Quantum State of a Gravitating Spacetime Region

Let $w_{\sigma}$ be a $(d+1)$-dimensional Lorentzian spacetime wedge with finite edge $\sigma$; see Fig. 1. That is, $w_{\sigma}$ is a causal domain of dependence whose $d$-dimensional Cauchy slices all have the $(d-1)$-dimensional spatial surface $\sigma$ as their boundary. Suppose that the classical metric and fields are real analytic functions of $d+1$ coordinates $(t,r,\ldots)$ on $w_{\sigma}$, so that we may view $w_{\sigma}$ as a complex manifold in a neighborhood of the real Lorentzian section. Let $\Sigma$ be a Cauchy slice of $w_{\sigma}$, and suppose for simplicity that $\Sigma$ corresponds to $t=0$ and $\sigma$ to $r=1$. Let $J(\sigma)$ be the embedded $d$-dimensional submanifold

where $s$ is a real analytic function on $\Sigma$ that vanishes only at $r=1$. Thus $\sigma$ is the boundary not only of $\Sigma$ but also of $J(\sigma)$.

Let $\{h_{ab},K_{ab}\}$ denote the induced metric and extrinsic curvature on $J(\sigma)$. Let $\{([h_{ab}],K)_{J(\sigma)\setminus\sigma},(\gamma_{pq})_{\sigma}\}$ be the data subset consisting only of the conformal class of $h_{ab}$ and the trace of $K_{ab}$, in the interior of $J(\sigma)$, as well as the full induced $(d-1)$ metric $(\gamma_{pq})_{\sigma}$ on $\sigma$. We shall refer to this data subset as mixed-conformal data. The manifold $J(\sigma)$, when equipped with mixed-conformal data, will be called an elliptic data manifold and denoted by $\mathcal{J}(\sigma)$.

We consider the vector space $\mathbf{V}_{\sigma}$ spanned by all manifolds that have boundary $(\sigma,\gamma_{pq}|_{\sigma})$ and are equipped with mixed-conformal elliptic data. Importantly, we do not restrict to $\mathcal{J}(\sigma)$ obtained by complex-deforming a Cauchy slice of a Lorentzian wedge, Eq. (31). We presented that procedure only to gain some intuition for how a classical-quantum correspondence might emerge. But the quantum state $\mathcal{J}(\sigma)$, not $w_{\sigma}$, is the primary object in our construction, and its mixed-conformal data can be specified freely. Whether $\mathcal{J}(\sigma)$ admits a classical approximation or classical limit is a secondary question to which we will return shortly.

A dual vector or “bra” $\mathcal{J}^{*}(\bar{\sigma})$ is obtained by reversing the orientation of $\mathcal{J}(\sigma)$ and complex conjugating its elliptic data. [Thus if $\mathcal{J}(\sigma)$ was obtained by Eq. (31), then the opposite deformation $t=-is(r,\ldots)$ yields $\mathcal{J}^{*}(\bar{\sigma})$.] A pre-inner product $\braket{\mathcal{K}(\sigma),\mathcal{L}(\sigma)}$ is defined by gluing $\mathcal{K}^{*}(\bar{\sigma})$ to $\mathcal{L}(\sigma)$ across $\sigma$ and evaluating the GPI (over $(d+1)$-manifolds with complex metric) with the resulting closed boundary condition. In particular, gluing $\mathcal{J}(\sigma)$ to $\mathcal{J}^{*}(\bar{\sigma})$ and evaluating the GPI computes the norm of $\mathcal{J}(\sigma)$. The pre-inner product is assumed to be positive semidefinite; quotienting by null states yields a Hilbert space with positive-definite inner product.

If the norm-GPI is dominated by a single classical saddle geometry $(M,g_{ij})$ in the $G_{N}\to\ 0$ limit, then we call $(M,g_{ij})$ the norm geometry of $\mathcal{J}(\sigma)$. The (complexified) norm geometry is invariant under a CRT transformation (i.e., an antiholomorphic involution) $\boldsymbol{\Theta}$ that swaps $\mathcal{J}(\sigma)$ and $\mathcal{J}^{*}(\bar{\sigma})$, or else $\boldsymbol{\Theta}(M,g_{ij})$ would furnish a second saddle with the same real part of the action. The fixed point set $\text{Fix}(\boldsymbol{\Theta})$ of $M$ under $\boldsymbol{\Theta}$ is a real $(d+1)$-manifold with real induced metric.

If the signature of the metric on $\text{Fix}(\boldsymbol{\Theta})$ is Lorentzian, then $w[\mathcal{J}(\sigma)]\equiv\text{Fix}(\boldsymbol{\Theta})$ is a Lorentzian wedge with edge $\sigma$. If in addition $\mathcal{J}(\sigma)$ can be smoothly deformed to a Cauchy slice $\Sigma[\mathcal{J}(\sigma)]$ of $w[\mathcal{J}(\sigma)]$, while preserving the appropriate subset of $(M,g_{ij})$ as the unique dominant saddle, then we call $w[\mathcal{J}(\sigma)]$ the classical spacetime limit of the quantum state $\mathcal{J}(\sigma)$.

Several remarks are in order. If no norm geometry exists, or if the metric on its CRT fixed point set is not Lorentzian, or if the smooth-deformation criterion fails, then $w[\mathcal{J}(\sigma)]$ is not defined. In this case, $\mathcal{J}(\sigma)$ has no classical spacetime limit. We are not currently aware of any example of a mixed-conformal elliptic data manifold with this property. The classical limit criterion could be broadened to assign superpositions of classical geometries to “sufficiently simple” superpositions of states with classical limits, but we will not attempt this here.

If $\mathcal{J}(\sigma)$ is constructed by a sufficiently small complex deformation of a Cauchy slice of a given $w_{\sigma}$, as in Eq. (31), then it is interesting to ask whether $\mathcal{J}(\sigma)$ “returns” $w_{\sigma}$ as its classical spacetime limit:

This relation does appear to hold broadly; we currently know of no counterexamples with $\sigma\neq\varnothing$ (except when Dirichlet data instead of mixed-conformal data are retained Bousso et al. (2026)).

However, Eq. (2) does not hold universally. With $\mathcal{J}(\varnothing)$ constructed as a complex deformation of a slice $\Sigma$ of Lorentzian de Sitter space $w^{\rm dS}_{\varnothing}$, we find in Sec. 4 that $w[\mathcal{J}(\varnothing)]=\varnothing\neq w^{\rm dS}_{\varnothing}$. This may have some bearing on the puzzling Hilbert space structure of closed universes Abdalla et al. (2026); see also Usatyuk et al. (2024); Usatyuk and Zhao (2025); Harlow et al. (2026); Abdalla et al. (2025); Harlow (2026); Zhao (2026); Nomura and Ugajin (2026). It also raises the question of whether de Sitter space exists as the classical limit of any quantum state. (Note that the existence of de Sitter space is not required from an empirical standpoint Dyson et al. (2002).)

Given a Lorentzian wedge $w_{\sigma}$, let $\mathbf{C}(w_{\sigma})=\{\mathcal{J}(\sigma)\in\mathbf{H}_{\sigma}:w[\mathcal{J}(\sigma)]=w_{\sigma}\}$ be the set of quantum states whose classical spacetime limit is $w_{\sigma}$. This defines a classical-quantum correspondence $w_{\sigma}\leftrightarrow\mathbf{C}(w_{\sigma})$. Unless $\mathbf{C}(w_{\sigma})$ is empty, as it appears to be for $w^{\rm dS}_{\varnothing}$, $\mathbf{C}(w_{\sigma})$ will contain infinitely many states. We interpret this as a close analog of the choice of polarization when a quantum wavepacket is associated to a point particle with phase space data $(x,p)$. Concretely, we expect that increasing $s$ in Eq. (31) increases the uncertainty in $([h_{ab}],K)$ around the values on $\Sigma$, and decreases the uncertainty in their canonical conjugates.

In order to test our proposal quantitatively, as we do in Sec. 6, it is useful to introduce additional structure. We shall assume that $\mathbf{H}_{\sigma}$ factorises over the connected components of $\sigma$. For example, suppose that $\sigma=\chi\sqcup\chi_{c}$; then we assume that the Hilbert space is bipartite: $\mathbf{H}_{\sigma}=\mathbf{H}_{\chi}\otimes\mathbf{H}_{\chi_{c}}$. With this assumption, partial traces can be defined and Rényi entropies computed. Given a state $\mathcal{J}(\sigma)$, a reduced state on $\chi$ is defined by the partial gluing of $\mathcal{J}(\sigma)$ to $\mathcal{J}^{*}(\bar{\sigma})$ across $\chi_{c}$. This yields an elliptic data manifold whose boundary is $\chi\sqcup\bar{\chi}$. It represents the unnormalized density operator $\tilde{\rho}(\chi,\bar{\chi})$; see Fig. 2. The full gluing followed by GPI evaluation defines its trace, $\tr\tilde{\rho}(\chi,\bar{\chi})$. (The full trace agrees with the norm of $\mathcal{J}(\sigma)$ by construction, as it should.)

Operator products are obtained by sequential gluing. For example, $[\tilde{\rho}(\chi,\bar{\chi})]^{2}$ is given by two copies of $\tilde{\rho}(\chi,\bar{\chi})$, partially glued to one another. Its trace is the two copies fully (cyclically) glued; and the trace is evaluated by the GPI; see Fig. 3. This allows us to compute the $n$-th Rényi entropy of the reduced state $\rho(\chi,\bar{\chi})$ as

Analytic continuation to $n=1$ yields the von Neumann entropy $S_{1}$.

A simplified version of the above framework was proposed and examined in Ref. Bousso et al. (2026). Some of the technical aspects of our proposal may evolve in response to further study of examples and applications.

Asymptotic applications of the GPI, including the seminal examples in Refs. Gibbons and Hawking (1977); Hawking and Page (1983), are special cases of our proposal. Though often presented as computations of a partition function, they can equivalently be viewed as norms of states $\mathcal{J}(\sigma)$. For example, a closed boundary of the form $S^{1}(\beta)\times S^{d-1}$ with real data can be cut in half to yield $\mathcal{J}(\sigma)$, $\mathcal{J}^{*}(\bar{\sigma})$ of the form $[0,\beta/2]\times S^{d-1}$.

Our approach also encompasses closed universe amplitudes in quantum cosmology, as the special case $\sigma=\varnothing$. For example, the unnormalized amplitude for the creation of a closed universe from nothing is an inner product $\braket{\mathcal{K}(\varnothing),\mathcal{L}(\varnothing)}$, where $\mathcal{K}$ is a closed $d$-manifold and $\mathcal{L}=\varnothing$ is the Hartle-Hawking state Hartle and Hawking (1983). (In a consistent application of the GPI, norms must also be computed from the GPI. This implies that historical evaluations of the Hartle-Hawking proposal were inconsistent, as noted in Ref. Abdalla et al. (2026).)

Our proposal is less than a nonperturbative quantum theory of gravity, but more than a formal Hilbert space construction. It is a tool for extracting quantum information theoretic properties of states of such a theory. We may hope that this tool will constrain, and help us construct, a theory that is directly applicable to the observed universe.

The assumption of Hilbert space factorisation is crucial for this aspect, as is our prescription for replicating and partially gluing elliptic data manifolds. This set of ideas is rooted mainly in the role played by the GPI in the proof Lewkowycz and Maldacena (2013) of the Ryu-Takayanagi prescription Ryu and Takayanagi (2006) for computing the von Neumann entropy of boundary subregions in AdS/CFT.

It is important, however, to distinguish between state preparation in a known quantum theory such as the CFT, and our proposal that elliptic data manifolds represent states. In AdS/CFT, the CFT state is the primary object in terms of which norms and Rényi entropies are defined, and a bulk-boundary dictionary Gubser et al. (1998); Witten (1998) must be invoked to determine the elliptic data for the GPI that computes the same quantities on the bulk side. By contrast, we define states directly as elliptic data manifolds — roughly, as half of a GPI boundary condition — while remaining agnostic about the theory that governs those states.

For example, suppose that in some alternate history, the AdS/CFT correspondence had not been discovered by 2013. Then the GPI “proof” Lewkowycz and Maldacena (2013) of the Ryu-Takayanagi prescription Ryu and Takayanagi (2006) would instead have been interpreted as a GPI prediction for the entanglement structure of the states of an unknown quantum theory that lives, in some sense, on the conformal boundary of Anti-de Sitter space. It would not have been difficult to recognize the entanglement entropies computed by the GPI as those of a CFT with a certain central charge. This would not have told us precisely which CFT is the correct dual theory, but it would have provided a helpful insight nevertheless. This alternate history is precisely the situation we find ourselves in when seeking a quantum gravity theory beyond AdS; and that is the setting in which our proposal may lead to some advances.

As mentioned above, state preparation in AdS/CFT is usually viewed as taking place in the CFT. The corresponding boundaries for the gravitational path integral necessarily reside in an asymptotic region, usually of a Euclidean spacetime geometry. An important change of perspective arrived with the introduction of MERA tensor networks Vidal (2007); Vidal (2008) as toy models for the bulk-to-boundary map of AdS/CFT Swingle (2012).

Certain tensor networks Pastawski et al. (2015); Hayden et al. (2016) were able to model the RT prescription (with min-cuts playing the role of minimal area surfaces) and the quantum-error-correcting properties Almheiri et al. (2015) of the bulk-to-boundary map. In these contexts, tensor networks were viewed both as models for CFT state preparation, and as discretizations of Lorentzian slices of AdS. Implicitly this supported a more direct notion of state preparation by Lorentzian slices.

Building on a reinterpretation of tensor networks as Euclidean evolution Caputa et al. (2021) in a $T\bar{T}$-deformed holographic CFT McGough et al. (2018); Hartman et al. (2019); Lewkowycz et al. (2020); Gorbenko et al. (2019); Belin et al. (2020), Ref. Araujo-Regado et al. (2023) proposed that a $T\bar{T}$-deformed CFT lives on any Cauchy slice of an asymptotically locally AdS spacetime. This “Cauchy slice holography” proposal is an important antecedent to our work, but our proposal differs significantly in several respects.

Our proposal is agnostic; its spirit is to deduce properties of quantum gravity from the properties of the states we define. In particular, we do not invoke a specific quantum mechanical theory that lives on the elliptic data manifold $\mathcal{J}(\sigma)$. As a result, we are able to operate in more general settings, both in that we do not require $\sigma$ to reside in an asymptotic region, and in that we need not base our construction on an asymptotically AdS spacetime in the first place.

Our $\mathcal{J}(\sigma)$ does not carry full initial value data but only elliptic data. As such, it would not be meaningful to say a priori that an elliptic data manifold $\mathcal{J}(\sigma)$ is associated with a particular Cauchy slice of a particular Lorentzian geometry. In our approach, however, it is critical that elliptic data are extracted from a given Lorentzian slice only after a small deformation of its initial data away from the real Lorentzian sector. The complexified classical solution furnishes a well-defined saddle-point candidate in the GPI that computes the norm of $\mathcal{J}(\sigma)$. The correspondence between $\mathcal{J}(\sigma)$ and its norm geometry defines the relation between quantum states and associated Lorentzian spacetime wedges. To our knowledge, this aspect is entirely novel to our proposal.

Among further technical differences to Ref. Araujo-Regado et al. (2023), the most significant is our choice of mixed-conformal elliptic boundary conditions. Dirichlet boundary conditions are not strongly elliptic Anderson (2008); Witten (2021), and the corresponding one-loop determinant need not therefore be well defined. Moreover, Dirichlet boundary conditions have been shown to lead to negative norm states Wall (2021). As shown in Appendix A, this problematic example does not survive with mixed-conformal boundary conditions.

Our geometric Hilbert space construction closely follows Marolf and Maxfield Marolf and Maxfield (2020) (see also Chen (2025)); our construction of operators follows Ref. Colafranceschi et al. (2024). These constructions have much older mathematical predecessors, especially the geometric assignment of state spaces and operators to boundaries and bordisms in the Atiyah–Segal formulation of topological and conformal field theory Atiyah (1988); Segal (2004).

In Marolf and Maxfield (2020), the closest recent work to ours, the focus is on the special cases where $\sigma=\varnothing$ (so that $\mathcal{J}(\sigma)$ is a closed universe or “baby universe” Marolf and Maxfield (2020)), or where $\mathcal{J}(\sigma)$ is asymptotically AdS in its norm geometry, so that $\mathcal{J}(\sigma)$ can be thought of as preparing a CFT state. A common feature key to our work is the use of the GPI to evaluate inner products and full traces. Here we consider more general elliptic data manifolds $\mathcal{J}(\sigma)$, whose boundaries need not be asymptotic; we establish a specific relation between deformed Lorentzian data and quantum states; and we explore the substructure of the gravitational quantum states by computing the Rényi entropies of subsets of $\sigma$.

In the family of examples studied in Sec. 6, we will find that the von Neumann entropy of a boundary subregion $\chi\subset\sigma$ is given by the maximin surface Wall (2014) homologous to $\chi$ in the norm geometry of $\mathcal{J}(\sigma)$, irrespective of all other parameters. This relates to extant work in two ways.

The maximin result shows that our approach, despite its superficial similarity, transcends the tensor network framework. The maximin surface does not lie on the Cauchy slice $\Sigma$ from whose deformation that elliptic data manifold $\mathcal{J}(\sigma)$ was obtained. It is not the minimal surface on $\Sigma$. A tensor network that discretizes $\Sigma$ cannot reproduce this result. There is presently no satisfactory notion of a relativistic causal development of a tensor network — i.e., there is no analogue of the Lorentzian spacetime wedge whose maximin surface defines the entropy.

The maximin result suggests, moreover, that the present approach may offer a path towards deriving a recent proposal for generalized entanglement wedges Bousso and Penington (2022); Bousso and Penington (2023). We will now turn to this and other interesting questions left for future work.

As described in Sec. 1.1, let $w_{\sigma}$ be a Lorentzian spacetime wedge with boundary $\sigma$. Let $\mathcal{J}(\sigma)$ be a quantum state whose complex norm geometry contains $w_{\sigma}$ as a fixed point set under CRT. Let $\chi\subset\sigma$, and let $S_{1}(\chi)$ be the von Neumann entropy of the reduced state on $\chi$.

In the language of the present paper, the generalized entanglement wedge proposal Bousso and Penington (2022); Bousso and Penington (2023) becomes the prediction

Here $A_{\rm maximin}$ is the area of the maximin surface Wall (2014) homologous to $\chi$ in $w_{\sigma}$. (We have stated the classical limit for simplicity.) Sec. 6 confirms this prediction in a class of examples.

Conversely, our proposal furnishes a framework in which a proof of the proposal Bousso and Penington (2022); Bousso and Penington (2023) may be constructed. (Our $w_{\sigma}$ is the spacelike complement of the “input region” of Bousso and Penington (2022); Bousso and Penington (2023); the latter will play no role in the proof. This distinguishes our approach from extant attempts Balasubramanian and Cummings (2023); Kaya et al. (2025) to prove the static limit Bousso and Penington (2022).) Indeed, when the maximin surface lies in the interior of $w_{\sigma}$, the proof Lewkowycz and Maldacena (2013) of the maximin prescription in AdS/CFT Ryu and Takayanagi (2006); Hubeny et al. (2007); Wall (2014) should generalize straightforwardly. In more general cases, the generalization would be nontrivial. We discuss this further at the beginning of Sec. 6.

Qualitatively novel challenges arise because in our general setting, $w_{\sigma}$ is finite and $\sigma$ is not asymptotic. Therefore, the maximin slice need not lie in the interior of $w_{\sigma}$. A simple example of this arises when $w_{\sigma}$ is the causal domain of a shell in Minkowski or AdS. We are currently analyzing examples of this type, along the lines of Sec. 6. This analysis may confirm Eq. (4), or it may necessitate a modification of Refs. Bousso and Penington (2022); Bousso and Penington (2023).

Our discussion so far has somewhat oversimplified the proposal Bousso and Penington (2022); Bousso and Penington (2023), which reduces to a maximin prescription only if max- and min-entanglement wedges of $\chi$, constructed from distinct prescriptions, coincide. This can fail even in the classical limit, for example if $w_{\sigma}$ has a trapped edge component. The framework presented here could potentially explain this enigmatic feature, by demonstrating that replica-symmetry-breaking configurations $(M,g_{ij})$ contribute significantly Akers et al. (2019) to the GPI that computes the entropies.

A crucial assumption in our work is that the quantum states we define have positive definite norm. This is not manifest when the inner product is defined by a GPI. Counterexamples are known when Dirichlet boundary conditions are used Wall (2021). The minimal assumption we need is that no such examples arise with some suitable choice of elliptic data (perhaps, but not necessarily, the mixed-conformal boundary conditions used here) — and in a class of gravity theories that should include, at the very least, General Relativity in 3+1 dimensions on the scales where it has been tested. As a first consistency check, we show in Appendix A that with conformal or mixed-conformal boundary conditions, the negative norm saddle does not persist as a leading norm contribution. It will be of great interest to search for counterexamples to norm positivity with these and other boundary conditions. Extra conditions on $\mathcal{J}(\sigma)$ in a neighborhood of $\sigma$ may also be necessary in order to ensure positivity.

The inner product, and hence the Hilbert spaces we construct, are defined only through the GPI. The GPI should be viewed as a perturbative asymptotic expansion about $G_{N}\to 0$, so Eq. (6) need not converge. A complete construction of $\mathbf{H}_{\sigma}$ would require a nonperturbative definition of the inner product. We must therefore regard the present construction as a GPI-approximation to an exact holographic Hilbert space in quantum gravity. This limitation is not unique to our work and is present in other GPI-based Hilbert-space definitions; see, for example, Ref. Marolf and Maxfield (2020). Alternatively, one may posit a nonperturbative completion with suitable axiomatic properties, including the reality and positivity properties above. This is the perspective taken in Ref. Colafranceschi et al. (2024) in the context of asymptotically locally AdS boundary data.

The involution map

acts on the elliptic data manifold $\mathcal{J}(\sigma)$ by complex conjugation of its data:

along with reversing its orientation, i.e. $[\omega_{J^{*}(\bar{\sigma})}]=[-\omega_{J(\sigma)}]$. Note that $\phi_{J^{*}}=\phi_{J}$ because the involution flips the orientation on both the manifold and its boundary, so the same boundary parametrization remains orientation-preserving.

We define an inner product on $\mathbf{V}_{\!\sigma}$ by acting on a vector with a dual vector obtained by involution,

and extending this map sesquilinearly. (We remind the reader of the index conventions introduced earlier; in particular, the repeated index $A$ denotes gluing across the argument $\sigma$.) By (5),

which implies that the inner product has conjugate symmetry:

We assume that the inner product is positive semidefinite: for any finite set of vectors $\{\mathcal{J}_{m}(\sigma)\}$ and complex coefficients $\{c_{m}\}$,

We do not assume positive-definiteness, because related, less general GPI constructions are known to contain nonzero states with zero norm Marolf and Maxfield (2020). Such states must be removed, by a standard process. Define the null subspace

Under the positivity assumption (25), $\mathbf{N}_{\sigma}$ is a subspace, and vectors in $\mathbf{N}_{\sigma}$ are orthogonal to all vectors. We define a pre-Hilbert space by identifying vectors in $\mathbf{V}_{\!\sigma}$ that differ by a null state:

$\mathbf{H}^{(0)}_{\sigma}$ is equipped with the induced inner product, which will be positive definite by assumption (25). Finally, we define the holographic Hilbert space $\mathbf{H}_{\sigma}$ to be the completion of $\mathbf{H}^{(0)}_{\sigma}$ in the norm $\|v\|:=\sqrt{\langle v,v\rangle}$. The special case $\mathbf{H}_{\sigma=\varnothing}$ is the closed universe (or “baby universe”) Hilbert space Marolf and Maxfield (2020); Abdalla et al. (2026).

In an abuse of notation, we will not distinguish between the vectors $\mathcal{J}^{A}(\sigma)\in\mathbf{V}_{\!\sigma}$ and the elements of $\mathbf{H}_{\sigma}$, even though the latter are equivalence classes under the above quotient and should strictly be denoted $[\mathcal{J}^{A}(\sigma)]$. The dual Hilbert space $\mathbf{H}^{*}_{\sigma}=\mathbf{H}_{\bar{\sigma}}$ is constructed analogously from $\mathbf{V}_{\bar{\sigma}}$.

Let $\mathcal{J}(\sigma)$ be a mixed-conformal elliptic data manifold. Recall that we use an upper abstract index to denote its role as a vector in Hilbert space, $\mathcal{J}^{A}(\sigma)$; that we reverse the orientation and complex conjugate the elliptic data of $\mathcal{J}^{A}(\sigma)$ to construct the corresponding bra-vector $\mathcal{J}^{*}_{A}(\bar{\sigma})$; that the norm of $\mathcal{J}^{A}(\sigma)$ is represented by gluing the two manifolds across $\sigma$, yielding a closed elliptic data manifold $\mathcal{J}^{A}(\sigma)\mathcal{J}^{*}_{A}(\bar{\sigma})$; and that the norm is evaluated by computing the GPI with boundary condition $\mathcal{J}^{A}(\sigma)\mathcal{J}^{*}_{A}(\bar{\sigma})$:

Suppose that the norm-GPI, Eq. (28), is dominated by exactly one leading saddle $(M,g_{ij})$. Then we call $(M,g_{ij})$ the norm geometry of $\mathcal{J}(\sigma)$, and we interpret it as the classical geometry that most closely corresponds to $\mathcal{J}(\sigma)$.

Note that $(M,g_{ij})$ is necessarily CRT invariant, i.e., its complexification $(M^{\mathbb{C}},g_{ij})$ admits an antiholomorphic involution $\boldsymbol{\Theta}:M^{\mathbb{C}}\to M^{\mathbb{C}}$ that exchanges $\mathcal{J}^{A}(\sigma)$ with $\mathcal{J}^{*}_{A}(\bar{\sigma})$ and thus preserves the elliptic data on $\mathcal{J}^{A}(\sigma)\mathcal{J}^{*}_{A}(\bar{\sigma})$. To see this, assume for contradiction that the complex manifold $\boldsymbol{\Theta}(M^{\mathbb{C}},g_{ij})$ obtained by reversing the orientation of $M^{\mathbb{C}}$ and complex conjugating its metric and fields is distinct from $(M,g_{ij})$. However, $\boldsymbol{\Theta}(M^{\mathbb{C}},g_{ij})$ would satisfy the same boundary condition $\mathcal{J}^{A}(\sigma)\mathcal{J}^{*}_{A}(\bar{\sigma})$. It would thus contribute as a second saddle, with the same real part of the action. This contradicts the assumption that only one leading saddle exists.

The fixed point set (or “fixed locus”) $\text{Fix}(\boldsymbol{\Theta})\subset M^{\mathbb{C}}$ defines a $(d+1)$-dimensional embedded real submanifold with real metric and field data. If the real metric of $\text{Fix}(\boldsymbol{\Theta})$ has Lorentzian signature,⁵⁵ 5 Some of the conditions in this and the previous paragraph could be adjusted or refined in response to further exploration of this framework. For example, it is not clear that Lorentzian signature needs to be imposed as a condition; it may be sufficient to restrict allowed elliptic data by requiring, e.g., that the real part of $h_{ij}$ have Euclidean signature and the real part of $K$ be nonvanishing. Similarly, we expect dominance of a single saddle to be generic in a broad class of states. However, a possible alternative to the stated dominance condition would require only that $(M,g_{ij})$ dominate among saddles. The smooth-approach condition should be either sharpened or omitted. We expect that explicit study of well-chosen examples will discriminate among these and other options. then we define

Suppose, moreover, that $\mathcal{J}(\sigma)$ can be smoothly deformed towards (a Cauchy slice of) $w[\mathcal{J}(\sigma)]$ in $M^{\mathbb{C}}$, defining a foliation $\mathcal{J}_{\eta}(\sigma)$, $0<\eta\leq 1$; and suppose, finally, that the norm geometries of the states $\mathcal{J}_{\eta}(\sigma)$ are given by a nested family of subsets of $M^{\mathbb{C}}$. Then we regard the Lorentzian wedge $w[\mathcal{J}(\sigma)]$ as the classical spacetime limit of $\mathcal{J}(\sigma)$.

Given a Lorentzian wedge $w_{\sigma}$, let

be the set of quantum states whose classical spacetime limit is $w_{\sigma}$. As noted in the introduction, this defines a classical-quantum correspondence $w_{\sigma}\leftrightarrow\mathbf{C}(w_{\sigma})$. By the deformation condition articulated after Eq. (29), $\mathbf{C}(w_{\sigma})$ is either empty or has an infinite number of elements. We expect that $s$ in Eq. (31) controls the tradeoff between the uncertainty in $([h_{ab}],K)$ and the uncertainty in their canonical conjugates around the classical initial data on $\Sigma$.⁶⁶ 6 The elementary quantum-mechanical analogue is a squeezed Gaussian centered on a fixed classical phase-space point $(x_{0},p_{0})$, $\psi_{\Delta}(x)\propto e^{-(x-x_{0})^{2}/(4\Delta^{2})+ip_{0}(x-x_{0})/\hbar}\propto e^{-(x-z_{\Delta})^{2}/(4\Delta^{2})}$, with $z_{\Delta}=x_{0}+2i\Delta^{2}p_{0}/\hbar$. Thus, at fixed $(x_{0},p_{0})$, changing the squeezing changes the displacement into complexified configuration space while leaving the classical state unchanged. $\mathbf{C}(w_{\sigma})$ can be made into a subspace of $\mathbf{H}_{\sigma}$ under a suitable definition of the classical limit of superpositions of elliptic data manifolds.

By definition, the elements of $\mathbf{C}(w_{\sigma})$ all have the same classical limit. But they differ as quantum states. In particular, we expect their entanglement structure to differ when $\sigma$ admits a bipartition $\sigma=\chi\sqcup\chi_{c}$. Correspondingly, we expect that the von Neumann entropy of the reduced state on $\chi$ will be the same for all elements of $\mathbf{C}(w_{\sigma})$; but the Rényi entropies may differ. Indeed, this is exactly the behavior we will find in the examples studied in Sec. 6.

Classically, unless $\sigma=\varnothing$, $w[\mathcal{J}(\sigma)]$ can be extended to a spacetime wedge $w^{\prime}[\mathcal{J}(\sigma)]\supsetneq w[\mathcal{J}(\sigma)]$ by adding extra initial data in a neighborhood of $\sigma$. But semiclassically, any smooth state in the larger spacetime wedge must have infinite entanglement across $\sigma$, and therefore the semiclassical state obtained by reducing the state of $w^{\prime}[\mathcal{J}(\sigma)]$ to $w[\mathcal{J}(\sigma)]$ will have divergent von Neumann entropy. It follows that $w[\mathcal{J}(\sigma)]$ cannot be so extended. In particular, modes of the quantum fields localized near $\sigma$ must be in an unentangled state similar to the Rindler vacuum Fulling (1973). Such a state should have quantum singularities on the boundary of $w_{\sigma}$ Candelas and Deutsch (1977), for example in the expectation value of the stress-energy tensor, similar to a “firewall” state Almheiri et al. (2013).

Let $(w_{\sigma},g_{ij})$ be a real-analytic manifold with real-analytic Lorentzian metric $g_{ij}$ and compact Cauchy slice $\Sigma$; we will refer to $\sigma=\partial\Sigma$ as the edge of $w_{\sigma}$. By allowing the coordinates to take complex values, $(w_{\sigma},g_{ij})$ can be complexified. Real analyticity implies that the extension $w_{\sigma}\hookrightarrow w_{\sigma}^{\mathbb{C}}$ is holomorphic in an open neighborhood of $w_{\sigma}$.

For simplicity, assume that $\Sigma$ corresponds to $t=0$, $r<1$, so $\sigma$ is given by $t=0$, $r=1$. Let $s(r,\ldots)$ be a smooth function on $\Sigma$ that is strictly positive for $r<1$ and vanishes for $r=1$. If $s$ is chosen small enough to lie in the holomorphic extension of $w_{\sigma}$, then

defines a real $d$-dimensional embedded submanifold $J_{s}(\sigma)$ with boundary $\sigma$. Complex phase space data $h_{ab},K_{ab}$ are induced on $J_{s}(\sigma)$ by $w_{\sigma}^{\mathbb{C}}$. We retain only the mixed-conformal data: $([h_{ab}],K)$ in the interior of $J_{s}(\sigma)$ and the induced metric $(\gamma_{pq})_{\sigma}$ on $\sigma$. This associates an elliptic data manifold, and thus a quantum state $\mathcal{J}_{s}(\sigma)\in\mathbf{H}_{\sigma}$, to $w_{\sigma}$.

The above deformation protocol $w_{\sigma}\to\mathcal{J}_{s}(\sigma)$ will have succeeded if $\mathcal{J}_{s}(\sigma)\in\mathbf{C}(w_{\sigma})$, i.e., if the resulting quantum state $\mathcal{J}_{s}(\sigma)$ has $w_{\sigma}$ as its classical spacetime limit:

where $w[\mathcal{J}_{s}(\sigma)]$ is defined by Eq. (29). In the family of examples studied in Sec. 6, we find that the above success criterion is met.

But this is not guaranteed; in fact there are several failure modes. By construction, the original complex solution $(M,g_{ij})$ that contains $w_{\sigma}$ as a real Lorentzian section is a saddle-point candidate for the norm-GPI of $\mathcal{J}_{s}(\sigma)$. However, it is not obvious that the saddle $(M,g_{ij})$ lies on the correct integration contour. Even if it does, the saddle $(M,g_{ij})$ need not dominate; and even if it dominates, there may not exist a smooth deformation of $\mathcal{J}_{s}(\sigma)$ to $w[\mathcal{J}_{s}(\sigma)]$ as described after Eq. (29). Let us exhibit an explicit example where Eq. (32) fails.

Let $w^{\rm dS}_{\varnothing}$ be Lorentzian de Sitter space in 3+1 dimensions, with metric

Let $\Sigma=\{t=0\}$, so $\sigma=\varnothing$; and let $s\in(0,\pi/2)$. Then Eq. (31) defines the state $\mathcal{J}_{s}(\varnothing)$ as a deformation of $\Sigma$. $\mathcal{J}_{s}(\varnothing)$ is a topological $S^{3}$ and carries the mixed-conformal data induced on $t=is$:

The norm-GPI has boundary conditions consisting of the disjoint union $\mathcal{J}_{s}(\varnothing)$ and $\mathcal{J}_{s}^{*}(\varnothing)$. There are two $S^{3}$-symmetric saddles. The first saddle is the equatorial belt $-is\leq t\leq is$ of Euclidean de Sitter (the round $S^{4}$). This geometry contains $w^{\rm dS}_{\varnothing}$ as the Lorentzian CRT fixed locus $\operatorname{Im}(t)=0$. The second saddle consists of two disconnected geometries, each larger than half of Euclidean de Sitter: $-i\pi/2\leq t\leq is$ and $-is\leq t\leq i\pi/2$. CRT exchanges the two components so $\text{Fix}(\boldsymbol{\Theta})=\varnothing\neq w^{\rm dS}_{\varnothing}$. The second saddle has real Euclidean action $I_{\rm first}-\sqrt{3}\pi/G_{N}$, and thus it dominates over the first saddle.

In the above example, the norm geometry of $\mathcal{J}_{s}(\sigma)$ fails to contain $w_{\sigma}$ regardless of the value of $s$. Suppose, however, that in some other example, we found that the norm geometry contains $w_{\sigma}$ but only for sufficiently large $s$. For these large $s$, it would be impossible to deform $\mathcal{J}_{s}(\sigma)$ to $w_{\sigma}$, or else we would have found small values of $s$ that succeed in preparing a quantum state with classical limit $w_{\sigma}$. Therefore large deformations, by themselves, cannot generate quantum states for $w_{\sigma}$ if small deformations cannot. One could consider eliminating the deformation criterion after Eq. (29). This would be physically difficult to justify, at least if we interpret the $s\to 0$ limit as a squeezing limit as outlined above.

If the assignment $w_{\sigma}\to\mathcal{J}_{s}(\sigma)$ succeeds, it will be far from unique. This immediately follows from the deformation criterion for the classical spacetime limit, and it is consistent with our discussion, in the previous subsection, of the cardinality of $\mathbf{C}(w_{\sigma})$. Indeed, there is considerable freedom in the choice of a real function $s$. Moreover, we could choose any other Cauchy slice of $w_{\sigma}$ as our starting point. All these different choices will in general define inequivalent quantum states on $\sigma$.