Tensor Networks and Discrete Geometry

Tensor networks turn a quantum state into a graph of finite-dimensional maps. Graph cuts bound entanglement, coarse-graining layers organize scale, and code properties make reconstruction redundant. These features provide precise models of geometric relations, but a network drawn like a spatial slice is not automatically a dynamical spacetime. The scientific task is to identify which statement follows from the tensors and which was placed in the graph by design.

Required background. Quantum Information supplies Schmidt decompositions, channels, entanglement entropy, and quantum error correction.

Helpful background. Flat-Spacetime Quantum Field Theory supplies continuum locality and renormalization; Einstein Equations supplies the gravitational target; Quantum Gravity supplies comparison standards; and Causality and Global Structure supplies the distinction between spatial connectivity and Lorentzian causal structure.

Bonds, cuts, and area-like quantities

A tensor network contracts internal indices and leaves physical indices uncontracted. If a cut $\gamma_A$ separating boundary region $A$ from its complement crosses bonds of dimensions $\chi_e$, the Schmidt rank obeys

$$\operatorname{rank}\rho_A\le \prod_{e\in\gamma_A}\chi_e, \qquad S(A)\le\min_{\gamma_A}\sum_{e\in\gamma_A}\ln\chi_e.$$

This is an exact network bound. In special random, perfect, or large-bond constructions it can be saturated at leading order, allowing the weighted minimal cut to model a geometric area. Bond dimension then controls area in entropy units; it does not fix a proper length until a reconstruction map and scale are supplied.

Entanglement-renormalization networks organize tensors in layers. Moving inward removes short-distance degrees of freedom, so depth resembles a renormalization scale and graph connectivity resembles a discrete hyperbolic geometry Swingle 2012, pp. 3–8. The relation is powerful because correlations and entropy can be computed from causal cones within the network. Yet the circuit’s directed contraction order is an algorithmic dependency, not necessarily physical time.

Reconstruction and quantum error correction

In a code tensor network, a logical bulk operator can be represented on more than one boundary region. This redundancy models subregion reconstruction and explains why erasure of some physical factors need not destroy the logical information. The HaPPY construction makes this relation exact for perfect tensors Pastawski et al. 2015, §§ 2–4.

Three structures must be distinguished:

• the graph, which declares possible contractions and often encodes adjacency;
• the state or code map, determined by tensor entries;
• the reconstructible region, determined by the chosen code subspace and boundary region.

Changing tensor entries at fixed graph can radically change correlations. Changing the code subspace can move the reconstruction boundary. Therefore neither graph distance nor a minimal cut is state-independent geometry in general. Approximate codes and finite bond dimensions introduce corrections that should be bounded rather than hidden.

Dynamics and a continuum limit

A static network represents a state or encoding map. Quantum gravity requires dynamics or a sum over structures, constraints selecting physical states, and a continuum regime. A network Hamiltonian can generate time, but its microscopic locality may already rely on the graph one hoped to derive. A path integral of tensors can model spacetime amplitudes, but then one must specify weights, boundary data, reflection positivity or Lorentzian contour, and the limit in which observables become regulator independent.

Recovering local QFT requires more than the correct entropy scaling. Low-energy excitations should have approximately Lorentz-invariant dispersion, commutators should decay outside one emergent cone, and correlation functions should exhibit the expected operator-product and cluster properties. Recovering gravity further requires a massless spin-2 mode, universal coupling, constraints, and nonlinear Einstein dynamics. No generic tensor-network theorem supplies this package.

Explicit networks inspired by established holographic dualities are useful existence models for encoding geometry and error correction. The duality, string construction, and detailed bulk reconstruction belong to the string-theory volume; here the network is assessed as a general pregeometric mechanism.

Results and open questions through 2026-08-20

This discussion reflects results available through 2026-08-20. Competing interpretations remain credible. One treats tensor networks as concrete microscopic candidates whose code structure explains geometric area and bulk locality. Another treats them as variational or pedagogical representations of states whose geometry was already chosen. A middle position regards network universality as evidence for information-theoretic organization while withholding a claim of dynamical spacetime emergence.

Network feature or result	Evidence at the cutoff	Scope and limitation
Weighted minimal cuts	Exact entropy bounds and saturation in selected tensor ensembles	Area normalization and geometric interpretation depend on state and construction
Renormalization layers	Efficiently reproduce multiscale entanglement and correlation causal cones	Circuit depth is not automatically radial proper distance or time
Error-correcting code subspaces	Exact and approximate redundant reconstruction in explicit models	Code subspace and background graph are usually inputs
Tensor dynamics or sums	Provide calculable discrete amplitudes and toy time evolution	Continuum Lorentz symmetry, gravitational constraints, and universality remain unresolved

The formulation axis matters: fixed networks, variationally optimized networks, random ensembles, and dynamical sums answer different questions. Evidence from one cannot be transferred without showing that its state space, measure, and continuum limit agree.

The conclusions could change if a network model derives rather than assumes its connectivity and reaches a continuum phase with a common Lorentzian cone, local QFT algebra, and universal spin-2 dynamics, or if a no-go theorem rules out that limit; the interpretation could also change with controlled finite-bond corrections that invalidate or confirm an area reconstruction.

Independent checks

First compare the minimal-cut entropy with a direct reduced-density-matrix calculation for small instances. Second reconstruct distance from both correlations and propagation; disagreement exposes a graph convention. Third perturb tensor entries and the code subspace to test whether geometric relations are phase-stable. Fourth verify that any claimed causal cone refers to operational signaling, not merely the set of tensors needed to compute an output.

Exercise

A network cut crosses three bonds of dimensions $2$, $3$, and $5$. What entropy bound follows in natural units? If a direct calculation gives $S(A)=\ln20$, is the proposed cut minimal and saturated?

Solution

The crossed bond space has dimension $2\times3\times5=30$, so

$$S(A)\le\ln30.$$

The value $\ln20$ respects the bound but does not show saturation. It also does not prove that this cut is minimal: another cut might have product bond dimension between $20$ and $30$, or below $20$—the latter would contradict the stated entropy and therefore cannot occur for the same exact network. To establish minimality one must enumerate or optimize all separating cuts; to establish saturation one needs $S(A)=\ln30$ for the minimizing cut.

References

• Pastawski, Fernando, Beni Yoshida, Daniel Harlow, and John Preskill (2015), “Holographic Quantum Error-Correcting Codes: Toy Models for the Bulk/Boundary Correspondence,” Journal of High Energy Physics 2015(6), 149, arXiv:1503.06237.
• Swingle, Brian (2012), “Entanglement Renormalization and Holography,” Physical Review D 86, 065007, arXiv:0905.1317.