Skip to main content
Back to timeline
arXivSource publication:

A black hole's interior, read back from entanglement: with one unknown function, a neural ODE recovers the metric behind the horizon

Synopsis

In a three-dimensional eternal black hole, the time-dependent entanglement entropy of a pair of half-lines depends on the bulk metric only through one function, the local surface gravity written in terms of the timelike Killing norm. A neural ordinary differential equation fed this entropy reconstructs, from random initial metrics, the whole region probed by the extremal surfaces in charged BTZ black holes, including part of the interior, and yields the Hawking temperature and the critical surface; with two unknown metric functions the data no longer fix the geometry uniquely.

Interpretation

Time-dependent entanglement entropy can be inverted for a black hole metric, not only checked against a known one. Earlier holographic reconstructions mostly used static entanglement entropy, whose extremal surfaces stop outside the horizon and constrain only the exterior geometry; Hartman–Maldacena surfaces threading the Einstein–Rosen bridge pass behind the horizon, so this data is sensitive to the interior. On a charged BTZ background, the blackening factor learned from random initial metrics agrees with the true one over the probed region, from the asymptotic boundary through the horizon to the deepest point reached within the data window.

The reconstruction works because half-line entanglement dynamics senses the bulk metric through a single function. Rewriting the area and boundary time in terms of the timelike Killing norm shows both depend only on the local surface gravity, splitting the inverse problem into invertibility and uniqueness. Learning a single unknown function, the network converges on charged BTZ, and the learned metric yields a Hawking temperature and critical-surface location close to their exact values, neither of which is a training target.

With two unknown functions, what the data leaves is a symmetry of the probe rather than a numerical artifact. The half-line HM surface sits at fixed transverse coordinate, so its length is computed entirely within the two-dimensional metric, which after boundary normalization of the Killing time carries one diffeomorphism-invariant function. Three trainings reproduce the same data equally well, with final losses comparable to the single-function case, yet the learned geometries differ from each other and from the true metric.

Perspective

The discussion is limited to a pair of half-lines in a three-dimensional asymptotically AdS spacetime, whose HRT surface is a Hartman–Maldacena surface threading the Einstein–Rosen bridge, with neutral BTZ and charged BTZ as backgrounds; the single-function dependence and the degenerate family are extended in an appendix to planar black holes in any dimension.

Convergence to the correct metric in the single-function case relies on the real-analytic parametrization from the softplus activation, which restricts the search to the class of metrics where the inverse problem is unique; the superluminal-data result is reported as a property of the present reconstruction rather than a general no-go statement.