20′ five-point function from AdS5× S5 supergravity

We develop new techniques to compute five-point correlation functions from IIB supergravity on AdS5 × S5. Our methods rely entirely on symmetry and general con- sistency conditions, and eschew detailed knowledge of the supergravity effective action. We demonstrate our methods by computing the five-point function of the 20′ operator, which is the superconformal primary of the stress tensor multiplet. We also develop systematic methods to compute the five-point conformal blocks in series expansions. Using the ex- plicit expressions of the conformal blocks, we perform a Euclidean OPE analysis of the 20′ five-point function. We find expected agreement with non-renormalized quantities and also extract new CFT data at strong coupling.


Introduction
Recent years have witnessed a resurgence of activity in studying holographic correlation functions using the AdS/CFT correspondence. An abundance of interesting new results has been obtained by leveraging modern techniques, thanks to an inflow of ideas and technologies from the conformal bootstrap and the scattering amplitude program. The progress is especially evident in the paradigmatic example of 4d N = 4 Super Yang-Mills theory, which is dual to IIB string theory on AdS 5 × S 5 . At the level of two-derivative supergravity, all JHEP10(2019)247 the four-point functions of one-half BPS operators have been obtained at subleading order in 1/N by solving an algebraic bootstrap problem in Mellin space [1,2]. 1 The complete set of tree-level four-point correlators contains a wealth of physical information. However, extracting the data is still highly nontrivial since the double-trace operators in the operator product expansion have degenerate contributions. One therefore needs to solve the associated mixing problem by exploiting the knowledge of all four-point functions. In [6][7][8] machinery for performing a systematic analysis was developed, and the complete anomalous dimension spectrum of double-trace operators has been obtained [9]. The tree-level data in turn allows one to further obtain one-loop results, by using "AdS unitarity methods" [6][7][8][10][11][12]. More precisely, this is achieved by feeding the anomalous dimensions and OPE coefficients into the crossing equation, and focusing on the double-discontinuity [13]. The one-loop four-point functions are rather cumbersome in position space. However, the corresponding Mellin amplitudes look remarkably simple [14], suggesting that the Mellin representation remains a natural language beyond tree level. Furthermore, four-point functions receive higher-derivative corrections from AdS string theory, which are suppressed by inverse powers of the 't Hooft coupling. These stringy effects have recently been studied in [15,16] at both tree level and at one loop, 2 showing an interesting interplay between Mellin amplitudes and flat space scattering amplitudes. These results shed new light on quantum gravity from the CFT perspective, and constitute new precision tests of the AdS/CFT correspondence. While the AdS 5 × S 5 background has attracted most attention, many interesting results have been obtained for other string theory/M-theory backgrounds as well. See [18][19][20][21][22][23][24][25][26][27][28][29] for some recent developments.
In this paper we will initiate a systematic study of five-point functions from tree-level IIB supergravity on AdS 5 × S 5 , as a first step towards extending the above program to arbitrary n-point functions. There are several motivations for considering higher-point correlators. First of all, a very practical reason to study holographic correlators is to extract CFT data at strong coupling. Considering higher-point correlation functions of one-half BPS operators allows us to access new unprotected data not contained in their four-point functions. This becomes especially clear when we look at OPE limits. For example, by taking the OPE limit for one pair of operators in the five-point function, we can obtain four-point functions with one unprotected double-trace operator. These four-point functions encode infinitely many new unprotected three-point functions, which can be extracted after taking another OPE limit. Secondly, previous studies of four-point functions suggested an intricate relation between holographic correlators and scattering amplitudes in flat space. Many aspects of holographic correlators appear to be analogous to the ones in flat space. We would like to further explore these connections and sharpen the analogies, by studying five-point functions. In particular, we will demonstrate how factorization, an important tool for flat space amplitudes, can be used to understand the structure of correlation functions from AdS supergravity. Finally, the study of correlators at strong coupling is motivated by the possibility of discovering unexpected structures.

JHEP10(2019)247
Recently, it was observed that tree-level one-half BPS four-point functions from AdS 5 × S 5 exhibit a hidden ten dimensional conformal symmetry [30]. 3 In terms of this symmetry, four-point functions of different conformal dimensions can all be related to each other. It is interesting to see if such a symmetry also exists in higher-point correlation functions. This may shed some light on its mysterious origin.
Results of five-point functions at strong coupling are scarce. To classify them, it is useful to grade the correlators by ascending extremality E, which is defined by 2E = 4 i=1 k i − k 5 . Here k i are the scaling dimensions of the operators and we have assumed k 5 is the largest. R-symmetry selection rules require E to take integer values. For E = 0 and E = 1, the five-point functions are called extremal and next-to-extremal. In these cases, it is known that the five-point functions are protected by non-renormalization theorems [31][32][33][34][35] and therefore can be obtained from the free theory. If we further increase E by one, the five-point functions are no longer protected and start to become nontrivial. It was argued in [36] that such five-point functions should have factorized structures and can be expressed in terms of lower-point correlators. These near-extremal correlators (E = 0, 1, 2) however are very special, and the derivation of these results (from the bulk side) rely heavily on the fact that extremal couplings vanish. 4 When the extremality is further increased, i.e., E ≥ 3, one encounters the generic case and no such simplification exists. One would imagine that examples of generic five-point functions may have been computed using the traditional algorithm of Witten diagram expansion. However the traditional algorithm is too complicated to be a practical recipe. Implementing this method requires inputting all the precise vertices, which can be in principle obtained from expanding the effective supergravity Lagrangian to the quintic order. Such an expansion is devilishly complicated and has never been attempted in the literature.
In this paper, we will develop new techniques for computing five-point correlators with arbitrary extremality. Since a brute force approach is not viable, our strategy is to avoid the details of the effective Lagrangian as much as possible. We accomplish this by using superconformal symmetry and self-consistency conditions, in the same spirit of [1,2]. Let us sketch the methods and state our main results. For simplicity and concreteness, we will focus on the five-point function of the 20 operator, which is the bottom component of the stress tensor multiplet. 5 Although the methods will be phrased in this particular context, it will be clear that they can be applied to general five-point correlators after some obvious modifications. The starting point of our method is an ansatz which splits into a singular part and a regular part. The singular part includes all possible exchange Witten diagrams, and the regular part contains all possible contact Witten diagrams. The coefficient of each diagram could be computed if the vertices were known, but we will leave them as undetermined coefficients. To solve this ansatz, we use superconformal symmetry and self- 3 See also [28] for an analogous story in AdS3 × S 3 where a hidden six dimensional conformal symmetry emerges in the tree-level supergravity four-point correlators. 4 The vanishing of extremal couplings is a self-consistency condition. This is because extremal contact Witten diagrams are divergent but the effective action should be finite. consistency conditions. The singular part can be uniquely fixed by using factorization in AdS space. Roughly stated, the factorization condition means that the "residue" of the five-point function at an internal bulk-to-bulk propagator is a "product" of three-point functions and four-point functions. This can be stated more precisely in Mellin space [37].
To fix the regular part, we use the chiral algebra twist [38], which predicts that the twisted five-point function is the same as in the free theory. This fixes all but one coefficient in the ansatz, which multiplies a structure insensitive to the chiral algebra constraint. The last coefficient can be determined by further using an independent topological twist [39] which involves the entire SO(6) R-symmetry group. This gives the complete answer to the 20 five-point function from AdS supergravity, and is one of the main results of this paper. We will also discuss a variation of this method which starts with an ansatz in Mellin space. The alternative method avoids certain position space calculations and is more suitable for generalizing to higher-weight five-point functions. The final result is expressed as a Mellin amplitude in (5.6), and takes a very compact form. As a technical development, we have also set up systematic methods to compute five-point conformal blocks in series expansions. This allows us to perform a conformal block decomposition for the 20 fivepoint function and extract new data. For simplicity, we looked at the Euclidean OPE and restricted our attention to the singular and leading regular terms. By taking a single OPE limit, we obtain a new four-point function with three 20 one-half BPS operators and one unprotected double-trace operator. The result can be compactly written as a combination of D-functions, which is presented in (6.23). By taking a double OPE limit, we extract various three-point functions. The protected three-point functions we found are in perfect agreement with their free theory values, which constitute nontrivial consistency checks of our result. We also extract a new unprotected three-point function (6.20) involving one 20 one-half BPS operator and two operators from semi-short multiplets. The unprotected three-and four-point functions give new predictions of N = 4 SYM at strong coupling. We hope these results can one day be compared with the integrability program.
The rest of the paper is organized as follows. In section 2 we discuss the superconformal kinematics of the five-point function. In section 3 we review the Mellin representation and the factorization of Mellin amplitudes. After these preparations, we introduce our position space method in section 4 and compute the five-point function of the 20 operator. In section 5 we point out an alternative approach using Mellin space, which simplifies some calculations in position space. The result for the five-point function is analyzed in section 6, where we perform consistency checks and extract new CFT data. Various technical details are relegated to the appendices.

Superconformal kinematics
The 20 operator O IJ 20 = tr(Φ {I Φ J} ) has protected conformal dimension ∆ = 2 and transforms in the rank-2 symmetric traceless representation of SO(6) R . It is the superconformal primary of the 1/2-BPS multiplet which also contains the R-symmetry current J It is convenient to absorb the R-symmetry indices by contracting with null R-symmetry The contraction automatically projects the operator into the symmetric traceless representation and turns the five-point correlator into a scalar function which depends not only on the spacetime coordinates but also the R-symmetry coordinates It is easy to see that the null vectors can only appear in where the λ i are independent. A basis of R-symmetry structures for G 5 is then given by all the monomials of t ij satisfying the homogeneity condition. There are 22 such terms, which can be dividided according to the lengths of their cycles, as explained in figure 1 The basis vectors are in one-to-one correspondence with the inequivalent Wick contractions of the Φ I fields in the free theory limit. We can parameterize the five-point function such that each R-symmetry structure is multiplied with a function of the spacetime coordinates. However, the full correlator is invariant under permutations of the five external operators. As we can easily check, crossing symmetry permutes separately the R-symmetry structures {A (ijklm) } and {A (ijk)(lm) }. Therefore the various functions multiplying the different Rsymmetry monomials in the same group are interrelated under crossing, and in the end there are only two independent functions of spacetime coordinates in the five-point function. We can further exploit the conformal covariance to extract a kinematic factor from the correlator The kinematic factor takes care of the covariance under conformal transformations, and the five-point correlator becomes a function of the five conformal cross ratios 6 .
(2.8) 6 Similarly, the correlator depends also on five R-symmetry cross ratios, which can be chosen as Figure 1. Each line between points i and j corresponds to a factor of t ij . We can then see that there are two types of structures, A (ijklm) and A (ijk)(lm) , with one and two closed cycles respectively. Note that each cycle is invariant under cyclic permutations and reflection, so the number of independent structures is given by the number of ways to distribute the points into the cycles, modulo those symmetries.
So far we have exploited only the bosonic part of the superconformal group PSU(2, 2|4). The fermionic charges impose further constraints on the correlator. One such constraint comes from the special properties of 1/2-BPS correlation functions under twisting a subalgebra su(1, 1|2) ⊂ psu(2, 2|4), known as the chiral algebra twist [38]. In order to perform the twist, we restrict all five operators to a two-dimensional plane inside R 4 . This allows us to parameterize the positions of these operators in terms of the holomorphic and antiholomorphic coordinates z i ,z i . The t 5 and t 6 components of the six-dimensional null vector t I are set to zero, reducing the null vector into a four-dimensional one denoted as t µ . The vector t µ can be further written as the product of a pair of spinors Rescaling v andv amounts to multiplying t with a number, which does not change the null vector since it is defined modulo rescaling. Both v α andvα therefore have only one degree of freedom, and can be written as When the R-symmetry orientations of the operators are correlated with the positions on the planeȳ the construction of [38] dictates that the twisted five-point function becomes a holomorphic function of the z i only Moreover, the twisted correlator is independent of the marginal coupling, and therefore equal to its free field value g(z i ; y i ) = g free (z i ; y i ) . (2.13)

JHEP10(2019)247
The holomorphic requirement of the twisted correlator imposes nontrivial constraints on the structure of the correlation function G 5 . A similar twist applies to the holomorphic variables.
Another important constraint comes from performing an independent topological twist which probes the full SO(6) R-symmetry group [39]. Unlike in the chiral algebra twist, operators are inserted at generic points x i ∈ R 4 , with position-dependent polarizations (2.14) Such twisted n-point correlation functions preserve two common supercharges. 7 Moreover, the twisted translations and the exactly marginal deformation are exact with respect to the preserved the supercharges. The SO(6) twisted correlators are therefore topological and protected. All in all, the topological twist imposes the constraint that where the two-point function of O 20 is unit normalized, and the twisted five-point function is computed in the free theory. A small comment is in order. For correlation functions of two, three and four 1/2-BPS operators, it is possible to show that the constraints derived from the chiral algebra twist have exhausted the full constraining power of superconformal symmetry. In particular, the requirement of the twisted four-point function being a holomorphic function is equivalent to the superconformal Ward identity [40,41]. On the other hand, the chiral algebra twist leads only to a subset of the full superconformal constraints for correlation functions with five points or more. The SO(6) twist of [39] imposes extra constraints which are not captured by the chiral algebra twist. It is an interesting question for the future to explore the full consequence of superconformal symmetry on five-point and higher-point correlation functions.

Mellin representation and factorization
The goal of this section is to give a brief review on the Mellin representation formalism [42][43][44] and the factorization properties of Mellin amplitudes [37,45]. Mellin amplitudes for scalar operators are defined as an integral transform of the correlation function where the integration variables satisfy the constraint i γ ij = 0, with γ ii = −∆ i , ensuring the correct scaling of the external operators. Correlation functions for operators with spin are more easily expressed using the embedding space formalism for CFTs, see, e.g., [46] for a detailed account. In this formalism, each point in R d is mapped to a null ray through the origin in R d+1,1 , and the action of the conformal group is linearized as the Lorentz rotations in the embedding space R d+1,1 .

JHEP10(2019)247
A primary operator with dimension ∆ and spin J in R d is mapped to a field in R d+1,1 depending on both null rays P and Z and satisfying and These properties guarantee that the operator is symmetric, traceless and transverse. Now we can define the Mellin amplitude for one operator with spin J and n scalar operators as where {a} stands for the set of indices a 1 . . . a J , and {a} i counts the number of times that the index i appears in the set a 1 , . . . a J Moreover, we impose In the case where the spinning operator is conserved, i.e., ∆ = d − 2 + J, the Mellin amplitude has to satisfy one further constraint 2J n a,b=1 The operator product expansion is one of the most important properties of a conformal field theory as it allows to write a product of k local operators at different positions in terms of an infinite sum of local operators

JHEP10(2019)247
where the position y is arbitrary as long as it stays within a sphere that encircles all k local operators. This expansion can be used inside a correlation function, effectively rewriting an n-point correlation function as a sum of products of (k + 1)-and (n − k + 1)-point functions. This property implies that the Mellin amplitude is an analytic function of the Mellin variables γ ij with at most simple poles where the residue Q J m depends on the product of lower-point Mellin amplitudes. Each pole is associated with the contribution of an exchanged primary operator (m = 0), or a descendant (m > 0) with twist τ = (∆ − J) + m. For instance, for an exchanged scalar operator it is given by where we only spelled out the m = 0 since it will be enough for this work. The Mellin amplitudes M L and M R are defined as where λ a = − and analogously for ρ. We also use the notation where a, b label the first k operators while i, j label the remaining n − k operators. The residue Q J m associated with the exchange of an operator with spin depends on mixed Mellin variables where both types of indices appear. For the exchange of a vector operator the residue is given by while for the exchange of a spin 2 operator it is given by The residues for any m and up to spin 2 as well as any spin and m = 0 have been obtained in [37] but they are not needed for this work.  The major difficulty of following this recipe is in obtaining the precise interaction vertices. To compute the five-point functions one needs to expand the supergravity effective action to the quintic order. This is extremely tedious and nonetheless unnecessary as we will see. Instead our plan is to use an "on-shell" approach which works directly with the five-point correlator. By working with the correlator, we can shortcut through the intermediate complexities that one encounters starting from the off-shell effective Lagrangian. Moreover, correlators are constrained by superconformal symmetry, and satisfy nontrivial self-consistency conditions. Among them is factorization, which relates higher-point correlation functions to the lower-point ones. By exploiting symmetries and self-consistency conditions we bootstrap the supergravity correlator and eschew the details of the effective Lagrangian altogether.
Our concrete line of attack comes in three steps. We outline the procedure below.  Step 1: computing the singular part of the correlator using factorization. We divide the five-point function into two parts according to their behaviors in the OPE limits  The various contributing Witten diagrams can be evaluated by generalizing the method of [47], however the coefficient of each diagram is not fixed. To fix these coefficients, the key ingredient of our method is the factorization of the supergravity five-point correlator. For example, we can collect all the exchange Witten diagrams with a scalar exchange in the 12 channel. The Mellin amplitude of this collection of diagrams has a simple pole at γ 12 = 1.  [48]. It requires some work to extract their Mellin amplitudes and we will discuss its details in appendix C. It turns out that factorization uniquely fixes the singular part of the correlator G sing 5 which contains all the double-exchange and single-exchange diagrams.
Step 2: computing the regular part of the correlator by taking the chiral algebra twist. Factorization is agnostic about the regular part of the correlator G reg 5 since the regular part does not contribute to factorization. To fix it, we first write down the most general ansatz for G reg 5 which contains contact Witten diagrams with all R-symmetry structures and up to two derivatives. The upper bound on the number of derivatives comes from the fact that 5d N = 8 gauged supergravity is a two-derivative theory. We then take the chiral algebra twist of the total correlator G sing 5 +G reg 5 . The requirement that the ansatz should reduce to the same holomorphic function as obtained from the free theory imposes nontrivial constraints on the unknown coefficients in G reg 5 . After the dust settles, we find that G reg 5 is fixed up to a single undetermined coefficient, which multiplies the following JHEP10(2019)247 The R-symmetry factor vanishes identically under the chiral algebra twist, and therefore the coefficient λ c remains unfixed at this stage.
Step 3: fixing the remaining coefficient by taking the SO(6) twist. To fix the remaining coefficient, we exploit the SO(6) twist which sets with generic insertion points x i ∈ R 4 . The analysis of [39] dictates that the twisted fivepoint function is topological and protected. Note that the combination (4.2), which vanishes under the chiral algebra twist, does not vanish under the SO(6) twist. This implies λ c can be fixed by comparing with the free theory. Our final results for G sing 5 and G reg 5 are given respectively by (4.36) and (4.49). 9 In the following subsections we spell out the details of the above procedure.

Singular part of the correlator
It is not difficult to see that the diagrams in figures 2 and 3, under permutations of the external labels, exhaust all the possibilities of exchange diagrams allowed by R-symmetry selection rules. Double-exchange diagrams involving one graviton and one graviphoton or two gravitons, for example, are not allowed. 10 The allowed exchange diagrams constitute the singular part of the correlator. In this subsection, we fix the coefficients of these exchange diagrams by using the factorization properties of the five-point function.

Factorization on an internal graviton line
We start from the factorization of the correlator on an internal graviton line. Without loss of generality, we choose the exchanged graviton field to be in the 12 channel. This isolates the diagrams of type 2d and 3c (see figure 5). Because there is a unique solution to the R-symmetry Casimir equation for exchanging the singlet representation in the 12 channel, all the exchange Witten diagrams have the same R-symmetry polynomial. We can therefore forget about the R-symmetry polynomial in intermediate steps and only multiply it back in the end.
Let us denote the double-exchange diagram 2d as W ϕ [12] ,s [34] . The other two doubleexchange diagrams can be obtained by permuting the external labels 3, 4, 5, and are denoted as W ϕ [12] ,s [35] and W ϕ [12] ,s [45] . In W ϕ [12] ,s [34] , the graviton field is minimally coupled to the scalar field, i.e., the cubic vertex has the form where T µν is the energy-stress tensor By using the AdS Feynman rules, the diagram W ϕ [12] ,s [34] takes the following form W ϕ [12] ,s [34] ρσ (x 5 ; y, w) are obtained from T µν by replacing the scalar field with the scalar bulk-to-bulk and bulk-to-boundary propagators as prescribed by the diagram The evaluation of this diagram has an important subtlety: the source coupled to one end of the graviton bulk-to-bulk propagator is not gauge invariant. In fact, by using the equation of motion identity of the bulk-to-bulk propagator we find that the source I ρσ has a nonzero divergence

JHEP10(2019)247
This seems to create problems because gauge fields can only couple to conserved sources, and also renders the method of [47] inapplicable. However we should notice that gauge invariance is not necessarily achieved by an individual diagram, but only the sum of diagrams. 11 To fix this problem, we must also include the single-exchange diagrams 3c. The sum of all double-exchange diagrams introduces a source with divergence The minimal choice to cancel this divergence is to introduce a single-exchange diagram which is derived from a quartic coupling of the form Here c IJK is an R-symmetry invariant tensor that makes the vertex a singlet. Denoting the single-exchange diagram as W ϕ [12] , we have 14) It is easy to verify that the sum of diagrams [12] ,s [34] + W ϕ [12] ,s [35] + W ϕ [12] ,s [45] is the overall R-symmetry factor defined in appendix B and λ ϕ is an overall coefficient.
The double-exchange diagrams and the single-exchange diagram can be evaluated using the method of [47], pretending the coupling to the graviton is conserved in each diagram. This prescription can be justified since the total coupling in (4.15) is conserved and the extra contributions from each non-vanishing divergence cancel in the end. Details of the evaluation are discussed in appendix A and the corresponding results of the exchange diagrams are given by (A.18) and (A. 19).
Having obtained the gauge invariant combination (4.15), it is straightforward to go to Mellin space and check that Mellin factorization on the internal graviton line is satisfied. We find the residue of the Mellin amplitude at γ 12 = 1 is correctly related to the Mellin . A more careful analysis of the normalizations could also fix λ ϕ , but we will leave it undetermined for the moment and fix it when we consider the factorization on an internal scalar line. We therefore have fixed all the exchange diagrams involving a graviton internal line up to an overall normalization. 11 More precisely, these are the diagrams with a graviton exchange in the 12 channel. We can require gauge invariance of this collection of diagrams because the factorization on the internal graviton line gives a physical three-point function and a physical four-point function, which are gauge invariant.

Factorization on an internal graviphoton line
Let us now proceed to the factorization on an internal graviphoton line. We first focus on the double-exchange diagrams, which turn out to consist of only two types.
One type of double-exchange diagrams is 2c which involves two internal graviphotons. The graviphoton couples to the scalars via the minimal coupling The graviphoton-graviphoton-scalar coupling is given by where F a µν is the field strength of the graviphoton field and d I,ab is a tensor that makes the vertex a singlet. Denoting 2c as W V [12] ,V [34] , we have W V [12] ,V [34] [12] ,V [45] (4.21) can be obtained from the first term via permuting the external labels.
The other type of double-exchange Witten diagram is 2b. We denote 2b as W V [12] ,s [34] . The diagram reads W V [12] ,s [34] where the source is It is easy to check that the source is not conserved therefore the diagram is not gauge invariant. However when we multiply the diagram with the R-symmetry polynomial R and sum over all the permutations of 3, 4, 5, the combination [12] ,s [45] (4.25) is gauge invariant, since  The diagram is easy to evaluate and the result is given by (A.17). We can check if these two gauge invariant combinations of diagrams can already reproduce the factorization. It turns out that gives the correct answer. The overall normalization could also be determined from factorization but we will defer it until later. From the factorization analysis, we can conclude that no graviphoton single-exchange diagrams of 3b appear. The factorization of the five-point function on an internal graviphoton line is illustrated in figure 6.

JHEP10(2019)247
The double-exchange diagram W s [12] ,s [34] is constructed from the cubic vertex s I s J s K c IJK , and is given by the integral W s [12] ,s [34] This diagram can be easily evaluated using the method of appendix A and the result reads W s [12] ,s [34] The diagram W s [12] ,s [34] is associated with an R-symmetry factor R , which can be found in appendix B.
There are also scalar single-exchange diagrams 3a, which can have zero or two derivatives in the quartic coupling. There cannot be more than two derivatives because the 5d N = 8 supergravity contains only two derivatives. The zero-derivative single-exchange diagram is denoted by W s [12] 0-der and evaluates to W s [12] 0-der = D 11222 4x 2 12 . (4.31) For the two-derivative type, we have a basis of diagrams where the pair of derivatives are on {3, 4}, {3, 5} and {4, 5}. These diagrams are denoted respectively by W s [12] , , and are related to each other by permuting the external labels 3, 4, 5. The diagram W s [12] , (5) 2-der reads W s [12] ,(5) The scalar single-exchange diagrams can have 6 independent R-symmetry structures (which can be seen by solving the 12 channel R-symmetry Casimir equation alone). We can pick a basis of solutions as (see appendix B for the definition of A i , D i , E i , H and I) Let us now collect all the exchange diagrams containing a scalar internal line in the 12 channel (figure 7). We have the following ansatz [12] ,s [34] [12] ,s [35] [12] ,s [45] [12] ,V [45] [12] ,ϕ [34] [12] ,ϕ [35] [12] ,ϕ [45] [12] 0-der and all the other coefficients are zero.
We have now computed the singular part of the five-point correlation function G sing 5 . The result is the following and G reg 5 , and their contribution can be combined into the ansatz for the latter. We have set these free parameters to zero without loss of generality.

Regular part of the correlator
We now solve the regular part of the ansatz. The regular part G reg 5 consists only of contact Witten diagrams with zero and two derivatives To fix the coefficients, we first use the chiral algebra twist as was reviewed in section 2. The five operators are now restricted on a plane, parameterized by the 2d coordinates z i , z i . The R-symmetry polarizations are restricted to rotate under only an SO(4) subgroup of SO(6) R , and the null vectors are parameterized as The chiral algebra twist amounts to settingȳ i =z i , and the non-renormalization of chiral algebra requires that where the G free 5 is the correlator computed in the free theory and is given by

JHEP10(2019)247
Note that the r.h.s. is a simple rational function of the holomorphic coordinates. On the other hand, the l.h.s. comes from a complicated sum of D-functions where each D-function has transcendental degree 2 and is far from being a rational function. This means that the unknown coefficients in the l.h.s. must be fine-tuned to reproduce a rational function, therefore imposing strong constraints on the unknown coefficients. The condition (4.41) is not yet in a form that is ready for use. Extracting the constraints on the coefficients from (4.41) still requires some nontrivial work. Our strategy is to find a basis to decompose the l.h.s. . Using the differential recursion relations in appendix D, all these D-functions can be related to the basic D-function D 11112 (and its permutations) by taking derivatives. The function D 11112 can be evaluated in closed form in terms of one-loop scalar box diagrams [49,50] Here η i5 , N 5 are rational functions of the conformal cross ratios, andÎ (i) 4 are one-loop box diagrams (also denoted as Φ in appendix D) where the i th point is omitted. When the insertion points x i are generic, i.e., not lying on a two-dimensional plane, the five box diagramsÎ (i) 4 , i = 1, . . . , 5 are independent. Taking derivatives with respect to x 2 ij , one can obtain D-functions of higher weights. Since the box diagrams obey differential recursion relations (D.13), one finds that all the D-functions can be uniquely decomposed into a basis spanned byÎ (i) 4 , logarithms and 1, with rational coefficient functions. Apparently, the ansatz G sing 5 + G reg 5 also admits such a unique decomposition under this basis with rational coefficient functions. However, to use the chiral algebra twist condition (4.41), we need to further restrict the five insertions on a plane. This gives rise to subtleties which require some extra care. The problem is that some elements of the basis develop relations. For example, the five one-loop box diagrams are now linearly dependent 13 which follows from the identity [51] (4.45) After properly taking care of the relations among the basis vectors, we find the following basis of independent functions 13 However the denominator N5 also becomes zero at the same rate so D11112 remains finite (and nonzero) when all the points are put on plane.

JHEP10(2019)247
where z and w are the complex coordinates of the two insertion points not fixed by conformal symmetry, withz,w being their complex conjugates. They are related to the cross ratios V i defined in (2.8) via Decomposing the supergravity ansatz into this basis gives coefficient functions which are rational in the cross ratios. Equating the coefficients in (4.41) gives a set of linear equations for the unknown coefficients. The constraints turn out to be remarkably constraining. We find that all the twoderivative vertices vanish λ Moreover, all but one of the zero-derivative coefficients is fixed, yielding Chiral algebra is incapable of fixing λ c because the multiplied R-symmetry polynomial vanishes automatically under twisting.
To determine the remaining coefficient, we use the SO(6) twist as we reviewed in section 2. This uniquely fixes the coefficient to be Before we end this section, let us make a comment about the contact Witten diagrams which contribute to G reg 5 . These five-point contact interactions in fact are not intrinsic in the sense that they can be absorbed into G sing 5 by redefining certain vertices of the exchange Witten diagrams. We have already noticed such an ambiguity in footnote 12. More precisely, we can rewrite G reg 5 in such a way that it can be absorbed in the scalar single-exchange Witten diagrams W scalar 12 tot,2 (and all other diagrams by permutations) while keeping the quartic vertices in 3a symmetric and with no more than two derivatives. To see this, let us define the scalar single-exchange Witten diagrams W s [12] ,(i) 2-der for which the two derivatives act on the external leg i = 3, 4, 5, and the internal leg I. By using the equation of motion identities of the propagators and integration by parts, one can show W s [12] ,(i) 2-der = W s [12] ,(i) 2-der These identities can be used to make the two ways of distributing the derivatives, i.e., (3,4), (4,5), (3,5) and (3, I), (4, I), (5, I), appear symmetrically in the solution at the cost of generating some new D 22222 . The total collection of D 22222 with different R-symmetry structures can then be reinterpreted as scalar single-exchange Witten diagrams with derivatives on the same leg. Let us note that when the derivatives are on the same external leg, the diagram is simply − 4W s [12] 0-der , s [12] 0-der − D 22222 . (4.53) The latter term allows us to absorb all the D 22222 into G sing 5 .
5 An alternative approach using Mellin space This motivates us to write down the following ansatz for the five-point Mellin amplitude which has the structure of a sum of simultaneous poles, single poles and a regular piece. The residues P ij,i j also polynomials in t ij , but we will suppress the R-symmetry dependence for the moment and focus on the γ ml dependence. We will assume that P ij,i j 2 are degree 2 polynomials of γ ml , while both P ij 1 and P 0 are of degree 1. The degrees of these polynomials can be justified as follows. For the residues P ij,i j 2 , the degree simply follows from the counting of the total number of derivatives in the cubic vertices. Just as in flat space, the residue has degree L if the vertices contain in total 2L derivatives. For the double-exchange diagrams involving one stress tensor, it follows from R-symmetry selection rule that the other exchange field can only be the scalar field. Such diagrams have only four derivatives in all vertices. In the vector-vector double-exchange diagrams, the total number of derivatives is also four. For the vector-scalar and scalar-scalar double-exchange diagrams, the total numbers of derivatives are two and zero. 14 This leads us to conclude that P ij,i j 2 are degree 2 polynomials of γ ml . It is tempting to apply the same argument on the single pole residues P ij 1 . However the counting holds only for the vector and scalar single-exchange diagrams where at most two derivatives are present. For the graviton single-exchange diagrams, the residue appears to have degree 2 since there could be in total four derivatives. The leading degree 2 terms would correspond to a constant piece in the term M ij 4,T of the factorization formula (5.1). 15 However, a closer look at M ij 4,T reveals that the contribution of the constant term to (5.1) vanishes after the summation. Therefore, the single pole residues P ij 1 are degree 1 polynomials. Finally, P 0 receives contribution from the five-point contact vertices. Since the gauged supergravity contains at most two derivatives, it follows that the degree of P 0 is at most 1.
Let us now be more explicit about the R-symmetry dependence. We write the residues as P ij,i j 2 (γ) =  14 Note the derivatives in the scalar cubic coupling have been removed by nonlinear redefinition of the scalar fields [52]. 15 The Mellin amplitude M ij 4,T contains simple poles with constant residues which are due to the exchange of scalar fields. There is also an additional constant piece which is due to the quartic interactions. As is shown in appendix C, this constant term is completely determined by the singular terms via transversality.

JHEP10(2019)247
to the single Casimir equation, of which a basis can be obtained from R is any other compatible channel and I 1 is over all possible R-symmetry representations. Note that the sums over the representations I 2 , I 2 in the simultaneous pole residues are restricted to 20 , 15 and 1, in correspondence to the R-symmetry representations of the exchanged single-trace fields. Similarly, the sums over I 1 in the single pole residues are also restricted to 20 , 15 and 1, which correspond to the single-exchange Witten diagrams. On the other hand, in the regular part we sum over all 22 R-symmetry structures A I 0 defined in (2.5), since all of them can appear.
The Mellin amplitude ansatz M ansatz is further constrained by three other consistency conditions. First, the ansatz M ansatz should be crossing symmetric. In implementing this constraint, it is important to take into account the linear constraints satisfied by γ ij which leaves only five independent variables. Second, the correlator needs to satisfy the chiral algebra condition (4.41). Unfortunately this condition is not straightforward to implement in Mellin space. This is essentially because the independent Mandelstam variables γ ij are dual to the independent conformal cross ratios for generic configurations. To perform the chiral algebra twist, one needs to restrict the five insertion points on a two-dimensional plane. This reduces the number of independent cross ratios to four, while the Mellin representation is oblivious to it. Therefore our strategy is to rewrite the Mellin amplitude ansatz as a sum of D-functions and then implement the chiral algebra twist in position space. However we should note that the rewriting is not unique. Different expressions in terms of D-functions with the same Mellin amplitude may differ in position space by a rational function or a logarithmic term. 16 On the other hand, the part with transcendental degree 2 does not suffer from such ambiguities. Therefore, we only use the chiral algebra constraints from the coefficient functions of the box diagrams. 17 Finally, the correlator satisfies the condition (2.15) imposed by the SO(6) twist. We also implement this condition in position space and focus on the pieces with transcendental degree 2.
Solving the above constraints fixes the ansatz up to an overall normalization. The leftover degree of freedom is expected because the twisted five-point functions in (4.41) and (2.15) are rational, and do not contribute to the box diagram coefficients. The conditions from the chiral algebra twist and the SO(6) twist are therefore homogenous and do not allow us to determine the overall coefficient. We can fix the remaining coefficient by, for example, looking at the factorization of the five-point Mellin amplitude on a scalar exchange.
The final result for the Mellin amplitude takes the following form where M sim are the simultaneous poles

OPE analysis
In this section we analyze the short-distance behavior of the supergravity five-point function and use the Euclidean OPE to extract new CFT data of strongly coupled N = 4 SYM.
To simplify the analysis, we restrict our attention to only the singular and the leading non-singular behavior of the correlator. A complete analysis of the supergravity five-point function is left to the future. In section 6.1 we discuss the kinematics of the Euclidean OPE. We discuss the decomposition of five-point functions in conformal blocks, and also explain how to take a single OPE to obtain four-point functions. This part can be read independently, and applies to generic CFTs with and without supersymmetry. In section 6.2 we introduce all the operators up to dimension four that contribute to the OPE of two 20 operators. The reader interested solely in the results of the OPE analysis might skip directly to section 6.3, where we present the new data obtained.

Euclidean OPE limit
The information of the CFT is encoded in five-point functions according to the principle of operator product expansion. By leveraging this expansion in different ways, we can extract various information from the five-point functions.
To extract the CFT data, it is most straightforward to use OPE in two different channels. The five-point function essentially becomes a sum of products of three-point functions, analogous to the case of four-point functions. More precisely, we send the points x 1 , x 3 and x 4 to 0, 1 and ∞ respectively, by using the global conformal symmetry. The Euclidean double coincidence limit (in the 12, 35 channel) is then obtained by taking both

JHEP10(2019)247
In Euclidean kinematics we have two small parameters, s 1 and s 2 , and three angle variables ξ 1 , ξ 2 and ξ 3 defined by In these variables, the cross ratios defined in (2.8) become 18 Operator product expansion dictates that the five-point function can be expanded in terms of conformal blocks The five-point conformal block G p k,k (s i , ξ i ) encodes all the contribution of the exchanged primaries O k , O k , as well as their conformal descendants. The label p is associated with the different structures of a three-point function with two spinning operators. We will refrain from giving here the explicit expressions for the conformal blocks. They will be given in appendix E, where we discuss how to compute them as series expansions in both s 1 and s 2 .
Similarly, we can apply a single OPE and obtain information about the full four-point functions. To achieve this let us consider the OPE of two external scalar operators where the function F (12k) (x, ∂ y , D z ) and the derivative D z are defined in appendix E. The exact coefficients in this expansion can be fixed by imposing the consistency of the OPE with the conformal structure of the three-point function. Applying F (ijk) on the spinning four-point function gives its contribution to the single OPE of the five-point function To proceed, we show how we can distinguish operators with different spins. Four-point functions with an external leg of spin J have J +1 conformal structures, where the coefficient 18 Note that if all the five points are restricted to the plane then only four of the five cross ratios are independent, as ξ3 = − cos(θ1 + θ2), and V5 simplifies to

JHEP10(2019)247
of each structure is a function of the cross ratios lim (6.8) For simplicity we set x 2 13 = 1 and rewrote the cross ratios defined in (2.8) with complex variables The derivatives from the OPE expansion (6.7) act not only on the conformal structures, but also on its coefficients α (p) l (w,w). This effect is important for the subleading terms in the expansion of s 1 . The leading spin one and spin two contributions to the five-point function are therefore given by where we introduce the new angle variable The dependence on ξ allows us to disentangle the different spinning four-point tensor structures in the single OPE of the scalar five-point function, just like ξ 3 parametrizes the contribution of different three-point tensor structures in the double OPE limit. Finally, when glueing the three-and four-point spinning correlators of 4d N = 4 SYM into fivepoint function contributions, we also need to perform the contractions of the R-symmetry structures. The details of this procedure can be found in appendix B.

Low-lying operators
From the representation theory of the 4d N = 4 superconformal algebra, we know that the tensor product of two stress tensor multiplets takes the following schematic form [41,53]  is the R-symmetry Dykin label of the super primary and (j,j) are the Lorentz spins. We will also use J = 2j when j =j, as the spin of the superconformal primary. The multiplets B and C are short (semi-short) multiplets satisfying (b,b) and (c,c) type shortening conditions, while A are generic long multiplets which do not satisfy any shortening condition. We refer the reader to [54] for details of the classification of superconformal multiplets. The multiplets JHEP10(2019)247 [0,0,0],(j,j) contain higher-spin currents and therefore should not appear in an interacting CFT. Moreover, the restriction to the singular and leading regular part of the Euclidean OPE leaves us with only a handful of contributing operators. Below we list the operators which appear in the OPE at the supergravity limit, and enumerate their properties. The operators responsible for the singular contributions are: • The operator O 20 from the 1/2-BPS multiplet B where Z and X are two complex scalar fields defined as Z = Φ 1 + iΦ 2 , X = Φ 3 + iΦ 4 . Note that Wick contraction of the two scalar fields vanishes, so that the operator Q is completely traceless. The OPE coefficients of 1/4-BPS and 1/2-BPS operators are also protected [57].

JHEP10(2019)247
• The bottom component of the multiplet C [1,0,1],(0,0) is also a scalar operator with ∆ = 4, but it has R = [1, 0, 1]. The total parity of R-symmetry and spacetime spins of this operator is therefore odd, which forbids it to appear in the OPE.

Extracting CFT data
We can already make some qualitative predictions about the result after taking the Euclidean OPE. For example, the only unprotected operator appearing in the OPE is the super primary of the long multiplet, which is an R-symmetry singlet. Therefore none of the other representations should contribute to the logarithmic singularities which are associated with anomalous dimensions. Even without decomposing into conformal blocks, an R-symmetry projection of the correlator expanded to order O(s 4 1 s 4 2 ) confirms this to be a feature of our supergravity five-point function.
Let us now consider the double OPE limit in more detail. In the following we always write the OPE coefficients for normalized operators, and we strip off the R-symmetry structures which are defined explicitly in appendix B. This notation follows naturally from the decomposition of the five-point function into the R-symmetry polynomials defined in (B.20). 19 To begin, let us first project the five-point function into the channel in the 20 representation and with ∆ = 2. This simply corresponds to the intermediate operator being the operator O 20 . We obtain several OPE coefficients with two chiral primaries, which were known previously from the analysis of the 20 supergravity four-point function [59] Note that our results have different normalizations as we use the tensor structures defined in appendix B. Except for C O 20 O 20 A , the three-point functions above are protected, and so they coincide with their free field theory values. 20 19 By contrast if we want to directly compute the normalized OPE coefficients in the free theory, we need to evaluate both two-and three-point functions where the former set the normalizations. 20 Three-point functions of half-BPS operators are known to be independent of the coupling, thanks to the non-renormalization theorems [52,[60][61][62][63][64][65][66][67], while three-point functions mixing half-and quarter-BPS operators were shown to be protected in [57]. Meanwhile, the non-renormalization of CO 20 O 20 C was observed in [59], and proved in [68] using superspace techniques.

JHEP10(2019)247
Focusing now on intermediate operators of dimension 4 in both channels, we are able to extract three-point functions which could not be obtained from the four-point function of single-trace operators. In particular, we find the following OPE coefficients are known to be protected. We reproduced these three-point functions from supergravity calculations, which gives nontrivial checks of our results. Moreover, the precise match of C O 20 CQ with the free theory value also strongly indicates that the three-point function is protected, supporting the claim from [68] using superspace arguments.
We also extract the OPE coefficient of one O 20 with two C operators at strong coupling, which reads (6.20) We find that this OPE coefficient does not match the free field theory computation, indicating that this type of three-point function is unprotected. 21 Further support for our claim can be obtained from perturbation theory. In [69] the authors obtained the fivepoint function at one loop, and a decomposition in conformal blocks reveals that the OPE coefficient receives a one-loop correction 22 Finally, we consider the singlet and 20 R-symmetry channels, from which we derive a new OPE coefficient involving the unprotected operator A and a semi-short operator C While the machinery developed in appendix E makes it convenient to directly extract the CFT data, there is still much to gain by performing just one single OPE. It allows us to obtain the complete four-point functions from the five-point function. To start with, we reproduce the known four-point functions of single-trace operators. We found that the projection of the singular part of the correlator on the [0, 2, 0] channel is exacly reproduced by 21 At first sight our result seems to contradict the protected nature of the chiral algebra. However, unlike the B-type multiplets where the Schur operators are the super primary, Schur operators in the Ctype are superconformal descendants. We strongly suspect that the three-point functions for one B-type multiplet and two C-type multiplets have more than one superstructure in superspace. The protected chiral algebra three-point function and O 20 CC are in different superstructures which are unrelated by the action of supercharges. The non-renormalization theorem applies only to the former case. We thank Carlo Meneghelli for discussions on this point. 22 The weak coupling analysis is more subtle, as there could be more operators appearing in the OPE.
Much new information is encoded in this correlator, with the OPE coefficient of (6.22) being just an example of the type of data that can be extracted. Also note that there are many possible rewritings of the correlator (6.23) in terms of D-functions, and we have only presented the simplest expression. It is also possible to write an expression which requires only D-functions of total conformal dimension 10.

Discussion and outlook
In this paper, we developed new systematic methods to compute five-point functions from AdS 5 × S 5 IIB supergravity. We also obtained five-point conformal blocks in series expansions, which allowed us to perform conformal block decompositions for five-point correlators. As a concrete example, we computed the five-point function of the 20 operator. We performed a number of consistency checks on the 20 five-point function and extracted new CFT data at strong coupling. There are many directions which one can pursue in the future.
• First of all, an immediate interesting extension is to apply our methods to more general five-point functions. As the complexity of the correlators grows with the extremality, the best starting point is correlators with the same extremality as the 20 five-point function. These correlators should have very similar structures, which is particularly manifest in Mellin space. Work in this direction is in progress and we hope to report the results in the near future.
• Second, we would like to better understand the general structure of the five-point correlation functions dictated by superconformal symmetry. For four-point functions, superconformal constraints boil down to the partial non-renormalization theorem of [40]. This theorem reduces the correlators to a free part and a "quantum correction" part, which has a much simpler form than the full correlator. For five-point functions the pressing issue is to find and solve the full set of constraints from superconformal symmetry, and the solution will constitute the five-point analogue of the "partial non-renormalization theorem". Such a solution will give us a more compact way to write the five-point function.

JHEP10(2019)247
• Relatedly, it has recently been observed that the correction part in four-point functions exhibits a hidden ten dimensional conformal symmetry [30]. Using this hidden symmetry, one can lift the lowest-weight four-point function into a generating function. Establishing the five-point "partial non-renormalization theorem" will be extremely useful for identifying the action of the hidden symmetry at the level of fivepoint functions. It should then also be possible to write down a generating function which gives five-point functions of arbitrary conformal dimensions.
• From our analysis, it is clear that there are close analogies between holographic correlators and flat space scattering amplitudes. For example, factorization in Mellin space played a crucial role in our position space approach of computing the five-point function. We also showed that the 20 five-point function has no intrinsic five-point contact interaction. This seems to suggest certain "constructibility" of the holographic correlators. It would be extremely interesting to develop such constructive approaches further and extend them to higher points, perhaps in the form of Mellin recursion relations similar to the famous BCFW relation [70].
• One aspect which we have not considered in detail is the flat space limit. We would like to examine this limit more carefully in the future. The flat space limit will also be important when we consider higher-derivative (stringy) corrections to the five-point functions, as has been emphasized in the four-point function case by, e.g., [14-17, 23, 25].
• Finally, the technology developed in this paper can be readily applied to eleven dimensional supergravity on AdS 7 ×S 4 . The chiral algebra in six dimensions [71] places strong constrains on the five-point functions. However, it is not clear if a twist similar to the SO(6) twist of [39] exists for the (2, 0) theories. It may be necessary to resort to the flat space limit, which gives extra constraints on contact interactions. It would be interesting to compute five-point correlators for this background, and extract new information about the (2, 0) theory in six dimensions.

A Integrating out an internal line
In this appendix we extend the method of [47] to higher-point Witten diagrams with more than one internal line, and evaluate the various diagrams that appear in section 4. The key point is that integrating out an internal line replaces the integrated cubic vertex by a sum of contact vertices. When the quantum numbers are fine-tuned to satisfy certain conditions (such as in N = 4 SYM and the 6d (2, 0) theory), the sum in the contact vertices truncates to finitely many terms. Repeated use of the vertex identities then allows us to write an exchange Witten diagram in terms of a finite sum of contact diagrams.

A.1 Vertex identities
In [47] the consequence of integrating out a bulk-to-bulk propagator for a four-point exchange Witten diagram was worked out. The upshot is that the exchange Witten diagram can be expressed in terms of a sum of four-point contact diagrams. For our purpose, we want to extract from their result the vertex identities that relate an integrated cubic vertex to a sum of contact vertices.

A.1.1 Scalar exchange
Let us start with the scalar internal line. The integrated cubic interaction is Using the result in [47], we can express this integral as with ∆ 1 + ∆ 2 − ∆ being a positive even integer.

A.1.2 Graviphoton exchange
We now consider the integral involving the exchange of a vector field of general dimension ∆. When ∆ = d − 1, the vector field is a massless gauge field and couples to a conserved current. Denoting ∆ 1 = ∆ 2 = ∆ ext , we will consider the coupling of the vector field to a conserved current

JHEP10(2019)247
where G ∆,1,µν BB (z, y) is the vector bulk-to-bulk propagator. This integral can be evaluated as a sum of contact vertices .
The truncation requires that k max −k min is a non-negative integer. Notice that in evaluating the cubic integral, vanishing divergence of the source is not required. Therefore this result holds even when the source coupled to I µ vector is not conserved.

A.1.3 Graviton exchange
Finally we consider the cubic integral involving a graviton field. Let ∆ 1 = ∆ 2 = ∆ ext , the cubic integral is Using the result of [47], we find that this integral reduces to the following sum of contact vertices .

JHEP10(2019)247
For the above expression to be valid, I µν graviton must be coupled to a conserved current. This is because in the derivation of [47] total derivative terms in I µν graviton are assumed to drop out, which is consistent only when coupled to divergence-less sources. On the other hand, when I µν graviton is coupled to a non-conserved source, there is a nonzero contribution from the total derivative terms in I µν graviton . The contribution of these terms cannot be determined using the techniques of [47].

A.2 Five-point exchange Witten diagrams
Using the vertex identities, we can evaluate the exchange diagrams that we encountered in section 4. We record here their explicit expressions.

2-der
The scalar single-exchange diagram with a two-derivative quartic vertex in which the two derivatives are on the external legs 3 and 4 is defined by W s [12] ,(5) Using the identity we find W s [12] , (

JHEP10(2019)247
A.2.4 W V [12] ,V [34] The double-exchange diagram W V [12] ,V [34] is defined by (4.19). Using the vertex identities, we have W V [12] ,V [34] A.2.5 W V [12] ,s [34] The double-exchange diagram W V [12] ,s [34] is defined in (4.22) and it evaluates to A.2.6 W ϕ [12] ,s [34] The definition of the double-exchange diagram W ϕ [12] ,s [34] is given by (4.6). As we commented before, the vertex identity in section A.1.3 does not hold because the source (4.9) is not conserved. A naive application of the identities leads to a wrong answer since the dropped total derivative terms have nonzero contributions. On the other hand, the total coupling to the graviton field is conserved when we sum up all the diagrams, and the extra contributions due to the ignored total derivative terms will vanish in the sum. Therefore it does not matter that we use the vertex identities of section A.1.3 to evaluate the diagrams so long as all the diagrams are added up correctly at the end of the day. With this caveat, we find that W ϕ [12] ,s [34] where we used • = to remind us that this expression only makes sense in the sum of all diagrams.
A.2.7 W ϕ [12] The single-exchange diagram W ϕ [12] is defined by (4.14). The coupling to the graviton is also non-conserved, but we will evaluate it with the same caveat for W ϕ [12] ,s [34] . Using the vertex identities we have

JHEP10(2019)247 B R-symmetry polynomials
An R-symmetry basis can be obtained by solving the two-particle Casimir equations in two compatible channels. In terms of the null vectors, the SO(6) R generators take the form L (i) (B.1) The two-particle Casimir operator, say for 1 and 2, is constructed from L When acting on the five-point correlator, which is a polynomial of t ij = t i · t j , the twoparticle Casimir C (1,2) can be written as where d R = 6 for SO (6) and D x ≡ x ∂ ∂x . Other two-particle Casimir operators C (i,j) are similarly defined and can be obtained from C (1,2) by permuting the labels.
We now consider the solution to the following Casimir equations is the R-symmetry structure associated with the process. R-symmetry selection rules at the vertices with a, b and c, d require 0 ≤ q ≤ p ≤ 2 and 0 ≤ q ≤ p ≤ 2, and the solutions are further restricted by the selection rule at the vertex with the remaining external point e. There are in total 22 solutions to the equations, which are in correspondence with the 22 R-symmetry structures and form a complete basis. Notice that when the two intermediate representations are [1,2,1] there are two solutions to the Casimir equations. This is because [1,2,1] appears twice in the tensor product of [1,2,1] with [0, 2, 0].
The OPE coefficients are meaningfully defined only when we set the conventions for the three-point SU(4) tensor structures. In the context of a double OPE analysis it is useful to fix the normalization of the polynomials R (p,q),(p ,q ) ab|cd such that they correspond to the product of three such structures. In order to do so, let us first introduce the tensors JHEP10(2019)247 where i, j, k, l are SO(6) vector indices, and J parametrizes the degrees of freedom of the representation. C 2 2 is completely symmetric, C 2 1 is symmetric in both i ↔ j and k ↔ l but antisymmetric in the exchange (ij) ↔ (kl), while C 2 0 is antisymmetric in both i ↔ k and j ↔ l, symmetric in the exchange (ik) ↔ (jl) and also obeys ijklmn (C 2 0 ) J klmn = 0. Similarly, for p = 1 we encode representations by traceless orthonormal tensors of rank 2, which we write as where C 1 1 is symmetric and C 1 0 is antisymmetric. We can now set our definitions for the tensor structures arising in all relevant threepoint functions. When two of the representations are the 20 , all possible tensor contractions can be related to the following three-point structures If only one of the representations is the 20 , then the independent three-point structures can be chosen as (B.14)

JHEP10(2019)247
Note that when both (p, q) and (p , q ) are equal to (2,1) there are two possible tensor structures, which reflects the fact that there are two independent singlets in the tensor product [1,2,1] It is natural to write the basis of four-point structures in the (ab) channel as the product of three-point structures introduced above which can be written as polynomials in t ij by performing the following contraction Analogously, the natural basis of five-point tensor structures in the double OPE analysis in (ab) and (cd) channels is then given by the following product of three three-point structures With these definitions, the normalization of the polynomials R (p,q),(p ,q ) ab|cd is fixed by requiring that Performing the contractions on the right-hand side we obtain the following basis suitable where we introduced the following short-hand notation for the monomials  When performing the single OPE, it is useful to know how the structures of the fourpoint function contract with the three-point structure, so that we can recognize their contribution to the five-point function. We will now see that we can do this easily with the It is useful to rewrite this tensor in terms of the variables t i , which we do as follows It is possible to recognize part of the definition for the four-point polynomials in this expression, from which we obtain ab|ce . (B.24)

C Spinning correlators
In order to use factorization we must obtain the Mellin representation for correlators of the chiral operator O 20 with a single insertion of the R-current J µ or the stress-tensor T µν , whose AdS duals are the graviphoton V µ and graviton ϕ µν , respectively. Threepoint functions of these fields are protected, and their Mellin transform is a constant, but factorization requires also the knowledge of the Mellin representation for the following four-point functions It is useful to think of these spinning operators as different components of the superfield which depends only on four chiral and four antichiral Grassmann variables, due to a shortening condition. Therefore, the four-point function depends on 16 chiral and 16 antichiral variables, which exactly matches the number of supercharges in N = 4 SYM. In [48] the superconformal symmetry was used to relate all elements of (C.4) to the lowest component, i.e., the four-point function of chiral primaries. The chiral primary four-point function can be split into a free part and an "anomalous" part which depends on the coupling manifest in (C.8) that the result can be expressed in terms of X µ 124 and X µ 134 alone, so the fact that (C.12) and (C.13) have this form provides a good consistency check of the result. Furthermore, γ(u, v) is such that the expression for the stress-tensor is traceless, and we observe also that, as expected, both correlators satisfy the equations for conserved currents 14) The expressions for the spinning four-point functions simplify greatly when we use the spinning Mellin formalism [37]. Here we can ignore the free piece because they are rational functions which do not contribute to the Mellin amplitude [2]. The Mellin amplitudes therefore come exclusively from the anomalous piece computed above. By inverting (3.4), we can extract the Mellin amplitudes by performing the following Mellin transforms where γ i are fixed in (3.6) and we define the remaining Mellin variables as Since the functions α (i) (u, v) and β (i,j) (u, v) are linear combinations of Φ(u, v) and its derivatives, it is useful to consider the relation . (C.18) Putting equations (C. 15), (C.17) and (C.18) together, we obtain the following Mellin representation for the R-current correlator (C.1) , , while for (C.2) we get

JHEP10(2019)247
We also obtained the other Mellin components, M 4 and M a4 , and verified that they satisfy transversality a γ a M a = 0 , As a side comment, we observe that the transversality condition (C.21) imposes nontrivial constraints on the Mellin amplitudes. For example, the Mellin amplitudes with one external stress tensor can be cast in the following form Imposing transversality fixes all the coefficients c ab,i in terms of just one, say c 22,1 .
It is now instructive to go back to position space once again. This is a subtle procedure, but when performed correctly we are able to recover both the anomalous and free pieces of the correlator. From the definition of the spinning Mellin amplitude (3.4), the integration variables must sit in the domain (C. 23) In order to translate our Mellin amplitudes (C. 19) and (C.20) into D-functions, we use the definition

JHEP10(2019)247
The contributions with the ladder integral and logarithms are exactly the same as in (C.12), and so we are able to extract the free correlator

D Properties of D-functions
In this appendix, we summarize some basic properties of the D-functions which we encountered in this paper. A general D-function with n-external points is defined as an integral in AdS d+1 After some standard manipulations, the D-function can be written as a Feynman integral where Σ = 1 2 n i=1 ∆ i . One should notice that the d-dependence only appears in the overall factor Γ(Σ − d 2 ), and therefore D-functions defined in all AdS d+1 are essentially the same. From the Feynman representation, we can derive a useful relation which relates D-functions of higher weights. Let us now focus on the special case with n = 5 and (∆ 1 , ∆ 2 , ∆ 3 , ∆ 4 , ∆ 5 ) = (1, 1, 1, 1, 2). This seed function D 11112 and its permutations generate all other D-functions in this paper via the differential recursion relations (D.3). To explicitly evaluate D 11112 , we need to compute the following integral

JHEP10(2019)247
In fact (D.4) is a special case of a more general class of integrals where P ({a i }) is a polynomial of a i . These integrals can be evaluated in terms of the scalar one-loop box integral I 4 [1] [49,50]. The result is Here N 5 and η i5 are defined via a matrix ρ  Figure 9. A five-point function can be written in terms of sum of four-point function where one of the operators appears in the OPE of (12) or in terms of a sum of three-point function where now one of the operator appears in OPE of (12) and the other in the OPE of (35).

E Five-point conformal blocks
A CFT correlator contains information about lower-point functions, which can be accessed through the OPE. For example, a five-point function can be either written in terms of a sum of four-point functions or a double sum of three-point functions, as illustrated in figure 9 where the sum is over both primary and descendant operators. 23 The conformal algebra fixes the contribution of the descendants in terms of the primary operators, in what is usually called the conformal block. In the following we will be interested in obtaining this kinematical contribution to the double OPE channel (12)(35) of the five-point function. This can also be viewed by inserting a complete basis of states labeled by their dimension and spin where s 1 = e −(τ 2 −τ 4 ) and s 2 = e −(τ 4 −τ 5 ) are two cross ratios and we have used the cylinder picture of figure 10, which is obtained with a Weyl transformation that maps R d to R×S d−1 .
The goal of this appendix is to lay out the strategy to obtain the conformal blocks for fivepoint functions in the double OPE 24 in the channels (12) and (35) as an expansion in terms of powers of s 1 and s 2 . The method used here is an adaptation of the one already implemented for four-point functions in [74,75]. The double OPE in the cylinder picture (E.2) will not be explicitly used in the derivation of the conformal blocks but its schematic form and simplicity makes it very appealing to explain the key points in the derivation. First notice that there are three three-point functions in this decomposition with two of them of the form scalar-scalar-spin while the third has two spinning operators and one scalar The spinning operators must transform in a symmetric traceless representation since they appear in the OPE of two scalar operators. A three-point function scalar−scalar−spin has only one structure, but on the other hand a three-point function with scalar − spin − spin has two fundamental structures [46] and the upper index m is used to label this property. The spinning operators carry with them Lorentz indices that should be contracted among each of the three-point functions. The most efficient way to do this contraction of indices is to introduce null polarization vectors z 1 and z 3 that are contracted with the indices of the J − J − scalar three-point function and use the differential operator in order to recover the tensor structures. The final formula for the contraction is given by (see [46] for more details) where h = d/2, and z 3 · x 13 , z 1 · x 13 and z 1 · z 3 x 2 13 − 2z 1 · x 13 z 3 · x 13 are related with the fundamental structures V i,jk and H ij of spinning three-point functions In the formula above we used conformal symmetry to put the point x 4 to infinity and x 2 13 = 1 to simplify the computation. The right-hand side of (E.5) depends only on the angles The coefficients a n,n ,j,j ,m are fixed by conformal symmetry alone and they could be obtained by working out precisely how the contribution of a primary operator is related with the corresponding descendant in (E.2). However we find it easier to use the property that the conformal block G satisfies two eigenvalue equations coming from applying the quadratic Casimir of the conformal group 12 s 2 1 ∂ s 1 +d 12 (s 1 s 2 ∂ s 2 −s 1 ∆ 1 )+d 12 s 1 +s 1 ξ 1 (s 2 1 ∂ 2 s 1 +s 1 ∂ s 1 (s 2 ∂ s 2 −(∆ 1 −1))) (E.17) d (1) 12 = (ξ 1 ξ 2 +ξ 3 )∂ ξ 2 −3(ξ 1 ξ 3 +ξ 2 )∂ ξ 3 +2(1−ξ 2 1 )∂ ξ 1 , d 12 = (ξ 3 ξ 1 +ξ 2 )∂ ξ 3 −(1−ξ 2 1 )∂ ξ 1 d (3) 12 = 2(ξ 1 ξ 3 −2(h−1)ξ 2 )∂ ξ 3 +(ξ 2 1 +2(h−1))∂ ξ 1 +(1−ξ 2 3 −ξ 1 ξ 3 ξ 2 −ξ 2 2 )∂ ξ 3 ∂ ξ 2 +2ξ 1 (ξ 2 3 −1)∂ 2 ξ 3 −ξ 1 (ξ 3 +ξ 1 ξ 2 )∂ ξ 1 ∂ ξ 2 +3(ξ 1 ξ 3 +ξ 2 )ξ 1 ∂ ξ 1 ∂ ξ 3 +ξ 1 (ξ 2 1 −1)∂ 2 ξ 1 . (E.18) The differential equation for D (0) 35 and D 35 can be obtained with the replacement ξ 1 ↔ ξ 2 . Notice that the differential operators D 35 keep the degree of the cross ratios s 1 and s 2 , while the other differential operators raise the degree of cross ratios by one.
The angular function H J,J ,m (ξ i ) plays an analogous role in the conformal block as the Gegenbauer polynomial for the case of four-point functions [75]. In particular it has to satisfy two eigenvalue equations coming from the leading order limit of (E.14) .

(E.29)
Obviously there are similar relations for the differential operators d 12 − C ∆ k ,J a n,n ,j,j ,m s ∆ k +n 1 s ∆ k +n 2 H jj m (ξ 1 , ξ 2 , ξ 3 ) = a n,n ,j,j ,m s ∆ k +n 2 s ∆ k +n 1 2s 1 d 12 H − 2(∆ + n)s 1 d 12 H − 2s 1 (∆ 1 − ∆ − n )d Now one can use that H functions with different indices jj m are orthogonal to each other to write a recurrence relation between the unknown coefficients a n,n ,j,j ,m in (E.10). We verified that this method gives the same result as the one where we use the formal expression for the OPE twice on the five-point function 26 l q=0 a n,m,q (x · D) l−q (x 2 ∂ y · D) q (x · ∂ y ) n (x 2 ∂ 2 y ) m , with a 0,0,0 = l!(h − 1) l . 26 The coefficients an,m,q associated with a given primary in this expression for the OPE can be obtained by solving the equation