thus, if the original theory aInfact,wecanextendthedefinitionsbeyondconformalfieldtheoriestopossiblymassivetheories obtained by relevant deformations of conformal field theories. However, the transformations of conformalfixedpointsareabasiccase,sowewillphraseourdiscussionintermsofconformalfield theories. September2,2004 9:14 WSPC/TrimSize: 9.75inx6.5inforProceedings witten SL(2,Z) action on three-dimensional conformal field theories 1175 has a Lagrangian description in terms of fields Φ and a Lagrangian L, the new theory has fields Φ and A, and a Lagrangian that is a gauge-covariant extension of L. (In the simplest case, this extension is just Le = L+AiJi.) One defines a new conformal field theory in which A is treated as a dynam- ical field, without adding any kinetic energy for A. The conserved current of the new theory is Je= ∗F/2π, where F = dA. The main result of Kapustin and Strassler was to show that (after mak- ing a supersymmetric extension of the definition) the S operation applied to a free hypermultiplet gives back a mirror free hypermultiplet, and more- over that this implies mirror symmetry of abelian gauge theories in three dimensions. „ « SL(2,Z) is generated along with S by the matrix 11 . The relations 01 they obey are that (ST)3 = 1 and that S2 is a central element whose square is 1 (this is often described by writing S2 = −1). In the present context, T corresponds to a rather trivial operation that merely shifts the two-point function of the U(1) current J by a contact term. Though this operation is not very interesting by itself, it gains interest because it does not commute with S. (This is analogous to four-dimensional gauge theories, where [12] the simple operation θ → θ+2π gains interest because it does not commute with electric-magnetic duality.) This paper is devoted to understanding from several points of view that the S and T operations defined for three- dimensional quantum field theories do generate SL(2,Z).b The T operation is closely related to the possibility of having a Chern– Simons interaction for gauge fields in three dimensions [15]; in fact, T sim- ply shifts the Chern–Simons level of the background gauge field. Kapustin andStrasslerin[10]consideredChern–Simonsinteractions(andalsocontact terms in two-point functions) in relation to the S operation. (In terms of the quantum Hall effect, T is understood [1] as a 2π shift in the statistics of the charge carriers.) Since the definition of S using a gauge field without a kinetic energy seems a bit hazardous at first sight, we may derive some encouragement from two dimensions, where gauging of WZW models without any kinetic energyforthegaugefieldshasbeenusedtodescribecosetmodels[16-18]. In bThe partition function computed on a three-manifold Q transforms under SL(2,Z) with a c- number phase factor that depends only on the topology of Q and not on the specific conformal fieldtheoryunderdiscussion,oritscurrents. Thisfactorpossiblycouldberemovedbymodifying thecouplingtogravityortheactionofSL(2,Z)bytermsinvolvingthegravitationalbackground only. SL(2,Z)dualityoffour-dimensionalgaugetheoryinvolvesasomewhatanalogoustopological effect[13,14]. September2,2004 9:14 WSPC/TrimSize: 9.75inx6.5inforProceedings witten 1176 Edward Witten addition, there is an illuminating three-dimensional situation in which the S operation can be made more concrete. This is the large N limit of a theory f of N free fermions with U(1) symmetry. The S operation applied to this f theory produces a strong coupling limit of three-dimensional QED that has been much studied some time ago [19-22] as well as recently [23-25]. In this example, the current J has almost Gaussian correlation functions. In such a situation, we show that SL(2,Z) acts by τ → (aτ+b)/(cτ+d), where τ will be defined later in terms of the two-point function of J. (As noted above, this essentially duplicates considerations in [1].) We also show that the SL(2,Z) symmetry has a closely related interpre- tation in the AdS/CFT correspondence. A U(1) global symmetry in three dimensions corresponds to a U(1) gauge symmetry in a dual description in AdS . In U(1) gauge theory in four dimensions, there is an SL(2,Z) ambi- 4 guity in what we mean by the “gauge field.” For each choice, we can pick an associated boundary condition and, by a familiar construction, define a conformal field theory on the boundary of AdS with a conserved current J. 4 The SL(2,Z) action on conformal field theories on the boundary is induced from SL(2,Z) duality transformations on the gauge fields in the bulk. This is analogous to the behavior of scalar fields in AdS space in a certain range of masses; they can be quantized in two ways [26] leading to two possible CFT duals on the boundary [27]. In the above summary, we omitted one interesting detail. The definition of T assumes that we are working on a three-manifold with a chosen spin structure. Intheabsenceofachosenspinstructure,onecoulddefineonlyT2, not T, and accordingly one gets only an action of the subgroup of SL(2,Z) thatisgeneratedbyS andT2. Thisisdualtothefact[14]thatfullSL(2,Z) duality for free abelian gauge fields on a four-manifold requires a choice of spin structure. This paper is organized as follows. In Section 2, we review some aspects of abelian Chern–Simons theory in three dimensions (for more information, see for example [28]), elucidating details such as the role of a spin structure. (The rest of the paper can be read while omitting most details in Section 2.) InSection3,wedefinetheT operationanddemonstrateSL(2,Z)symmetry. In Section 4, we consider the case that the current is an almost Gaussian field. And in Section 5, we discuss the duality with gauge fields in AdS . 4 September2,2004 9:14 WSPC/TrimSize: 9.75inx6.5inforProceedings witten SL(2,Z) action on three-dimensional conformal field theories 1177 2. Abelian Chern–Simons Interactions In Three Dimensions 2.1. Generalities ForanabeliangaugefieldA,letF = dAbethefieldstrength,andx = F/2π. R On a compact oriented four-manifold M, in general x∧x is an integer. M R If M is spin, then x∧x is even. M So J = R F ∧F/4π2 is integral in general and is even if M is spin.c In M physical notation, this would often be written Z 1 J = d4x(cid:15)ijklF F . (1.1) 16π2 ij kl Now consider an abelian gauge field A on an oriented three-manifold Q. If A is topologically trivial, the Chern–Simons functional of A is simply Z 1 I(A) = d3x(cid:15)jklA ∂ A . (1.2) j k l 2π Q It is important to extend the definition so that it makes sense when A is a connection on a topologically nontrivial line bundle L, and hence is not really defined as a one-form. There is a standard recipe to do so. We find a four-manifold X with an extension of L and A over M. In four dimensions, (cid:15)ijkl∂ (A ∂ A ) = 1(cid:15)ijklF F , so we replace (1.2) by i j k l 4 ij kl Z 1 I (A) = d4x(cid:15)ijklF F . (1.3) X ij kl 8π X Now we must understand to what extent this is independent of the choice of X. If Y is some other four-manifold with an extension of A, and M is the closed four-manifold built by gluing X and Y along their common boundary Q (after reversing the orientation of Y so the orientations are compatible), then Z 1 I (A)−I (A) = 2π· d4x(cid:15)ijklF F = 2πJ(M). (1.4) X Y 16π2 ij kl M In particular, I (A)−I (A) is an integer multiple of 2π. X Y Thus, exp(iI(A)) is independent of the choice of X and the extension of A. ThisisgoodenoughforconstructingaquantumfieldtheorywithI(A)as cFor spin manifolds, a typical example to keep in mind is M = T ×T0 with T and T0 being two-toriandR F =R F =2π. Forexample,ifT andT0 aremadebyidentifyingboundariesof T T0 unit squares in the x1−x2 and x3−x4 planes, respectively, we take F12 =F34 =2π and other componentszero. Onereadilycomputesthatinthisexample,J =2. Thisisthesmallestnon-zero valueofJ thatispossibleonaspinmanifold. ForasimpleexampleinwhichM isnotspinand J =1,takeM =CP2 andtakeF suchthatR F =2π,whereU isacopyofCP1⊂CP2. U September2,2004 9:14 WSPC/TrimSize: 9.75inx6.5inforProceedings witten 1178 Edward Witten a term in the action. I(A) is called the abelian Chern–Simons interaction at level1. Weoftenwriteitasin(1.2)eventhoughthisformulaisstrictlyvalid onlyforthetopologicallytrivialcase. (Allmanipulationswemakelater,such as integrations by parts and changes of variables in path integrals, are easily checked to be valid using the more complete definition of the Chern–Simons functional.) If,however,Qisaspinmanifold(bywhichwemeanathree-manifoldwith a chosen spin structure), we can do better. In this case, we can pick X and Y sothatthechosenspinstructureofQextendsoverX andY. Accordingly, M is also spin and hence J(M) is an even integer. Consequently, we can divide I(A) by two and define Z I(A) 1 Ie(A) = = d3x(cid:15)jklAj∂kAl, (1.5) 2 4π Q which is still well-defined modulo 2π in this situation. On a three- dimensional spin manifold, the “level one-half” Chern–Simons interaction Ie(A) is the fundamental one. 2.2. A Trivial Theory There are a few more facts that we should know about abelian Chern– Simons gauge theory in three dimensions. Consider a theory with gauge group U(1)×U(1) and two gauge fields A and B, and with action Z 1 I(A,B) = d3x(cid:15)jklA ∂ B . (1.6) j k l 2π Q The extension to a topologically non-trivial situation in such a way that I(A,B) is well-defined mod 2π is made just as we did above for the case of a single gauge field. No choice of spin structure is required here. This theory has no framing anomaly because the quadratic form used in writ- ing the kinetic energy has one positive and one negative eigenvalue; if it is diagonalized, the two fields with opposite signs of the kinetic energy make opposite contributions to the anomaly. We claim, in fact, that this precise theory with two gauge fields is com- pletelytrivial. OneaspectofthistrivialityisthattheHamiltonianquantiza- tion of the theory is trivial, in the following sense: if the theory is quantized on a Riemann surface Σ of any genus, then the physical Hilbert space is one-dimensional, and the mapping class group of Σ acts trivially on it. We will call this property Hamiltonian triviality. September2,2004 9:14 WSPC/TrimSize: 9.75inx6.5inforProceedings witten SL(2,Z) action on three-dimensional conformal field theories 1179 Assuming Hamiltonian triviality for the moment, we will prove another property that we might call path integral triviality: the partition function Zb(Q) of the theory on an arbitrary closed three-manifold Q is 1. (We call this partition function Zb as we will use the name Z for a different partition function presently. The partition function is a well-defined number because theframinganomalyvanishes,asnotedabove.) First,forQ = S2×S1,Zb(Q) is the dimension of the physical Hilbert space on S2, and so is 1 according to Hamiltonian triviality. Since S2 can be built by gluing together two copies of a disc D along their boundary, S2 ×S1 can be built by gluing together two copies E and 1 E of D×S1. The boundary of E is a two-torus F = S1×S1. By making a 2 1 diffeomorphism S of F before gluing E to E (the requisite diffeomorphism 1 2 corresponds to the modular transformation S : τ → −1/τ exchanging the two S1 factors in F), one can build S3. From these facts we can prove that Zb(S3) is also 1. In fact, the path integral on E computes a vector ψ in the physical Hilbert space H of F, 1 1 F andthepathintegralonE2 likewisecomputesavectorψ2 inHF. Zb(S2×S1) is the overlap hψ1|ψ2i, and Zb(S3) is the matrix element hψ1|ρ(S)ψ2i, where ρ(S) is the linear transformation that represents the diffeomorphism S on the physical Hilbert space H . The assumption of Hamiltonian triviality F says that ρ(S) = 1, so S3 and S2 × S1 have the same partition function. Hence Zb(S3) = 1. The same argument can be extended, using standard facts about three- manifolds, to show that Zb(Q) = 1 for any Q. Consider a genus g Riemann surface embedded in S3. It has an interior and also an exterior. They are equivalent topologically and are called “handlebodies.” So S3 can be made bygluingtogethertwogenusg handlebodiesH andH . Anythree-manifold 1 2 can be made by gluing together H and H (for some g) after first making a 1 2 diffeomorphismσ oftheboundaryofH . AssumingHamiltoniantriviality,σ 1 actstriviallyandtheargumentofthelastparagraphshowsthatthepartition function of the three-manifold obtained this way is independent of σ. Hence all three-manifolds have the same partition function; in view of the special case S2×S1, the common value is clearly 1. Toillustrate,wewillcomputemoreexplicitlyZb(Q)forthecaseofathree- manifold Q with b (Q) = 0 (and by Poincar´e duality, hence also b (Q) = 0). 1 2 We simply perform the path integral Z (cid:18) Z (cid:19) i DADB exp d3x(cid:15)jklA ∂ B . (1.7) j k l 2π Q September2,2004 9:14 WSPC/TrimSize: 9.75inx6.5inforProceedings witten 1180 Edward Witten To do the path integral, we write A = A +A0, where A is a connection triv triv on a trivial line bundle and A0 has harmonic curvature, and similarly B = B +B0. In the case at hand, as b (Q) = 0, A0 and B0 are both flat and triv 2 the action is a simple sum I(A,B) = I(A ,B )+I(A0,B0). The path triv triv integral is a product of an integral over A and B and a finite sum over triv triv A0 and B0. The path integral over A and B gives a ratio of determinants in triv triv a standard way. As shown by A. Schwarz in his early work on topological fieldtheory[29],thisratioofdeterminantsgivesexp(T),whereT istheRay- Singer torsion. Schwarz considered a slightly more general model in which A and B are twisted by a flat bundle (and one gets the torsion of the given flat bundle), butinthepresentinstancethisisabsent, sowewantexp(T)forthe trivialflatbundle. Inthreedimensions,thisisequalto1/#H (Q;Z),d where 1 for a finite group F, #F denotes the number of elements of F. Poincar´e duality implies in three dimensions that #H (Q;Z) = #H2(Q;Z), a fact 1 that we will use in the next paragraph. Now let us perform the sum over A0 and B0. They are classified by x ,x ∈ H2(Q;Z), the first Chern classes of the line bundles on which A A B and B are connections. However, it is convenient to break the symmetry between A and B and to look at x differently. We use the exact sequence B 0 → Z → R → U(1) → 0 and the associated exact sequence of cohomology groups ··· → H1(Q;R) → H1(Q;U(1)) → H2(Q;Z) → H2(Q;R) → .... Since we are assuming that b (Q) = b (Q) = 0, we have Hi(Q;R) = 0, 1 2 i = 1,2 and hence H2(Q;Z) ∼= H1(Q;U(1)). So naturally associated with x is an element η ∈ H1(Q;U(1)). Instead of summing over x and x B B A B to get the contribution of non-trivial topologies to the path integral, we can sum over x and η . The sum we want is loosely speaking A B (cid:18) Z (cid:19) X i exp d3x(cid:15)ijkA0∂ B0 , (1.8) 2π i j k xA,ηB but here we are considering topologically non-trivial gauge fields so we need to use a more precise definition of the action. Actually, the cup product dThiscanbeseenasfollows(theargumentwassuppliedbyD.Freed). Athree-manifoldQwith b1 =0 is called a rational homology sphere and has a cell decomposition with a single 0-cell, N 1-cells and 2-cells for some N, and a single 3-cell. The associated chain complex given by the boundary operator looks like Z → ZN → ZN → Z. The first and last differentials vanish (as b0=b3=1)andafterpickinganorientation,H0 andH3 havenaturalbases(givenbyapointin Q,andQitself). Thereremainsthemiddledifferential;itisinjective,asb1=0,anditscokernel isH1(Q;Z). Thedeterminantofthisdifferentialisthus|H1(Q;Z)|,andthisistheexponentialof thetorsionforthetrivialconnection,relativetothenaturalbasesonH0 andH3. September2,2004 9:14 WSPC/TrimSize: 9.75inx6.5inforProceedings witten SL(2,Z) action on three-dimensional conformal field theories 1181 and Poincar´e duality give a perfect pairing T : H2(Q;Z)×H1(Q;U(1)) → H3(Q;U(1)) = U(1). The action (for flat gauge fields) can be written in terms of T, and the sum we want is really X T(x ,η ). (1.9) A B xA,ηB Perfectness ofthe pairing T says in particular that foranyfixed andnonzero P η , we have T(x ,η ) = 0. On the other hand, if η = 0, then TB= 1 for all xxAand Awe Bget P T(x ,0) = #H2(Q). So tBhe sum over A xA A topologicalclassesofgaugefieldgivesafactor#H2(Q) = #H (Q),canceling 1 the factor that comes from the torsion. Thus, the partition function equals 1, as claimed. Another View For future reference, we can look at this situation in another way. The following remarks do not require assuming that b = b = 0 and hold on 1 2 any three-manifold. Go back to the path integral (1.7) and perform first the integral over A. The integral over A gives us a delta function setting the triv curvature of B to zero. Being flat, B defines an element η ∈ H1(Q;U(1)). B Given that B is flat, the action depends on A only through its characteristic class x , and the sum over x (for fixed B) is the one encountered in the A A P last paragraph, T(x ,η ). As we have just noted, this sum is a multiple A A B of δ(η ). Altogether then, if we perform first the path integral over A, B we get a multiple of δ(B), that is a delta function saying that B must vanish up to a gauge transformation. Since we have found that the partition function is 1, the multiple of the delta function is precisely 1. This result is a three-dimensional analog of an a result used by Rocek and Verlinde [30] in understanding T-duality in two dimensions. The fact that we have just explained – the integral over A equals δ(B) – is the fact that really will be used in section 3. What we gained by the preceeding derivation of triviality of the theory is the not entirely trivial fact that the coefficient of δ(B) is 1. Hamiltonian Triviality To complete the story, it remains to establish Hamiltonian triviality of the theory. We consider first the case of quantization on a Riemann surface Σ of genus 1. We pick two one-cycles C , C giving a basis of H1(Σ;Z). The 1 2 gauge invariance and Gauss law constraints imply that the physical Hilbert space is constructed by quantizing the moduli space of flat connections mod gauge transformations. (For example, see [31].) A flat gauge field A or B is September2,2004 9:14 WSPC/TrimSize: 9.75inx6.5inforProceedings witten 1182 Edward Witten H H determined mod gauge transformations by α = A, β = B. For flat i Ci i Ci gauge fields, the action is Z (cid:18) (cid:19) 1 dβ dβ 2 1 dt α −α . (1.10) 1 2 2π dt dt Thesymplecticstructureisthereforeω = (dα ∧dβ −dα ∧dβ )/2π. Theα 1 2 2 1 i andβ allrangefrom0to2π, sothephasespacevolumeforeachcanonically j conjugate pair α ,β or α ,−β is 2π. The quantization of each pair thus 1 2 2 1 leads to precisely one quantum state. The physical Hilbert space is obtained by tensoring together the spaces made by quantizing the commuting pairs of conjugate variables, and so is also one-dimensional. Moreover, we can carry out the quantization by regarding the wavefunctions as functions of (say) the β . Since the β are mapped to themselves by modular transformations, i i the mapping class group acts in the natural way in this representation: a modular transformation τ that maps β to τ(β ) maps a wavefunction ψ(β ) i i i to ψ(τ(β )). (If we carry out the quantization in a way that is not mani- i festly modular invariant, the action of the modular group is more difficult to describe.) In the β representation, the unique physical state is a delta function i supported at β = 0. (For example, the operator exp(iβ ) shifts α by 2π i 2 1 and hence must act trivially on physical states – so they have their support at β = 0. Similarly exp(iβ ) shifts α by −2π and must act trivially.) 2 1 2 This state is clearly modular-invariant. So we have established Hamiltonian triviality in genus one. More generally, for Σ of genus g, we simply have g independent pairs of variables α , β , each governed by the same action as above. The same i j argumentsgothroughtoshowthatthereisauniquequantumstate, givenin the β representation by a delta function supported at β = 0; the mapping j j class group acts trivially on this state for the same reasons as in genus one. This establishes Hamiltonian triviality in general. 2.3. An Almost Trivial Theory Finally, on a spin manifold Q, let us consider a U(1) gauge field U with the level one-half Chern–Simons action Z 1 IU = Ie(U) = d3x(cid:15)ijkUi∂jUk. (1.11) 4π Q Its partition function Z is a topological invariant of the framed manifold Q U (the framing is needed because of a gravitational anomaly [32]) and changes