Equivalent Equations of Motion for Gravity and Entropy

We demonstrate an equivalence between the wave equation obeyed by the entanglement entropy of CFT subregions and the linearized bulk Einstein equation in Anti-de Sitter pace. In doing so, we make use of the formalism of kinematic space [arXiv:1505.05515] and fields on this space, introduced in [arXiv:1604.03110]. We show that the gravitational dynamics are equivalent to a gauge invariant wave-equation on kinematic space and that this equation arises in natural correspondence to the conformal Casimir equation in the CFT.


I. INTRODUCTION
The recent direct detection of gravitational waves [3] adds to an impressive list of observational tests of general relativity, the theory that describes long-distance dynamics of spacetime. At the quantum level, however, the fate of spacetime requires a more careful assessment; indeed, it is not even clear what the fundamental degrees of freedom in a theory of quantum gravity are. The holographic principle [4,5] suggests that these degrees of freedom should in fact be nonlocal.
This notion is made explicit in the AdS/CFT duality, an equivalence between d−dimensional conformal field theories (CFTs) and (d + 1)-dimensional gravitational systems with Anti de Sitter (AdS) asymptotics. In particular, the proposal of Ryu and Takayangi (RT) [6,7] relates the entanglement entropy S of a CFT region B to the area of a bulk extremal surfaceB, Developments over the past few years [8][9][10] [11,12] have shown that the dynamics of this nonlocal CFT quantity is closely related to the bulk Einstein equation. In recent work [2] we demonstrated an extension of the RT proposal to other bulk scalar fields. We introduced the OPE block, the contribution to the CFT * czech@stanford.edu † llamprou@stanford.edu ‡ smccandlish@stanford.edu § bmosk1@stanford.edu ¶ jsully@slac.stanford.edu operator product expansion (OPE) from a single conformal family, and equated it to the Radon transform, the integral of a bulk scalar field over a minimal surface. Both of these objects obey an equation of motion in kinematic space, a space which geometrizes the set of bulk surfaces of a given dimension. We showed that this kinematic equation of motion emerges directly from the bulk Klein-Gordon equation.
In this paper, we will demonstrate the relationship between the Ryu-Takayanagi proposal and our results on kinematic space. Specifically, we will show that the wave equations obeyed by a perturbation to entanglement entropy [11][12][13][14] correspond directly to the linearized bulk Einstein equation with matter via an intertwining relation of the Radon transform: ( K + 2d) δS = 0 ↔ Einstein equations ( dS + d) δS = 0 ↔ Hamiltonian constrant where the Laplacians K and dS on kinematic space arise from conformal Casimir equations. This clarifies and connects the results of [8][9][10][11][12].
Before stating our result explicitly, we begin by outlining the results of [2] on kinematic space, the OPE block, and the Radon transform.

II. SCALAR KINEMATIC DICTIONARY
The kinematic dictionary, introduced in [2] and recently also explored in [15], connects Radon transforms of AdS-fields with OPE-blocks in the dual CFT. Here we outline the basic formalism.
Radon transform The Radon transform is a map from functions f (x) on some manifold to functions on the space of n-dimensional totally geodesic submanifolds B. It is defined via the integral transform where dA is the induced area element on the surfaceB. Though [2] also considered the case n = 1 of geodesics, we will focus here on the case n = d−1 of codimension-2 minimal surfaces in AdS d+1 . It is useful to define an auxiliary space K, which we call kinematic space, to organize information about bulk surfaces. A point in K denotes equivalently any of the following: • a particular bulk minimal surface,B, • the boundary sphere where that surface ends, B, • the two timelike separated boundary points x 1 , x 2 at the tips of the causal development of this sphere (see Fig. 2), 12 .
The points of K are most conveniently parameterized by the two points (x 1 , x 2 ). When the context is clear, we will often denote any of the above three objects by the pair (x 1 , x 2 ). The conformal group SO (d, 2) then endows kinematic space with a metric structure of (d, d) signature (see [2]), where I µν (x) = η µν −2 xµxν x 2 is the CFT inversion tensor. It will also be useful to note that if we consider only spheres living in a particular equal-time slice of AdS d+1 , preserved by an SO (d, 1) subgroup, the corresponding slice of K has the structure of a d-dimensional de Sitter (dS) space [1,2,14,16] (see Fig. 1). There is one such dS slice for each time slice of AdS. For instance, if we consider the t = 0 slice of AdS and parameterize the spheres it contains by their radius R and center x, the induced metric on this slice is given by A particularly useful feature of the kinematic space K is that the domain of the CFT, including the time direction, appears as a spacelike surface at x 1 = x 2 . This allows us to impose boundary conditions and choose a causal propagator in the usual way when solving wave equations in kinematic space. Figure 1. The kinematic space for spherical regions that lie on a single time slice is given by d-dimensional Lorentzian de Sitter space. The de Sitter space is a corresponding slice of the larger kinematic space for all (boosted) spherical regions, which is 2d-dimensional with signature (d, d).

OPE-blocks
We now introduce a CFT object whose domain is also K. Recall that in a CFT, the product of two identical scalar local operators can be expanded in terms of the global primary operators of the theory as The coefficients C OOk are known as the OPE coefficients and are theory-dependent, while the coefficients b i depend only on the scaling dimension ∆ k of the primary operator O k . We have grouped the contributions from a single primary to the OPE into an object B k (x 1 , x 2 ), which we call the OPE-block. In [2], we showed that operators σ (x 1 , x 2 ) localized on a CFT sphere can be expanded in terms of the same OPE blocks as where C σk are theory-dependent "OPE" coefficients. The OPE-block also has a compact integral expression [2] over the causal diamond formed by the points x 1 , x 2 . For a scalar, this is just: x 13 x 23 x 12 A point in kinematic space labels two timelike-separated boundary points, x1, x2, or equivalently a codimension-2 sphere formed by their causal domain. Operators σ(x1, x2) localized on the sphere have an OPE expansion in terms of OPE blocks for primaries O k , where contributions can be identified as bulk surface operators φ k . OPE-blocks are useful as CFT objects in their own right; however, they become even more powerful in the presence of an AdS dual, as we describe next.
Kinematic Dictionary At leading order in the N → ∞ limit, the OPE-block B k (x 1 , x 2 ) and the Radon transform R [φ k ] (x 1 , x 2 ) of the dual AdS field are directly related: where c ∆ is a constant depending only on the dimension d and the scaling dimension ∆ k . To prove this, in [2] we showed that both sides obey the same equation of motion with the same boundary conditions: The OPE block equation of motion comes from a conformal Casimir equation, where L 2 is the Casimir element of the conformal group SO (d, 2), with eigenvalue The Casimir element is represented on kinematic space scalar fields by the Laplacian K , yielding the equation of motion. The boundary condition as x 1 → x 2 comes from inspecting the definition 5.
To find the equation of motion for the Radon transform we use an intertwining property: Together with the Klein-Gordon equation, this implies an equation of motion for the Radon transform: The boundary condition then comes from the AdS/CFT Since kinematic space has signature (d, d), an additional d − 1 equations are required to fix a solution uniquely. These take the form of constraint equations, explained in detail in [2,15]. These equations of motion together with the boundary conditions establish the validity of 8.
In the remainder of this paper, we will use the same techniques to extend the kinematic dictionary to the CFT stress tensor, which is dual to the bulk metric perturbation. We will extend this dictionary to first subleading order in the 1/N expansion, finding that the correction is precisely that found in [17]. This will allow us to prove an equivalence between the linearized Einstein's equations in the bulk and a simple equation satisfied by the stress tensor OPE-block, which we show is equal to the modular Hamiltonian.

III. TENSOR RADON TRANSFORMS AND EINSTEIN'S EQUATIONS
In this section, we will show that the linearized Einstein equations are equivalent to a set of equations obeyed by the fluctuation in the area of the minimal surfaces. We will do this in a similar way as in Eq. 13, by using an intertwining relation. The goal will be to find an analog of the field equation and boundary conditions of Eq. 9; we will then match to CFT quantities in the following section.
Since the bulk field of interest, the metric perturbation δg µν , is a tensor, we must first introduce a tensor analog of the Radon transform Eq. 2. For a symmetric 2-tensor field s µν , we define the longitudinal and transverse Radon transforms, denoted R and R ⊥ respectively, as Here, h µν denotes the induced metric on the surfaceB 12 .
As before, these transforms output a scalar function on kinematic space; we write indices on the left side only to indicate that the input is a tensor. We now note some useful identities for the tensor Radon transform. First, note that the sum of the two tensor transforms is just a scalar Radon transform of trs. This implies that the two are related by a trace-reversal of the input tensor, Before we continue, let us pause to note a striking simplification of the full nonlinear Einstein equation when written in terms of tensor Radon transforms. The Einstein equation takes the form where we have defined T µν to include the cosmological constant term. If we apply the transverse transform R ⊥ to both sides, using the identity 15, we obtain This remarkable simplification occurs only when we integrate over a codimension-2 surface, due to the appearance of the coefficient 1 2 in the Einstein tensor 1 . Let us now consider the linearized version of Eq. 17.
where δA is the first order change in the area of the surface of integration. In terms of the tensor Radon transform, the area perturbation can be written as We would like to specialize to minimal surfaces in AdS d+1 , and recast the linearized equation 19 as an equation of motion in kinematic space. To do this, we make use of an intertwining relation analogous to 12, The right side of this equation is given by the action of 1 However, note for a dimension-2 (rather than codimension-2) surface, the longitudinal transform of the Einstein equations yields the equation the casimir L 2 SO(d,2) on s µν , as shown in Appendix A. In fact, for the case of δg µν , Eq. 21 can be rewritten as where δR µν is the variation of the Ricci tensor due to the variation δg µν in the metric; this is shown explicitly in Appendix A. Together with 19, this implies the equation of motion We have thus shown that the area perturbation δA obeys an equation of motion in kinematic space as a consequence of the linearized Einstein equation about AdS.
To show complete equivalence between the kinematic equation of motion and Einstein equations, it remains only to show that the tensor Radon transform is invertible (up to diffeomorphisms). Unfortunately, while reasonable, we are not aware of a proof of this fact in the literature, and our statement of equivalence must carry a technical asterisk awaiting further input from the mathematical community.
To avoid the technical problem in the preceding paragraph, we will now prove an additional equation of motion for the area perturbation, but this time restricting ourselves to surfaces on a time slice of AdS; this corresponds to a particular de Sitter slice of kinematic space. Using the same techniques as above, we prove in Appendix A that The right hand side is a tt component of the Einstein tensor; hence, using the Einstein equation, we find Here, T 00 denotes the energy density relative to the particular AdS time slice we are considering; there is a separate de Sitter equation for each time slice. Hence, the Hamiltonian constraint of the Einstein equation implies a de Sitter equation of motion for the area perturbation. In thise case, since the scalar Radon transform is known to be injective [18], Eq. 25 is equivalent to the Hamiltonian constraint on a time slice, and the collection of de Sitter equations for every slice is equivalent to the full linearized Einstein equation, because knowing E 00 = T 00 for every choice oft implies E µν = T µν .
To complete the description of the Cauchy problem for the area perturbation δA in kinematic space, we must fix boundary conditions. This can be done using the extrapolate dictionary for the metric perturbation [19][20][21], and was shown in [9] to be Now that we have formulated the Cauchy problem for the area perturbation in the form of Eqn. 26 along with either 23 or 25, we can proceed to match with CFT variables.

IV. MODULAR HAMILTONIAN AND THE TENSOR KINEMATIC DICTIONARY
In this section we use the Radon-transformed Einstein equations we have just derived to give a novel derivation of the quantum-corrected Ryu-Takayanagi formula.
To begin, we take a moment to review the entanglement first law and the form of the vacuum modular Hamiltonian. Given a quantum mechanical system in a certain state |ψ , the state of a subsystem B is described by the reduced density matrix ρ B , obtained from |ψ ψ| by tracing over the degrees of freedom of the complement B c . For such a subsystem, the entanglement entropy is defined as: The modular Hamiltonian H mod of the state ρ B is then defined implicitly by Using this expression, the change in the entanglement entropy of A due to a small perturbation of the state can be compactly expressed as: This equation is known as the first law of entanglement entropy. When B is a ball of radius R in the vaccum state of a CFT, the modular Hamiltonian can be written as where T 00 is the energy density in the CFT. The form of the vacuum modular Hamiltonian was computed in [22]. The fact that H mod is an OPE block was pointed out in [2]; in Appendix C we give the details for general dimension.
We can now make contact with the equations of motion (23,25). It was pointed out by [14] that the vacuum modular Hamiltonian, when viewed as a field on kinematic space, obeys a de Sitter wave equation It in fact obeys a separate equation for each CFT time slice, each of which has a corresponding de Sitter slice of the full kinematic space K. To see this, note that the SO (d, 1) subgroup of the conformal group that preserves a time slice has the Casimir Since with eigenvalue 2d coming from Eq. 11 [2].
Having written an equation of motion for H mod , we would now like to check the boundary conditions in kinematic space for H mod . Taking the limit R → 0 of Eq. 30, we find We can now compare directly with the results of Sec. III. First, let us consider the leading order in 1/N behavior, for which δT µν = 0. In that case, Eqns. 25 and 31 match, and the boundary conditions 26 and 34 differ only by a constant 4G N . This gives us the leading-order kinematic dictionary: Of course, this is just the linearized Ryu-Takayanagi formula 1. To find the O N 0 correction to the dictionary, we must find an object X which satisfies Using Eq. 25, this will the guarantee that δA 4G N + X satisfies the same EOM as H mod , Eq. 31. The solution can be written as Figure 3. The retarded de Sitter propagator corresponds to an integral over all the geodesics that lie on the t = 0 slice of the causal wedge. This integral reconstructs the bulk modular Hamiltonian, which is also an integral over the same spatial slice.
where the integration region is the past light cone of (x 1 , x 2 ), and where G dS is a bulk-to-bulk kinematic space causal propagator. The result is where ξ µ is the Killing vector corresponding to modular flow [15], and where dΣ ν is the timelike unit normal vector to Σ (see Figure 3). We immediately recognize that X is none other than H bulk , the bulk vacuum modular Hamiltonian for a Rindler wedge. (In Appendix B, we show via an intertwining relation that H bulk indeed satisfies Eqn. 36.) We thus arrive at the corrected kinematic dictionary, which matches the FLM correction to the Ryu-Takayanagi formula [17,23]. We hence recognize the Ryu-Takayanagi formula arises as a special case of our more general kinematic dictionary.

V. DISCUSSION
We have given a simple and elegant demonstration of the equivalence of the Einstein equations and the quantum-corrected Ryu-Takayanagi formula for linearized perturbations about the vacuum. In doing so, we exploited the fact that both the bulk Einstein equations and the boundary modular Hamiltonian obey simple dynamical equations in an auxiliary kinematic space.
The derivation of the quantum corrected Ryu-Takayanagi formula from the Einstein equations was already described in [17]. While the previous work is more general, the bulk quantum contribution depends on a generally non-local and unknown modular Hamiltonian. Our approach, on the other hand, makes more explicit how quantum corrections arise from bulk interactions. We hope that these techniques will prove insightful when extended away from simple regions of the vacuum state.
This paper was also primarily focused on demonstrating how the bulk Einstein equations imply the Ryu-Takayanagi formula, but our work equally leads to the reverse statement. This is most mathematically rigorous when we make use of the kinematic space for a singletime slice and the scalar Radon transform to derive the tt-component of the Einstein equations (all components then follow by appropriate boosts, as in [8]). Here, we can derive the local bulk EOM because the scalar Radon transform on Hyperbolic space has been proven to be invertible [18]. However, to exploit the full kinematic space and directly derive any component of the Einstein equations, we must invert the tensor Radon transform. While the invertibility of these transforms (up to diffeomorphism) is well-motivated, we are not aware of a mathematical proof. Thus, a stickler for rigor will conclude that only an integrated version of the Einstien equations has been derived by this second approach.
As the gravitational equations of motion intertwine to become a kinematic equation of motion which is fixed by conformal invariance, it may be confusing to wonder what happens to the gravitational equations of motion in a generalized theory of gravity. However, the particular gravitational equation of motion is determined not by the fixed kinematic EOM, but by the choice of entropy functional. In particular, it was shown in [8][9][10] using the Wald-Iyer formalism [24] that the entanglement first law 29 can be written for perturbations of the vacuum as In this equation, δE µν is the linearized equation of motion for a general theory of gravity, and the point-wise equation of motion follows from considering all surfaces Σ.
In upcoming work, one of us [25] will show the relationship of the Wald-Iyer formalism to the present work, where an integral over a surfaceB rather than a time slice Σ appears. Applying the differential operator ( dS + d) to both sides of 40 and using the intertwining relation of [APPENDIX] along with the equation of motion 31 yields the equation [25] Applying ( K + 2d) instead yields the equation Hence, the equation of motion for δS is equivalent to the linearized gravity equation integrated over a bulk surface, and both vanish due to the entanglement first law.
The localization of the equation of motion onto the Ryu-Takayanagi surface after applying these differential operators, a somewhat surprising fact from the Wald-Iyer point of view, was required by the kinematic space formalism. It would be interesting to study whether such a localization occurs more generally away from the vacuum. If so, it may be more natural to consider whether the bulk gravitational equations can be derived not from entanglement equations, but from one-point entropy equations [26][27][28].
It is also possible to assume both the Ryu-Takayanagi formula and Einstein equations, and then derive the kinematic space entropy equations [11,12]. One can understand this as a consistency check of the approach, as the entropy equations are pre-determined by conformal invariance.
The techniques we used to derive the quantum corrections for holographic entanglement entropy link this story with a more general program of including interactions in the dynamics of both kinematic space and local bulk operators operators [29][30][31][32]. In particular, we can think of δA as a kinematic field, whose interactions with the stress tensor generate quantum corrections to the kinematic operator. We will report on interacting kinematic operators in upcoming work.
This type of relation has been In this section, we extend the intertwinement relation (A1) to symmetric two-tensors. We will present a short group theoretic derivation.
The only property of the tensor transform that we will use is that it transforms as a scalar field on kinematic space under the relevant isometry group. Hence, our proof will hold for a general transformR with this transformation property. In particular,R can denote the longitudinal transform R , the transverse transform R ⊥ , or even the wedge transform used in Appendix B.
Consider a tensor transformR [s] of a symmetric 2tensor field s µν . Under some isometry g ∈ SO (d, 2) of AdS d+1 , this field transforms as s → g AdS · s, while a point in kinematic space transforms asB → g K ·B. The scalar transformation property then implies that In infinitesimal form equation A2 becomes where the L AB are the differential operators representing the generators of the conformal algebra, and the superscript denotes the representation. It follows that where L 2 = L AB L AB is the conformal Casimir. To find the intertwining relation for the Laplacian, we must now find the quadratic differential operator representing L 2 on the tensor s µν and its transformR [s]. To do this, we make use of the fact that AdS and K are both coset spaces G/H, where G = SO(d, 2) and H is the stabilizer of a bulk point or bulk surface respectively: 1) .
(A5) The Casimir operator on G is represented by the Laplacian with respect to the Cartan-Killing metric. For a coset space G/H, the Laplacian G can be written as [33]: A scalar function on AdS-spacetime is in the kernel of the Casimir of the little group SO(d, 1), so the conformal Casimir is represented on functions on AdS-spacetime by the AdS-Laplacian, up to a constant proportionality factor. A similar argument holds for functions on kinematic space. The relative proportionality factor in the intertwining relation A1 is fixed by a choice of the Cartan Killing form on the Lie-algebra of the conformal group G = SO (d, 2). 2 For general tensors on AdS-spacetime, there will be an additional term from the non-trivial representation of the little group H = SO(d − 1, 1). The tensor Radon transform maps symmetric (two-) tensors on AdS-spacetime to functions on kinematic space, so there will be no additional contributions from the Casimir of the kinematic space little group. Tensors on AdSspacetime do receive a contribution from the Casimir of the little group H=SO(d, 1). One can decompose a general tensor on AdS-spacetime in terms of irreducible representations of the little group H = SO(d−1, 1). The irreducible representations can be labeled by the spin , and the conformal Casimir is represented by [33,34]: where ∇ 2 denotes the covariant Laplacian. We recover the representation of the conformal Casimir on functions on AdS-spacetime by setting = 0. The traceless part of a symmetric two-tensor corresponds to the = 2 representation, whereas the trace-part of a tensor corresponds to the = 0 representation. We decompose a general symmetric two-tensor s µν into the traceless symmetric and trace parts Then, using Eqns. A4, A7, and A8, we find the following intertwinement rule for symmetric two-tensors:

Einstein Equations from Intertwinement
We would now like to verify Eqn. 22. First recall from Eqn. 20 that the area perturbation can be written as 2 For details on the factor of proportionality, see [2].
From here, we can see that the longitudinal Radon transform annihilates total derivatives: where v ν is a vector field falling off sufficiently quickly at infinity that the longitudinal transform is well defined. This follows from the fact that δA vanishes at first order for small deformations of the surface, which correspond to small coordinate transformations δg µν = ∇ (µ v ν) . Applying the Laplacian K to δA and using the intertwining relation A9, we find (A12) Now, note that the variation of the Ricci tensor is given by By commuting covariant derivatives, this can be rewritten for AdS d+1 as The last three terms match those in A12, while the longitudinal transform annihilate the first three terms, yielding the desired result Let us now proceed to verify Eqn. 24. The Hamiltonian constraint equation, the tt component of the Einstein equation, can be written as where R and K µν denote the Ricci scalar and extrinsic curvature tensors on the time-slice of interest. For an equal-time slice of AdS, we have K µν = 0, so that the linearized equation takes the simple form Hence, we can prove 24 by showing If we denote the perturbation of the induced metric on the time-slice of interest by δw µν , then we can write δA = 1 2 R [δw µν ]. Then, both sides of Eqn. A18 depend only on δw µν , and we can restrict ourselves to considering perturbations of the metric of hyperbolic space H d . Generalizing A9 to hyperbolic space, we can write the intertwining relation (A19) Using the same methods as above, we find that (A20) where n µ denotes the unit normal vector toB in H d .
To proceed, we must use the fact that the extrinsic curvature tensor onB is zero, which implies that ∇ µ n ν = n µ n α ∇ α n ν . Repeatedly using this along with Eqn. A11, we find that the right-hand side of A20 vanishes, proving the result 24.

Appendix B: Wedge Integral Relations
In this appendix, we will prove that the equation 36 is satisfied by the bulk modular Hamiltonian 38. The object of interest will be a transformation R ∧ [s µν ] of a conserved symmetric 2-tensor s µν , defined by where Σ is an equal-time slice of a causal wedge, dΣ ν is the unit normal vector to that slice, and ξ µ is a Killing vector given in Poincaré coordinates by [15] Here the capital letters denote bulk coordinates X = (x, z), whose indices are contracted using the Minkowski metric. The bottom and top points of the corresponding causal diamond are denoted by X 1 = (x 1 , 0) and X 2 = (x 2 , 0) respectively, with ξ representing a flow from X 2 to X 1 . In particular, we have H bulk = 2πR ∧ [T µν ]. Note that since s µν is conserved, R ∧ is independent the choice of time slice Σ.
We can now restrict ourselves to a constant time slice of AdS d+1 , which we take to be the t = 0 slice. Then, ξ points only in the time direction, and we have R ∧ [s µν ] = Σ s 00 |ξ| dΣ.
Hence, this restriction of the wedge transform is really a scalar transform, and is subject to the scalar intertwining relation considered in [2]. In particular, we have Integrating by parts, this becomes where n µ is the outward-pointing unit normal vector tõ B within the specified time slice. The first term vanishes because |ξ| = 0 atB, and because s 00 goes to zero sufficiently quickly at ∂AdS; the second vanishes since ∇ 2 H |ξ| = d |ξ|, as can be checked explicitly from B2. Finally, we can check from B2 that n µ ∇ µ |ξ| = −1. This gives us our result, wheret is the timelike unit vector to the time slice correponding to the chosen de Sitter slice of kinematic space.
With some more effort, we can also prove a similar relation using the Laplacian on the full kinematic space, rather than a de Sitter slice. First, note that the wedge integral can be written as where we have defined the conserved current j µ = s µν ξ ν .
Now, since R ∧ transforms as a scalar in the full kinematic space, we can use the tensor intertwining relation A9 to obtain wherej µ = − ∇ 2 + 2 s µν + 2g µν trs ξ ν . Using the fact that s µν is conserved and ξ µ is Killing, it can be shown with significant effort [25] thatj = ∆j, where ∆ is the Hodge Laplacian. Then, conservation of j implies that where µν is the antisymmetric tensor in the two directions perpendicular to ∂Σ, defined such that 01 = −1.
Next, note that since ξ vanishes onB, we can plug in j to find Finally, it can be checked explicitly from B2 that µα ∇ α ξ ν = g µν − h µν , where h µν is the induced metric onB. This yields the result This relation was required for consistency between equations 23, 33, and 39.