The bulk
A spatial geometry containing points and curves. A point answers: where?
CONTEXT JAMMING / KINEMATIC SPACE
THE NETWORK
IS NOT THE SPACE
How a tensor network became a map of every possible way to probe a universe
For years, physicists looked at MERA and saw a crude lattice of curved space. Czech, Lamprou, McCandlish, and Sully proposed a more radical interpretation: the network is a map of intervals and geodesics—the questions from which space can be reconstructed.
01 · The cast
Three objects that look alike
The visual rhyme is real. Identity is not. The paper argues that MERA and kinematic space share an organizing structure strongly enough that one can discretize the other.
Kinematic space
A space in which each point represents an entire geodesic or boundary interval.
MERA
A tensor network that organizes information by boundary position and scale.
02 · Coordinate reversal
One geodesic becomes one point
A location inside ordinary geometry is a point. Kinematic space changes what qualifies as a coordinate: choose two boundary endpoints, and the complete geodesic between them becomes one point in a new space.
θ = 1.10 · α = 0.75This point does not mark a place inside the disk. It marks the complete geodesic anchored between u and v. This interval probes a mesoscopic scale.
Interactive reconstruction after Fig. 2 and Eqs. 2.4–2.5 of Czech et al. (2015).
03 · Integral geometry
Measure a curve by counting crossings
One can measure the length of a curve without walking along it. Count how many lines intersect it, weight the lines correctly, and length emerges from crossing statistics. In AdS, the lines become boundary-anchored geodesics; their density is determined by interval entanglement entropy.
CROSSINGS24/42CROFTON ESTIMATE377SCHEMATIC LENGTH405
Finite-sample geometric demonstration—not a precision numerical integrator.
04 · Ordering without a clock
Causality without time
In kinematic space, “future” means larger containment. A broad interval can contain the information domain of a narrower one. This ordering supplies a Lorentzian causal structure without describing physical time.
TIMELIKEOne interval contains the other: larger containment is the kinematic “future.”
The same partial order appears in MERA: each tensor can influence only a particular ultraviolet interval. Fix one interval endpoint and move the other, and the corresponding path is lightlike in kinematic space.
05 · The network
How MERA compresses a state
A generic many-body quantum state is exponentially expensive to represent. Tensor networks exploit structure. In MERA, disentanglers change basis to separate short-range correlations; isometries project locally simplified degrees of freedom into a smaller effective space. Moving upward means coarse-graining—not traveling into the bulk.
A MERA tensor is not a neural-network neuron. Its unitary and isometric constraints give exact cancellations that ordinary learned layers do not inherit.
6 schematic bond crossingsInclusive cone: tensors that may affect the interval. Exclusive cone: the subset not shared with its complement.
Cut count approximates entropy in an optimized network; it is not an unconditional theorem for every tensor network or state.
06 · The central identification
A node is not a place. It is a way of asking.
The traditional picture treats MERA as a pixelated slice of anti-de Sitter space. The paper proposes a different map: each tensor organizes a boundary interval—and therefore a family of measurements or geodesics.
VIEW MERA AS
The paper’s proposed identification
- A tensor maps to an interval (u,v).
- Location and scale become coordinates.
- Causal order becomes interval containment.
- Lightlike directions hold one endpoint fixed.
- Causal diamonds carry conditional information.
THE PAPER DOES NOT MOVE MERA DEEPER INTO SPACETIME.
IT MOVES MERA ONE CONCEPTUAL LEVEL ABOVE SPACETIME.
07 · A volume made of dependence
Information becomes volume
Place three adjacent intervals A, B, and C on the boundary. Four entropy terms overlap. Their shared contributions cancel, leaving a causal diamond whose volume measures the dependence A and C retain after the information available through B is removed.
ILLUSTRATIVE CFT MODEL0.058
I(A,C|B)
S(AB) + S(BC) − S(B) − S(ABC)Illustrative CFT entropy model. This is not empirical data from a physical system or language model.
Cut counting is a discrete flux law
Entropy lives on a causal cut; compression operations fill its interior. Their relation resembles a discrete Stokes theorem: changing a valid boundary changes the volume it encloses and the bonds it crosses.
SURVIVING DOF12BOUNDARY FLUX10ENCLOSED COMPRESSION18
08 · The stress test
The black-hole quotient
A strong interpretation should survive a nontrivial geometry, not merely reproduce the vacuum’s superficial shape. The paper compares a quotient MERA for the thermofield-double state with the geodesic space of the two-sided BTZ black hole—and finds three matching sectors.
GEODESIC SECTOR
Connects opposite asymptotic boundaries, crosses the horizon, and prepares two-sided correlations.
This relation belongs to the thermofield-double construction. It is not a general result about neural networks.
09 · Beyond the vacuum
Geometry as compression
For excited states, a fixed vacuum MERA is insufficient. The network should adapt to the state’s particular entanglement pattern while retaining causal organization and an approximate link between cuts and entropy. Geometry, in this proposal, records where and how information can be compressed.
SELECTED DIAMOND 94 bits
- Incoming dimension
- 96
- Outgoing dimension
- 52
- Local compression
- 4
Conceptual compression model—not a simulation of the original tensor optimization.
That is the forward-looking move in Section 5: not geometry drawn first and populated later, but an effective geometry induced by successful, state-dependent compression.
10 · Context Jamming extension
The LLM bridge
STRUCTURAL ANALOGY, NOT IDENTITY
What changes if we analyze a language model not only as a cloud of activation vectors, but as an adaptive map of which spans, scales, and transformations carry irreducible information?
PAPERBoundary interval↔LLM ANALOGYToken span or contextual region
PAPERLocation + scale↔LLM ANALOGYToken position + abstraction or receptive scope
PAPERDisentangler↔LLM ANALOGYLearned basis transformation separating correlated features
PAPERIsometry↔LLM ANALOGYProjection, bottleneck, pruning, routing, or effective compression
PAPERCausal cone↔LLM ANALOGYComputation materially influencing a selected span or output
PAPERConditional mutual information↔LLM ANALOGYResidual dependence across separated contextual regions
PAPERState-dependent geometry↔LLM ANALOGYInput-, task-, or state-dependent computation graph
PAPERCompression network↔LLM ANALOGYSparse routing, expert selection, low-rank flow, or adaptive token compression
Where the analogy breaks
- MERA uses explicit unitary and isometric structure; standard transformers generally do not.
- Full self-attention is globally connected, unlike a strictly local MERA lattice.
- Transformer attribution does not enjoy MERA’s exact cancellation property.
- Quantum entropy and model feature dependence are mathematically distinct.
- Current LLMs have not been shown to instantiate AdS/CFT or kinematic space.
Five questions an experiment could answer
These are research prompts, not conclusions smuggled in from physics.
Can one define a span-based kinematic space?
- Measurable object
- Contextual intervals across layers
- Possible experiment
- Measure conditional dependence while varying span and scale.
- Falsifying result
- No stable geometry across prompts.
- Primary caveat
- Coordinates may depend on tokenizer and architecture.
11 · Epistemic ledger
What the paper actually establishes
- MERA and kinematic space share interval-based causal ordering.
- Lightlike directions hold one endpoint fixed.
- Conditional information localizes in causal diamonds.
- Optimized MERA cut counting mirrors geometric entropy calculations.
- The BTZ quotient creates matching network and geodesic sectors.
- MERA is best viewed as discretizing kinematic space rather than the bulk slice.
- Conditional mutual information acts as a network volume.
- Excited-state geometry may be represented by adaptive compression networks.
- Similar tools may illuminate LLM span dependence, routing, low-rank structure, and adaptive compression.
- This remains a research program—not an implication proved by the paper.
12 · Field guide
A compact glossary
AdS
Anti-de Sitter space: a negatively curved spacetime that supplies the gravitational side of the correspondence used here.
CFT
Conformal field theory: a quantum field theory unchanged by local changes of scale.
AdS/CFT
A proposed duality relating gravity in AdS to a conformal field theory on its boundary.
Boundary interval
A connected segment on the one-dimensional CFT boundary, specified by endpoints u and v.
Bulk
The higher-dimensional gravitational geometry enclosed by the boundary.
Geodesic
A locally shortest or extremal path; in the static AdS slice, a boundary-anchored arc.
Kinematic space
The space whose points label boundary intervals, or equivalently the associated bulk geodesics.
MERA
Multi-scale Entanglement Renormalization Ansatz, a tensor network organized by position and scale.
Tensor network
A factorized representation of a many-body state built from contracted multidimensional arrays.
Disentangler
A unitary tensor that changes basis to separate short-range correlations.
Isometry
A tensor that maps a locally simplified space into a smaller effective one while preserving inner products on its image.
Causal cone
The exact subset of a MERA network able to influence a selected boundary interval.
Causal cut
A valid spacelike path through MERA that defines a coarse-grained state.
Renormalization
A systematic change of description that integrates out short-scale detail.
Entanglement entropy
A measure of quantum entanglement between a region and its complement.
Mutual information
The total correlation shared by two subsystems.
Conditional mutual information
The dependence between A and C remaining after conditioning on B.
Crofton formula
An integral-geometric rule that measures curve length by counting intersections with lines or geodesics.
de Sitter space
A Lorentzian constant-curvature geometry; half of two-dimensional de Sitter describes vacuum kinematic space.
Thermofield double
A pure entangled state of two copies whose individual sides look thermal.
BTZ black hole
A black-hole solution in three-dimensional AdS, obtainable as a quotient of AdS.
Entwinement
A generalized entanglement quantity associated here with winding geodesics.
Bond dimension
The size of an internal tensor-network index; it limits representational capacity.
Compression network
The paper’s proposed broader, state-adaptive network retaining MERA-like causal and entropic organization.
13 · Conclusion
Geometry may be the shape of successful compression
The deepest move in the paper is not a new drawing of spacetime. It is a change in what counts as a coordinate. Instead of labeling places, kinematic space labels intervals—the probes through which a system becomes knowable. MERA then becomes a discrete atlas of those probes. Its volume records information. Its causal order records containment. Its geometry records compression.
Perhaps the useful geometry of an intelligent system will likewise be found not in where its activations sit, but in which questions, spans, and transformations its computation makes irreducible.
14 · Source notes
The argument, pinned to the paper
- Main thesis: Abstract, Introduction, and Section 6; MERA as half of two-dimensional de Sitter/kinematic space appears in Fig. 1.
- Geodesic coordinates: Section 2.1, Fig. 2, Eqs. 2.4–2.5.
- Crofton form and kinematic metric: Sections 2.1.1–2.1.3.
- Conditional information: Fig. 3, Eqs. 2.12–2.14; localization in MERA, Figs. 10–11 and Eq. 3.2.
- Causal cones and cut counting: Figs. 6–8; spacelike and timelike cuts, Fig. 9.
- Entropy, causal wings, complementary intervals: Figs. 12–14.
- BTZ quotient: Section 4 and Figs. 15–17, including the entwinement sector.
- Excited states: Section 5, “Toward excited states: Geometry as compression.”
Czech, Bartłomiej, Lampros Lamprou, Samuel McCandlish, and James Sully. “Tensor Networks from Kinematic Space.” arXiv:1512.01548, 2015. Published in JHEP 07 (2016) 100.
Open the primary source ↗