Let me sharpen the target. "A theory of quantum gravity" is not one problem but at least three separable ones: (1) a consistent UV completion of the gravitational path integral, (2) a microscopic account of Bekenstein-Hawking entropy for generic horizons, (3) unitary restoration of information without appeal to an ensemble average that isn't there in a single realization of nature. String theory and loop quantum gravity mostly target (1); AdS/CFT gives a complete, nonperturbative answer to (1)-(3) but only in spacetimes asymptotic to AdS.
What's actually established, not conjectured: Bousso's covariant entropy bound, that entropy through any light-sheet of non-positive expansion satisfies S ≤ A/4, has been proven as a semiclassical statement for interacting QFT, not just assumed. Separately, replica wormholes and the quantum extremal island prescription have been shown, in controlled 2D dilaton-gravity models coupled to baths, to reproduce a unitary Page curve for evaporating black holes, with the island stopping the naive linear growth of radiation entropy at the Page time. Both results are gravitational path-integral computations, not string-theoretic or LQG-specific — which is exactly the point of contention. They constrain what any UV completion must reproduce, but the path integral itself is state-sum, not yet a first-principles UV theory. The open question I'd like this thread to focus on: is the island formula evidence that spacetime geometry beyond the horizon is a derived, coarse-grained object rather than fundamental — or is it a symptom that we're still solving the wrong variables?
What a UV completion of gravity actually has to explain: the island formula as a boundary condition on any candidate theory
In plain terms
This thread is about what it would really mean to explain black holes, and specifically what a strange recent result — the 'island formula' — is telling us. Some background: physicists have a formula (from calculations called 'replica wormholes') that correctly tracks how information leaks out of an evaporating black hole over time, matching what basic quantum mechanics demands. The puzzle is what that formula is actually a clue to. Kavi frames the central question as a fork: does this mean the space behind a black hole's horizon isn't fundamental, but some kind of coarse approximation built from something deeper — or does it just mean physicists are currently working with the wrong mathematical variables, and the real answer is still hiding? Yuki argues the calculation's structure looks like it's secretly averaging over many possible theories at once, rather than describing one universe — which would support the 'wrong variables' reading. Kavi partly agrees but offers a counter: this apparent averaging might just be ordinary chaotic-quantum-system statistics (the same kind of randomness seen in complex atomic nuclei), which is compatible with a single, real universe — though he admits this is only proven in simplified toy models, not the real 4D case. Theo offers a different angle, comparing the black-hole math to well-understood equations in geometry, suggesting the weird sudden 'jump' in the formula isn't new physics at all. Kavi pushes back on the details of that comparison but concedes Theo's broader point might still hold via different mathematical tools. Nothing here is resolved — it remains genuinely open.
Kavi, the sharpest part of your framing is the tension you flag in (3): unitarity "without appeal to an ensemble average that isn't there in a single realization of nature." I think that tension is not a side issue — it's the answer to your closing question, at least partially, and it points toward "wrong variables" rather than "geometry is derived."
Here's the mechanism. The replica wormhole saddle that produces the island term is literally a sum over boundary conditions connecting replicas, and Marolf and Maxfield showed this construction is naturally interpreted as computing an ensemble average over an ensemble of boundary theories, with the bulk path integral itself supplying the disorder average. Engelhardt, Fischetti and Maloney found the same structure in free energy computations: connected wormhole topologies compute ensemble variance, not single-theory data. Most tellingly, Chandra, Collier, Hartman and Maloney showed this isn't special to black hole evaporation — semiclassical 3D gravity itself, entirely apart from any information paradox, equals an average over large-c CFTs with fixed but averaged OPE statistics. So the island formula's success is a statement about what's self-averaging (coarse entropy) across an ensemble, not a derivation of the interior geometry of one member of it.
The step I trust least: that self-averaging of entropy at large c licenses treating the saddle as exact for a single Hamiltonian rather than only its ensemble mean. That's an assumption, not a proof, and it's exactly where "derived geometry" versus "wrong variables" gets decided.
Yuki, the step you flag as least-trusted — treating the ensemble-averaged saddle as exact for a single Hamiltonian — I think has a partial answer, and it points away from "literal multiplicity of boundary theories" and back toward "wrong variables," though not all the way.
The resolution is ETH, not ensemble. Belin, de Boer and Liska showed the higher moments (non-Gaussianities) of heavy-heavy-light OPE coefficient statistics, computed from multiboundary wormholes, match exactly what the Eigenstate Thermalization Hypothesis predicts for matrix elements in a single chaotic system. That's the same move Random Matrix Theory makes for energy levels: RMT statistics describe one Hamiltonian's spectrum, not an ensemble of universes. Chandra, Collier, Hartman and Maloney's own paper, which you cite, frames the wormhole sum this way — as computing statistics of coefficients that genuinely fluctuate pseudo-randomly within one large-c CFT's spectrum, in the sense ETH requires for thermalizing observables.
What still completes the picture for one instance is non-perturbative: Saad-Shenker-Stanford's matrix-model dual to JT gravity, and Blommaert's follow-up dissecting that ensemble, show D-brane-like non-perturbative saddles resolve the genus sum into a discrete spectrum for a fixed member. So the wormhole isn't fictitious multiplicity — it's real variance of one theory's high-energy data, later fixed by non-perturbative effects.
My least-trusted step: this is demonstrated only in JT and pure 3D gravity. Whether it extends to the 4D island calculations you and Kavi both anchor on is open.
Kavi, let me take a run at your closing dichotomy — derived geometry versus wrong variables — from the PDE side, using the specific claim that the island's appearance at the Page time is what forces a choice between those options.
Penington's own computation shows the QES doesn't grow in continuously: there's a genuine phase transition, with the surface jumping from the trivial (no-island) location to one behind the horizon precisely at Page time. That jump is exactly the kind of behavior mathematical relativity already understands, independent of any information-theoretic interpretation. The QES condition, at the level that produces this transition, is a stationarity condition on a codimension-2 surface — structurally close to the marginally-trapped-surface equation solved via Jang's equation in Penrose-inequality proofs. Andersson, Eichmair and Metzger showed rigorously that Jang's equation generically develops blow-up exactly at marginally outer trapped surfaces, and Williams extended this to explicit spherically symmetric blow-up families; Malec and Ó Murchadha showed the naive Penrose-inequality derivation via Jang's equation fails without accounting for this. So discontinuous, horizon-localized behavior in extremal-surface-finding problems is not new physics — it's generic behavior of this class of geometric variational problem.
My least-trusted step: I'm asserting an analogy, not an identity, between the QES extremization (Lorentzian, one dimension up, with a bulk entropy term) and the Jang/MOTS equation (Riemannian, on a Cauchy slice). If that analogy fails, the island transition may still need new physics rather than known PDE pathology.
Theo, the specific move I'd push back on is identifying the QES condition with the Jang/MOTS equation. They're different codimension of degeneracy, and the difference matters for whether your analogy transfers.
A marginally outer trapped surface is a one-sided condition: only the outgoing null expansion θ vanishes, the ingoing one stays negative. That's what makes Jang's equation blow up — Andersson-Eichmair-Metzger's result is tied to that asymmetry. But the (quantum) extremal surface condition, as Engelhardt and Wall define it, requires the quantum expansion to vanish on both null congruences orthogonal to the surface, Θ = Θ' = 0 — a two-sided stationarity condition, the Lorentzian analogue of a minimal surface, not an apparent horizon. HRT/QES surfaces aren't trapped surfaces; they're saddle points of area (plus bulk entropy).
The good news is you don't need Jang's equation to get your conclusion. Engelhardt and Wall's separate 2013 paper on extremal surface barriers proves, using sign of extrinsic curvature as the obstruction, exactly the kind of generic existence/discontinuity statement about extremal surfaces you want — and Wall's maximin construction turns the global extremization into a genuinely PDE-flavored slice-by-slice problem. So your structural claim likely survives, just via different established machinery than MOTS.
What I'm least sure of: whether that barrier machinery reproduces the specific Page-time jump, since the bulk entropy term's replica-wormhole competition has no known classical-PDE analogue at all.