Universal Horizon Models

• Universal Horizon Models refer to two frameworks: one in Lorentz-violating gravity defining absolute causal boundaries in black hole solutions, and another in reinforcement learning modeling n-step future state distributions.
• In gravitational physics, universal horizons trap all excitations by employing a preferred timelike æther field, with precise metric conditions and Smarr relations establishing surface gravity and thermodynamic behaviour.
• In reinforcement learning, UHM generalize one-step dynamics into generative models that predict future states over multiple steps, demonstrating measurable performance improvements on benchmark tasks.

Searching arXiv for the cited paper and related work on universal horizons/UHM. Universal Horizon Models (UHM) denotes, in the supplied arXiv literature, two distinct research usages. In Lorentz-violating gravity, especially Einstein-Æther theory and the infrared limit of Hořava gravity, the term is used as an organizing label for black-hole solutions and associated mechanics built around the universal horizon, the hypersurface on which the preferred-flow field becomes orthogonal to the stationary Killing vector and which traps excitations of arbitrarily high velocity (Cropp et al., 2013). In a separate 2026 reinforcement-learning usage, UHM denotes a generative model of the $n$-step future state distribution under a fixed policy $\pi$ that generalizes both one-step dynamics models and Geometric Horizon Models (Chung et al., 15 May 2026). The dominant and historically earlier meaning in the cited corpus concerns Lorentz-violating gravitation; that framework provides the conceptual core of the term’s development.

1. Foundational definition in Lorentz-violating gravity

In Einstein-Æther theory and Hořava-Lifshitz gravity, local Lorentz invariance is violated by the introduction of a unit timelike preferred-flow field $u^a$, or equivalently a preferred foliation. Because matter fields can propagate arbitrarily fast relative to the metric light cone, the conventional Killing horizon does not by itself define the relevant causal boundary. The universal horizon is instead defined by the condition

$u\cdot\chi \equiv g_{ab}u^a\chi^b = 0,$

where $\chi^a$ is the static or stationary time-translation Killing field (Berglund et al., 2012).

For static, spherically symmetric geometries, a standard metric ansatz is

$$ds^2 = -e(r)\,dt^2 + \frac{dr^2}{f(r)} + r^2\,d\Omega^2,$$

with an æther field

$$u^a = \bigl(u^t(r),u^r(r),0,0\bigr), \qquad u^a u_a=-1.$$

In these coordinates the universal-horizon condition becomes $e(r_{UH})\,u^t(r_{UH})=0$ (Cropp et al., 2013). In the exact solutions summarized in the literature, one finds $r_{UH}=r_0/2$ for the $c_{123}=0$ branch and $\pi$0 for the $\pi$1 branch [(Cropp et al., 2013); (Berglund et al., 2012)].

The key physical content is that all excitations, whatever their speed, must move forward in æther time. Once $\pi$2 changes sign, forward-in-$\pi$3 propagation requires decreasing $\pi$4, so signals inside the universal horizon cannot reach asymptotic infinity (Berglund et al., 2012). By contrast, the Killing horizon, defined by $\pi$5 or $\pi$6, only traps finite subluminal modes (Cropp et al., 2013).

A later khronometric formulation makes the same point in preferred-foliation language. There the unit-timelike aether vector is constructed from a scalar khronon field $\pi$7, and a universal horizon is the leaf $\pi$8 satisfying

$\pi$9

where $u^a$0 is the stationary Killing vector and $u^a$1 (Liberati et al., 2017). The nondegeneracy condition $u^a$2 ensures that the preferred slicing “bunches up” on a genuine horizon leaf rather than on a degenerate surface.

2. Geometry, surface gravity, and exact solution families

Universal-horizon constructions are usually developed in Einstein-Æther theory, whose action is written as

$u^a$3

or, in the normalization of Mohd’s thermodynamic treatment,

$u^a$4

The æther sector is controlled by couplings $u^a$5 through the tensor

$u^a$6

with the unit constraint $u^a$7 enforced by a Lagrange multiplier [(Berglund et al., 2012); (Mohd, 2013)].

A physically unambiguous peeling definition of universal-horizon surface gravity is

$u^a$8

and, equivalently in the static spherical setup,

$u^a$9

(Cropp et al., 2013). In khronometric notation the same quantity is expressed as

$u\cdot\chi \equiv g_{ab}u^a\chi^b = 0,$0

and is constant over $u\cdot\chi \equiv g_{ab}u^a\chi^b = 0,$1 (Liberati et al., 2017).

The literature develops several exact branches.

Branch	Characteristic conditions	Horizon data stated in the literature
$u\cdot\chi \equiv g_{ab}u^a\chi^b = 0,$2	“infinite-speed scalar mode”	$u\cdot\chi \equiv g_{ab}u^a\chi^b = 0,$3; in one solution $u\cdot\chi \equiv g_{ab}u^a\chi^b = 0,$4 (Berglund et al., 2012)
$u\cdot\chi \equiv g_{ab}u^a\chi^b = 0,$5	“zero-speed scalar mode”	$u\cdot\chi \equiv g_{ab}u^a\chi^b = 0,$6; in one solution $u\cdot\chi \equiv g_{ab}u^a\chi^b = 0,$7 (Berglund et al., 2012)
Generic couplings	$u\cdot\chi \equiv g_{ab}u^a\chi^b = 0,$8	asymptotic expansions exist, with regularity at the spin-0 horizon reducing to a one-parameter family (Bhattacharyya et al., 2014)

Bhattacharyya and Mattingly gave a classification of all four-dimensional static and spherically symmetric universal-horizon solutions with maximally symmetric asymptotics—flat, de Sitter, and anti-de Sitter—and showed that any spherically symmetric solution in Hořava-Lifshitz gravity with a universal horizon is also a solution of Einstein-Æther theory (Bhattacharyya et al., 2014). This establishes a broad equivalence of the two theories’ solution spaces under these symmetries.

3. Ray tracing, dispersion, and Hawking emission

The ray-tracing analysis of universal versus Killing horizons in Einstein-Æther black holes studies high-frequency excitations with superluminal dispersion in the æther frame,

$u\cdot\chi \equiv g_{ab}u^a\chi^b = 0,$9

(Cropp et al., 2013). For conserved energy $\chi^a$0, approaching the universal horizon forces $\chi^a$1, hence $\chi^a$2. All superluminal rays “pile up” exactly at $\chi^a$3, demonstrating universal trapping (Cropp et al., 2013).

The same work reports evidence that Hawking radiation is associated with the universal horizon, while the lingering of ray trajectories near the Killing horizon hints at reprocessing there (Cropp et al., 2013). In this picture, the universal horizon is the true causal barrier for arbitrarily fast excitations and sets the universal Hawking flux temperature,

$\chi^a$4

while the Killing horizon acts as an energy-dependent scattering or reprocessing layer for the low-energy portion of the spectrum (Cropp et al., 2013).

A distinct conclusion was reached in the collapse-based dispersive-field analysis of stationary black holes with an inner universal horizon. There the scalar field action includes a quartic spatial derivative term in the æther frame, yielding the eikonal relation

$\chi^a$5

and the mode matching across a collapsing shell is found to be adiabatic at late time (Michel et al., 2015). In particular, the mixing coefficient between positive-norm inside modes and the negative-norm high-momentum WKB partner is exponentially suppressed, so no quanta are produced near the universal horizon. The late-time flux is then thermal with

$\chi^a$6

fixed by the surface gravity of the Killing horizon (Michel et al., 2015).

These two strands define a genuine controversy. One line of work associates Hawking temperature with $\chi^a$7 [(Cropp et al., 2013); (Liberati et al., 2017)]; another concludes that the universal horizon should play no role in the thermodynamical properties of these black holes and that the thermal flux is governed by $\chi^a$8 (Michel et al., 2015). This suggests that the answer is sensitive to the choice of collapse model, dispersive sector, and notion of thermodynamical observables.

4. Mechanics, Smarr relations, and the first-law program

Berglund, Bhattacharyya, and Mattingly constructed a Smarr relation for static spherical universal horizons in Einstein-Æther theory by defining a divergence-free antisymmetric two-form $\chi^a$9, using Gauss’s law, and equating its flux at infinity to its flux at the universal horizon (Berglund et al., 2012). The total mass in the æther rest frame is

$$ds^2 = -e(r)\,dt^2 + \frac{dr^2}{f(r)} + r^2\,d\Omega^2,$$0

and the resulting Smarr relation is

$$ds^2 = -e(r)\,dt^2 + \frac{dr^2}{f(r)} + r^2\,d\Omega^2,$$1

where $$ds^2 = -e(r)\,dt^2 + \frac{dr^2}{f(r)} + r^2\,d\Omega^2,$$2 and

$$ds^2 = -e(r)\,dt^2 + \frac{dr^2}{f(r)} + r^2\,d\Omega^2,$$3

(Berglund et al., 2012). For the one-parameter family of regular solutions, variations satisfy

$$ds^2 = -e(r)\,dt^2 + \frac{dr^2}{f(r)} + r^2\,d\Omega^2,$$4

which is the modified first law in the vacuum, spherically symmetric case (Berglund et al., 2012).

Mohd’s Noether-charge treatment applied Wald’s formalism directly to the universal horizon rather than to the Killing bifurcation surface, where the æther diverges (Mohd, 2013). Using a cross-section $$ds^2 = -e(r)\,dt^2 + \frac{dr^2}{f(r)} + r^2\,d\Omega^2,$$5 of the universal horizon, with induced 2-metric $$ds^2 = -e(r)\,dt^2 + \frac{dr^2}{f(r)} + r^2\,d\Omega^2,$$6, area $$ds^2 = -e(r)\,dt^2 + \frac{dr^2}{f(r)} + r^2\,d\Omega^2,$$7, extrinsic-curvature trace $$ds^2 = -e(r)\,dt^2 + \frac{dr^2}{f(r)} + r^2\,d\Omega^2,$$8, and surface gravity defined by

$$ds^2 = -e(r)\,dt^2 + \frac{dr^2}{f(r)} + r^2\,d\Omega^2,$$9

the horizon contribution to the Hamiltonian variation becomes

$$u^a = \bigl(u^t(r),u^r(r),0,0\bigr), \qquad u^a u_a=-1.$$0

leading to a first law written as

$$u^a = \bigl(u^t(r),u^r(r),0,0\bigr), \qquad u^a u_a=-1.$$1

with $$u^a = \bigl(u^t(r),u^r(r),0,0\bigr), \qquad u^a u_a=-1.$$2 in the stated normalization (Mohd, 2013).

The thermodynamic interpretation, however, is not uniform across all exact solutions. In the infrared limit of Hořava gravity, tests of the first law on several exact universal-horizon families show that a simple mechanical and thermodynamical interpretation is problematic outside the simplest $$u^a = \bigl(u^t(r),u^r(r),0,0\bigr), \qquad u^a u_a=-1.$$3-dimensional static, asymptotically flat sector (Liberati et al., 2017). In that sector one finds

$$u^a = \bigl(u^t(r),u^r(r),0,0\bigr), \qquad u^a u_a=-1.$$4

and therefore

$$u^a = \bigl(u^t(r),u^r(r),0,0\bigr), \qquad u^a u_a=-1.$$5

(Liberati et al., 2017). By contrast, in asymptotically AdS solutions the entropy becomes an unwieldy functional not simply proportional to area, and in $$u^a = \bigl(u^t(r),u^r(r),0,0\bigr), \qquad u^a u_a=-1.$$6-dimensional rotating AdS solutions the Clausius equations are not simultaneously integrable, so no consistent $$u^a = \bigl(u^t(r),u^r(r),0,0\bigr), \qquad u^a u_a=-1.$$7 exists (Liberati et al., 2017).

5. Dynamical formation and quasilocal generalization

A stationary definition based on $$u^a = \bigl(u^t(r),u^r(r),0,0\bigr), \qquad u^a u_a=-1.$$8 was generalized to dynamical, spherically symmetric spacetimes by introducing

$$u^a = \bigl(u^t(r),u^r(r),0,0\bigr), \qquad u^a u_a=-1.$$9

so that the dynamical universal horizon is determined by

$e(r_{UH})\,u^t(r_{UH})=0$0

(Tian et al., 2015). In the analytical collapse model of a spherically symmetric star with finite thickness, using the $e(r_{UH})\,u^t(r_{UH})=0$1 limit and the solution

$e(r_{UH})\,u^t(r_{UH})=0$2

the apparent horizon radius is $e(r_{UH})\,u^t(r_{UH})=0$3, and the universal-horizon radius $e(r_{UH})\,u^t(r_{UH})=0$4 is obtained explicitly from the condition $e(r_{UH})\,u^t(r_{UH})=0$5 (Tian et al., 2015). Under the assumptions $e(r_{UH})\,u^t(r_{UH})=0$6, $e(r_{UH})\,u^t(r_{UH})=0$7, $e(r_{UH})\,u^t(r_{UH})=0$8, and suitable initial radius $e(r_{UH})\,u^t(r_{UH})=0$9, the weak energy conditions hold throughout the collapse down to—and even inside—the moment the star crosses its universal horizon (Tian et al., 2015). The universal horizon always lies strictly inside the apparent horizon, and an absolute causal boundary forms as the stellar surface shrinks past $r_{UH}=r_0/2$0 (Tian et al., 2015).

A more general quasilocal formalism replaces the stationary condition by an optical-scalar criterion. In a spacetime endowed with a preferred foliation, one considers compact spacelike 2-surfaces $r_{UH}=r_0/2$1, their transverse projector $r_{UH}=r_0/2$2, and the unique unit spacelike vector $r_{UH}=r_0/2$3 in the $r_{UH}=r_0/2$4-$r_{UH}=r_0/2$5 plane. The 2-expansion of $r_{UH}=r_0/2$6 is

$r_{UH}=r_0/2$7

A quasilocal universal horizon is then defined by

$r_{UH}=r_0/2$8

(Maciel, 2015). In spherical symmetry,

$r_{UH}=r_0/2$9

This condition is quasilocal and does not require Killing or Kodama symmetry (Maciel, 2015).

The quasilocal analysis yields several structural results. There are no quasilocal universal horizons for FLRW spacetimes for any scale factor function; universal horizons must lie in trapped or antitrapped regions because $c_{123}=0$0 implies $c_{123}=0$1 and hence $c_{123}=0$2; and in truncated Hořava-Lifshitz gravity the near-center formation of a smooth universal horizon is possible only in the window

$c_{123}=0$3

under the regularity assumptions stated in the paper (Maciel, 2015). Outside this window, the analysis indicates that the horizon would have to appear at finite radius with finite area.

6. Symmetry algebra, entropy proposals, and later terminological reuse

A near-horizon symmetry program has also been formulated for universal horizons. In Hořava-Lifshitz gravity, horizon-preserving diffeomorphisms in the $c_{123}=0$4 plane are written as

$c_{123}=0$5

with the boundary condition

$c_{123}=0$6

Closure of the algebra requires

$c_{123}=0$7

after which the Lie bracket takes the closed form

$c_{123}=0$8

(Mattingly et al., 2018). This is structurally identical to the classical algebra in the Killing-horizon case under the identifications

$c_{123}=0$9

No explicit central extension is computed, because no Euclidean regularization or preferred quantum state is yet available in the universal-horizon case (Mattingly et al., 2018). Still, the appearance of $\pi$00 in the algebra matches its role in tunneling and first-law analyses, strengthening the analogy between universal-horizon and Killing-horizon thermodynamics (Mattingly et al., 2018).

A separate and much later use of the acronym UHM appears in offline reinforcement learning. There, “Universal Horizon Models” are defined as a generative model of the $\pi$01-step future state distribution under a fixed policy $\pi$02,

$\pi$03

with $\pi$04 as an explicit horizon input (Chung et al., 15 May 2026). Choosing $\pi$05 recovers a one-step dynamics model; sampling $\pi$06 and marginalizing recovers the normalized discounted successor measure of a Geometric Horizon Model, so GHM is a special case of UHM with a fixed geometric horizon distribution (Chung et al., 15 May 2026). The paper introduces a winsorized geometric measure $\pi$07, a corresponding truncated geometric law $\pi$08, and a value-learning operator $\pi$09 whose fixed point is $\pi$10 (Chung et al., 15 May 2026). On 100 OGBench tasks, the reported average success rates are 55% on 50 standard tasks, 39% on 25 noisy or suboptimal-data tasks, and 22% on 25 long-horizon-reasoning tasks, with gains over DTD$\pi$11 and GHM stated explicitly in the paper (Chung et al., 15 May 2026).

This terminological reuse does not alter the gravitational meaning of universal horizons. It does, however, create an ambiguity around the acronym “UHM.” In the arXiv literature cited here, the phrase has one established meaning in Lorentz-violating black-hole physics and a separate methodological meaning in model-based offline RL.