Ensemble Interpretation in Solvable Models

In selected low-dimensional models, a gravitational topology expansion is reproduced by a random-matrix ensemble. Connected wormholes then match connected ensemble correlators, and universal spectral statistics arise after the appropriate averaging or smoothing. This is a controlled mathematical correspondence for JT gravity; it does not imply that every gravitational theory averages over Hamiltonians or that a finite set of matched observables uniquely determines an ensemble.

Required background. Quantum Information supplies mixed classical–quantum descriptions, moments, and conditioning.

Helpful background. Semiclassical Gravity supplies the genus expansion; Quantum Fields in Curved Spacetime supplies thermal partition functions; Black-Hole Information supplies wormhole correlators; Quantum Gravity supplies the microscopic-completion question.

What the ensemble averages

Let $d\mu(H)$ be a probability measure on Hermitian Hamiltonians. For each sample,

$$Z_H(\beta)=\operatorname{Tr}e^{-\beta H}$$

is a number and products factorize ordinarily. Ensemble moments are different observables:

$$\left\langle\prod_{i=1}^{b}Z_H(\beta_i)\right\rangle_\mu = \int d\mu(H)\prod_{i=1}^{b}Z_H(\beta_i).$$

Their connected cumulants need not vanish because every factor depends on the same random $H$. In the spectral-density language,

$$Z_H(\beta)=\int dE\,\rho_H(E)e^{-\beta E},$$

so multi-boundary amplitudes probe connected density correlations. The probability measure, matrix symmetry class, scaling limit, spectral normalization, and averaging operation are part of the definition.

Annealed and quenched quantities must be distinguished. $\log\langle Z\rangle$ is not generally $\langle\log Z\rangle$; Jensen’s inequality gives $\langle\log Z\rangle\le\log\langle Z\rangle$. A gravitational calculation of ensemble moments does not automatically supply the quenched free energy of a typical sample.

The same warning applies before computing entropy. If sample $H$ prepares an unnormalized density operator with moments $Z_{n,H}$, its normalized moment is $M_{n,H}=Z_{n,H}/Z_{1,H}^{n}$. In general,

$$\left\langle\frac{Z_{n,H}}{Z_{1,H}^{n}}\right\rangle \ne\frac{\langle Z_{n,H}\rangle}{\langle Z_{1,H}\rangle^{n}}.$$

For positive integer $n\gt1$, the entropy of the averaged moment, $(1-n)^{-1}\log\langle M_{n,H}\rangle$, also differs from the average Rényi entropy $(1-n)^{-1}\langle\log M_{n,H}\rangle$. Equality needs extra conditions such as concentration or fixed normalization. If every sample is normalized and differentiation can pass through the average, the $n\to1$ derivative of $\log\langle M_{n,H}\rangle$ does equal the average von Neumann entropy. It still need not equal the entropy of the averaged density operator. Keeping these operations in order is essential when interpreting replica amplitudes.

The controlled JT correspondence

JT gravity sums orientable surfaces weighted by $e^{S_0\chi}$, with boundary lengths $\beta_i$ and Weil–Petersson moduli volumes. The resulting genus expansion is reproduced order by order by a double-scaled matrix integral whose leading spectral density is fixed by the disk amplitude. Connected genus-$g$, $b$-boundary geometries match connected $b$-trace ensemble correlators.

Schematically,

$$Z_{g,b}^{\mathrm{JT}}(\beta_1,\ldots,\beta_b) \longleftrightarrow \left\langle \prod_{i=1}^{b}\operatorname{Tr}e^{-\beta_iH} \right\rangle_{\mathrm c,g}.$$

The spectral curve, trumpet gluing, topology recursion, and matrix-integral match are derived in Saad, Shenker, and Stanford 2019, §§2–5. The treatment of matrix symmetry classes and nonorientable variants is developed in Stanford and Witten 2020, §§2–5.

“JT gravity is a matrix integral” therefore has a precise perturbative meaning. A chosen matrix integral also supplies nonperturbative information, but the asymptotic genus series alone need not select that completion uniquely. Nor does the dictionary identify a particular Hamiltonian sample unless extra data are supplied.

The large-$S_0$ ordering clarifies the match. A connected surface with genus $g$ and $b$ boundaries carries $e^{S_0(2-2g-b)}$, while a matrix connected correlator has the corresponding large-matrix suppression. Disk data determine the leading smooth density, cylinders determine the leading connected two-level correlations, and higher boundaries test higher cumulants. Agreement at one topology is therefore weaker than agreement of the full recursive hierarchy.

State preparation remains part of the observable. Thermal traces weight the spectrum with $e^{-\beta E}$; microcanonical windows, real-time continuations, and operator insertions probe different transforms. An ensemble claimed from one thermal curve can fail on these sharper tests even when its mean density agrees.

Comparison axes

Axis	Fixed Hamiltonian	Matrix ensemble	Perturbative JT gravity
Basic object	$Z_H(\beta)$ for one spectrum	Probability distribution $d\mu(H)$	Sum over geometries with stated boundaries
Two-boundary factorization	Exact product of two numbers	Connected covariance can be nonzero	Connected wormhole topology can contribute
Fine spectral structure	Sample-specific fluctuations	Universal after scaling and averaging	Reproduces ensemble correlators order by order
Nonperturbative definition	Given once $H$ is specified	Given once the measure and contour are specified	Not fixed by the genus series alone

This comparison prevents a category error. A gravitational answer can equal an ensemble moment without equalizing every individual sample. Conversely, a fixed chaotic Hamiltonian can display local random-matrix statistics without its full theory being an ensemble.

The same distinction applies to uncertainty. Sample-to-sample variance is a property of $d\mu(H)$, whereas numerical error estimates uncertainty in evaluating a chosen object. Smoothing a fixed spectrum suppresses fluctuations but does not manufacture a probability measure unless the averaging prescription is explicitly promoted to one.

Results and open questions as of 2026-08-19

This discussion reflects results available through 2026-08-19.

Evidence or result	Controlled scope	Limitation or alternative interpretation
JT/matrix-integral genus match, Saad, Shenker, and Stanford 2019, §§2–5	Double-scaled matrix integral and perturbative JT topology expansion	Does not select one Hamiltonian or a unique completion from asymptotics alone
Symmetry-class extension, Stanford and Witten 2020, §§2–5	JT variants with specified time-reversal and pin structures	Model-specific classification, not a theorem that higher-dimensional gravity is averaged
Spectral universality, Cotler et al. 2017, §§2–4	Random matrices and selected chaotic spectra with averaging or smoothing	Universal local correlations do not uniquely reconstruct the global ensemble or a wormhole

The literal ensemble position takes the matrix measure as the boundary definition of JT gravity. An emergent-statistics position treats random-matrix behavior as a universal approximation to spectral correlations of one complex system. A sector-conditioned position attributes averaging to unobserved baby-universe data and expects factorization after conditioning. These competing positions can reproduce overlapping observables while differing on ontology and exact factorization; extension beyond the solvable models remains unresolved.

An adverse check changes the ensemble while preserving a finite number of moments: many probability measures share those moments but differ in tails and nonperturbative spectral density. Matching several wormhole amplitudes therefore does not uniquely determine the microscopic measure. Another adverse check selects one sample; its connected ensemble covariance disappears although its smoothed spectral form factor may retain a ramp.

The conclusions could change after a theorem establishes uniqueness of a nonperturbative completion from gravitational data, after a fixed microscopic theory reproduces all multi-boundary amplitudes without averaging, or after a controlled observable distinguishes literal ensemble averaging from universal single-sample spectral statistics.

Exercises

An ensemble has $Z=1$ or $Z=4$ with equal probability. Compute the annealed quantity $\log\langle Z\rangle$ and the quenched quantity $\langle\log Z\rangle$. Which one can be inferred from the one-boundary average alone?

Solution

$\log\langle Z\rangle=\log(5/2)$, whereas $\langle\log Z\rangle=(\log1+\log4)/2=\log2$. Only the annealed quantity follows from $\langle Z\rangle=5/2$. Computing the quenched answer needs the distribution, a replica limit, or equivalent additional data. The strict inequality $\log2\lt\log(5/2)$ is the expected Jensen check.

References

• Cotler et al. 2017, Jordan S. Cotler et al., “Black Holes and Random Matrices,” Journal of High Energy Physics 05 (2017) 118.
• Saad, Shenker, and Stanford 2019, Phil Saad, Stephen H. Shenker, and Douglas Stanford, “JT Gravity as a Matrix Integral,” arXiv:1903.11115 (2019).
• Stanford and Witten 2020, Douglas Stanford and Edward Witten, “JT Gravity and the Ensembles of Random Matrix Theory,” Advances in Theoretical and Mathematical Physics 24 (2020) 1475–1680.