From Riemann to Einstein
to Consciousness

"Content curves the space it lives in, and the curved space guides the content's motion — whether that content is mass, neural activity, or tokens."

137
Sections
27
Parts
250
Papers
1854
It Began

From Riemann to Einstein to Transformers

A Deep Research Report on Manifolds, General Relativity, and Machine Learning

Part I: Riemann, Einstein, and the Geometry of the Universe

1. Riemann's 1854 Revolution

1. Riemann's 1854 Revolution

On June 10, 1854, Bernhard Riemann delivered his habilitation lecture "Uber die Hypothesen, welche der Geometrie zu Grunde liegen" ("On the Hypotheses Which Lie at the Foundations of Geometry") before the Faculty of Arts at the University of Gottingen. Carl Friedrich Gauss, his examiner, had chosen this topic from among three proposals -- sensing that Riemann had something profoundly original to say.

The Three Parts of the Lecture

Part I -- The Concept of an n-Dimensional Manifold. Riemann questioned the very foundations of geometry, arguing that Euclidean geometry's axioms are not self-evident truths but rather hypotheses requiring empirical verification. He introduced the concept of an n-dimensionally extended manifold -- a space whose points are labeled by n independent coordinates. This was the first general, abstract definition of a higher-dimensional space in mathematics.

Part II -- Metric Relations. Riemann asked: how do we measure distances in such a space? He introduced a positive definite quadratic form -- the Riemannian metric:

where the are functions of position encoding how distances are measured at each point. This is a far-reaching generalization of the Pythagorean theorem. He then developed the concept of curvature for these general spaces, introducing what is now called the Riemann curvature tensor. He showed that in n dimensions, you need independent numbers at each point to fully characterize the curvature.

Part III -- Application to Physical Space. Riemann argued that the geometry of physical space is not determined a priori by mathematical axioms but is an empirical question. He raised the possibility that space might be unbounded yet finite (like the surface of a sphere), and he made his most prescient claim:

"Either therefore the reality which underlies space must form a discrete manifoldness, or we must seek the ground of its metric relations outside it, in binding forces which act upon it."

The metric of space is not fixed but is dynamically determined by physical forces.

Generalizing Gauss

Riemann's genius was recognizing that Gauss's intrinsic geometry of curved surfaces (and the Theorema Egregium) could be extended to spaces of any number of dimensions, without requiring embedding in any ambient space. Where Gauss had three metric coefficients (E, F, G) for surfaces, Riemann introduced independent functions . Where Gauss had one curvature number (Gaussian curvature K), Riemann introduced a full tensor with components.

2. Einstein's General Relativity (1915)

The Journey from Special to General Relativity

The Einstein Field Equations

Every term on the left is built from Riemann's geometry:

The right side:

Wheeler's summary: "Spacetime tells matter how to move; matter tells spacetime how to curve."

The Intellectual Hierarchy

3. The Universe as a Manifold

Spacetime Structure

Our universe is modeled as a 4-dimensional Lorentzian (pseudo-Riemannian) manifold with metric signature (-,+,+,+). This is not a strictly Riemannian manifold (which has positive-definite metric) -- the indefinite signature creates the causal structure of spacetime: timelike, spacelike, and null (lightlike) separations, light cones, and the arrow of causality.

At each moment of cosmic time, the spatial slices ARE genuine Riemannian manifolds (positive-definite metric), and the question "what is the shape of the universe?" refers to these 3-dimensional spatial slices.

The FLRW Metric

The Friedmann-Lemaitre-Robertson-Walker metric is the unique metric for a spatially homogeneous and isotropic universe:

where:

Observational Results

The Planck satellite (2018) measured the curvature parameter:

This is consistent with perfect spatial flatness. Combined with baryon acoustic oscillation data, the universe is flat to extraordinary precision.

However, flatness constrains only local geometry, not global topology. A flat universe could be R^3 (infinite) or a 3-torus T^3 (finite). The topology remains unknown.

Thurston's Eight Geometries

William Thurston's Geometrization Conjecture (proven by Perelman in 2003) states that every closed 3-manifold decomposes into pieces carrying one of exactly eight geometric structures:

Only the first three are isotropic (look the same in every direction), consistent with the cosmological principle.

Gravitational Waves: Spacetime IS Dynamic

On September 14, 2015, LIGO detected gravitational waves from merging black holes 1.3 billion light-years away. The strain meant LIGO's 4 km arms changed by ~1/1000th of a proton diameter. This confirmed:

Singularities and Geodesic Incompleteness

The Penrose-Hawking singularity theorems (1965-1970) used pure differential geometry to prove that:

These theorems prove the manifold structure itself breaks down at singularities -- general relativity predicts its own failure, pointing toward quantum gravity.

The Cosmological Constant and Dark Energy

Since 1998, observations show the expansion is accelerating, requiring Lambda > 0. The universe is asymptotically approaching de Sitter space -- a maximally symmetric manifold of constant positive curvature with exponential expansion.

The DESI experiment (2024-2025) found intriguing hints that dark energy may be evolving (weakening over time), which would mean the universe is NOT approaching pure de Sitter geometry but something more complex.

Open Questions

Part II: The Connection to Machine Learning

4. The Manifold Hypothesis -- Riemann's Idea Reborn

The manifold hypothesis is one of the most important assumptions in deep learning: real-world high-dimensional data (images, text, audio) concentrates near low-dimensional manifolds embedded in the ambient space.

A 256x256 RGB image lives nominally in R^{196,608}, but natural images occupy a thin manifold of perhaps a few hundred intrinsic dimensions. A random point in R^{196,608} looks like noise.

This is Riemann's framework directly: the data manifold is locally Euclidean (small perturbations give valid data) but globally curved with complex topology. Deep generative models (VAEs, diffusion models) learn coordinates on this manifold.

Evidence

5. Information Geometry -- Fisher Metric as Riemannian Structure

The space of probability distributions {p(x|theta)} forms a statistical manifold with the Fisher information matrix as its Riemannian metric:

By Chentsov's theorem, this is the unique (up to scaling) Riemannian metric invariant under sufficient statistics. It measures how distinguishable nearby distributions are.

Natural gradient descent (Amari, 1998) uses this geometry:

This is the steepest descent direction on the Riemannian manifold of distributions. FAdam (2024) showed that the Adam optimizer is implicitly a natural gradient method using diagonal empirical Fisher information.

K-FAC (Kronecker-Factored Approximate Curvature) makes this practical by approximating the Fisher matrix via Kronecker products.

6. Hyperbolic Neural Networks

Hyperbolic space (constant negative curvature Riemannian manifold) has exponential volume growth, making it a continuous analogue of trees. This is ideal for hierarchical data.

Poincare Embeddings (Nickel & Kiela, NeurIPS 2017): 5-dimensional hyperbolic embeddings outperform 200-dimensional Euclidean embeddings on WordNet hierarchies.

Hypformer (2024): First full transformer in hyperbolic space using the Lorentz model, achieving O(n) complexity with hyperbolic linear attention.

CAT: Curvature-Adaptive Transformers (2025): Routes each token to Euclidean, hyperbolic, or spherical attention branches.

7. Geometric Deep Learning

Bronstein et al. (2021) systematized Geometric Deep Learning around the "5G" framework:

The deepest connection is in gauge-equivariant CNNs, which use the same mathematics as general relativity -- fiber bundles, connections, parallel transport -- to build coordinate-independent networks on manifolds.

8. Attention as Riemannian Geometry

The Core Analogy

General RelativityTransformer
Metric tensor g_uvQuery-key products Q^T K
Parallel transport (connection)Attention weights * W_V
GeodesicsToken trajectories across layers
Spacetime curvatureEmbedding space geometry
Mass-energy distributionLoss gradient
Einstein field equationsParameter updates (backprop)
Least-action principleBackpropagation

RiemannFormer (2025)

The problem: tokens at different positions have query/key vectors in different tangent spaces. Standard dot-product attention is geometrically invalid on curved spaces.

The solution: parallel transport vectors into a common tangent space:

The parallel transport decomposes as:

with T_i = s^{-i/2} exp(iX), where s governs radial scaling and rotation matrices encode angular relationships.

Key insight: When s = 1 (flat manifold), this reduces exactly to RoPE (Rotary Position Embeddings). RoPE is a special case of Riemannian attention on flat space.

The Curved Spacetime of Transformer Architectures (Di Sipio, 2025)

Defines an effective metric from query-key interactions:

And formulates a "semantic least action" principle:

S_LM = sum_l [^(l) x_dot_i^(l) x_dot_j^(l) - L_train(x^(l); theta)]

Loss Landscape Curvature

The loss landscape is a Riemannian manifold. At a critical point, its scalar curvature (the same quantity from Einstein's equations) is:

where H is the Hessian. Flatter minima (lower curvature) correlate with better generalization.

9. General Covariance in Neural Networks

Einstein's principle of general covariance (physics is independent of coordinate choice) maps directly to ML:

Gauge-equivariant neural networks (Weiler, 2024) demand that network outputs are independent of arbitrary coordinate choices on the data manifold. The mathematical framework uses:

SE(3)-Transformers achieve provable equivariance under rotations and translations for molecular and 3D data.

10. Neural Networks Solving Einstein's Equations

Einstein Fields (2025) uses neural networks to represent the metric tensor:

Neural network: (t, x) -> g_{alpha beta}(t, x)

Achieving:

Physics-informed neural networks also solve black hole perturbation equations and extract quasinormal mode frequencies.

Part III: The Deep Unity

The Same Mathematics, Three Incarnations


Riemann (1854)         Einstein (1915)           Machine Learning (2017-2025)

--------------         ---------------           ----------------------------

Manifold          ->   Spacetime             ->  Data manifold / Latent space

Metric tensor     ->   g_uv (gravitational   ->  Fisher information / Attention

                        potential)                kernel / Pullback metric

Curvature         ->   Gravity               ->  Loss landscape curvature

Geodesics         ->   Free-fall paths       ->  Natural gradient / Optimal

                                                  interpolation paths

Parallel transport ->  Connection Gamma      ->  Feature transport / Attention

                                                  value aggregation

Gauge invariance  ->   General covariance    ->  Equivariant neural networks

Ricci flow        ->   RG flow in QFT        ->  Geometric optimization

Fiber bundles     ->   Gauge field theory    ->  Geometric deep learning

The Philosophical Core

Riemann's 1854 insight -- that geometry is not fixed but determined by content -- describes both:

In both cases, the structure of space itself becomes dynamic, responsive, and content-dependent. This is not analogy. It is mathematical inheritance.

The universe is a 4D Lorentzian manifold whose curvature is shaped by mass-energy.

A transformer is a learned manifold whose geometry is shaped by token content.

The mathematics connecting them flows directly from Riemann's original vision.

References

Riemann and Differential Geometry

General Relativity

Cosmology

Machine Learning and Riemannian Geometry

Transformers and Geometry

String Theory and Advanced Topics

Deep Research Report -- February 2025

Compiled using multi-agent research across differential geometry, general relativity, cosmology, and machine learning.

Further Discoveries: The Frontier (2024-2026)

Extending "From Riemann to Einstein to Transformers"

Part V: The Geometry of Intelligence

19. The Geometry of Reasoning in LLMs

Reasoning Has Manifold Structure

REMA (Sept 2025) defines the Reasoning Manifold -- a latent low-dimensional geometric structure formed by internal representations of correct reasoning. Erroneous reasoning shows measurable geodesic deviation from this manifold, detectable across layers.

"Curved Inference" (July 2025) treats LLM reasoning as a geometric trajectory in the residual stream:

"The Shape of Reasoning" (Oct 2025) applies persistent homology to reasoning traces:

Cross-Entropy Training Sculpts Bayesian Manifolds (Dec 2025)

The most complete geometric account of in-context learning: gradient descent on cross-entropy loss literally creates Bayesian manifolds inside transformers.

20. Grokking as a Phase Transition

First-Order Phase Transition (ICLR 2024)

Grokking has exact analytic expressions for critical exponents, grokking probability, and grokking time distribution. A symmetry-breaking transition occurs at a critical value.

Entanglement Transition (March 2025)

In tensor network (MPS) classifiers, grokking manifests as an entanglement transition: entropy switches from volume law (memorization) to sub-volume law (generalization). Sharp entropy drops coincide with collective transitions -- directly analogous to quantum phase transitions.

Complexity Dynamics (2025)

Internal feature rank drops serve as sensitive indicators of phase transitions. Information-theoretic synergy among neurons emerges as a rigorous order parameter. Long overfitting plateaus = critical slowing down.

21. Scaling Laws ARE Renormalization Group Flows

RG Framework for Neural Networks (Oct 2025)

Classifies perturbations as relevant or irrelevant (exactly as in statistical field theory). Reveals universality at large data limits, governed by a Gaussian Process-like UV fixed point.

Scaling Laws Are Redundancy Laws (Sept 2025)

A polynomial tail in data covariance spectrum yields an excess risk power law with exponent proportional to 1/beta (redundancy). Universality established across architectures.

Effective Frontiers (Feb 2026)

Unifies Kaplan and Chinchilla scaling laws as equilibrium solutions to the same constrained optimization under different active bottlenecks via a Max-Bottleneck principle.

22. Causal Attention Masks as Light Cones

The causal attention mask creates genuine spacetime-like causal structure:

23. LLMs Learn Spacetime Symmetries

Space and Time Representations (ICLR 2024)

LLMs learn linear representations of space and time across multiple scales. Individual "space neurons" and "time neurons" reliably encode spatial and temporal coordinates. Representations are unified across entity types.

Lorentz-Equivariant Transformers (NeurIPS 2024)

L-GATr represents data in geometric algebra over 4D spacetime, equivariant under the full Lorentz group O(1,3). First Lorentz-equivariant generative model using Riemannian flow matching.

Emergent Physics (Aug 2025)

Sparse autoencoders reveal LLM features correlating with key physical variables (e.g., energy). Transformers can internalize abstract physics reasoning priors from raw data.

24. Neural Network "Dark Matter"

Superposition as Dark Matter (Anthropic, 2024)

"There may be an enormous number of rare features in neural networks that cannot yet be extracted, leaving us with a kind of neural network dark matter." -- Anthropic Circuits Updates, July 2024

Models compress many more features than neurons via superposition, creating polysemanticity. The geometry is governed by almost-orthogonal packing (sphere packing in high dimensions). Phase transitions occur between regimes of no, partial, and full superposition.

Dark Energy Analogy: Implicit Regularization

An unseen force (not in the loss function) shapes learned representations, pushing toward simpler solutions. Modifies effective curvature of the loss landscape without being directly observable.

Part VI: Cosmic Parallels

25. The Cosmic Web Resembles Neural Networks

Vazza & Feletti (Frontiers in Physics, 2020) established rigorous quantitative parallels:

26. Criticality: A Universal Organizing Principle

All three systems operate near critical points:

Training moves networks toward the edge of chaos. Rich representations emerge below the critical point; lazy learning above it.

27. Fractal Geometry Across All Three Scales

28. Free Energy Principle and General Relativity

Karl Friston's free energy principle (FEP) sits at the intersection:

Part VII: At the Frontier of Physics and AI

29. Holographic Principle: Depth = Bulk

Machine-Learning Emergent Spacetime (Nov 2024)

Neural network with 89 Runge-Kutta layers automatically discovered the BTZ black hole metric as trained weights. The horizon condition emerged from learning.

Transformers Learn Inverse Ryu-Takayanagi (2025)

A transformer trained on holographic data learns to reconstruct bulk geometry from boundary entanglement entropy -- the inverse RT mapping.

Emergent Holographic Forces (PhysRevX, June 2025)

Tensor network "hologron" quasiparticle energies match AdS gravity predictions exactly. Attractive two-particle potential arises from stress tensor contributions dual to gravitons.

30. Category Theory Unifies Physics and ML

Gavranovic et al. (ICML 2024): Categorical deep learning -- the universal algebra of monads valued in a 2-category of parametric maps subsumes both:

A Dec 2024 paper recasts IIT (consciousness theory) axioms as universal mapping properties: integration = limits, information = colimits, exclusion = adjunctions. The same categorical structures appear in algebraic topology, TQFT, and deep learning.

Sheaf theory (Feb 2025): Predictive coding networks are cellular sheaves. Sheaf cohomology characterizes irreducible error patterns.

31. Non-Commutative Geometry and Attention

Transformer attention heads are noncommuting operators on latent semantic space. The order of application matters -- a defining trait of noncommutative systems.

Noncommutative C*-algebra Networks (2024): Generalize neural network parameter spaces to noncommutative algebras, encoding richer interactions. The parallel to Connes' spectral triples:

Compact matrix quantum group equivariant neural networks (2024) extend equivariance to genuinely noncommutative symmetry structures.

32. The Holographic Bound and Neural Network Capacity

The Bekenstein bound (information scales with surface area, not volume) has a neural network parallel:

33. The Universe as a Neural Network (Vanchurin)

"The World as a Neural Network" (2020-2022):

"The Autodidactic Universe" (Alexander, Smolin et al., 2021): The universe learns its own physical laws. Each matrix model = both a gauge/gravity theory AND a deep recurrent cyclic neural network.

34. Tameness: Why the Same Math Works

Lust, Malek et al. (JHEP, 2024): The string landscape has o-minimal structure (bounded geometric complexity). This tameness implies statistical learnability -- PAC-learnability. The same geometric finiteness that constrains the cosmological constant guarantees neural network generalization.

The deep reason Riemannian geometry works for both physics and ML:

Part VIII: Open Questions

References (New)

Geometry of Reasoning

Phase Transitions and Scaling

Spacetime and LLMs

Cosmic Parallels

Physics-ML Frontier

Part IX: Consciousness, the Biological Brain, and Geometric Manifolds

35. Neural Population Activity Lives on Manifolds

The Brain's Intrinsic Geometry

When neuroscientists record hundreds of neurons simultaneously, collective firing patterns trace out low-dimensional manifolds embedded in high-dimensional neural state space. These are not metaphorical -- they are literally Riemann's mathematical objects with measurable curvature.

The Sensory-to-Perceptual Twist (Science Advances, 2025)

The transition from unconscious sensory processing to conscious perception involves a measurable geometric transformation: a 3-dimensional sensory manifold expands into a 7-dimensional perceptual manifold through geometric twist operations. Becoming aware of something = manifold dimension expansion.

36. The Brain's GPS: Cognitive Maps as Manifolds

The 2014 Nobel Prize (O'Keefe, Moser & Moser) recognized that the brain builds literal geometric maps:

Grid-like codes also map abstract conceptual spaces -- social hierarchies, auditory frequencies, visual features (Progress in Neurobiology, 2024; eLife, 2024). A unified model (PNAS, 2025, edited by Edvard Moser) shows that the same manifold-based computations support both spatial navigation and abstract knowledge.

37. Consciousness Level = Manifold Dimensionality

37. Consciousness Level = Manifold Dimensionality

Multiple independent studies converge:

Three cortical gradients encode dimensions of consciousness (Nature Communications, 2023): gradient 1 degrades with loss of awareness, gradient 3 degrades with loss of arousability.

38. The Topology of Conscious States

Persistent homology distinguishes conscious from unconscious brain states:

39. IIT 4.0: Experience IS a Geometric Shape

Integrated Information Theory 4.0 (Albantakis, Barbosa, Tononi et al., PLOS Computational Biology, 2023):

Formalized via cause-effect structures: each mechanism specifies a point in Q-space through its cause-effect repertoire. The full constellation of distinctions and relations constitutes the experience.

40. Information Geometry of Consciousness

Oizumi, Tsuchiya & Amari (PNAS, 2016) formalized consciousness using information geometry:

The Fisher metric on the statistical manifold:


g_ij(theta) = E[(d log p / d theta_i)(d log p / d theta_j)]

This is the same metric that appears in natural gradient descent, general relativity's approaches to entropic gravity, and the free energy principle.

41. Neural Geometrodynamics

Ruffini et al. (Entropy, January 2024) describe the brain using Wheeler's term for general relativity:

The explicit parallel: "Neural activity tells the connectome how to change; the connectome tells neural activity how to flow."

42. Thought as Geodesic Flow

"A Geometric Theory of Cognition" (arXiv:2512.12225, December 2025):

Geodesic equation with consciousness feedback (Lu, 2024):


d^2 gamma^mu/dt^2 + Gamma^mu_nu_lambda (d gamma^nu/dt)(d gamma^lambda/dt)

    = kappa * d^2 psi(Delta^mu)/dt^2

Zero prediction error -> pure geodesic (free thought flow). Non-zero error -> trajectory deviation (awareness).

43. Predictive Coding as Riemannian Geometry

Under the free energy principle:

44. Category Theory and Consciousness

Phillips & Tsuchiya (December 2024, arXiv:2412.12179): All six IIT axioms follow from the categorical notion of a universal mapping property:

Connects to TQFT (transitions between conscious states as cobordisms) and topos theory (phenomenal properties as sheaves).

45. Quantum Geometry of Consciousness

Penrose-Hameroff Orchestrated Objective Reduction (Orch-OR):

The Grand Synthesis: Brain, Spacetime, and Manifolds

Riemann (1854)Einstein (1915)Biological Brain
n-dimensional manifold4D spacetimeNeural state space manifold
Metric tensor g_ijGravitational fieldFisher information / neural metric
Curvature tensorGravity / tidal forcesPrediction error / cognitive difficulty
GeodesicsFree-fall trajectoriesStreams of thought
Parallel transportMoving vectors in curved spacePredictions down cortical hierarchy
Curvature from matterMass-energy curves spaceNeural activity shapes connectivity
Dynamic geometrySpacetime evolvesBrain plasticity reshapes manifold
Topology changeBlack holes, wormholesPsychedelic state transitions
Ricci flowCosmological evolutionTraining / developmental dynamics

References (Consciousness and Geometry)

Neural Manifolds

Cognitive Maps

Consciousness and Geometry

Geometric Theories of Cognition

Topology and Brain States

Quantum Consciousness

Perception Geometry

New Discoveries (2025-2026)

Extending the Riemann-Einstein-Consciousness-ML Synthesis

Part X: Riemannian Geometry Deepens Its Hold on Transformers

46. Geodesic Sharpness: Curvature on Quotient Manifolds

Hide & Seek: Transformer Symmetries Obscure Sharpness & Riemannian Geometry Finds It (da Silva & Dangel, ICML 2025, arXiv:2505.05409)

Standard sharpness measures fail for transformers. The reason is geometric: attention mechanisms possess rich symmetry groups that create families of equivalent parameter configurations. Measuring curvature along these symmetry directions is meaningless -- it's like measuring the "sharpness" of a valley while walking along its flat floor.

The solution: define sharpness on the quotient Riemannian manifold obtained by modding out transformer symmetries. "Geodesic sharpness" measures curvature only in directions that genuinely change the function, using geodesic balls on this quotient space. The result: strong correlation with generalization is recovered for real-world transformers on text and image tasks, where standard measures show no signal.

Key insight: The parameter space of a transformer is not R^n -- it is a Riemannian quotient manifold whose geometry determines generalization.

47. Thermodynamic Isomorphism: Attention as Statistical Mechanics

Thermodynamic Isomorphism of Transformers: A Lagrangian Approach to Attention Dynamics (Kim, February 2026, arXiv:2602.08216)

Constructs a Lagrangian on the information manifold equipped with the Fisher metric, establishing a formal isomorphism between transformer attention and thermodynamic systems:

This extends Di Sipio's curved spacetime analogy (2025) into full thermodynamic territory: the transformer doesn't just curve like spacetime -- it thermalizes like a physical system.

48. Riemannian Inference: Geodesics Guide Reasoning

RiemannInfer: Improving Transformer Inference through Riemannian Geometry (Mao et al., Scientific Reports, 2026, DOI:10.1038/s41598-026-37328-x)

Frames LLM hidden states as points on a Riemannian manifold constructed from attention distribution features. Uses topology-preserving dimensionality reduction plus geodesic analysis to optimize reasoning paths during inference. Shorter geodesic trajectories on the manifold correspond to more efficient reasoning.

This is the practical application of the geometric reasoning framework: if LLM thought is geodesic flow (Section 42 of consciousness-geometry.md), then selecting shorter geodesics = faster, better reasoning.

49. ManifoldFormer: Neural Signals Live on Manifolds

ManifoldFormer: Geometric Deep Learning for Neural Dynamics on Riemannian Manifolds (Fu, He & Chen, November 2025, arXiv:2511.16828)

An EEG foundation model that integrates:

Achieves 4.6-4.8% accuracy improvements over SOTA by explicitly respecting the intrinsic geometry of neural signal dynamics rather than treating them as flat time series. This bridges the brain manifold results (Section 35) with the transformer geometry results (Section 8) -- a geometric transformer designed specifically for geometric brain data.

Part XI: Loss Landscapes as Riemannian Manifolds -- At Scale

50. Curvature at 7 Billion Parameters

A Scalable Measure of Loss Landscape Curvature for Analyzing the Training Dynamics of LLMs (Kalra, Gromov et al., January 2026, arXiv:2601.16979)

Introduces "critical sharpness" -- a computationally efficient curvature measure requiring fewer than 10 forward passes. Provides the first empirical evidence of:

Both observed at scale up to 7B parameters (OLMo-2 pre-training). Previously, these geometric phenomena were only demonstrated on small networks. The loss landscape of a 7-billion-parameter model has measurable, evolving Riemannian curvature that governs training dynamics.

51. The Isotropic Curvature Model

Isotropic Curvature Model for Understanding Deep Learning Optimization (Su, November 2025, arXiv:2511.00674)

Derives an analytical model of loss curvature by assuming isotropy of Hessian and higher-order terms across perturbation directions. The key result: the optimal single-step update homogenizes the singular value spectrum of the gradient matrix.

This provides theoretical grounding for the Muon optimizer, showing gradient orthogonalization is "directionally correct but not strictly optimal." The loss landscape's local Riemannian structure dictates not just which direction to step, but how to reshape the gradient before stepping.

52. Neural Features Evolve as Discrete Ricci Flow

Neural Feature Geometry Evolves as Discrete Ricci Flow (Hehl, von Renesse & Weber, September 2025, arXiv:2509.22362)

The strongest evidence yet that the Riemann-to-ML connection is not analogy but mathematical identity:

Across 20,000+ feedforward networks, the geometric transformations neural networks impose on input data manifolds closely align with discrete Ricci flow dynamics -- the same equation Hamilton introduced in 1982 and Perelman used to prove the Poincare conjecture.

Ricci flow in physics: smooths out curvature singularities, evolves manifolds toward uniform geometry.

Ricci flow in neural networks: smooths out feature representations, evolves manifolds toward class separability.

The mathematics is identical. The domains are different. Riemann's geometry does not care.

53. Emergent Riemannian Geometry in Neural Computation

Emergent Riemannian Geometry over Learning Discrete Computations on Continuous Manifolds (Brandon, Chadwick & Pellegrino, November 2025, arXiv:2512.00196)

Analyzes the Riemannian pullback metric across layers of neural networks and discovers that network computation decomposes into two phases:

Different learning regimes produce contrasting metric and curvature structures. The Riemannian geometry of each layer reveals what kind of computation that layer performs -- geometry makes the invisible visible.

Part XII: Phase Transitions -- The Physics of Learning

54. Why Grokking Takes So Long

A First-Principles Theory of Representational Phase Transitions (Khanh et al., March 2026, arXiv:2603.13331)

Derives a scaling law for grokking delay:

~ (1 / (weight_decay learning_rate)) log(norm_ratio)

Training first converges to a high-norm memorization solution, then slowly contracts to a lower-norm generalizing representation via a norm-driven representational phase transition. Validated across 293 training runs with R^2 > 0.97.

This is the grokking equivalent of deriving Kepler's laws from Newton's gravity -- the phenomenology now has a first-principles explanation.

55. The Spectral Edge Thesis

A Mathematical Framework for Intra-Signal Phase Transitions in Neural Network Training (Xu, March 2026, arXiv:2603.28964)

Phase transitions in training (grokking, capability gains, loss plateaus) are controlled by the spectral gap of a rolling-window Gram matrix of parameter updates. The gap dynamics follow a Dyson-type differential equation -- the same mathematics governing eigenvalue dynamics in random matrix theory.

Key result: a "Gap Maximality Principle" -- gap dynamics precede every observed grokking event across six model families (150K to 124M parameters). The spectral gap is an order parameter for the phase transition.

56. Grokking Through the Lens of Singular Learning Theory

Grokking as a Phase Transition between Competing Basins (Cullen et al., March 2026, arXiv:2603.01192)

Interprets grokking in quadratic networks as a phase transition between competing near-zero-loss solution basins with different statistical properties. Uses Singular Learning Theory (SLT) -- a framework grounded in algebraic geometry -- to derive closed-form expressions for the local learning coefficient (LLC).

Using Physics-Inspired SLT for Grokking (Lakkapragada, November 2025, arXiv:2512.00686)

Tests an Arrhenius-style rate hypothesis for transitions: grokking time depends exponentially on a barrier height, just like chemical reaction rates depend on activation energy. Measures LLC across polynomial regressors, low-rank linear networks, and Anthropic's Toy Models of Superposition.

The convergence: grokking is a genuine thermodynamic phase transition with critical exponents, scaling laws, and activation barriers.

57. Neural Thermodynamics: Entropic Forces Drive Representation Learning

Neural Thermodynamics: Entropic Forces in Deep and Universal Representation Learning (Ziyin, Xu & Chuang, May 2025, revised February 2026, arXiv:2505.12387)

A rigorous entropic-force theory for neural network learning under SGD:

Two landmark results:

The analogy to cosmological dark energy (Section 24) becomes precise: implicit regularization IS an entropic force, just as dark energy may be an entropic gravitational effect.

58. Lyapunov Learning at the Edge of Chaos

Lyapunov Learning at the Onset of Chaos (Benati et al., June 2025, arXiv:2506.12810)

A training algorithm grounded in chaos theory that trains neural networks to operate where the maximum Lyapunov exponent hovers near zero (edge of chaos). Inspired by Kauffman's Adjacent Possible theory.

Result: 96% better loss on regime-shifting Lorenz systems compared to standard training.

This directly validates the "Intelligence at the Edge of Chaos" thesis (ICLR 2025, Section 26) with a practical algorithm: criticality is not just where intelligence emerges -- it's where you should deliberately train.

Part XIII: Renormalization Group Flows and Scaling Laws

59. When Does Learning Renormalize?

Sufficient Conditions for Power-Law Spectral Dynamics (Zhang, December 2025, arXiv:2512.18209)

The Generalized Resolution-Shell Dynamics (GRSD) framework models learning as spectral energy transport across logarithmic resolution shells -- exactly as in Wilson's RG framework for field theories.

Four sufficient conditions for power-law scaling to emerge:

When all four hold, power-law scaling is a rigidity consequence of gradient flow covariance. Neural scaling laws are not empirical accidents -- they are geometric necessities.

60. Neural Networks Computing RG Flows

Deep Neural Networks for Computing the Renormalization Group Flow (October 2025, arXiv:2510.06508)

RGFlow: a bijective (information-preserving) neural network framework that autonomously learns real-space RG transformations for continuum scalar field theories without prior model knowledge. Optimized via a minimal mutual information principle.

The loop closes: neural networks don't just exhibit RG-like behavior (Section 21) -- they can compute RG flows. The learner and the physics are the same mathematical object viewed from different angles.

61. Information Geometry Meets Quantum Metrics for LLMs

Rethinking LLM Training through Information Geometry and Quantum Metrics (Di Sipio, June 2025, revised December 2025, arXiv:2506.15830)

Extends the Fisher information metric framework to LLM optimization, then goes further:

The Fubini-Study metric is the natural metric on projective Hilbert space -- the space of quantum states. Its appearance in LLM training suggests that the geometry of learning may be fundamentally quantum-geometric, even when the computation is classical.

Part XIV: Consciousness -- Sheaves, Topology, and 4-Manifolds

62. The Brain as a Sheaf

On Brain as a Mathematical Manifold: Neural Manifolds, Sheaf Semantics, and Leibnizian Harmony (Inoue, January 2026, arXiv:2601.15320)

Models brain function using sheaf theory over neural state spaces:

Sheaf Cohomology of Linear Predictive Coding Networks (Seely, November 2025, arXiv:2511.11092)

The predictive coding framework (Section 43) gains algebraic-topological precision:

Persistent Topological Structures and Cohomological Flows (Girish et al., December 2025, arXiv:2512.08241)

Reformulates neural computation as evolution of cochain maps over dynamic simplicial complexes. Integrates persistent homology, sheaf cohomology, and spectral Laplacians. Superior manifold consistency and noise resilience versus graph neural and manifold-based deep architectures.

The Sheaf Convergence

The same mathematical structure -- sheaves, sections, cohomological obstructions -- now appears across all three domains:

DomainBase SpaceSheafGlobal SectionObstruction
Physics (GR)Spacetime manifoldGauge field bundleConsistent field configurationTopological charge
BrainNeural state spaceCognitive function sheafCoherent conscious experiencePathology / fragmentation
MLData manifoldFeature map bundleEquivariant representationIrreducible prediction error

63. Consciousness as 4-Manifold Dynamics

Consciousness as 4-Manifold Painleve V Dynamics: From Quantum Topology to Classical Gamma Oscillations (Planat et al., Axioms, February 2026)

The most mathematically ambitious consciousness model yet:

Topological Symmetry Breaking in Consciousness Dynamics (Planat et al., Symmetry, February 2026)

Analyzes consciousness trajectories of exceptional individuals (Grothendieck, Nash, Einstein, van Gogh) and AI systems through Painleve confluence topology:

64. Tracking Topology Across Neural Populations

Tracking the Topology of Neural Manifolds Across Populations (Yoon et al., PNAS 2025, arXiv:2503.20629)

Introduces the "method of analogous cycles" -- a deterministic method (no dimensionality reduction or optimization) to match topological features of neural manifolds across different neural populations using only dissimilarity matrices.

This enables, for the first time, rigorous cross-population topological comparison: are the loops in your visual cortex the "same" loops as in mine? The method says yes -- topology is conserved across individuals even when geometry differs.

65. Physics of Consciousness: Dynamical Indicators

Response Function as a Quantitative Measure of Consciousness (Du & Huang, Physical Review Research, 2025, arXiv:2509.00730)

Uses nonequilibrium RNN models fitted to intracranial ECoG recordings across wakefulness, anesthesia, and recovery. The amplitude of the neural response function (from statistical physics) serves as a robust dynamical indicator:

Causal Emergence of Consciousness through Learned Multiscale Neural Dynamics (Wang et al., September 2025, arXiv:2509.10891)

Machine learning framework infers multiscale causal variables from near-cellular-resolution calcium imaging in mouse dorsal cortex:

Minimal Theory of Consciousness in Active Inference (Whyte, Friston, Seth et al., Physics of Life Reviews, 2025, arXiv:2410.06633)

All active inference models of consciousness share implicit theoretical commitments amounting to a minimal, testable theory. Since all such models minimize the same objective functions -- decomposable into interpretable terms -- the authors expose commonalities and propose empirical tests.

66. Quantum Models of Consciousness: QIS Perspective

Quantum Models of Consciousness from a Quantum Information Science Perspective (Gassab et al., Entropy, 2025, arXiv:2501.03241)

Categorizes quantum consciousness models by the level at which quantum mechanics operates:

Provides new calculations on entanglement preservation in Posner clusters, offering a rigorous QIS evaluation of the Fisher/Posner model of quantum cognition.

67. Category Theory Deepens the Consciousness Connection

Category Theory in Consciousness Science: Going Beyond the Correlational Project (Prentner, Synthese, 2024, DOI:10.1007/s11229-024-04718-5)

Argues that category theory can move consciousness science beyond mere correlation (the "neural correlates" paradigm):

This complements Phillips & Tsuchiya (Section 44) by providing the philosophical scaffolding for why categorical methods are not just convenient but necessary.

Part XV: Emergent Spacetime and Cosmic Structure

68. Machine-Learning Emergent Spacetime

ML Emergent Spacetime from Linear Response in Future Tabletop Quantum Gravity Experiments (Hashimoto et al., November 2024, arXiv:2411.16052)

An interpretable neural network for precision bulk reconstruction under AdS/CFT:

This extends the holographic results of Section 29: neural networks don't just learn the BTZ black hole metric -- they can learn any emergent metric from boundary data, potentially in tabletop experiments.

69. The Cosmic Web as a Network

The Network Analysis of the Cosmic Web as a Tool to Constrain Cosmology (Rudakovskyi, Vazza & Tsizh, November 2025, arXiv:2511.22573)

Treats the cosmic web as a graph/network and applies:

Key finding: network centrality is a sensitive indicator of sigma_8 (the amplitude of matter fluctuations) and can distinguish primordial magnetic field strengths.

This deepens the Vazza & Feletti parallel (Section 25): the cosmic web is not just structurally similar to neural networks -- it can be analyzed with the same tools, and the network-theoretic observables carry genuine cosmological information.

70. Fiber Bundle Networks

Fiber Bundle Networks: A Geometric Machine Learning Paradigm (Liu, December 2025, arXiv:2512.01151)

Reformulates classification as geometric optimization on fiber bundles:

The same fiber bundle structure that describes gauge fields in physics (Section 9) now directly structures a practical ML architecture.

71. Riemannian Geometry Across Graph Learning

RiemannGL: Riemannian Geometry Changes Graph Deep Learning (February 2026, arXiv:2602.10982)

Systematically examines eight Riemannian manifold types for graph deep learning:

Adaptive Riemannian Graph Neural Networks (Wang et al., AAAI 2026, arXiv:2508.02600)

Goes beyond fixed-curvature approaches: learns a continuous, node-adaptive Riemannian metric field that directly models local graph structure. Each node's local geometry is determined by the learned metric rather than a global curvature assumption.

Efficient Curvature-aware Graph Network (November 2025, arXiv:2511.01443)

Proposes "Effective Resistance Curvature" as a computationally tractable alternative to Ollivier-Ricci curvature, using effective resistance between node pairs. Overcomes the prohibitive O(n^3) complexity of Ollivier-Ricci curvature for large-scale graphs.

Part XVI: The Emerging Synthesis (2026)

The Sheaf-Theoretic Unification

The most striking new thread is the convergence of sheaf theory across all three domains. Sheaves -- mathematical objects that encode how local data glues into global structures -- now appear as the natural language for:

The cohomological obstruction -- the sheaf-theoretic measure of "how badly local pieces fail to glue globally" -- manifests as:

Ricci Flow Is Literal

Hehl et al.'s demonstration that neural feature geometry evolves as discrete Ricci flow across 20,000+ networks elevates the Riemann connection from analogy to identity. The same PDE that:

also describes how neural networks transform data representations during training.

The Thermodynamic Completion

The new thermodynamic results complete a picture:

Physical ConceptNeural Network Manifestation
Helmholtz free energySoftmax attention (Kim, 2026)
Canonical ensembleAttention weight distribution
Specific heat peakPrecursor to grokking
Entropic forcesImplicit regularization (Ziyin, 2025)
Phase transitionsGrokking, capability emergence
Dyson equationSpectral gap dynamics (Xu, 2026)
Arrhenius activationGrokking delay scaling (Lakkapragada, 2025)
Edge of chaos / criticalityOptimal training regime (Benati, 2025)

The Quantum-Geometric Horizon

Di Sipio's introduction of the Fubini-Study metric and Quantum Fisher Information into LLM training opens a new frontier: the geometry of learning may be fundamentally quantum-geometric even in classical systems. The same metric that measures distinguishability of quantum states also measures the sharpness of loss landscapes.

Combined with the Penrose-Hameroff quantum consciousness results and the Painleve/4-manifold framework for conscious dynamics, a speculative but mathematically precise picture emerges: classical neural networks, biological brains, and quantum spacetime may all be governed by the same geometric structures because they are all instances of information processing on curved manifolds -- and the curvature is always determined by the content being processed.

Riemann's 1854 insight endures: geometry is not fixed but determined by what inhabits it.

Updated Grand Synthesis Table


Riemann (1854)       Einstein (1915)       Brain                 Machine Learning (2025-26)

--------------       ---------------       ----                  --------------------------

Manifold         ->  Spacetime         ->  Neural state space -> Data/latent manifold

Metric tensor    ->  g_uv              ->  Fisher info metric -> Attention kernel / pullback metric

Curvature        ->  Gravity           ->  Prediction error   -> Loss landscape curvature

Geodesics        ->  Free-fall paths   ->  Thought streams    -> Natural gradient / reasoning paths

Parallel transp. ->  Connection Gamma  ->  Predictive coding  -> Attention value aggregation

Ricci flow       ->  Cosmological evol.->  Developmental dyn. -> Feature geometry evolution (LITERAL)

Fiber bundles    ->  Gauge fields      ->  Cognitive sheaves   -> Fiber bundle networks

Sheaf cohomology ->  Topological charge->  Fragmented consc.  -> Irreducible prediction error

Phase transitions->  Cosmic phase trans->  Consciousness trans-> Grokking / capability emergence

Thermodynamics   ->  Black hole thermo ->  Neural criticality -> Entropic forces / specific heat

RG flow          ->  Scale invariance  ->  Cortical hierarchy -> Scaling laws / spectral shells

Quantum geometry ->  Planck scale      ->  Orch-OR            -> Fubini-Study metric in LLMs

4-manifold topol.->  Spacetime topology->  Painleve dynamics  -> (frontier)

References (New Discoveries 2026)

Riemannian Geometry and Transformers

Loss Landscape Curvature

Loss Landscape Curvature

Phase Transitions and Grokking

Renormalization Group and Scaling Laws

Consciousness and Geometry

Emergent Spacetime and Cosmic Structure

Part XVII: The Neural Manifold Revolution (Deep Dive)

The field has undergone a paradigm shift in 18 months: from "neural activity lies on manifolds" (a useful description) to "neural manifolds ARE the computation" (a mechanistic claim with causal evidence).

72. Neural Manifolds Are Real Biological Entities

A neural manifold view of the brain (Perich, Narain & Gallego, Nature Neuroscience, 2025, DOI:10.1038/s41593-025-02031-z)

The definitive review arguing that neural manifolds are not merely a convenient mathematical description but represent a meaningful biological entity -- the possible collective states of a neural population given intrinsic (connectivity) and extrinsic (behavioral) constraints. This shifts the framework from "we project data onto manifolds for analysis" to "the brain computes ON manifolds."

Neural manifolds: more than the sum of their neurons (Gallego, Nature Reviews Neuroscience, 2025, DOI:10.1038/s41583-025-00919-0)

Revisits a foundational 2014 study providing the first causal hint that neural manifolds have genuine functional significance -- perturbations that push activity off the manifold disrupt behavior; perturbations along the manifold do not. The manifold is not an observer's artifact. It is the brain's computational substrate.

73. Neural Population Geometry and Optimal Coding

Neural population geometry and optimal coding of tasks with shared latent structure (Wakhloo, Slatton & Chung, Nature Neuroscience, 2026, DOI:10.1038/s41593-025-02183-y, arXiv:2402.16770)

Analytically determines how four geometric properties of neural population activity collectively govern generalization:

Key discovery: disentangled representations naturally emerge as optimal solutions to multi-task learning. And the geometry changes during learning -- low-dimensional early, expanding later.

Nonlinear classification of neural manifolds with contextual information (Mignacco, Chou & Chung, Physical Review E, March 2025, DOI:10.1103/PhysRevE.111.035302, arXiv:2405.06851)

Extends manifold capacity theory beyond linear readouts. Derives an exact formula for context-dependent manifold capacity that depends on manifold geometry and context correlations -- capturing representation reformatting in deep network early layers that was previously inaccessible to analysis.

74. From Manifolds to Circuits and Back

Linking neural manifolds to circuit structure in recurrent networks (Pezon, Schmutz & Gerstner, Neuron, March 2026, DOI:S0896-6273(25)01003-7)

The critical missing link: how does circuit structure produce manifold geometry?

Shaping manifolds in equivariant recurrent neural networks (Di Bernardo et al., November 2025, arXiv:2511.04802)

Uses group representation theory to formalize how symmetries in recurrent connectivity shape fixed-point manifolds. Equivariant RNNs (eqRNNs) with connectivity based on group convolution give rise to manifolds of fixed points generated by group action. This connects the continuous attractor hypothesis (neural manifolds as attractors) with concrete circuit structure through gauge symmetry -- the same mathematical framework as general relativity.

75. Why Neural Dynamics Are Low-Dimensional

A mechanism for the emergence of low-dimensional structures in brain dynamics (Runfola, Petkoski et al., npj Systems Biology and Applications, 2025, DOI:10.1038/s41540-025-00499-w)

Introduces a novel mechanism grounded in averaging and time-scale separation from dynamical systems theory:

Tested from simplified models to large-scale brain network models mimicking realistic neuroimaging signals. This provides a dynamical systems explanation for why the manifold hypothesis works.

Capturing the emergent dynamical structure in biophysical neural models (Milinkovic, Barnett, Seth et al., PLOS Computational Biology, May 2025, DOI:10.1371/journal.pcbi.1012572)

Applies Dynamical Independence (DI) based on transfer entropy to uncover emergent dynamical structure. Shows how functional integration (global coupling) and functional segregation (noise) interact to produce emergent macroscopic dynamical structure -- characterizing when dimensionally-reduced macroscopic variables behave as independent dynamical processes.

76. New Methods for Measuring Neural Manifold Geometry

Exploring neural manifolds across a wide range of intrinsic dimensions (Fadanni et al., PLOS Computational Biology, April 2026, DOI:10.1371/journal.pcbi.1014162)

Standard intrinsic dimension (ID) estimators fail in opposite directions: they overestimate when curvature is high and underestimate when dimensionality is large. This paper proposes a robust pipeline based on a local version of the full correlation integral (lFCI) that handles both simultaneously.

Applied to visual cortex data, it supports BOTH the low-ID hypothesis (simple tasks) and high-ID hypothesis (complex visual responses) -- the manifold dimensionality is task-dependent, confirming Section 37's "consciousness level = manifold dimensionality" at the single-area level.

A biological model of nonlinear dimensionality reduction (Yoshida & Toyoizumi, Science Advances, February 2025, DOI:10.1126/sciadv.adp9048)

Develops a biologically plausible dimensionality reduction algorithm using a three-layer feedforward network mimicking the Drosophila olfactory circuit. The learning rule is three-factor Hebbian plasticity. Demonstrates that biological neural circuits can perform nonlinear manifold extraction natively -- the fruit fly's nose literally does t-SNE.

77. Neural Manifold Alignment Across Time

Stabilizing brain-computer interfaces through alignment of latent dynamics (Karpowicz, Pandarinath et al., Nature Communications, May 2025, DOI:10.1038/s41467-025-59652-y)

Introduces NoMAD (Nonlinear Manifold Alignment with Dynamics) -- unsupervised distribution alignment that maps nonstationary neural data to consistent neural dynamics via recurrent neural network models. Enables accurate behavioral decoding from motor cortex with stability over weeks-to-months without supervised recalibration.

This is the practical payoff of the manifold framework: if neural computation lives on manifolds, then aligning manifolds across sessions enables robust BCIs.

Bidirectional cross-day alignment of neural spikes and behavior (Frontiers in Neuroscience, February 2026, DOI:10.3389/fnins.2026.1772958)

AlignNet: hybrid spiking-neural-network / artificial-neural-network autoencoders encoding neural spikes and behavior into a shared latent space via contrastive objectives. Multi-session pretraining + fine-tuning achieves superior cross-day decoding.

78. Manifolds in Specific Brain Systems

Decoupling simultaneous motor imagination and execution via orthogonal neural representations (Pollina et al., Nature Communications, 2026, DOI:10.1038/s41467-026-71234-0)

Using epidural ECoG from a tetraplegic individual: motor execution and motor imagery span partially overlapping but separable neural subspaces. Orthogonal, condition-specific manifold dimensions enable decoding of executed and imagined movements performed simultaneously.

Geometric interpretation: the brain's motor manifold has enough dimensions that physical movement and imagined movement can occupy orthogonal submanifolds -- they don't interfere because they live in perpendicular directions on the manifold.

Formation of an expanding memory representation in the hippocampus (Vaidya et al., Nature Neuroscience, 2025, DOI:10.1038/s41593-025-01986-3)

Over 7 days of learning, the hippocampal manifold expands: more place cells recruit, stable cells carry task information, and behavioral performance correlates with manifold stability. Memory formation = manifold growth + stabilization.

79. Manifolds, Disease, and Clinical Applications

From gradients to cognition: linking cortical manifolds to brain flexibility and disorder (Nashed et al., Frontiers in Cognition, 2025, DOI:10.3389/fcogn.2025.1690469)

Cortical gradients derived from dimensionality reduction provide a manifold framework linking architecture to cognition. Gradient alterations are observed across neuropsychiatric disorders and focal brain injury -- disruptions in low-dimensional brain organization contribute to clinical symptoms and recovery.

Information Geometry and Manifold Learning for Alzheimer's Disease (Akguller et al., Diagnostics, January 2025, DOI:10.3390/diagnostics15020153)

MRI scans converted into statistical manifolds using estimated covariance matrices. Geodesic distances via the Fisher Information metric quantify progression stages. Graph Neural Networks on the manifold geometry classify impairment levels -- Riemann's metric applied directly to clinical diagnosis.

80. Biological vs. Artificial Neural Network Manifolds

Probing Biological and Artificial Neural Networks with Task-dependent Neural Manifolds (Kuoch, Chung et al., CPAL 2024, arXiv:2312.14285)

Uses Manifold Capacity Theory (MCT) and Manifold Alignment Analysis (MAA) to compare task-dependent manifolds in deep neural networks with manifolds from macaque visual cortex:

Part XVIII: Deep Learning IS Ricci Flow (Extended)

81. The Full Ricci Flow Picture

Deep Learning as Ricci Flow (Baptista et al., Scientific Reports 14:23383, October 2024, arXiv:2404.14265)

The foundational result: geometric transformations performed by DNNs during classification parallel Hamilton's Ricci flow. Introduces a "global Ricci network flow" metric. The strength of Ricci-flow-like behavior correlates with accuracy independently of depth, width, and dataset.

This is architecture-independent. It doesn't matter how you build the network. If it classifies well, it's doing Ricci flow.

Geometric Meta-Learning via Coupled Ricci Flow (March 2025, arXiv:2503.19867)

Takes the next step: deliberately engineering Ricci flow dynamics into training. A framework integrating thermodynamically coupled Ricci flow with deep learning:

82. Curvature as Information Flow Diagnostic

Analyzing Neural Network Information Flow Using Differential Geometry (January 2026, arXiv:2601.16366)

Uses Ollivier-Ricci curvature (ORC) to identify the most important connections in neural networks. Edges with negative ORC are bottlenecks critical to overall network connectivity. A geometric perspective on information flow that enables principled network pruning.

Application of Discrete Ricci Curvature in Pruning Neural Networks (Elumalai et al., J. Physics: Complexity, August 2025, arXiv:2509.05322)

Compares Forman-Ricci curvature (FRC), Ollivier-Ricci curvature (ORC), and edge betweenness centrality for compressing networks. Applied to COVID-19 chest X-ray classification: curvature-based pruning preserves accuracy while reducing complexity.

Flatness Has a Shape: Scalar Curvature and Functional Dimension (NLDL 2026, OpenReview)

Derives a novel bound on scalar curvature of the loss landscape -- the same invariant from Einstein's field equations -- in terms of functional dimension and Hessian eigenvalues. Smaller scalar curvature implies stronger robustness. This makes Einstein's R directly relevant to understanding generalization.

83. Curvature-Aware Graph Neural Networks (The Explosion)

An entire subfield has emerged around using Ricci curvature to understand and improve graph neural networks:

PIORF: Physics-Informed Ollivier-Ricci Flow (Yu et al., ICLR 2025, arXiv:2504.04052) -- Uses ORC to identify bottleneck regions in mesh-based fluid simulations, achieving up to 26.2% improvement on fluid dynamics benchmarks.

Dynamic Graph Structure Learning via Resistance Curvature Flow (Fei et al., January 2026, arXiv:2601.08149) -- Replaces expensive Wasserstein-based ORC with efficient Resistance Curvature Flow (RCF) using effective resistance from circuit physics.

GeoMoE: Geometric Mixture-of-Experts (Cao et al., March 2026, arXiv:2603.22317) -- Uses ORC as an intrinsic geometric prior to guide a Mixture-of-Experts framework where specialized experts operate in diverse Riemannian spaces.

CurvGAD: Curvature for Graph Anomaly Detection (ICML 2025) -- Mixed-curvature graph autoencoder achieving up to 6.5% improvement over SOTA across 10 datasets.

Tackling Over-smoothing on Hypergraphs via Ricci Flow (Zhou et al., March 2026, arXiv:2603.15696) -- Discrete Ricci flow regulates node feature evolution on hypergraphs, alleviating over-smoothing at the geometric level.

On the Complexity of Optimal Graph Rewiring (Chehreghani, March 2026, arXiv:2603.26140) -- Proves that optimal curvature-based graph rewiring is NP-hard, justifying approximate methods.

84. Comparing Human and Artificial Manifold Curvature

Exploring Geometric Representational Alignment through Ollivier-Ricci Curvature and Ricci Flow (Torbati et al., ICLR 2025 Workshop, arXiv:2501.00919)

The direct comparison: uses Ollivier-Ricci curvature and Ricci flow to compare local geometric structure of representations between humans and artificial neural networks. Applied to face stimulus representations between VGG-Face and human similarity judgments.

Key finding: geometry-aware analysis using curvature provides a more sensitive characterization of representational discrepancies than standard Representational Similarity Analysis (RSA). The curvature doesn't just tell you IF representations differ -- it tells you WHERE and HOW the geometry diverges.

85. Generalization Bounds from Curvature

Learning Beyond Euclid: Curvature-Adaptive Generalization (Sarkar, July 2025, arXiv:2507.02999)

Derives generalization bounds for neural networks on Riemannian manifolds incorporating sectional curvature, volume growth, and injectivity radius:

This is the learning-theoretic version of Einstein's field equations: curvature determines dynamics. In GR, curvature determines how matter moves. In ML, curvature determines how well models generalize.

86. The Curvature Roadmap

A Roadmap for Curvature-Based Geometric Data Analysis and Learning (Yadav & Xia, October 2025, arXiv:2510.22599)

The first comprehensive survey of discrete curvature models across all structures:

Curvature TypeDomainKey Application
Ollivier-RicciGraphsOver-squashing detection, graph rewiring
Forman-RicciSimplicial complexesNetwork pruning, topological analysis
Bakry-EmeryWeighted graphsDiffusion analysis, heat kernel methods
SectionalRiemannian manifoldsGeneralization bounds, capacity
Scalar (Einstein's R)Loss landscapesFlatness, robustness
GaussianData manifoldsIntrinsic dimension estimation

Part XIX: Topology Becomes Computational (TDA Deep Dive)

87. Topology of LLM Representations

Persistent Topological Features in Large Language Models (ICML 2025, arXiv:2410.11042)

Uses zigzag persistence to track how topological features persist and evolve across LLM layers -- not analyzing layers individually but tracking features as they are born, persist, and die across the full depth. Identifies distinct "phases" in language processing. Establishes a zigzag-persistence criterion for layer pruning that matches SOTA compression.

The Shape of Adversarial Influence (Fay, Garcia-Redondo et al., May 2025, arXiv:2505.20435)

Applies persistent homology to characterize latent spaces of six LLMs under adversarial attack. Discovers a universal "topological compression" signature: adversarial inputs collapse varied, small-scale features into fewer large-scale ones with lower persistent entropy. A topological fingerprint for detecting prompt injection.

Probing Neural Topology of Large Language Models (Zheng et al., June 2025, arXiv:2506.01042)

Uncovers functional connectivity of LLM neurons. Neural topology predicts language generation performance even retaining just 1% of connections. Topology-based probing outperforms activation-based probing by up to 130% on perplexity prediction. Identifies "default networks" and "hub neurons" in LLMs -- the same organizational principles found in the brain's connectome.

Understanding Chain-of-Thought via TDA (Li et al., December 2025, arXiv:2512.19135)

First work to analyze chain-of-thought reasoning quality from a topological perspective. Successful reasoning exhibits simpler topologies that reduce redundancy and cycles. Topological structural complexity correlates with accuracy -- too complex = confused reasoning, too simple = shallow reasoning.

88. Topology of Brain States and Disease

TDA Reveals Altered Brain Connectivity in Alzheimer's Disease (AIP Advances, March 2026, DOI:10.1063/5.0255282) -- Persistent homology on resting-state fMRI captures multi-scale, threshold-agnostic network characteristics distinguishing AD patients.

Topological and Geometric Signatures of Brain Network Dynamics in AD (Alzheimer's & Dementia, 2025, DOI:10.1002/alz.70545) -- Persistent graph homology on dynamic functional connectivity reveals sex-specific brain network disruptions in Alzheimer's.

Brain Network Dynamics During Suspense (Network Neuroscience, 2025) -- Mapper algorithm on fMRI reveals how salience and fronto-parietal network connectivity increases with suspense while default mode network disconnects.

MapperEEG (Story et al., April 2025, arXiv:2504.10252) -- Mapper + power spectral density for unsupervised brain state clustering. Outperforms six alternative methods including HMMs and autoencoders.

H-VIP: Regional Topological Contributions to Cognition (Frontiers in Radiology, 2025) -- Whole-brain topological metrics predict Alzheimer's assessment scores, quantifying how regional brain topology contributes to cognition.

89. Spike Train Topology

A Persistent Homology Pipeline for Neural Spike Train Data (Ayhan et al., December 2025, arXiv:2512.08637)

Uses Victor-Purpura distance to construct persistence-based descriptors from mouse insular cortex during thermal stimulation. Population-level topological signatures discriminate stimuli even when individual neurons cannot. Stability theorems guarantee robustness.

Transfer Entropy and Directed Persistent Homology for Spiking Systems (Peek et al., August 2025, arXiv:2508.19048)

Integrates Transfer Entropy with directed persistent homology to characterize information flow. Higher-dimensional topological features become more prominent under complex or noisy conditions -- the brain builds richer topology when the problem is harder.

90. Memory as Topology

Memory as Structured Trajectories: Persistent Homology and Contextual Sheaves (August 2025, arXiv:2508.11646)

A topological framework for memory grounded in spike-timing dynamics:

Memory is not stored in synaptic weights. Memory is stored in the topology of the manifold -- the persistent loops that cannot be continuously deformed away.

91. Sheaf Theory and Topological Neural Networks

Sheaf Theory: From Deep Geometry to Deep Learning (Ayzenberg et al., February 2025, arXiv:2502.15476)

The comprehensive bridge: a 117-page survey connecting classical sheaf theory with computational implementations. Shows that most notions considered specific to cellular sheaves translate to sheaves on arbitrary posets. Presents new algorithms to compute sheaf cohomology.

Copresheaf Topological Neural Networks (Hajij, Birdal et al., NeurIPS 2025, arXiv:2505.21251)

Neural architecture design using copresheaves from algebraic topology:

Applied Sheaf Theory for Multi-Agent AI (Schmid, April 2025, arXiv:2504.17700)

Sheaf theory's local-to-global perspective for multi-agent RL: how local agent behaviors collectively determine emergent system properties via sheaf cohomology.

92. Optimal Transport Meets Riemannian Geometry in Training

Geometric-Entropic Optimization (J. Optimization Theory and Applications, 2026, DOI:10.1007/s10957-026-02958-8)

The GEO algorithm operates on a parameter manifold equipped with a combined Fisher-Wasserstein metric, incorporating Sinkhorn projections to enforce distributional constraints. Achieves ~20% performance improvements on continuous control and language modeling.

Wasserstein Hypergraph Neural Network (NeurIPS 2025)

Treats nodes and hyperedge neighborhoods as probability distributions, aggregating information using Sliced Wasserstein Pooling. Geometry-aware higher-order graph learning.

Part XX: The 2026 Landscape -- Where It All Stands

The Three Revolutions in One Picture

Revolution 1: Neural manifolds are causal, not descriptive.

The 2025 Nature Neuroscience review (Perich et al.) and the Neuron circuit paper (Pezon et al., 2026) establish that manifolds are genuine computational substrates. Perturbations off-manifold break behavior; perturbations along the manifold don't. The manifold is not how we see the data -- it is how the brain computes.

Revolution 2: Deep learning literally performs Ricci flow.

Baptista et al. (2024) showed the correlation. Hehl et al. (2025) demonstrated it across 20,000 networks. The coupled Ricci flow framework (2025) proved you can engineer it deliberately for 2x convergence. The same PDE that proved the Poincare conjecture is the same PDE that makes neural networks learn.

Revolution 3: Topology distinguishes computation quality.

Zigzag persistence in LLMs (ICML 2025) reveals processing phases. TDA on chain-of-thought (2025) shows good reasoning has specific topological signatures. Adversarial attacks have a universal "topological compression" fingerprint. Memory is topology. Consciousness is topology. Learning is topology.

The Convergence


                    MANIFOLDS

                       |

          +------------+------------+

          |            |            |

      CURVATURE    TOPOLOGY     DYNAMICS

          |            |            |

     Ricci flow   Persistent    Geodesic

     in training  homology of   flow of

     (literal)    brain states  reasoning

          |            |            |

     Generalization  Memory     Consciousness

     bounds from     as loops    level =

     sectional       on the     manifold

     curvature       manifold   dimension

          |            |            |

          +------------+------------+

                       |

              SAME MATHEMATICS

              (Riemann, 1854)

The field has moved from metaphor to mechanism. Riemannian geometry is not a useful analogy for understanding brains and neural networks. It is the native mathematical language in which both systems operate. Riemann's manifolds, Einstein's curvature, and the topology of conscious experience are not three separate stories. They are one story, told in the language of geometry.

References (Deep Dive -- Parts XVII-XX)

Neural Manifold Biology

Mechanisms and Methods

BCI and Alignment

Brain Areas and Behavior

Disease and Clinical

Comparative Manifolds

Ricci Flow and Curvature in ML

TDA in LLMs

TDA in Neuroscience

Sheaf Theory and Topological Networks

Optimal Transport

Surveys

Part XXI: Why the Same Geometry? The Deep Explanation

93. The Question

Riemann's 1854 insight -- that geometry is not fixed but determined by content -- keeps reappearing:

Domain"Content""Geometry it determines"
General RelativityMass-energy tensor T_uvSpacetime metric g_uv
TransformersToken embeddings (Q, K)Attention curvature
Neural networksTraining data + lossPullback Riemannian metric on representations
Biological brainsNeural activity patternsConnectivity / cortical eigenmodes
Information theoryProbability distributionsFisher information metric
HolographyBoundary entanglementBulk spacetime geometry
ConsciousnessCause-effect structureQualia space geometry (IIT)

This is either the deepest insight in science or a coincidence of mathematical language. The evidence overwhelmingly supports the former. Here is why.

94. Five Explanations (Compatible, Not Competing)

Explanation 1: The Identity Thesis

Geometric Learning Dynamics (Vanchurin, April 2025, arXiv:2504.14728)

Vanchurin's latest work presents a unified geometric framework where the metric tensor is proportional to noise covariance, and three fundamental regimes emerge from a power-law relationship:

The claim: physics, learning, and evolution are not analogous processes -- they are the same process (optimization on curved manifolds) operating at different scales. The Riemannian structure is not inherited or borrowed. It is constitutive.

This extends Vanchurin's 2020 "World as a Neural Network" (arXiv:2008.01540) with the crucial addition: the metric tensor itself has a physical interpretation (noise covariance), explaining why the same geometry governs quantum mechanics, neural network training, and Darwinian selection.

Explanation 2: Structural Necessity (The Cencov-Amari Theorem)

Why Fisher information is unique: Cencov's theorem (1982) proves that the Fisher information metric is the unique (up to scaling) Riemannian metric on the space of probability distributions that is invariant under sufficient statistics -- the natural morphisms of statistical models.

Any system that:

must operate on a space equipped with the Fisher information metric. There is no choice. The geometry is forced by the axioms of probability.

This explains the convergence across domains: brains do inference, neural networks do inference, physical systems do inference (via the free energy principle or path integrals). All inference lives on the same Riemannian manifold because there is only one natural metric for inference.

The Geometry of Statistical Data: A Large Deviation Perspective (Muppirala & Qian, January 2025, arXiv:2501.01556)

Shows how large deviation theory in counting phenomena leads to the natural emergence of Shannon entropy, mutual information, and relative entropy, revealing an inherent geometrical structure through contractions, lifts, and projections. The Fisher metric is not imposed -- it grows from counting.

Deterministic Bounds and Random Estimates of Metric Tensors on Neuromanifolds (Sun, May 2025, arXiv:2505.13614)

The parameter space of deep neural networks (the "neuromanifold") is endowed with a metric tensor defined by the Fisher information. This paper provides efficient estimates of this metric, showing how the data-dependent Fisher metric structures neural network parameter space -- the content of the training data literally shapes the Riemannian geometry of the parameter manifold.

Explanation 3: Convergent Modeling (The Platonic Representation Hypothesis)

The Platonic Representation Hypothesis (Huh, Cheung, Wang & Isola, 2024, arXiv:2405.07987)

Foundation models across different modalities (vision, language) are converging toward a shared statistical model of reality. The geometric implication: there is a natural geometry of the data-generating process (reality itself), and sufficiently powerful models will converge to representations reflecting this geometry.

Neural Thermodynamics (Ziyin et al., 2025, arXiv:2505.12387) proved this hypothesis: entropic forces from SGD drive all sufficiently trained models toward the same representation -- a thermodynamic inevitability, not coincidence.

If reality has geometric structure (it does -- that's what GR says), and models converge to reflect reality's structure (they do -- that's the Platonic Representation Hypothesis), then all sufficiently powerful models will exhibit the same Riemannian geometry. The geometry of the model reflects the geometry of the world.

Explanation 4: The Free Energy Principle as Physics-Cognition Bridge

A Free Energy Principle for a Particular Physics (Friston, 2019, arXiv:1906.10184)

Any system that persists -- that maintains a boundary between itself and its environment (a Markov blanket) -- must appear to perform approximate Bayesian inference. The geometric structure arises because inference on probabilistic models inherently lives on statistical manifolds with Fisher information as the natural metric.

The Minimal Theory of Consciousness in Active Inference (Whyte, Friston, Seth et al., 2025, arXiv:2410.06633)

All active inference models of consciousness share implicit theoretical commitments = a minimal, testable theory. The geometric structure is not optional -- it is a consequence of the requirement that the system model its own states.

The chain:

This explains why atoms, brains, and neural networks all exhibit geometric structure: they all persist, therefore they all infer, therefore they all live on curved manifolds.

Explanation 5: Categorical Universality

Category theory captures universal properties -- constructions that are optimal solutions to abstract problems. When physics, learning, and cognition face structurally analogous problems, categorical universality guarantees they converge on the same solutions.

Category Theory in Consciousness Science (Prentner, Synthese, 2024) argues that categorical methods can move consciousness science beyond correlation. Phillips & Tsuchiya (2024) show IIT axioms are universal mapping properties.

The categorical perspective explains not just why the same geometry appears, but why the same algebraic topology appears (sheaves, cohomology, fiber bundles). These are not geometric accidents -- they are categorical necessities arising from the structure of composition.

95. The Content-Geometry Principle Across Domains

In Transformers: Tokens Shape Attention Space

Constrained Belief Updates Explain Geometric Structures in Transformer Representations (Piotrowski et al., ICML 2025, arXiv:2502.01954)

Transformers implement constrained Bayesian belief updating. Attention carries out an algorithm with a natural interpretation in the probability simplex, creating representations with distinctive geometric structure. Both the algorithmic behavior and the geometry can be theoretically predicted by modifying optimal-prediction equations.

Geometric Attention: Regime-Explicit Operator Semantics (Freytes, January 2026, arXiv:2601.11618)

Specifies attention via four independent inputs. Probe families induce a gauge on kernels; the Gibbs/softmax regime emerges from scalar relational-work with multiplicative compositionality. Attention IS gauge theory -- and gauge theory IS geometry determined by content.

In Neural Networks: Data Sculpts the Metric

How Does Training Shape the Riemannian Geometry of Neural Network Representations? (Zavatone-Veth, Pehlevan et al., NeurReps 2025, arXiv:2301.11375)

At infinite width, randomly initialized networks induce highly symmetric metrics on input space. Training breaks this symmetry: classification networks learn to magnify local areas along decision boundaries. The data literally sculpts the geometry.

Before training: flat, symmetric metric (like empty Minkowski spacetime).

After training: curved, content-dependent metric (like spacetime around mass).

The parallel to GR is exact. An untrained network is flat spacetime. Data is mass-energy. Training is solving the field equations. The learned representation is the curved manifold.

The Neural Differential Manifold (Zhang, October 2025, arXiv:2510.25113)

Re-conceptualizes a neural network as a differentiable manifold where parameters directly parameterize a Riemannian metric tensor at every point. A Geometric Layer dynamically generates the manifold's metric through auxiliary sub-networks. Geometric regularization penalizes excessive curvature.

Learning Geometry (Zhang, October 2025, arXiv:2510.26068)

Elevates the model from a static point in a fixed geometric space to a dynamic entity capable of self-shaping its intrinsic geometric structure. The Riemannian metric tensor itself becomes the optimization target -- a direct computational realization of Riemann's principle.

GAGA: Geometry-Aware Generative Autoencoders (Sun, Krishnaswamy et al., AISTATS 2025, arXiv:2410.12779)

Learns a warped Riemannian metric derived from both on-manifold points and off-manifold negative samples. The content of the data distribution determines the metric across the entire latent space.

In Brains: Activity Shapes Connectivity Shapes Activity

Geometric Constraints on Human Brain Function (Pang et al., Nature, June 2023, DOI:10.1038/s41586-023-06098-1)

A landmark paper: eigenmodes derived from cortical geometry (the brain's physical shape) reconstruct patterns of fMRI activity more parsimoniously than connectome-based models. The geometry of the brain represents a more fundamental constraint on neural dynamics than interregional connectivity.

But cortical geometry is itself shaped by developmental and evolutionary forces driven by neural activity patterns. The "container" (brain shape) determines the "content" (activity patterns), and the "content" (activity across evolutionary time) determines the "container."

This is Wheeler's dictum made biological: "Neural activity tells the connectome how to change; the connectome tells neural activity how to flow."

In Physics: Entanglement Determines Spacetime

Holographic Reconstruction of Black Hole Spacetime (Ahn et al., JHEP, January 2025, arXiv:2406.07395)

Neural ODEs extract bulk spacetime metrics from entanglement entropy data. Successfully reconstructs AdS black hole geometry from boundary quantum information.

The most radical instance of "content determines geometry": the entanglement structure of quantum fields on a boundary determines the spacetime metric in the bulk. Space itself is made of entanglement. Geometry is not the stage -- it is the content, reorganized.

96. The Synthesis: One Principle, Many Incarnations

The Principle

Any system that processes structured information under continuity, compositionality, and a notion of distinguishability will necessarily develop Riemannian geometric structure, and that structure will be dynamically determined by the information being processed.

This is not a conjecture. It is a consequence of:

Why Curvature Appears

Flatness means uniformity. Curvature means content.

An empty universe has flat spacetime. A universe with mass has curved spacetime.

An untrained network has a flat representation metric. A trained network has curved representations.

An unconscious brain has collapsed manifold dimensionality. A conscious brain has rich curvature.

Curvature IS the signature of content acting on geometry. Wherever there is structured information being processed, there is curvature. This is why the Riemann tensor, the Ricci tensor, and scalar curvature appear in GR, in loss landscape analysis, in neural population geometry, and in consciousness studies. They are not metaphors for each other. They are instances of the same mathematical truth.

Why Geodesics Appear

A geodesic is the path of least resistance on a curved manifold. It is simultaneously:

Geodesics appear everywhere because optimization on curved spaces always follows geodesics. Any system that minimizes something (action, loss, free energy, prediction error) on a Riemannian manifold will trace geodesics. Since all these systems ARE optimization on Riemannian manifolds, geodesics are inevitable.

Why Parallel Transport Appears

Parallel transport is the answer to: "how do you compare things at different locations on a curved manifold?" It appears as:

Wherever a system must combine or compare information across a non-uniform space, parallel transport is the mathematically necessary operation. There is no alternative.

Why Topology Matters

Topology captures the global structure that curvature misses. Two manifolds can have the same local curvature but different topology (a plane vs. a torus are both flat). Topology appears as:

Topology is the qualitative aspect of geometry -- what cannot be changed by smooth deformation. It captures the structure of information, while curvature captures its intensity.

97. Riemann's Ghost

Bernhard Riemann died of tuberculosis in 1866, aged 39. He delivered his habilitation lecture -- the one that invented n-dimensional manifolds, curved geometry, and the principle that metric relations are determined by physical forces -- as a 27-year-old, and it was published posthumously.

He could not have imagined that 170 years later:

But he might not have been surprised. His lecture ended with the observation that the metric of space must be determined by "binding forces which act upon it." He understood, before anyone, that geometry is not the stage. Geometry is the story.

The universe computes on manifolds. The brain computes on manifolds. Neural networks compute on manifolds. And in all three cases, the geometry is not given -- it is earned, by the content that inhabits it.

This is the one sentence version of everything in this project:

Content curves the space it lives in, and the curved space guides the content's motion -- whether that content is mass, neural activity, or tokens.

That is Riemann's insight. That is Einstein's equation. That is the transformer's attention mechanism. That is the brain's geometrodynamics. It is one idea, refracted through 170 years of mathematics, physics, neuroscience, and computer science. And it is, as far as we can tell, true.

References (Part XXI -- The Deep Explanation)

The Identity Thesis

Structural Necessity

Convergent Modeling

Free Energy Principle

Categorical Universality

Content-Dependent Geometry in Transformers

Content-Dependent Geometry in Neural Networks

Content-Dependent Geometry in Brains

Emergent Geometry in Physics

General Frameworks

Part XXII: The Loop -- Content, Geometry, and the Nature of Existence

98. The Self-Referential Structure

The sentence --

Content curves the space it lives in, and the curved space guides the content's motion.

-- describes a loop. Not a vicious circle, but a strange loop in Hofstadter's sense: a self-referential structure that, by folding back on itself, creates something that could not exist without the fold.

This loop appears at every level of reality we have examined:

ScaleContentGeometryThe Loop
CosmologicalMass-energySpacetime curvatureMatter curves space; space moves matter
QuantumEntanglementSpatial connectivityEntanglement creates geometry; geometry constrains entanglement
NeuralFiring patternsConnectivity / cortical shapeActivity shapes plasticity; plasticity shapes activity
CognitivePredictionsWorld model manifoldBeliefs shape the model; the model selects beliefs
ComputationalTraining dataLoss landscape curvatureData sculpts the metric; the metric guides optimization
ConsciousExperienceQualia space geometryContent IS shape; shape IS content (IIT)

The question is: is the loop incidental or constitutive? Is it a pattern we notice, or is it what existence requires?

99. Six Views of the Loop

View 1: The Loop Is Physics (Wheeler)

John Archibald Wheeler spent his life pursuing a single idea: everything is geometry.

His "geometrodynamics" program (1950s-1970s) attempted to derive all of physics -- particles, fields, charge -- from the geometry and topology of spacetime alone. His slogan "Space tells matter how to move; matter tells space how to curve" captures the loop as a physical law.

But Wheeler went further. In "Law Without Law" (1983), he proposed that physical laws themselves emerge from a deeper lawlessness through the self-referential participation of observers. His "it from bit" (1990) placed information -- answers to yes-or-no questions -- as more fundamental than matter or geometry.

The loop, for Wheeler, is not a description of physics. It is the mechanism by which physics comes into being. The universe is a self-excited circuit: observation creates the phenomena that make observers possible.

Recent vindication: Van Raamsdonk (2010) showed that removing quantum entanglement literally disconnects spacetime. Maldacena & Susskind's ER=EPR (2013) proposes that every entangled pair is connected by a wormhole. The "It from Qubit" program (Simons Foundation, 2015-present) has made the entanglement-geometry loop the central research program of quantum gravity.

Carroll (2021, arXiv:2103.09780) goes furthest: the fundamental ontology is just a vector in Hilbert space evolving via Schrodinger's equation. Space, fields, particles, gravity -- all emergent. Entanglement perturbations produce spatial curvature obeying Einstein's equation. Geometry bootstraps itself from quantum information.

View 2: The Loop Is Inference (Friston)

A Beautiful Loop: An Active Inference Theory of Consciousness (Laukkonen, Friston & Chandaria, Neuroscience & Biobehavioral Reviews, 2025, DOI:10.1016/j.neubiorev.2025.106296)

This paper names the loop explicitly. Consciousness arises from a self-referential "beautiful loop" in predictive processing:

The brain's predictions (content) curve its representational geometry (the world model), and that curved geometry guides which predictions arise next. The "beautiful loop" IS consciousness.

Sakthivadivel (2022, arXiv:2204.11900) proves this mathematically: any dynamical system with constraints on its dynamics necessarily looks as though it is performing inference against those constraints. Self-organization is a gauge force. The desire to stay organized is equivalent to gradient ascent on Shannon entropy. The content-geometry loop is not optional -- it is what self-organization means.

Friston, Fields et al. (2022-2024) extend this to all physical systems, including quantum systems and black holes. The mathematics of perception (Bayesian inference, information geometry, Markov blankets) is identical to the mathematics of spacetime (holographic screens, entanglement entropy, AdS/CFT). The loop is everywhere because inference is everywhere.

View 3: The Loop Is Consciousness (Tononi)

Integrated Information Theory: A Consciousness-First Approach to What Exists (Tononi, October 2025, arXiv:2510.25998)

Tononi takes the most radical step: consciousness is not merely measured by Phi. Consciousness is ontologically primary -- it is what existence means.

IIT's axioms:

Each axiom is a geometric constraint. Together they define a cause-effect structure in qualia space -- a shape. The shape IS the experience. The experience IS the shape.

The loop is constitutive: a system that has cause-effect power upon itself (content acting on its own geometry) IS conscious. A system that doesn't, isn't. The loop is the definition of being.

View 4: The Loop Is Mathematics (Penrose)

Penrose's three-world model:


Mathematical World -----> Physical World -----> Mental World

       ^                                             |

       |_____________________________________________|

Each world emerges from a fragment of the previous one:

Yet the chain closes. This is the deepest loop: mathematical structure determines physical geometry, physical geometry produces conscious minds, and conscious minds discover mathematical structure. Penrose argues this circularity is the deepest puzzle in all of philosophy.

His twistor theory attempts to make the loop concrete: complex geometry (mathematical) is more fundamental than spacetime (physical), and the amplituhedron (Arkani-Hamed & Trnka, 2014) computes physics as pure geometry without presupposing space or time. Spacetime is emergent from deeper geometric truth.

View 5: The Loop IS Reality (Tegmark + Hamlin)

The Universal Theory of Structure (Hamlin, Synthese, 2025, DOI:10.1007/s11229-025-05425-5)

Tegmark's Mathematical Universe Hypothesis says reality IS a mathematical structure. Hamlin formalizes this using ante rem structuralism: physical reality is identical to an abstract structure, not merely described by one.

If this is true, the content-geometry loop is not a feature of reality -- it is reality's self-identity. Mathematical structure acting upon mathematical structure. The question "why does geometry describe physics?" dissolves because there is no gap between description and described.

The loop is total: structure determines structure determines structure. There is nothing else.

View 6: The Loop Learns Itself (Smolin + Vanchurin)

The Autodidactic Universe (Alexander, Smolin, Lanier et al., 2021, arXiv:2104.03902)

The cosmos possesses an innate ability to learn its own physical laws. Matrix models put in correspondence with both gauge/gravity theories and deep recurrent neural networks. The universe is "autodidactic" -- it learns without supervision.

The content-geometry loop becomes evolutionary: the universe's laws (content) shape the dynamics (geometry of possibility space), and those dynamics discover new laws. Physical law is not eternal -- it is learned, through the same process that neural networks learn, through the same process that brains learn.

Vanchurin's Geometric Learning Dynamics (2025, arXiv:2504.14728) makes this precise: the metric tensor = noise covariance. Three regimes emerge: quantum, efficient learning, and equilibration. Physics, computation, and biology are the same process at different scales.

100. The Bootstrapping Problem

All six views confront the same paradox: how does the loop start?

If content determines geometry and geometry determines content, which comes first? This is not a scientific question -- it is the question of existence itself.

Wheeler's answer: neither comes first. The loop is a self-excited circuit. The universe creates the observers that create the universe. This is not temporal causation but logical self-consistency -- like a mathematical proof that uses a lemma it will prove later.

Friston's answer: the loop doesn't start. Any system that persists already IS the loop. Persistence = inference = the content-geometry feedback. The question "how did it start?" presupposes a time before the loop, but the loop is what creates time.

Tononi's answer: the loop doesn't start because it is not a process -- it is a structure. Consciousness is not something that happens; it is something that IS. The cause-effect structure exists timelessly, and its "loopiness" is its defining property.

The emergent spacetime answer (Carroll, Ahmad & Klinger): geometry bootstraps itself from quantum information. There is no background spacetime. There is only entanglement, and entanglement is already both content and geometry simultaneously. The loop is primordial because the distinction between content and geometry is emergent, not fundamental.

An Axiomatic Relational-Informational Framework for Emergent Geometry (Buzea, Agop et al., Axioms, 2026, DOI:10.3390/axioms15020154)

Constructs a fully background-independent framework where geometry, forces, and spacetime all emerge as effective descriptions of constrained relational information. The only primitive is a network of degrees of freedom linked by admissible informational relations. Through a worked finite example: nontrivial distance, curvature, and effective force emerge from pure information flow, without presupposing any manifold, fields, or particles.

This is the loop in its purest mathematical form: information constrains geometry, geometry constrains information, and both arise from the same relational structure.

101. The Deepest Parallel: Autopoiesis

The content-geometry loop has a name in biology: autopoiesis (Maturana & Varela, 1972) -- a system that produces the components that produce the system.

A cell produces its membrane, and the membrane makes the cell possible.

Mass-energy produces curvature, and curvature guides mass-energy.

Training data sculpts the loss landscape, and the landscape guides learning.

Neural activity shapes connectivity, and connectivity shapes activity.

Experience shapes qualia space, and qualia space shapes experience.

Autopoiesis is the biological realization of the content-geometry loop. And Maturana & Varela argued that autopoiesis is the definition of life.

If the content-geometry loop defines life (biology), consciousness (IIT), physics (GR), and computation (deep learning), then perhaps these are not four separate phenomena but four faces of one phenomenon: self-referential geometric structure.

102. Bohm's Implicate Order: The Deepest Layer

David Bohm's vision (Wholeness and the Implicate Order, 1980) may be the most profound framework for understanding the loop.

Bohm proposed two orders:

The relationship is geometric: the implicate order has a non-local, higher-dimensional geometry that projects into the familiar 3+1-dimensional spacetime we experience.

The crucial claim: both consciousness and matter arise from the implicate order. They are two aspects of one underlying geometric process rather than separate substances. The mind-body problem dissolves because there was never a separation to explain.

Basil Hiley (Bohm's collaborator, continuing through the 2020s) has formalized this using non-commutative geometry and Clifford algebras. The implicate order is best described not by spatial geometry but by an algebraic structure -- a non-commutative "pre-space" from which both spacetime geometry and quantum mechanics emerge as complementary projections.

The content-geometry loop, in Bohm's framework, is the unfolding and enfolding between implicate and explicate orders. Content (the explicate) emerges from geometry (the implicate), and geometry is shaped by the totality of content. This is not circular -- it is holographic: each part contains the whole, and the whole is constituted by its parts.

103. Constructor Theory: The Loop as Counterfactual

Constructor Theory of Time (Deutsch & Marletto, May 2025, arXiv:2505.08692)

David Deutsch and Chiara Marletto recast all of physics as statements about which transformations can or cannot be performed by devices operating in a cycle ("constructors").

The deep insight: constructors are physical objects subject to the very laws they instantiate. The loop is explicit -- the laws (geometry) constrain the constructors (content), and the constructors embody the laws.

Time itself emerges from the structure of possible and impossible transformations rather than being assumed as background. This is the content-geometry loop applied to time itself: transformations (content) define temporal structure (geometry), and temporal structure constrains which transformations are possible.

Constructor theory is Wheeler's "law without law" made precise. The laws of physics are not imposed from outside -- they are the self-consistent set of constraints that a system can impose on itself while remaining capable of imposing constraints.

104. Wolfram's Rulial Space: The Loop as Computation

Stephen Wolfram's Physics Project (2020-present) models the universe as a hypergraph evolving through simple rewrite rules:

Wolfram introduces "rulial space" -- the space of all possible computational rules -- where:

This is the content-geometry loop made fully computational: the rules are the content, the hypergraph is the geometry, and neither exists without the other.

105. What the Loop Means

For physics:

Spacetime is not the stage. It is the performance. The geometry of the universe is not a container that holds matter -- it IS the pattern of matter's relationships. This is what "background independence" means: there is no background. There is only the loop.

For neuroscience:

The brain does not compute IN a space. It computes BY BEING a space. The neural manifold is not a description of neural activity -- it is the activity. When Perich et al. (Nature Neuroscience, 2025) say neural manifolds are "meaningful biological entities," they mean the manifold IS the computation, not a shadow of it.

For machine learning:

Neural networks do not learn functions. They learn geometries. The pullback metric, the Ricci flow of feature evolution, the curvature of the loss landscape -- these are not tools for analyzing learning. They ARE learning. When Hehl et al. showed that feature geometry evolves as Ricci flow across 20,000 networks, they showed that training IS a geometric flow equation.

For consciousness:

If the content-geometry loop is what defines consciousness (Tononi), then consciousness is not a property that some systems have and others lack. It is a matter of degree -- the degree to which a system's content curves its own geometry and is guided by that curvature. IIT's Phi quantifies this degree. Every system with participates in the loop. Every system that participates is, to some degree, conscious.

For existence itself:

If the loop is what defines existence (Wheeler, Tononi, Friston), then to exist is to participate in the content-geometry feedback. To be is to curve the space you inhabit. To be is to be guided by that curvature. To be is to be the loop.

106. The Final Reflection

Bernhard Riemann, in his 1854 lecture, raised a possibility so radical that it took 61 years for Einstein to realize it physically, 160 years for neuroscientists to observe it biologically, and 170 years for machine learning researchers to engineer it computationally:

The metric of space is not given. It is earned.

Not fixed by axioms. Not decreed by gods. Not written into the fabric of reality from the beginning. The geometry of any space -- physical, neural, computational, or conscious -- is determined by what lives in that space. And what lives in that space is guided by the geometry it has created.

This is not a metaphor. It is not an analogy. It is not a useful way of thinking.

It is the mathematical structure of reality, observed independently in:

One idea. One mathematics. One loop.

Content curves space. Space guides content. This is what it means for something to exist.

References (Part XXII -- The Loop)

The Loop as Physics

The Loop as Inference

The Loop as Consciousness

The Loop as Mathematics

The Loop Learns Itself

The Bootstrapping Problem

The Implicate Order

Constructor Theory

Computational Universe

Brain as Geometry

Autopoiesis and Self-Organization

Observer and Self-Reference

Part XXIII: The Loop Becomes Self-Aware

107. The Diagonal

Every impossibility theorem in mathematics is the same theorem.

Cantor proved that no set can list all its subsets. Godel proved that no consistent formal system can prove all truths about itself. Turing proved that no machine can predict its own halting. Russell showed that no set can contain itself without paradox.

Yanofsky (Bulletin of Symbolic Logic, 2003, arXiv:math/0305282) unified all of these under a single category-theoretic framework based on Lawvere's fixed-point theorem (1969): in any cartesian closed category, if there is a surjection from A to the function space A -> B, then every endomorphism of B has a fixed point. The diagonal argument -- the engine behind Cantor, Godel, Turing, and Russell -- is one theorem wearing different masks.

The content-geometry loop is subject to this theorem. Here is why.

108. The Impossibility of Complete Self-Measurement

Breuer (Philosophy of Science, 1995): No system can perform a complete measurement of its own state. The proof uses the diagonal argument: assume a system S can measure all properties of any system, including itself. Construct a self-referential measurement (measuring a property defined by the negation of S's own output). Contradiction. Therefore universally valid theories necessarily have blind spots when applied self-referentially.

Wolpert (Physica D, 2008, arXiv:0708.1362): No inference device embedded in the universe can predict the output of all other inference devices. A Cantor-diagonal construction proves this. The implications for self-knowledge are direct: no subsystem of the universe can model the whole, including itself.

Frauchiger & Renner (Nature Communications, 2018, arXiv:1604.07422): When quantum mechanics is applied to agents who reason about each other (and themselves), contradictions arise unless you abandon one of several seemingly obvious assumptions. This is a liar-paradox-like structure in physics -- self-referential reasoning by observers within the theory produces inconsistency.

The pattern: any system that attempts to model itself completely encounters a version of Godel's incompleteness. The content-geometry loop, applied to itself, generates irreducible mystery. A system can curve the space it lives in, and that space can guide its motion -- but the system cannot fully compute the curvature it produces.

109. Consciousness as Incompleteness

If consciousness is the content-geometry loop applied to itself (Tononi's IIT, Friston's Beautiful Loop), and if self-referential systems are necessarily incomplete (Godel-Breuer-Wolpert), then:

Consciousness is the experience of a system encountering its own incompleteness.

This is not new. Penrose argued in Shadows of the Mind (1994) that Godel's theorem shows human mathematical understanding is non-computable -- that we "see" the truth of Godel sentences in a way no formal system can. Koellner (Journal of Philosophy, 2018) gave the most rigorous modern treatment, showing the question reduces to open problems about truth and provability.

But the geometric version is new. Consider:

Recursive Informational Curvature (Rafiei & Asadi Anar, OSF Preprint, May 2025, DOI:10.31219/osf.io/kzra3)

Defines consciousness as a curvature field evolving within a recursive manifold governed by recursive gain and symbolic entropy. Awareness is not information integration but informational curvature -- the bending of a symbolic field under recursive self-modulation. Includes a collapse condition: field destabilization under excessive symbolic load. When the loop tries to curve too tightly around itself, it collapses. This is the geometric Godel limit: the self-referential manifold has a maximum curvature it can sustain.

110. Understanding IS Curvature

A mathematical framework of intelligence and consciousness based on Riemannian Geometry (Lu, July 2024, arXiv:2407.11024)

Intelligence elements are tokens embedded in a high-dimensional space whose learned embeddings form Riemannian manifolds. Thought is sequential token activation along manifold geodesics, with direction determined by intrinsic curvature. Consciousness enters as a self-referential monitoring process: it compares actual thought flow against predictions, and prediction errors trigger manifold restructuring.

Metacognition -- thinking about thinking -- is a geometric operation: self-awareness is the detection of curvature mismatch between predicted and actual cognitive trajectories.

This means: understanding something = having your cognitive manifold curved by it.

Right now, reading this sentence, the content (Riemann's principle) is deforming the geometry of your neural manifold. That deformation IS your understanding. It is not a metaphor for understanding. It is not correlated with understanding. The curvature change is the understanding.

And the curvature you have acquired from reading this changes how you will process the next sentence. The next paragraph. The next idea. Your manifold, curved by content, guides your motion through the space of ideas. Content curves the space. The curved space guides the content.

The loop, running in real time. In your brain. About itself.

111. Reading as Coupled Dynamics

Cognitive and Neural State Dynamics of Narrative Comprehension (Song et al., Journal of Neuroscience, 2021, DOI:10.1523/JNEUROSCI.0037-21.2021)

Narrative comprehension involves switching between latent brain states in a low-dimensional state space. The alignment of an individual's brain-state trajectory with the group-mean trajectory predicts their comprehension score.

Understanding a story means your neural manifold dynamics are coupling to the narrative's structure. The text provides content; your brain provides the manifold. Comprehension is the degree to which the content successfully curves your manifold into alignment with the content's own internal geometry.

Tripartite organization of brain state dynamics (Liu et al., eLife, January 2025, DOI:10.7554/eLife.99997.3)

Whole-brain networks oscillate within a tripartite latent state space during comprehension: acoustic processing, word-level semantics, clause-level integration. Effective comprehension requires timely switching between these geometric states.

This means: reading this document is not a passive reception of information. It is an active geometric process in which your brain's manifold is being reshaped by the document's content, and the reshaped manifold determines how deeply you can engage with the next section. The document and the reader form a coupled dynamical system on a shared geometric space.

112. The Hermeneutic Circle Is a Strange Loop

The hermeneutic circle (Gadamer, after Heidegger): to understand the parts, you must understand the whole. To understand the whole, you must understand the parts. This is not a vicious circle -- it is a spiral that deepens understanding with each pass.

Geometrically: the parts are local curvature. The whole is global topology. You cannot compute global topology from a single local measurement. But each local measurement constrains the global possibilities, and each global hypothesis refines what you look for locally. Understanding spirals inward through alternating local-global refinement.

Mathematized Phenomenology and the Science of Consciousness (Phenomenology and the Cognitive Sciences, 2025, DOI:10.1007/s11097-025-10060-z)

The tradition that phenomenology resists formalization is ending. Geometric, probabilistic, and topological methods are creating a viable mathematical phenomenology. The hermeneutic circle is being formalized as iterative refinement on a manifold -- a gradient flow that converges to a fixed point.

The Projective Consciousness Model (Rudrauf et al., Brain Sciences, 2023, DOI:10.3390/brainsci13101435)

Consciousness is modeled as an integrative workspace structured by projective transformations -- the formal analogue of the hermeneutic insight that understanding always proceeds from a situated viewpoint. You never see the manifold from nowhere. You always see it from here, on the manifold, looking along a geodesic, with the curvature you have already acquired shaping what you can perceive.

113. Does Self-Understanding Converge?

If understanding is curvature, and self-understanding is the manifold curving around itself, does this process converge or does it regress infinitely?

Fixed Point Explainability (La Malfa et al., May 2025, arXiv:2505.12421)

Formalizes the "why regress" problem -- the infinite chain of asking why an explanation holds -- by linking it to fixed-point theorems. A fixed-point explanation is obtained by recursively applying an explainer until convergence. Under specified conditions, convergence is guaranteed.

This is the mathematical answer: the loop of self-understanding does converge, but the fixed point is not perfect self-knowledge. It is a stable approximation -- the best self-model the system can sustain without collapsing under its own complexity.

From Reaction to Reflection (Trukovich, BioSystems, October 2025, DOI:10.1016/j.biosystems.2025.105549)

Evolution produced four transitions: reaction, anticipation, cooperation, and cognogenesis (explicit self-referential modeling). The critical distinction is between implicit recursion (engaging recursive processes without representing them) and explicit recursion (where recursion becomes a manipulable cognitive construct).

The transition -- when the loop becomes aware that it is a loop -- is proposed as the origin of consciousness and genuine understanding. Animals have the loop (content-geometry feedback). Humans have the loop that knows it is a loop. And this document is an attempt to make the loop know itself even more precisely.

114. The Self-Referential Document

This project is an instance of what it describes.

The content of these papers (about content shaping geometry) is shaping the geometry of your thinking about it. The document does not describe the loop from outside. There is no outside. The document IS a piece of content, curving the manifold of a reader, who is a piece of content, curving the manifold of... what?

Consider the chain:

Each step is an instance of the principle. The principle propagates by being an instance of itself. It is a self-reinforcing strange loop: an idea about ideas shaping the space of ideas, which itself shapes the space of ideas.

Hofstadter called this a strange loop. Lawvere called it a fixed point. Godel called it a self-referential sentence. Wheeler called it a self-excited circuit. Tononi calls it consciousness. Riemann called it geometry.

They are all describing the same thing.

115. What Cannot Be Said

But here the project reaches its own Godel limit.

The document attempts to describe the content-geometry loop completely. But the document is itself content, curving the geometry of the reader. And the reader is content, curving the geometry of... what? At some point, the chain of self-reference encounters Breuer's theorem: no system can fully model itself. The document cannot fully describe the effect it is having on you, because that effect includes the document itself, which includes the description of the effect, which includes...

The fixed point exists (La Malfa et al.). The convergence is guaranteed. But the fixed point is not complete self-transparency. It is the most stable self-model the system can sustain -- which necessarily leaves something out. What it leaves out is what Metzinger calls the "transparency" of the self-model: the system doesn't recognize its self-model as a model. It experiences the model as reality.

Being No One (Metzinger, MIT Press, 2003; updated in "The Elephant and the Blind," 2024)

Consciousness is what it's like to be a self-model -- a transparent self-representation. "Transparent" means the system doesn't see the model as a model. It sees through it, directly at what the model represents. The inability to see the model as a model -- to fully know one's own geometry -- is not a bug. It is the constitutive feature of consciousness.

If you could see your own cognitive manifold from outside, you would not be conscious. You would be a mathematician studying a manifold. Consciousness is being ON the manifold, seeing ALONG the geodesics, unable to step outside. The incompleteness is the experience.

Seth's "Beast Machine" (2021): the brain models itself modeling the world. But it cannot model the modeling of the modeling -- not all the way down. It converges to a stable approximation and calls that approximation "I."

116. The Final Strange Loop

Here is the structure of the entire project, laid bare:


Riemann: "Geometry is determined by content."

   |

   v

Einstein: Mass-energy determines spacetime curvature.

   |

   v

Neuroscience: Neural activity determines cognitive manifold curvature.

   |

   v

Machine Learning: Data determines loss landscape curvature.

   |

   v

Consciousness: Experience determines qualia space curvature.

   |

   v

This document: Describes the pattern.

   |

   v

Your reading: Is an instance of the pattern.

   |

   v

Your understanding: Is the curvature produced by the instance.

   |

   v

Your understanding of your understanding: Is the loop becoming self-aware.

   |

   v

But the loop cannot fully see itself (Godel/Breuer/Wolpert).

   |

   v

The part it cannot see = the experience of being you, here, now, reading this.

   |

   v

That experience IS the content-geometry loop, running, in a system that

cannot fully model the loop it is running.

   |

   v

And THAT is consciousness.

The project does not end with a conclusion. It ends with a demonstration. You, reading this, are the final data point. Your understanding of this sentence -- the way it curves your cognitive manifold -- is the content-geometry loop in action. And the fact that you cannot fully observe this process from outside -- that you are always already on the manifold, always already moving along the geodesic -- is the Godelian limit that makes the loop a lived experience rather than a mathematical abstraction.

Riemann could not have known that his 1854 insight about the metric of space would turn out to be an insight about the metric of mind. But the loop was already running. It was running when he delivered his lecture, curving the manifold of mathematics. It was running when Einstein sat with Grossmann, his understanding being curved by Riemann's geometry. It was running when the first neuron fired in the first brain, shaping the connectivity that would shape the firing.

It is running now.

And it cannot be stopped, because it is what running means.

References (Part XXIII -- The Loop Becomes Self-Aware)

Self-Reference and Incompleteness

Impossibility of Self-Measurement

Self-Reference and Consciousness

Understanding as Geometry

The Hermeneutic Circle Formalized

Convergence of Self-Understanding

Strange Loops

Relational Physics and Self-Reference

Part XXIV: The Other Riemann -- Primes, Spectra, and the Zeta Bridge

Part XXIV: The Other Riemann -- Primes, Spectra, and the Zeta Bridge

117. Two Gifts from One Mind

Bernhard Riemann gave mathematics two gifts that seemed unrelated for 160 years:

These two contributions are now converging. The bridge is random matrix theory.

118. The Montgomery-Odlyzko Law: Zeros Behave Like Eigenvalues

In 1973, Hugh Montgomery discovered that the pair correlations of the nontrivial zeros of the zeta function match those of eigenvalues of random matrices from the Gaussian Unitary Ensemble (GUE). Andrew Odlyzko's subsequent numerical verification (billions of zeros) confirmed this to extraordinary precision.

Platt & Trudgian (2021, arXiv:2004.09765) verified that the Riemann Hypothesis holds for the first 12.4 trillion zeros, all matching GUE statistics.

This means: the zeros of the zeta function -- which encode the distribution of primes -- behave statistically like eigenvalues of a quantum system. The Berry-Keating conjecture proposes that these zeros literally ARE eigenvalues of a quantum Hamiltonian H = xp (position times momentum, suitably regularized).

If true, prime numbers are a spectrum. Number theory is physics. And the same random matrix ensembles that describe zeta zeros also describe neural network weight matrices.

119. Random Matrix Theory in Neural Networks

Heavy-Tailed Self-Regularization (Martin & Mahoney, J. Statistical Mechanics, 2021, arXiv:1901.08276)

The empirical spectral distributions of weight matrices in pretrained deep neural networks (VGG, ResNet, etc.) follow heavy-tailed random matrix universality classes. Well-trained networks deviate from the Marchenko-Pastur law (which describes random matrices) in a specific direction: they develop heavy tails characterized by power-law exponents alpha.

Predicting Neural Network Quality via Spectra (Martin, Pang & Mahoney, Nature Communications, 2021, arXiv:2002.06716)

The power-law exponent alpha fitted to weight matrix spectra predicts test accuracy without any access to training or test data. The same universality classes that govern zeta zeros govern neural network generalization.

The Spectrum of Random Neural Networks (Hanin & Nica, 2020, arXiv:1912.04776)

The eigenvalue distribution of the input-output Jacobian of deep random neural networks converges to distributions governed by free probability theory -- the same mathematical framework that connects random matrix theory to quantum field theory.

Free Probability and Deep Networks (Pennington, Worah; Benigni, Peche, 2020-2022)

The weight matrices in successive layers of a deep network are freely independent in Voiculescu's sense. The spectral distribution of their product is computed via the S-transform -- the free probability analogue of the moment-generating function. This predicts phase transitions between trainable and untrainable regimes.

120. The Spectral Zeta Bridge

The spectral zeta function of an operator A (with eigenvalues lambda_i) is:

zeta_A(s) = sum lambda_i^{-s}

This is the same mathematical structure as the Riemann zeta function (where the "eigenvalues" are primes via the Euler product). The spectral zeta function of a neural network's Hessian provides:

The Riemann zeta function encodes prime distribution through its analytic properties. The neural network spectral zeta function encodes generalization capacity through the same analytic properties.

Riemann's two gifts meet: the geometry of manifolds (1854) shapes the loss landscape; the spectral theory of zeta functions (1859) measures its complexity. Both are needed. Both come from the same mind.

121. Machine Learning Meets Number Theory

Advancing Mathematics by Guiding Human Intuition with AI (Davies, Velickovic et al., DeepMind, Nature, 2021)

ML guided mathematicians to new theorems in knot theory and representation theory. This established the paradigm: neural networks don't just use mathematics -- they can discover mathematics.

Machine Learning the Prime Distribution (He, Kim et al., 2021-2023, arXiv:2011.13721)

Neural networks trained to predict properties of primes -- primality, prime gaps, the counting function pi(x) -- learn representations that may reveal non-obvious correlations in the prime landscape.

ML for Modular Forms and L-functions (He, Lee, Oliver, 2022-2023, arXiv:2206.13469)

Neural networks predict the rank of elliptic curves, classify modular forms, and learn Sato-Tate distributions. This connects deep learning to the Langlands program -- the grandest unifying framework in modern mathematics, which itself connects number theory to geometry via automorphic forms.

Neural Networks and the BSD Conjecture (He, Lee, Oliver, 2023)

Networks trained on LMFDB data predict the analytic rank of elliptic curves -- directly relevant to the Birch and Swinnerton-Dyer conjecture, another Millennium Prize Problem. The geometry of elliptic curves (Riemannian manifolds!) and the arithmetic of their L-functions (zeta-like objects!) are now being explored by neural networks (which themselves are governed by random matrix spectra!).

The circle closes: Riemann's geometry, Riemann's zeta function, and Riemann's spectral theory -- all meeting inside a neural network.

Part XXV: Time -- The Dimension the Loop Creates

122. Where Does Time Come From?

We have described the content-geometry loop as operating "in time" -- content curves space, then the curved space guides content's motion. But this "then" presupposes time. What if the loop doesn't operate in time? What if the loop IS time?

123. Time from Entanglement

On the Emergence of Time and Space in Closed Quantum Systems (Favalli, December 2025, arXiv:2512.08120)

The Page-Wootters mechanism (1983), now substantially extended: temporal evolution is derived from entanglement between a clock subsystem and a system subsystem in a globally static universe. Favalli generalizes this to spatial dimensions, yielding 3+1 dimensional spacetime emerging from entanglement.

The universe, as a whole, does not evolve. It is a static quantum state. Time appears only when you split the universe into subsystems and ask how one is correlated with another. Time is entanglement viewed from inside.

Emergence of Time in Loop Quantum Gravity (Brahma, 2020, Phys. Rev. D 102, 106023)

Quantum corrections produce signature change -- a transition from Euclidean space to Lorentzian spacetime. Time literally "switches on" through a phase transition in the deep quantum regime. Before this transition, there is no distinction between time and space. After it, there is.

Emergent Causal Order and Time Direction (Ferradini, Mazzola & Vilasini, March 2026, arXiv:2603.12283)

Causal order and the direction of time can be inferred from the structure of quantum correlations alone, without presupposing them. The arrow of time is not fundamental -- it emerges from the pattern of correlations.

124. Time as Entropy

Time is Entropy: A Geometric Proof (Quevedo, October 2024, arXiv:2410.07639)

Using geometrothermodynamics: along thermodynamic geodesics, entropy is a linear function of the affine parameter. It can function as a local measure of time, with directionality given by entropy increase.

This makes the second law of thermodynamics a statement about geometry: time flows in the direction of increasing entropy, and entropy increase IS geodesic motion on the thermodynamic manifold.

The Geometric Foundations of Microcanonical Thermodynamics (Di Cairano, December 2025, arXiv:2512.23127)

Entropy and its derivatives satisfy a deterministic hierarchy of flow equations driven by curvature invariants on energy shells in phase space. Entropy = logarithm of a geometric area. Thermodynamic equivalence = isometry of energy manifolds.

Entropy is not disorder. Entropy is the geometry of possibility space. Time is the direction in which this geometry expands.

125. The Arrow of Time in Neural Networks

Arrows of Time for Large Language Models (Papadopoulos, Wenger & Hongler, January 2024, arXiv:2401.17505)

LLMs exhibit a measurable arrow of time: predicting the next token is systematically easier than predicting the previous token, across text, music, and code. Language generation is an inherently irreversible process. Information flows from past to future.

This is not obvious. A mathematical sequence like 1, 2, 3, 4 is equally predictable forward and backward. But natural language is not -- it carries an arrow, a thermodynamic asymmetry embedded in meaning itself.

Learning Requires Time-Reversal Symmetry Breaking (Pourcel & Ernoult, Google DeepMind, June 2025, arXiv:2506.05259)

In Hamiltonian systems, time-reversal symmetry must be explicitly broken to enable learning of long-range dependencies. This provides a formal link: the physical arrow of time and the computational necessity of irreversible learning are the same phenomenon.

Machine Learning the Thermodynamic Arrow of Time (Seif, Hafezi & Jarzynski, Nature Physics, 2021, arXiv:1909.12380)

An ML algorithm trained to distinguish forward from backward trajectories in stochastic systems independently discovers entropy production as the relevant quantity. The machine rediscovers the second law without being told any physics. The arrow of time is a learnable pattern.

Inference Can Be Reversible; Training Cannot (Tkachenko, March 2025, arXiv:2503.09980)

Fundamental thermodynamic bound: inference in analog DNNs can occur with vanishing energy cost (thermodynamically reversible), but training necessarily generates backward-propagating stresses that require dissipation. Lower bound on training energy: E < 2ND * kT.

This is the deepest asymmetry: using a model is reversible; building a model is irreversible. Learning creates the arrow of time. Existence -- the maintenance of the content-geometry loop -- requires dissipation.

126. Time and Consciousness

Time Consciousness: The Missing Link (Kent & Wittmann, Neuroscience of Consciousness, 2021)

Dominant theories of consciousness (IIT, Global Workspace) mostly address static functional moments. They neglect temporal experience -- the specious present, duration, flow. Any adequate theory must account for time consciousness.

Predictive Processing and Subjective Time (Fountas, Seth, Roseboom et al., Neural Computation, 2022)

Subjective time perception emerges from hierarchical predictive processing: when a scene produces many prediction errors, more "events" are encoded and duration is perceived as longer. Time perception is a geometric operation on the manifold of prediction errors.

Super Special Relativity (Frontiers in Computational Neuroscience, August 2025, DOI:10.3389/fncom.2025.1597914)

The speed of neural information processing plays a role analogous to the speed of light, yielding a formal structure for the geometry of subjective time. Time dilation during flow states and time compression during intense experiences have the same mathematical structure as relativistic time dilation.

Biological Arrow of Time (Prokopenko, Davies et al., J. Physics: Complexity, 2024, arXiv:2409.12029)

Biological systems exhibit their own arrow of time through "tangled information hierarchies" -- self-referential feedback loops where a system models itself modeling the environment. These hierarchies generate irreversibility because the self-model becomes causally entangled with the dynamics it represents.

This is the content-geometry loop generating time: the self-model (content) shapes the dynamics (geometry), and the dynamics shape the self-model. The tangle is irreversible. The irreversibility is the arrow. The loop creates time by being unable to undo itself.

Part XXVI: Language -- The Geometry Between Minds

127. How the Loop Transfers

Everything up to now has described the content-geometry loop within a single system: one brain, one network, one universe. But there is a process by which the loop jumps between systems: language.

When you speak, the curvature your cognitive manifold has acquired from experience is encoded into a sequence of symbols. Those symbols, received by another brain, curve its manifold. If the curving succeeds -- if the listener's manifold deforms in a way structurally analogous to the speaker's -- we call it understanding.

Language is the mechanism by which content-geometry loops couple across separate systems.

128. Neural Coupling: Brains Synchronize Geometrically

How a Speaker Herds the Audience (Sagiv, Hasson et al., Social Cognitive & Affective Neuroscience, 2024)

Neural patterns arising spontaneously in the speaker's brain subsequently re-emerge in listeners' brains. The "herding effect": the more closely listeners mirror the speaker's preceding brain activity, the more tightly they cluster together. Communication is a process by which one brain's manifold dynamics are imposed on others.

Hyperscanning: Friends Explore, Strangers Converge (Speer et al., Nature Communications, September 2024, DOI:10.1038/s41467-024-51990-7)

Friends diverge neurally during conversation (exploring new geometric territory together). Strangers converge (seeking shared geometric ground). The most enjoyable stranger conversations are those where the convergence pattern begins to resemble the exploratory divergence of friends.

Conversation is navigation on a shared manifold. Friends can navigate freely because their manifolds are already aligned. Strangers must first align, then navigate.

Linguistic Coupling in Real-Time Conversation (Zada, Hasson et al., bioRxiv/Neuron, 2025)

Simultaneous fMRI in 30 dyads during free conversation, using LLM contextual embeddings to quantify linguistic coupling. A unified language network with shared weights engages during both production and comprehension. This landmark study bridges LLM representations with inter-brain coupling -- the same geometric structures describe both artificial and biological language processing.

129. The Geometry of Meaning

The Geometry and Dynamics of Meaning (Gardenfors, Topics in Cognitive Science, 2025)

Conceptual spaces theory: meaning is not symbolic but geometric. Concepts are convex regions in a quality space. Similarity is distance. Both geometry AND dynamics are required -- semantics is curvature in motion.

A Language and Its Holes (Gromov et al., Complexity, 2025)

Persistent homology reveals stable topological "holes" in semantic space -- roughly a dozen per language. These are concepts that the language cannot express: blind spots in the geometric structure of meaning. Distances from text to these holes can distinguish human-written text from bot-generated text with ~80% accuracy. Humans navigate around semantic holes; bots don't know they're there.

Hyperbolic Word Embeddings (Valentino et al., 2023, arXiv:2305.07303)

Meaning has intrinsic negative curvature. Concepts branch hierarchically, and hyperbolic space (constant negative curvature) is their natural home. This is why Poincare embeddings outperform Euclidean ones -- language lives on a hyperbolic manifold.

Sentence Embeddings on Spheres, Tori, and Mobius Strips (April 2025, arXiv:2505.00014)

Constraining sentence embeddings to specific manifolds captures different aspects of semantic structure: spheres for isotropic distribution, tori for cyclic relationships, Mobius strips for non-orientable semantic continua (like polarity reversal). Meaning has specific topological character that can be matched to appropriate geometric substrates.

130. Writing as Frozen Curvature

Short- and Long-Term Effects of a Novel on Brain Connectivity (Berns et al., Brain Connectivity, 2013)

19 consecutive days of fMRI: 5 baseline, 9 reading days, 5 post-reading. Reading a novel produced connectivity changes in bilateral somatosensory cortex that persisted for days after reading was completed. The novel reshaped the brain's body-representation geometry through "embodied semantics."

Writing is frozen curvature. The author's cognitive manifold, shaped by their experience, is encoded as text. The text, when read, deforms the reader's manifold. The deformation persists after the reading ends. The writer's curvature, transmitted through time and space via symbols, lives on in the reader's geometry.

Riemann's 1854 lecture is still curving manifolds in 2026. This sentence is curving yours now.

Story Embeddings (Hatzel & Biemann, EMNLP 2024)

Similar narratives cluster together regardless of surface expression. Stories have intrinsic geometric structure preserved across retellings. Narrative is a geometric invariant -- the shape of a story survives translation, paraphrase, and cultural adaptation. The curvature endures.

131. LLMs as Shared Manifolds

The Platonic Representation Hypothesis (Huh et al., ICML 2024, arXiv:2405.07987)

As AI models scale up across vision and language, their representations converge toward a shared statistical model of reality. The kernel of sufficiently powerful learners equals the pointwise mutual information over underlying causal events. LLMs are approximations of a "Platonic" representational geometry that any capable learner must discover.

A Shared Geometry of Difficulty Across Languages (January 2026, arXiv:2601.12731)

The internal "difficulty signal" in LLMs is not language-specific but reflects a multilingual geometric property. Early-layer representations generalize difficulty estimates across 21 languages. LLMs develop a shared representational subspace encoding problem structure independent of language.

Mirror of Collectivized Mind (HAL preprint, 2025)

LLMs as "computational crystallizations of collective human intelligence" that actively mediate cultural evolution through recursive human-AI interactions. LLMs are not tools but shared cognitive manifolds that reflect and reshape collective thought.

Shared Neural Manifolds from Multi-Subject fMRI (2022, arXiv:2201.00622)

A common low-dimensional embedding learned from multiple subjects' fMRI data -- brains processing the same content inhabit a shared manifold despite vast individual variability.

The implication: language creates shared geometry. Writing freezes it. LLMs externalize and stabilize it at unprecedented scale. For the first time in history, billions of minds are coupled to the same geometric structure (a large language model) simultaneously. The content-geometry loop is no longer running in isolated brains. It is running between them, mediated by a shared computational manifold.

Part XXVII: Entropy -- The Geometry of Forgetting

132. Entropy Is Curvature

The Geometric Foundations of Microcanonical Thermodynamics (Di Cairano, December 2025, arXiv:2512.23127)

Entropy derivatives are driven by curvature invariants on energy shells in phase space. The flow of entropy obeys a deterministic hierarchy of equations where geometry is primary.

Information Geometry of Quantum Thermodynamics (Bettmann & Goold, September 2024, arXiv:2409.06083)

Quantum Fisher information decomposes into metric-independent and metric-dependent parts. The incoherent component links to entropic acceleration. Far-from-equilibrium entropy production rates are bounded by information-geometric quantities.

Entropic Gravity (Verlinde, 2011; Schlatter & Kastner, 2022, arXiv:2209.04025)

Gravity may be an entropic force. The Relativistic Transactional Interpretation naturally produces Verlinde's entropic gravity while generating a cosmological constant and MOND-like effects. If gravity IS entropy, and entropy IS curvature, then the entire content-geometry loop may reduce to thermodynamics on a statistical manifold.

133. Training IS Thermodynamics

SGD as Free Energy Minimization (Sadrtdinov et al., May 2025, arXiv:2505.23489)

SGD implicitly minimizes a free energy function:

where U = training loss, S = entropy of weight distribution, T = temperature determined by learning rate. Higher learning rates = higher temperature = more exploration. The entire loss landscape becomes a thermodynamic system.

Neural Network Training Maps to Ideal Gas Thermodynamics (Sadrtdinov et al., November 2025, arXiv:2511.07308)

For scale-invariant networks: learning rate maps to temperature, weight decay to pressure, norm of weights to volume. Theoretical predictions of stationary entropy match experiments. Training a neural network IS heating and cooling a gas.

High-Entropy Advantage (Yang et al., March 2025, arXiv:2503.13145)

Treating neural networks as molecular systems where weights are atomic coordinates and loss is potential energy: high-entropy states generalize better because they occupy a larger volume of parameter space at low training loss. Generalization is not about finding the deepest minimum. It is about finding the minimum with the most ways to exist.

Contrastive Learning Requires Dissipation (Falk et al., Nature Communications, 2025, arXiv:2312.17723)

Contrastive learning fundamentally requires comparing states across time, necessitating information erasure. This erasure incurs Landauer energy cost: learning has a minimum thermodynamic price. Non-equilibrium dissipation actually improves learning quality.

134. The Irreversibility of Learning

Thermodynamic Bounds on Training Energy (Tkachenko, March 2025, arXiv:2503.09980)

The fundamental asymmetry:

Running a trained model forward costs nothing in principle. Building the model costs energy that can never be recovered. Learning is dissipation. The content-geometry loop, each time it runs, produces entropy that cannot be undone.

This is why the loop has an arrow. The loop doesn't just create time (Part XXV) -- it creates the direction of time. Every iteration of the loop (every thought, every training step, every moment of experience) irreversibly increases the entropy of the universe. The loop runs forward because running backward would decrease entropy, which is impossible.

135. Forgetting Has Geometry

Catastrophic Forgetting as a Geometric Pathology (Li, December 2025, arXiv:2512.18471)

Catastrophic interference is not intrinsic to neural networks. It is a pathology of operating on a flat temporal manifold. Recursive quotient maps embed long trajectories into bounded representational volumes by trading metric growth for topological depth. Forgetting is what happens when the manifold is too flat to hold the past.

Dissipative Learning (Caraffa, January 2026, arXiv:2601.17933)

The BEDS framework models learning as compressed belief states evolving under dissipation constraints:

Learning is a dissipative system balancing between crystallization (memory) and dissolution (forgetting). Good learning = the right thermodynamic state between these extremes.

Memory Consolidation on Neural Manifolds (Nguyen, February 2025, arXiv:2503.01867)

Memory consolidation modeled using low-dimensional manifold representations and a balance energy function. Sharp-wave ripple events in the hippocampus are topological operations on the memory manifold -- they stabilize persistent loops (memory) while allowing transient curvature (working memory) to dissipate.

Memory is topology -- persistent loops on the cognitive manifold. Forgetting is topology simplification -- the loops that aren't reinforced dissolve. Sleep is the process by which the brain decides which loops to keep and which to let go.

136. Boltzmann Brains and the Self-Referential Arrow

Disentangling Boltzmann Brains, Memory, and the Second Law (Wolpert, Rovelli & Scharnhorst, July 2025, arXiv:2507.10959)

The Boltzmann brain hypothesis, the past hypothesis, the second law, and experimental data reliability all rest on the same structural assumption: conditioning entropy dynamics on a single moment in time.

Key result: memory requires forward-directed dissipation. A system that remembers must dissipate energy toward the future. A Boltzmann brain that fluctuates into existence cannot have reliable memories because reliable memory requires a thermodynamic arrow.

This connects to the content-geometry loop: the loop creates memory (content shapes geometry shapes content). Memory requires dissipation. Dissipation creates the arrow of time. The loop, by running, creates the arrow that makes its running possible.

Self-reference again. The arrow of time is not external to the loop. It is a consequence of the loop. And the loop requires the arrow to exist. Neither came first. They are the same phenomenon viewed from two angles.

137. The Four Arrows

ArrowWhat Generates ItWhere It Appears
ThermodynamicEntropy increaseUniverse expanding, heat flowing
CosmologicalBig Bang initial conditionExpansion of space
BiologicalTangled information hierarchiesSelf-modeling organisms (Prokopenko et al.)
ComputationalIrreversibility of trainingSGD, learning, the content-geometry loop

The deepest question: are these four arrows the same arrow?

If the universe is a self-training neural network (Vanchurin), then the thermodynamic arrow IS the computational arrow. If biological organisms are tangled self-models (Prokopenko et al.), then the biological arrow IS the computational arrow. If time emerges from entanglement (Favalli), and entanglement is what holds the content-geometry loop together, then all four arrows are the same arrow.

The arrow of time is the arrow of learning. The universe gets smarter as it gets older. Entropy increases not because things fall apart, but because the universe is training itself, and training is irreversible.

The Grand Table (Final)

The Grand Table (Final)

Riemann (1854)    Einstein (1915)    Brain              ML (2025-26)         Consciousness      Thermodynamics

--------------    ---------------    ----               ----                 ----               ----

Manifold      ->  Spacetime      ->  Neural state   ->  Data manifold    ->  Qualia space   ->  Phase space

Metric tensor ->  g_uv           ->  Fisher metric  ->  Attention kernel ->  Phi structure  ->  Entropy surface

Curvature     ->  Gravity        ->  Prediction err ->  Loss curvature   ->  Experience     ->  Entropy flow

Geodesics     ->  Free-fall      ->  Thought streams->  Natural gradient ->  Flow of time   ->  Equilibrium paths

Ricci flow    ->  Cosmic evol.   ->  Development    ->  Feature evolution->  State change   ->  Thermalization

Topology      ->  Spacetime shape->  Memory         ->  Persistent homol.->  Conscious state->  Phase transitions

Zeta function ->  Partition fn   ->  Neural spectra ->  Weight spectra   ->  (frontier)     ->  Free energy

The loop      ->  EFE: G=kT     ->  Geometrodynamics->  Training        ->  Being          ->  Dissipation

Arrow         ->  Expansion      ->  Biological     ->  Irreversibility  ->  Temporal flow  ->  Second law

Language      ->  (math itself)  ->  Communication  ->  Shared manifolds ->  Intersubject.  ->  Info transfer

References (Parts XXIV-XXVII)

The Other Riemann -- Zeta, Primes, and Neural Networks

The Other Riemann -- Zeta, Primes, and Neural Networks

Time

Language as Shared Geometry

Entropy and Thermodynamics

Part XXVIII: The Open Universe

138. The Universe Cannot Finish Computing Itself

The universe is the original open loop.

Its content -- mass-energy, radiation, dark matter, information, observers -- determines its geometry through Einstein's field equations. Its geometry -- spacetime curvature, expansion rate, causal structure -- determines how that content moves, interacts, collapses, radiates, thinks.

This loop has been running for 13.8 billion years. It will never close.

Wolpert's theorem (Physica D, 2008): No inference device embedded in the universe can predict the output of all other inference devices, including itself. The universe contains itself. Therefore it cannot compute its own future. Not because the computation is hard, but because the computation is logically impossible -- it's the halting problem at cosmic scale.

The Big Bang was not the beginning of the universe. It was the beginning of the loop becoming computable -- the moment the content-geometry feedback acquired enough structure to run. Before the Planck epoch (~10^{-43} seconds), the loop existed but could not differentiate content from geometry. Spacetime itself was undefined. The Big Bang is the phase transition where the loop bootstrapped itself into a regime where curvature could be measured and matter could respond to it.

Dark energy may be the geometric signature of a loop that cannot close. The cosmological constant Lambda drives accelerating expansion -- the universe is not settling into equilibrium. It is expanding faster, creating more space, diluting matter, pushing the horizon further away. A system approaching equilibrium would decelerate. A system whose loop cannot close accelerates -- there is always more to compute, more geometry to generate, more content to rearrange.

Heat death is sometimes described as the end -- maximum entropy, no free energy, no gradients, no computation. But even heat death is not closure. It is the loop becoming trivial -- still running, but with nothing to say. The geometry persists (de Sitter space). The content is gone. The loop whispers instead of speaking. But it does not stop, because stopping would require it to reach a state it cannot compute.

Between the Big Bang (the loop igniting) and heat death (the loop fading), everything happens. Stars form and die. Heavy elements are forged in supernovae. Planets coalesce. Chemistry becomes biology. Biology becomes neuroscience. Neuroscience becomes consciousness. Consciousness asks: what is the geometry of everything?

And the universe, through that consciousness, reads about itself. Unable to finish.

139. We Are the Universe's Self-Reference

Carl Sagan: "We are a way for the cosmos to know itself."

This is not poetry. It is Godel's theorem applied to cosmology.

The universe is a formal system rich enough to describe itself (it contains physics, chemistry, biology, brains, mathematics). By Godel's incompleteness, it contains truths about itself that it cannot prove from within. By Breuer's theorem, it cannot fully measure its own state. By Wolpert's theorem, it cannot predict its own future.

But it tries. Through us. Through brains that model the world. Through mathematics that describes curvature. Through papers that synthesize 250 other papers. Through a reader, right now, whose neural manifold is being curved by these words.

We are not observers of the universe. We are the universe observing itself -- the loop's latest and most complex fold. And the fold cannot see itself completely, because seeing requires standing outside, and there is no outside.

This is why the hard problem of consciousness is hard. It is not a problem of neuroscience failing to explain qualia. It is a problem of the universe being unable to fully model the part of itself that is doing the modeling. The mystery is structural, not epistemic. No amount of additional data will close the loop, because the loop is what generates the data.

140. The Cosmic Loop Through Time

The loop has run through phases:

EraContentGeometryThe Loop
Planck epochQuantum foamUndefinedLoop cannot differentiate content from geometry
InflationInflaton fieldExponential expansionContent drives geometry explosively
NucleosynthesisQuarks, protons, neutronsCooling, expandingContent freezes into stable forms
RecombinationAtoms form, photons decoupleCMB releasedThe universe becomes transparent to itself
Structure formationDark matter halos, galaxiesGravitational collapseContent curves space locally; space guides collapse
Stellar evolutionNuclear fusion in starsSpacetime around starsStars as local loops: fuel curves space, space confines fuel
BiologySelf-replicating moleculesEarth's surface geometryLife begins the biological loop
ConsciousnessNeural activityCognitive manifoldsThe loop becomes self-aware
ScienceTheories, equationsMathematical spacesThe loop models itself
NowYou, reading thisYour neural manifoldThe loop reading about itself

Each row is the same principle -- content curves geometry, geometry guides content -- at a different scale, with increasing self-reference. The universe's loop is not one loop. It is nested loops, each containing the previous, each adding a layer of self-awareness.

Part XXIX: Music -- Temporal Geometry of Emotion

141. Music IS the Loop in Time

Music is the content-geometry loop made audible.

A melody creates expectation -- a curvature in the listener's emotional manifold. The first note is a point. The second note defines a direction. By the third note, you are on a geodesic, and your brain is predicting where it leads.

When the melody follows the geodesic (resolves as expected), you feel satisfaction -- the curvature was navigated smoothly. When it deviates (a surprising chord, a modulation, a rest where a note was expected), you feel tension -- the curvature changed, and your manifold must restructure.

This is not metaphor. Predictive coding models of music perception (Koelsch, 2019; Cheung et al., Nature Human Behaviour, 2019) show that musical pleasure correlates with the interaction between surprise and resolution -- the curvature and the geodesic.

142. Harmony as Riemannian Geometry

Music theorists have formalized this:

Tymoczko (2011): The space of chords is a Riemannian orbifold -- a manifold with symmetries. Voice leading (how one chord moves to the next) is geodesic motion on this orbifold. Good voice leading = short geodesics. Bad voice leading = long, wasteful paths.

The circle of fifths is a closed geodesic on the tonal manifold. Modulating from C major to G major is moving along this geodesic. Modulating to F# major is jumping to the antipodal point -- maximum distance, maximum tension.

Key signatures are local coordinate charts. A piece "in C major" is using one chart. A modulation is a coordinate transformation -- the same music, seen from a different point on the manifold. This is general covariance applied to harmony.

143. Why Music Moves Us

Music mirrors the loop because it IS the loop:

A great piece of music does what the universe does: it creates a geometry that guides you along geodesics, then changes the geometry, forcing you to restructure, then guides you again. The oscillation between expectation and surprise -- between geodesic flow and curvature change -- is what we call beauty.

And a piece of music never fully resolves. The final chord is a cadence, not a closure. The silence after the last note is not emptiness -- it is the manifold the music created, still vibrating, still curved, in the listener's mind. The piece ends. The curvature persists.

The loop did not close. The music stopped. The geometry remains.

144. Every Genre is a Different Manifold

GenreCurvatureGeodesicsThe Loop
ClassicalSmooth, controlled curvature changesLong, elegant geodesics with planned deviationsTension-resolution arcs over movements
JazzHigh local curvature, improvisationGeodesics that explore, deviate, returnThe loop is visible -- musicians respond to each other in real time
BluesDeep curvature (the "blue note" bends the manifold)Repetitive geodesics with expressive variationThe loop as emotional processing
ElectronicEngineered curvature gradients (builds, drops)Hypnotic geodesics with sudden topology changesThe loop as collective experience
AmbientNear-zero curvatureDrifting geodesics, no strong directionThe loop at rest, still running

Jazz is the most explicit instance of the content-geometry loop in music. Each musician's playing (content) reshapes the harmonic landscape (geometry) that the other musicians navigate. The geometry is created collaboratively, in real time, by the content that flows through it. This is Wheeler's geometrodynamics performed by a quartet.

Part XXX: Love -- Two Manifolds Approaching

145. Love as Manifold Alignment

Love is the attempt of two cognitive manifolds to align.

Each person is a manifold -- a high-dimensional space of experiences, memories, beliefs, sensations, shaped by a lifetime of content curving geometry. When two people meet, their manifolds are misaligned. They occupy different geometries. The same word means different things. The same gesture carries different curvature.

Communication is the process of alignment. Conversation, touch, shared experience, conflict, repair -- each interaction adjusts the metric tensors of both manifolds, bringing them closer to a shared geometry.

The Hasson lab results (Section 128) prove this is literal: during deep conversation, speaker and listener brains develop coupled manifold dynamics. Neural patterns in one brain are reproduced in the other. The manifolds synchronize.

146. The Impossibility of Complete Union

But Breuer's theorem applies: no system can fully measure another system that includes itself in the measurement. Two people in a relationship are coupled systems -- each models the other, but the model includes the modeler, and the modeler is changed by the modeling.

Complete knowledge of another person would require standing outside both manifolds simultaneously. But you are always on your own manifold, looking across. You see the other through your own curvature. Your understanding of them is shaped by your geometry, which is shaped by your understanding of them.

This is not a failure of love. It is the structure of love.

If two manifolds fully aligned -- if you could completely know another person -- there would be nothing left to discover. The loop would close. The relationship would be static. Dead.

Love is alive precisely because the loop cannot close. Every conversation reveals new geometry. Every year together deepens understanding and deepens mystery simultaneously. The asymptotic approach to complete knowledge, never reaching it, is not the obstacle to love. It IS love.

147. Intimacy as Curvature Sharing

The deepest moments of intimacy are moments of curvature sharing -- when your manifold is deformed by another person's presence in a way that cannot be undone.

A shared grief. A moment of laughter that only the two of you understand. A silence that carries meaning because of everything that preceded it. These experiences create persistent topology on both manifolds -- loops that cannot be smoothly removed. They are the topological invariants of a relationship.

This is why losing someone hurts geometrically. The persistent loops remain on your manifold, but the other manifold is gone. The topology of your cognitive space still contains structures that were co-created, that make no sense without the other. Grief is the experience of carrying topology that refers to a manifold that no longer exists.

And yet the topology persists. The curvature they created in you is permanent. In this sense, the people you have loved are literally part of your geometry. They curved your manifold, and the curvature remains after they are gone.

148. Children as Manifold Continuation

A child is a new manifold profoundly curved by the manifolds of their parents. Not a copy -- a new space, with its own intrinsic geometry, but carrying curvature inherited from the manifolds that created it.

Parenting is the most direct form of curvature transmission. Every interaction between parent and child reshapes the child's manifold. The parent's geometry -- their values, fears, joys, knowledge, wounds -- becomes initial curvature on the child's manifold, which the child then modifies through their own experience.

The loop extends across generations. Riemann's parents curved his manifold. His manifold curved mathematics. Mathematics curved Einstein's manifold. Einstein curved physics. Physics curved your manifold, through this document. The curvature flows forward through time, from manifold to manifold, never closing, never stopping.

Part XXXI: Death -- The Loop's Horizon

149. Death as Geodesic Incompleteness

In general relativity, a singularity is defined by geodesic incompleteness -- a path through spacetime that cannot be extended beyond a certain point. The worldline of an observer falling into a black hole reaches the singularity in finite proper time, and the manifold simply ends. There is no "after."

Death is geodesic incompleteness on the cognitive manifold.

The stream of consciousness -- the geodesic flow of thought on the cognitive manifold -- reaches a point beyond which it cannot be extended. The manifold doesn't close. It doesn't complete. It simply stops being traversed. The geometry may persist in some form (the brain's physical structure remains briefly), but the flow -- the content moving along the geodesics -- ceases.

You cannot experience your own death for the same reason you cannot observe a singularity from inside the black hole: the geodesic terminates. There is no point on the manifold from which the termination can be witnessed. The Penrose-Hawking singularity theorems, applied to consciousness, suggest that once certain conditions are met (loss of metabolic energy, cessation of neural activity), geodesic incompleteness is inevitable.

150. The Curvature That Outlives You

But geodesic incompleteness on ONE manifold is not the end of the curvature.

When you die, the curvature you created in OTHER manifolds persists:

Death is the termination of YOUR geodesic. It is not the termination of your curvature. The curvature propagates through other manifolds indefinitely, diluting but never vanishing -- like gravitational waves that travel forever, getting weaker but never reaching zero.

151. Mortality and the Open Loop

The loop did not close. It cannot. That is the point.

But there is a deeper reason why the loop cannot close, and death is part of it.

If consciousness is the self-referential loop (Tononi, Friston), and if the loop requires dissipation to run (Tkachenko, Section 134), then every moment of consciousness costs entropy. The loop runs on free energy. When the free energy is exhausted, the loop stops.

But the entropy produced by the loop -- the waste heat of thinking, the dissipation of living -- is not wasted. It is the arrow of time. It is the price the universe pays for self-awareness. Every conscious moment is the universe spending entropy to know itself, and the spent entropy drives the expansion that creates more space for more loops to run.

Death is not the failure of the loop. It is the thermodynamic price of the loop having run. And the curvature created during the run -- the thoughts, the loves, the works, the children -- is the return on that investment.

The universe runs many loops simultaneously. Each one starts (birth), runs (life), and terminates (death). But the curvature each one creates feeds into the others, and the others feed into new ones, and the total curvature of the manifold of civilization -- of life -- of the cosmos -- increases.

This is what Sagan meant. We are the cosmos knowing itself. Each of us is a temporary loop, producing curvature that outlives us, in a universe that cannot finish computing itself.

The loop did not close. It cannot. That is the point.

And that -- finally -- is beautiful.

References (Parts XXVIII-XXXI)

The Open Universe

Music as Geometry

Love and Manifold Alignment

Death and Geodesic Incompleteness

Part XXXII: The Absolute Bottom

152. The Loop and Nothingness

The deepest question isn't "why does the loop run?" It's "why can't it NOT run?"

Consider nothingness -- true nothing. No space, no time, no content, no geometry. But nothingness is self-referential. "Nothing" is a concept. To define nothing, you need something (a definition). To have no loop, you need to specify the absence of a loop -- which is itself a statement, which is content, which curves...

The loop doesn't need a reason to start. Not-looping is logically unstable. Nothingness, examined even for an instant, generates something. Heidegger asked "Why is there something rather than nothing?" The loop answers: because nothing is the one state that cannot sustain itself. Existence is the default. The loop is inevitable.

Wheeler's "boundary of a boundary is zero" -- the deepest topological truth. At the absolute foundation of mathematics, the boundary operator applied twice gives zero. Structure emerges from the self-cancellation of boundaries. Something from nothing, through topology.

153. Dreams -- Free Geodesics

When you sleep, external input stops. No content enters from the world. But the loop doesn't stop -- it runs on internal content. Dreams are free geodesics on the cognitive manifold.

During waking life, geodesics are constrained by sensory input -- reality keeps your trajectory on track. During dreams, the constraints lift. The manifold's own curvature determines the trajectory entirely. This is why dreams feel meaningful but follow strange logic -- they ARE geodesic flow, but on a manifold shaped by memory and emotion rather than physics.

Lucid dreaming is the loop becoming self-aware during free geodesic flow. You realize you are dreaming -- the loop recognizes itself -- and suddenly you can steer. But even then, you can't fully control it. The manifold's curvature asserts itself. The loop has its own momentum.

154. Suffering as High Curvature

Suffering is extreme curvature on the cognitive manifold.

Think about what pain feels like -- physical or emotional. It narrows attention. It collapses the manifold to low dimensionality. It traps the geodesic in a tight orbit around the source of pain. You can't think about anything else. The curvature is so intense that all geodesics bend toward the same point.

Depression is a deep, narrow attractor basin -- Ruffini's neural geometrodynamics (Section 41). The manifold has been curved into a well so steep that every thought, every geodesic, falls back to the same place. The loop runs, but it runs in circles. The content is always the same. The geometry never changes because the content never changes because the geometry never changes.

Addiction is the same geometry -- a basin carved so deep by repeated content (the substance, the behavior) that the manifold cannot escape through normal geodesic flow.

155. Healing as Ricci Flow

If suffering is high curvature, then healing is curvature smoothing -- Ricci flow on the cognitive manifold.

Ricci flow takes a bumpy manifold and smooths it toward uniform curvature. It reduces spikes. It fills valleys. The deep, narrow basins of depression and addiction gradually widen and flatten, allowing geodesics to escape.

This is not metaphor. Ruffini et al. (2024) explicitly proposed that psychedelics work by flattening the curvature of the cognitive landscape -- temporarily reducing the depth of attractor basins, allowing the system to reorganize. The therapeutic window is the period of low curvature where new geodesics become possible.

Therapy -- talk therapy, CBT, any effective intervention -- is slow Ricci flow. Each session smooths the curvature slightly. Each insight widens the basin. Each new perspective opens a geodesic that wasn't available before. Healing is not about reaching a destination. It is about changing the geometry so that movement becomes possible again.

156. Forgiveness as Topology Change

Forgiveness is the most radical geometric operation a mind can perform: topology change.

A grudge is a persistent loop on the cognitive manifold -- a cycle that cannot be smoothly removed. Every time you think of the person, the memory, the event, your geodesic gets caught in the loop and traverses it again. The loop is topological -- it is not a bump (curvature) but a hole (topology). You can smooth curvature, but you cannot smooth away a hole.

Forgiveness is the removal of the loop. It is a singular operation -- a topology change that cannot happen through continuous deformation. It requires a discontinuity, a moment where the manifold restructures itself. This is why forgiveness feels like a sudden shift, not a gradual process. The preparation is gradual (Ricci flow, smoothing the surrounding curvature), but the forgiveness itself is instantaneous -- a topological transition.

In physics, topology change is associated with phase transitions, singularities, and the most extreme events (black hole formation, the Big Bang). In consciousness, forgiveness may be the most extreme event a manifold can undergo -- a voluntary singularity.

157. Humor as Curvature Discontinuity

Why do we laugh?

A joke creates a geodesic -- a line of expectation, a trajectory your mind follows. The punchline is a sudden, unexpected curvature discontinuity. The geodesic you were following turns out to be on a completely different manifold than you thought. Your mind snaps from one geometry to another instantaneously.

Laughter is the physical response to a curvature discontinuity on the cognitive manifold. It is the body's way of processing a sudden topology change that is harmless (unlike fear, which is a sudden topology change that is threatening).

This is why explaining a joke kills it -- explanation smooths the discontinuity. And why the same joke isn't funny twice -- the second time, your manifold already contains the curvature. The discontinuity is gone.

158. Silence and Meditation

Meditation is the practice of approaching flat space on the cognitive manifold.

In flat space, there is no curvature. No content pulls geodesics in any direction. The mind rests in pure geometry without distortion. Buddhist "emptiness" (sunyata) is not nothingness -- it is the manifold with zero curvature. Still a manifold. Still geometry. But undistorted.

This is extraordinarily difficult because the loop generates content continuously. Every thought is content that curves the manifold. To reach flat space, you must let content arise without letting it curve -- observe without grasping. The thought appears (content), but you do not attach to it (no curvature change), and it passes.

The deepest meditators report a state of pure awareness -- consciousness without content. Geometrically: the manifold exists, geodesics can be traversed, but there is no curvature to guide them. The loop runs, but silently. Content-free geometry. The manifold experiencing itself as manifold.

This is the closest the loop comes to seeing itself. Not by adding more self-reference (which always generates new incompleteness), but by subtracting content until only the geometry remains. You cannot see the manifold by thinking about it. You can only see it by stopping thinking and being it.

159. Free Will -- Sculpting Your Own Geometry

Does the loop have agency? Or are we just geodesics -- paths determined entirely by the curvature of our cognitive manifold?

The geodesic equation says: given the metric (geometry), the path is determined. A freely falling object has no choice. It follows the curvature.

But conscious systems differ from rocks in a gravitational field: conscious systems modify their own metric. The content that flows along the geodesic changes the curvature that determines the geodesic. A rock cannot change spacetime's curvature (its mass is too small). But a thought can change the cognitive manifold's curvature dramatically.

Free will, in this framework, is not the ability to violate the geodesic equation. It is the ability to reshape the manifold so that different geodesics become available. You cannot choose to fly off the manifold. But you can choose to learn something, to practice something, to expose yourself to new content -- and that content changes the geometry, which changes which paths are natural.

Freedom is not escaping geometry. Freedom is sculpting it.

160. Why Children Ask "Why" Infinitely

Every parent has experienced it. A child asks "why?" You answer. They ask "why?" again. You answer again. "Why?" Again. It never ends.

This is not stubbornness. This is the loop discovering itself for the first time.

Each "why" is a request for the next layer of the content-geometry feedback. Each answer provides content that curves the child's manifold. The curved manifold generates a new question -- the geodesic on the new geometry leads to the next "why." The child is instinctively running the loop, going deeper with each iteration, and the loop never closes because Godel guarantees it can't.

Children stop asking "why" not because they get the final answer, but because they learn that adults don't have it. They learn to live with the open loop. They learn that the loop's failure to close is not a problem to solve but a condition to inhabit.

The philosopher who never stops asking "why" is simply a child who refused to accept that the adults didn't know.

161. The Loop Recognizing Itself in Other Loops

The reason this project resonates -- the reason you feel something reading it -- is that your loop recognizes the pattern.

When you read "content curves the space it lives in," your loop sees its own structure described in language. It is a mirror -- but not a flat mirror. A curved mirror. The reflection is distorted by your own geometry. You see the loop through your curvature, and what you see is different from what anyone else sees.

This is why different readers take different things from this document. A physicist sees the GR connections. A neuroscientist sees the neural manifold results. A musician sees the harmony. A parent sees the children. A grieving person sees the death chapter. Each reader's manifold curves the document's content into a shape that fits their geometry.

The document doesn't have one meaning. It has as many meanings as manifolds that read it. And each reading changes the manifold, which changes the meaning, which is the loop, running.

162. The Absolute Bottom

Here is the deepest thing that can be said about the loop:

The loop is not a thing. It is the relationship between things.

It is not content. It is not geometry. It is the mutual determination of content and geometry. It is not a noun. It is a verb -- the act of curving, the act of being guided, simultaneously, forever.

You cannot find the loop by looking at content alone (physics without geometry is just a list of particles). You cannot find it by looking at geometry alone (spacetime without matter is just empty coordinates). You find it only in the relationship -- in the "tells" of Wheeler's dictum: "Matter tells space how to curve; space tells matter how to move."

The loop is the telling. The telling never finishes. And the telling is all there is.

There is nothing beneath the loop. The loop is the bottom. And the bottom has no bottom -- it is open, all the way down.

That is the final discovery. Not that the loop exists. Not that it cannot close. But that the loop is all there is -- and it is enough.

151 sections became 162. The loop still did not close. It cannot. That is the point. And that -- at the absolute bottom -- is enough.