The Fractal Gas

Volume II: Interacting Particles, Convergence, and Emergent Geometry

by Guillem Duran-Ballester and Sergio Hernández Cerezo

A Fractal Gas distributes a search across a population of interacting walkers. Each walker moves, measures its surroundings, and compares itself with selected companions. Cloning reallocates walkers toward higher-fitness states; noise and motion let them explore further. The mathematical problem is to understand the population produced by these repeated operations.

This volume begins with the update rule and develops its probability theory: operator estimates, finite-particle convergence, mean-field limits, and entropy bounds. It then studies the geometric record of a run, called the Fractal Set, and the field models constructed from that record. Computational chapters explain how to measure the resulting observables and compare simulations.

Keep three questions separate as you read. What happens when a fixed swarm runs for a long time? What happens when its population increases? What happens when a discrete geometric construction is refined toward a continuum? Each question has its own hypotheses and its own notion of convergence.

The algorithm and its variants

Imagine watching a cloud of walkers on a landscape. Motion changes their positions. Cloning changes where the population spends its computational effort: a walker can be replaced by a perturbed copy of a companion. The population therefore changes through both transport and selection.

Reward and diversity enter the fitness calculation. Reward describes how well a walker performs the task; diversity measures separation according to the chosen comparison rule. Their normalization and relative weights matter. A diversity term can oppose concentration, but its presence alone does not establish a lower bound on the population’s spread.

The common sequence is:

1. Select companions. Sample other walkers according to the specified interaction rule.

2. Measure and compute fitness. Combine regularized reward and diversity measurements.

3. Clone and revive. Apply the replacement rule, including its treatment of walkers that have left the alive region.

4. Advance the kinetic dynamics. Apply motion, friction, noise, and any interaction forces included in the chosen model.

The order of these operations, the noise law, and the boundary rule define the transition kernel. They are part of the mathematical model.

Three words get used for three different things, and it is worth separating them before they start colliding. The Fractal Gas is the family: every algorithm in this volume is one choice of components for a single step operator. The Fragile Gas and its instantiated Fragile Swarm are the abstract Markov chain and axioms of the framework chapter — a set of assumptions you verify for a model, not a model itself. And the Algorithmic Gas is the Rust engine of the architecture chapter; it executes a variant, and it is never a variant.

A variant is what you get when all twelve components are fixed: the Euclidean Gas, the Viscous Euclidean Gas, the Einstein–Hilbert Gas, the Geometric Gas, the Latent Fractal Gas, and the Environment Gas. Mean-field and continuum limits are limits of a variant, never variants of their own. Variants of the Fractal Gas gives each one as a component tuple and states, variant by variant, which theorems of this volume have been proved for it and which have not.

The algorithm introduction gives the operational description, and the variants chapter fixes the named variants and the vocabulary just described. The general framework specifies the operators and their assumptions. The Euclidean model provides the main setting for the convergence analysis, and the single-particle chapter isolates the underlying kinetic behavior. The Latent Fractal Gas introduces the geometric choices needed when positions and velocities live on a manifold.

Note

An estimate for one variant applies to another when the latter satisfies the estimate’s hypotheses. For example, momentum conservation during cloning depends on the specified velocity update, and revival depends on the replacement rule and the presence of a surviving companion. These properties must be checked for the actual transition kernel.

What convergence means here

A swarm that settles into a stable statistical pattern can still lose every walker through an extinction event. To describe the runs that remain alive, we condition on survival. This is why quasi-stationary distributions appear in the analysis. A model with revival or a prescribed source can instead have a stationary balance between loss and replacement. Its normalized alive profile and its alive fraction are determined together; that source balance is distinct from conditioning a killed process on survival.

There is another distinction to keep in view. The distribution of the entire swarm contains correlations between walkers. The distribution of one selected walker contains less information. A smooth density used in a mean-field equation is a third object, obtained only after justifying a limit. Confusing these objects can turn an estimate about one particle into an unsupported claim about the whole population.

Definition 403 (Finite-particle quasi-stationarity)

Let \(S_k\) be the state of the \(N\)-walker swarm after \(k\) updates, let \(\tau_{\dagger}\) be its extinction time, and let \(\nu_N\) be a probability measure on surviving swarm configurations with \(\mathbb{P}_{\nu_N}(\tau_{\dagger}>k)>0\) for every integer \(k\geq 0\). The measure \(\nu_N\) is a quasi-stationary distribution (QSD) if, when \(S_0\) has law \(\nu_N\),

\[ \mathbb{P}_{\nu_N}\!\left(S_k\in A\mid \tau_{\dagger}>k\right) =\nu_N(A) \]

for every measurable set \(A\) and every integer \(k\geq 0\).

Remark 85 (Proved results and their regimes)

The volume develops the following complete analytical arguments.

1. Composition and finite-particle relaxation. The comparison-matrix theorem Theorem 204 constructs positive Lyapunov weights when the nonnegative full-step comparison matrix has spectral radius below one. The conservative Harris argument and the killed-block QSD argument are proved separately in Theorem 206 and Theorem 209. Their full-kernel hypotheses give existence, uniqueness, and geometric relaxation to the appropriate invariant law or QSD.

2. Positive mean-field evolution and a stationary limit. For the specified mass-conserving gain-loss model, bounded loss rates and a globally Lipschitz reaction give a unique global positive mild solution in thm-chaos-mild-wellposedness. If the kinetic semigroup contracts zero-mass differences by \(Ke^{-at}\) and the reaction Lipschitz constant is \(L_{\mathcal R}\), the regime \(a-KL_{\mathcal R}>0\) gives a unique stationary density and global attraction at rate \(a-KL_{\mathcal R}\) in thm-uniqueness-uniqueness-stationary-solution. The finite-time consistency, tightness, and stationary-chaos proofs then identify particle limits under their stated concentration or global-attraction hypotheses.

3. Population-uniform logarithmic Sobolev inequalities. Corollary 94 proves four routes for specified continuous joint laws: tensorized kinetic references, uniformly bounded whole-law density tilts, uniform joint curvature, and contractive additive-noise invariant flows. These control the full position–velocity gradient. A law with discrete alive/dead strata also requires its status entropy form. The criteria apply to an actual QSD when that law meets the stated structural condition.

4. Hypocoercive entropy convergence. The complete kinetic commutator and modified-Fisher calculation proves Theorem 285, including nonconvex product references with the required LSI and bounded Hessian. The exact cloning, killing-normalization, and boundary identities feed Theorem 287: an actual-law LSI constant \(C_N\) and full dissipation bound \(\dot\Phi_G\leq-\delta_N I\) give the rate \(\delta_N/(C_N/2+g_{+,N})\). Uniform input constants give a uniform rate.

5. All-orders regularity of the actual companion laws. Normalized derivative calculus controls sampled fitness, fitness formed from expected measurements, and expected sampled fitness as separate objects. For bounded pair distances, positive scales and regularizers, and uniform analytic reward bounds, Theorem 272 proves block-derivative bounds \(C B^n n!\) independent of population size. The actual sequential greedy pairing law has its own full-history proof in Lemma 223; it is not replaced by an idealized matching law. Smooth inputs give \(C^\infty\) on the specified strata without automatically giving analytic bounds.

6. Observable errors and Gaussian boundary limits. A uniform total entropy \(D_{\mathrm{KL}}(\pi_N\Vert\rho^{\otimes N})\) gives \(O(N^{-1})\) mean-squared error for bounded empirical observables in Theorem 245; full-gradient LSI gives corresponding Lipschitz variance bounds. The Gaussian perimeter and sampling proofs Theorem 448 and Theorem 449 recover \(N^{(d-1)/d}\) raw cut scaling at range \(\varepsilon\propto N^{-1/d}\) for independent samples in dimension \(d>1\), under their geometric hypotheses. Corollary 141 transfers the area limit to dependent laws when total relative entropy is \(o(\log N)\). Identifying a genealogical count with that boundary uses the separate boundary-cell correspondence.

These results connect because each supplies a concrete input to another proof. For example, a uniform joint entropy bound controls empirical observables and also transfers the Gaussian cut limit to a dependent population. The stationary mean-field theorem identifies the limiting density when its contraction and consistency conditions hold. The conditions specify where the results apply; they are part of the mathematical conclusions, not a substitute for their proofs.

A physical interpretation adds an identification of variables and observables. The geometric chapters prove their area, curvature, and variation formulas, and state separately the additional conditions needed to interpret a cut as quantum entropy or a constructed field as the continuum limit of a sampled metric.

The structure of the volume

The chapter order follows the objects being studied. First define the swarm, then estimate its motion, then consider large populations. The geometric and field chapters come after these analytic foundations so that their additional assumptions can be stated precisely. The computational chapters provide a place to examine what a finite simulation actually measures.

Part I: Algorithms and foundations

Part I contains the algorithm intuition, the catalogue of variants, the general framework, the Euclidean Gas, the single-particle analysis, and the Latent Fractal Gas. Read it to identify the state space, the update order, and the assumptions attached to each variant. The variants chapter is the place where the Euclidean, Viscous Euclidean, Einstein–Hilbert, Geometric, Latent and Environment gases are written as component tuples and their established results are listed.

Part II: Finite-particle convergence

Part II develops the cloning estimates, Wasserstein control, kinetic estimates, and composed convergence analysis. The central task is to combine component inequalities into a statement about the full swarm transition.

Part III: Mean-field limits and quasi-stationarity

Part III develops the mean-field equation, propagation-of-chaos analysis, discrete QSD structure, and exchangeability theory. It proves positive mean-field well-posedness, stationary existence and attraction in the stated contractive regime, and the consistency and concentration arguments for stationary particle limits. The equilibrium-profile chapter also turns the residual of a proposed stationary density into a quantitative error bound.

Part IV: Entropy, regularity, and quantitative bounds

Part IV studies hypocoercive entropy estimates, Hellinger–Kantorovich convergence, third-order regularity, smooth regularity, and LSI-based entropy estimates. It also contains the Geometric Gas, finite-population error bounds, and parameter constraints. The complete LSI and kinetic hypocoercivity proofs supply explicit sufficient regimes. The regularity chapters preserve probability derivatives for each actual companion law, and the error chapter separates population bias, fluctuations, time-step error, and transient mixing. Their constants retain the dependence on population size, dimension, and regularization scales.

Part V: The Fractal Set and continuum geometry

Part V starts with the Fractal Set: the record of walker histories, cloning events, and interactions. It develops emergent geometry and scutoid constructions, then examines continuum conditions and causal-set constructions. The existence of a discrete record and convergence to a smooth geometric model are separate results.

Part VI: Fields and emergent physics

Part VI develops the proposed field descriptions: lattice QFT, Standard Model correspondences, Yang–Mills and Noether analysis, and the twistor formulation. The geometric developments continue through curvature, field equations, graph cuts and boundary thermodynamics, and cosmology. Read each physical correspondence together with its hypotheses and its stated mathematical status.

Part VII: Computation and experiments

Part VII connects definitions to measurements through computational proxies, Voronoi Wilson loops, and empirical studies. The calibration notebook, calibration report, and calibration guide present the computational workflow and its interpretation. Calibration choices belong to the specification of a numerical comparison.

Reference material

The reference section contains the FAQ and literature guide. The analytic arguments appear in the chapters where their results are developed.

How the arguments fit together

The first estimates concern individual operations. Cloning changes positions and correlations; kinetic motion changes velocities and transports the population. An estimate for one operation can contain a term that the other operation must control. This is the reason for studying their composition explicitly.

Beyond that composition, the argument branches. The positive mean-field proof controls the nonlinear evolution. Joint-law entropy and concentration estimates control fluctuations and stationary limits. The regularity and perimeter proofs then supply specific geometric limits. The diagram shows how those completed arguments enter the subsequent constructions.

        flowchart TD
    A["I: Define the transition kernel"] --> B["II: Cloning and kinetic estimates"]
    B --> C["II: Composed drift and full-kernel QSD criteria"]
    C --> D["III: Positive mean-field flow and stationary chaos"]
    A --> E["IV: Joint-law LSI, regularity, and entropy proofs"]
    E --> D
    A --> F["V: Construct the Fractal Set"]
    D --> G["V–VI: Sampling and continuum limits"]
    E --> G
    F --> G
    G --> H["VI: Fields with stated observable identifications"]
    F --> I["VII: Discrete observables and experiments"]
    H --> I
    

The arrows connect proof inputs; the hypotheses of individual theorems give the precise implications. In particular:

• A Lyapunov drift estimate supplies moment control; a convergence theorem also needs the appropriate accessibility, mixing, and survival estimates.

• Uniform moment bounds support compactness as \(N\) grows. Identifying a deterministic mean-field limit additionally requires control of correlations and passage through the nonlinear interaction terms.

• A rate for a fixed observable, a marginal law, and an empirical measure in Wasserstein distance are different estimates. Population-size rates must be quoted with their metric, moment assumptions, and dimension dependence.

• A continuum identification requires the complete set of geometric and scaling hypotheses used in its proof. The preceding probabilistic estimates establish only the inputs that their statements explicitly cover.

Reading guide

For a first pass, read the algorithm introduction, the variants chapter, and the Euclidean Gas before following the convergence chapters. Keep one question beside each theorem: which part of the update rule does this hypothesis control? That question makes the constants easier to interpret.

The probabilistic development starts from the definitions in this volume. The Latent Fractal Gas connects to the learned representations of Volume I; readers working in Euclidean space can begin directly with the Euclidean Gas.

Your aim	Suggested route	What to track
Implement a swarm	Part I, then parameter constraints in Part IV	Update order, boundary rules, regularization, and noise
Follow the finite-particle proofs	Foundations in Part I, then Part II	The full transition kernel and the hypotheses used in composition
Understand the mean-field model	Parts II–III, then finite-population bounds in Part IV	Marginals, correlations, uniformity in time, and the order of limits
Study geometric observables	The Fractal Set in Part V, then Part VII	Which quantities are defined on a finite record
Study continuum and field claims	Relevant estimates in Parts III–IV, then Parts V–VI	Geometric assumptions, scaling regimes, and conditional statements
Assess numerical evidence	Part VII and the reference material	Estimators, sampling dependence, uncertainty, and calibration choices

Historical context and empirical interpretation

The historical starting points are General Algorithmic Search (2017) [Hernández et al., 2017] and Fractal AI: A fragile theory of intelligence (2018) [Hernández Cerezo and Durán Ballester, 2018]. The literature guide collects these works and subsequent studies. This volume develops explicit particle models and their mathematical analysis; the transition rule and hypotheses of each chapter determine which results apply to an implementation.

A successful run gives information about the configuration that was run. To compare it with a theorem, identify the same observable on both sides: a population variance, a conditional distribution, a correlation function, or a geometric quantity. Then check the sampling procedure and the parameter regime.

The same care applies to calibration. If a measured quantity fixes the unit of mass or length, that choice is an input to the comparison. Predictions concern the other quantities computed after that input is fixed. The computational chapters provide the definitions and records needed to inspect these comparisons.

References

[HDA17]

Sergio Hernández, Guillem Durán, and José M. Amigó. General algorithmic search. 2017. Also published in: Recent Trends in Chaotic, Nonlinear and Complex Dynamics, World Scientific, 2017. arXiv:1705.08691.

[HCDB18]

Sergio Hernández Cerezo and Guillem Durán Ballester. Fractal AI: a fragile theory of intelligence. 2018. arXiv:1803.05049.