Python's Lunch Conjecture in Holography

• Python's Lunch Conjecture is a holographic conjecture linking exponential reconstruction cost to specific geometric features, such as bulges and constrictions, in entanglement wedges.
• It distinguishes between efficient and information-theoretic reconstructability by emphasizing the exponential gap when reconstruction is restricted to a single subregion.
• The conjecture is supported by tensor-network models, explicit spacetime examples, and cryptographic tests that demonstrate geometry-induced jumps in complexity.

Python’s Lunch Conjecture (PLC) is a conjectural relation in holography between the restricted computational complexity of reconstructing bulk information and the geometry of certain nonminimal extremal or quantum extremal surfaces. In its standard form, for a boundary region $\mathcal R$ whose entanglement wedge contains a constriction or appetizer $X^c_{\mathcal R}$ and a deeper bulge $X^b_{\mathcal R}$, the reconstruction cost is expected to scale as

$$\mathcal C(\mathcal R)\sim \exp\!\left(\frac{S_{\mathrm{gen}}(X^b_{\mathcal R})-S_{\mathrm{gen}}(X^c_{\mathcal R})}{2}\right),$$

with the exponent interpreted as the amount of post-selection that must be overcome when reconstructing the wedge beyond the outermost extremal surface (Brown et al., 2019). The conjecture was formulated to give a geometric account of the Harlow–Hayden hardness of decoding Hawking radiation and has since been elaborated through Petz-map reconstruction, explicit AdS and JT geometries, cryptographic tests, and tensor-network model building (Zhao, 2020).

1. Origin in decoding Hawking radiation and restricted complexity

The original motivation for PLC is the distinction between information-theoretic reconstructability and efficient reconstructability in black hole evaporation. After the Page time, the Hawking radiation contains the information needed to purify late modes and, in the entanglement-wedge sense, encodes interior data. The conjectural difficulty is not whether reconstruction exists, but whether it can be carried out efficiently when operations are restricted to a single side, such as the radiation system alone (Brown et al., 2019).

This distinction is central. In the unrestricted setting, one may use gates acting across both sides of the black hole–radiation system, and the relevant complexity can remain comparable to wormhole volume or action. In the restricted setting, one is allowed to act only on the radiation or only on one boundary subregion. PLC asserts that these two notions of complexity can differ exponentially. The proposed geometric reason is a specific obstruction in the Einstein–Rosen bridge: instead of a monotone throat, the bridge contains a min-max-min structure, with a narrow bottleneck, a wider interior bulge, and then another narrowing. That bulge is the “python’s lunch” (Brown et al., 2019).

In the strong form used in later work, operators in the simple or outer wedge are reconstructible with polynomial complexity, whereas operators behind the appetizer are exponentially complex to reconstruct. In semiclassical regimes, the generalized-entropy gap is typically of order $\mathcal O(1/G_N)$, so the conjectured reconstruction cost is exponentially large in $1/G_N$ (Engelhardt et al., 2024).

2. Geometric formulation

The conjecture is naturally stated in the language of RT/HRT and quantum extremal surfaces. For a boundary region $A$, the generalized entropy of an extremal surface $\gamma$ is

$$S_{\mathrm{gen}}(\gamma)=\frac{A[\gamma]}{4G_N}+S_{\Sigma[\gamma]},$$

and the entanglement wedge is determined by the minimal such surface (May et al., 2024). In the classical time-reflection-symmetric setting, one often distinguishes three surfaces associated with a boundary region $\mathcal R$: the globally minimal RT surface $X^c_{\mathcal R}$0, the constriction $X^c_{\mathcal R}$1, which is the outermost locally minimal surface, and the bulge $X^c_{\mathcal R}$2, an index-1 extremal surface lying deeper in the wedge (Arora et al., 2024).

In the QES language used in covariant treatments, the outer locally minimal nonminimal surface is often called the appetizer or throat, while the deeper nonminimal extremal surface is the bulge. The lunch is the region between the minimal QES and the appetizer, with the bulge controlling the complexity barrier beyond the appetizer (May et al., 2024). In this formulation, PLC is the claim that the complexity of reconstructing past the appetizer is exponential in the bulge-to-appetizer generalized-entropy gap (Brown et al., 2019).

Later work broadened the conjecture beyond the original spacelike-separated setup. Explicit Lorentzian examples were constructed with timelike-separated bulges and timelike-separated throats, and a third local QES type, the bounce, was identified. In that analysis, the Hessian of the generalized entropy yields an exhaustive trichotomy of nondegenerate QESs into throats, bulges, and bounces, and the conjecture was reformulated so that the relevant bulge and throat need not lie on a single Cauchy slice (Engelhardt et al., 2023). A notable implication is that the gravitational analogue of the tensor network governing reconstruction need not coincide with the time-reflection-symmetric slice, even when such a slice exists (Engelhardt et al., 2023).

3. Complexity mechanism: post-selection, Grover scaling, and the Petz map

The tensor-network motivation for PLC is a projective network picture. The network narrows at the appetizer, and the narrowing is modeled by projections onto fixed states. Reconstruction of information hidden behind the narrowing then requires inverting post-selection. If $X^c_{\mathcal R}$3 qubits are projected, naive measure-and-retry costs $X^c_{\mathcal R}$4, while a Grover-style protocol reduces this to $X^c_{\mathcal R}$5. This provides the microscopic rationale for the factor of $X^c_{\mathcal R}$6 in the exponent of the PLC formula (Brown et al., 2019).

A complementary perspective comes from entanglement-wedge reconstruction via the Petz map. In the code-subspace setup, with isometric embedding $X^c_{\mathcal R}$7, the boundary reconstruction of a bulk operator $X^c_{\mathcal R}$8 on a subsystem $X^c_{\mathcal R}$9 is

$X^b_{\mathcal R}$0

Under perfect recovery, this reconstruction can be written as a precursor,

$X^b_{\mathcal R}$1

depending on whether the entire relevant boundary system or only a subsystem is available (Zhao, 2020).

This reformulation makes the complexity comparison explicit. When the whole boundary is accessible, the reconstruction complexity is controlled by the horizontal circuit and tracks the volume or action of the wormhole segment from the bulk operator to the boundary. When one loses access to $X^b_{\mathcal R}$2 qubits, the reduced reconstruction circuit acquires a python’s-lunch structure with $X^b_{\mathcal R}$3, so

$X^b_{\mathcal R}$4

The same analysis gives a post-selection success probability of order $X^b_{\mathcal R}$5 and hence an implementation cost scaling like $X^b_{\mathcal R}$6 (Zhao, 2020). In this sense, PLC is not merely an abstract statement about bulk depth; it is a claim that partial loss of boundary access can induce an exponential jump in reconstruction cost.

4. Explicit geometric realizations

Several papers constructed explicit spacetimes exhibiting python’s-lunch structure. In Jackiw–Teitelboim gravity coupled to a massless scalar field, fully back-reacted solutions were found in which the minimal QES, bulges, and appetizers all lie inside the horizon. The null focusing equation implies that any extremum of the dilaton lies behind the horizon when the null energy condition holds. In this framework, Python’s lunches arise for nonchiral deformations, while purely chiral deformations produce no classical lunch (Bak et al., 2021).

In $X^b_{\mathcal R}$7-dimensional AdS gravity with positive-energy dust, a broad class of exact solutions was constructed, culminating in a localized dust profile for which the wormhole width function develops two local minima and an intermediate local maximum. This gives a direct metric realization of a python’s-lunch geometry in asymptotically AdS spacetime without exotic matter (Bao et al., 2020).

Systematic studies of bulges then revealed that they differ sharply from familiar RT surfaces. Bulges can spontaneously break continuous or discrete spatial isometries, can be sensitive to the infrared regulator, can self-intersect, and can probe entanglement shadows, orbifold singularities, and compact internal spaces such as the sphere in $X^b_{\mathcal R}$8. Within the PLC interpretation, these features imply that holographic complexity can behave qualitatively differently from entanglement entropy. The same analysis found that extended black brane interiors have a non-extensive complexity and that multi-boundary wormholes exhibit plateau behavior, where adding more boundaries does not further reduce the reconstruction cost once an optimal subsystem has been included (Arora et al., 2024).

An asymptotically flat realization was later developed using multi-boundary Brill–Lindquist wormholes. There, index-1 extremal surfaces in the wormhole interior were interpreted as candidate bulges, and their area differences were used to compute the restricted complexity of decoding Hawking radiation. The $X^b_{\mathcal R}$9 and $$\mathcal C(\mathcal R)\sim \exp\!\left(\frac{S_{\mathrm{gen}}(X^b_{\mathcal R})-S_{\mathrm{gen}}(X^c_{\mathcal R})}{2}\right),$$0 models reproduce a Page curve, while in the $$\mathcal C(\mathcal R)\sim \exp\!\left(\frac{S_{\mathrm{gen}}(X^b_{\mathcal R})-S_{\mathrm{gen}}(X^c_{\mathcal R})}{2}\right),$$1 case the “lunch transition” occurs near

$$\mathcal C(\mathcal R)\sim \exp\!\left(\frac{S_{\mathrm{gen}}(X^b_{\mathcal R})-S_{\mathrm{gen}}(X^c_{\mathcal R})}{2}\right),$$2

close to the benchmark value $$\mathcal C(\mathcal R)\sim \exp\!\left(\frac{S_{\mathrm{gen}}(X^b_{\mathcal R})-S_{\mathrm{gen}}(X^c_{\mathcal R})}{2}\right),$$3 emphasized in the Brown-et-al. story (Gupta, 7 Jul 2025).

5. Tests, refinements, and related frameworks

PLC has also been probed indirectly through information-theoretic and cryptographic constructions. A 2024 study used conditional disclosure of secrets to derive a boundary-information consequence of the projective tensor-network model: the mutual information between suitable boundary subregions should be lower bounded linearly by an area difference associated with the lunch. Weakened versions of this statement were proven in asymptotically $$\mathcal C(\mathcal R)\sim \exp\!\left(\frac{S_{\mathrm{gen}}(X^b_{\mathcal R})-S_{\mathrm{gen}}(X^c_{\mathcal R})}{2}\right),$$4 spacetimes satisfying the null energy condition, and explicit examples were checked. At the same time, the paper found counterexamples to a stronger unrestricted linear mutual-information bound in vacuum $$\mathcal C(\mathcal R)\sim \exp\!\left(\frac{S_{\mathrm{gen}}(X^b_{\mathcal R})-S_{\mathrm{gen}}(X^c_{\mathcal R})}{2}\right),$$5 with two intervals, showing that naive forms of the conjectural bound cannot be universal (May et al., 2024).

Another line of work connected PLC to entanglement spoofing. For geometric large-$$\mathcal C(\mathcal R)\sim \exp\!\left(\frac{S_{\mathrm{gen}}(X^b_{\mathcal R})-S_{\mathrm{gen}}(X^c_{\mathcal R})}{2}\right),$$6 EFI pairs, statistical far-ness together with computational indistinguishability implies that at least one state must contain a python’s lunch. For pseudoentangled state ensembles with a semiclassical bulk dual, the maximally mixed state over the low-entanglement ensemble must contain a lunch. Because a python’s lunch must lie behind an event horizon in this analysis, black holes become the exclusive semiclassical gravitational source of entanglement spoofing (Engelhardt et al., 2024).

Tensor-network model building has also complicated the simplest version of PLC. A 2026 paper argued that generic random tensor networks do not naturally realize the PLC exponent, because the conjecture tacitly assumes a property dubbed computational covariance: for every foliation by everywhere non-contracting cuts, each elementary step should be an approximate isometry or unitary of subexponential complexity in $$\mathcal C(\mathcal R)\sim \exp\!\left(\frac{S_{\mathrm{gen}}(X^b_{\mathcal R})-S_{\mathrm{gen}}(X^c_{\mathcal R})}{2}\right),$$7. To address this, the paper introduced twirled perfect tensor networks, which preserve the perfect-tensor isometric structure while avoiding stabilizer triviality. These networks satisfy a PLC-type upper bound and obey a lattice RT formula for arbitrary boundary subregions, but they also exhibit a discrete limitation from local postselection that the authors argue should be absent in smooth gravity (Arora et al., 22 May 2026).

6. Status, caveats, and overloaded terminology

PLC remains a conjecture rather than a theorem of semiclassical gravity. The geometric, tensor-network, and cryptographic results give multiple forms of evidence, but they do not amount to a general proof. Even in explicit tensor-network models, the precise exponent can depend on structural assumptions that are not generic. The distinction between unrestricted and restricted complexity, the role of code-subspace choice, and the validity of projective tensor-network heuristics are all essential to interpreting the conjecture correctly (Brown et al., 2019).

Several common misconceptions follow from omitting these qualifiers. First, PLC is not merely a restatement of complexity-volume or complexity-action duality: those proposals address unrestricted complexity, whereas PLC concerns one-sided or subsystem-restricted reconstruction (Brown et al., 2019). Second, bulges are not simply “larger RT surfaces”: they are nonminimal extremal or QES objects with distinct Morse-theoretic and covariant properties, and later work shows that they need not preserve symmetries or even lie on a common Cauchy slice with the relevant constrictions (Arora et al., 2024). Third, not every tensor-network realization of a lunch reproduces the original exponent; the random-tensor-network critique and the introduction of computational covariance make this point explicit (Arora et al., 22 May 2026).

The acronym PLC is also overloaded outside holography. In TFNP complexity theory, PLC denotes polynomial long choice, the class of search problems reducible to Long Choice; this use is unrelated to Python’s Lunch in gravity (Pasarkar et al., 2022, Ishizuka, 2023). An unrelated 2018 number-theoretic paper also used the phrase “Python’s Lunch Conjecture” for a conjecture about LCM-closed sets equivalent to Frankl’s union-closed sets conjecture (Fischer, 2018). In contemporary high-energy usage, however, “Python’s Lunch Conjecture” refers to the holographic reconstruction-complexity conjecture tied to bulges, constrictions, and generalized-entropy gaps.