Kinetic limits and fluctuations over the Boltzmann lifespan

Source and verification

This article explains family 364 of the October 6, 2026 OpenAI collection. Research conclusions are attributed to its manuscripts. The review covers statements, introductions, and selected proof mechanisms; it is not an independent verification of the complete proofs.

Overview

Family 364 joins two questions: when deterministic particle dynamics converge to the Boltzmann equation, and what random fluctuations remain around that limit. Both manuscripts work throughout an assumed regular kinetic interval. Neither supplies global regularity of the limiting equation.

🏷️ Preliminaries: the dilute-gas scaling

In three dimensions, let the interaction range be $\epsilon$ and the particle activity be ${\mu }_{\epsilon }={\epsilon }^{-2}$. A collision cross section is of order ${\epsilon }^2$, so the product remains of order one. The occupied-volume parameter is instead of order ${\mu }_{\epsilon }{\epsilon }^3=\epsilon$, which tends to zero.

The initial ensemble is grand canonical: particle number is random, and a product density is conditioned by exclusion of pairs closer than $\epsilon$. After initialization, dynamics are deterministic. The uncertainty in observations comes from the initial ensemble, not from adding random collisions.

The scaled factorial marginal $F_s^{\epsilon }$ describes $s$ ordered distinct particles. Propagation of chaos means that for fixed $s$ it approaches $f^{\otimes s}$, even though the full particle system develops correlations.

🏷️ The two manuscript claims

Regular-lifespan kinetic limit

For the manuscript’s stable finite-range radial potentials in three dimensions, fixed factorial marginals converge to $f(t{)}^{\otimes s}$ in full phase-space $L^1$, uniformly on every fixed interval $[0,T]$ on which the Boltzmann solution is classical and has a uniform Gaussian velocity bound. Initial density and its spatial gradient satisfy spatially summable Gaussian bounds. The potential class permits attractive wells and a repulsive singular core. (OpenAI, 2026)

Stability means that the total pair potential energy of any $N$-particle configuration is bounded below by $-CN$. It does not mean that the potential is everywhere repulsive. Long scattering delays and simultaneous interactions therefore have to be addressed.

For hard spheres, define an empirical observable and its fluctuation by

$$\begin{array}{rl}{\pi }_t^{\epsilon }(\varphi )& ={\mu }_{\epsilon }^{-1}\sum _{i=1}^N\varphi (z_i(t)),\\ {\zeta }_t^{\epsilon }(\varphi )& =\sqrt{{\mu }_{\epsilon }}({\pi }_t^{\epsilon }(\varphi )-\mathbb{E}{\pi }_t^{\epsilon }(\varphi )).\end{array}$$

The companion fluctuation claim

Under its smooth Gaussian-decay hypotheses and regular-lifespan assumption, every finite collection of these observables at fixed times converges jointly to a centered Gaussian field governed by the linear fluctuating Boltzmann equation. (OpenAI, 2026a)

This is finite-dimensional convergence, not a path-space tightness theorem. The fluctuation claim concerns hard spheres; the more general potential theorem establishes the kinetic limit.

🏷️ Why exact centering matters

An error tending to zero in the law of large numbers need not disappear after multiplication by $\sqrt{{\mu }_{\epsilon }}={\epsilon }^{-1}$. An order-$\epsilon$ correction to the mean becomes an order-one shift at the fluctuation scale.

The companion computes precisely such an excluded-volume correction at time zero. Thus replacing the exact microscopic expectation by $\int f_t\varphi$ would generally change the limiting mean. This is a useful lesson whenever a central limit theorem is built on a deterministic limit: the centering requires a sharper estimate than convergence alone.

Random particle number also changes the initial covariance. The stated limit is

$$\mathbb{E}[{\zeta }_0(\varphi ){\zeta }_0(\psi )]=\int f_0\varphi \psi .$$

There is no fixed-sample-size subtraction of $(\int f_0\varphi )(\int f_0\psi )$. For a Poisson sum, the fluctuation in the number of summands contributes exactly the term that would otherwise be subtracted.

🏷️ Retaining correlations across time layers

One cannot repeatedly restart a short-time chaos theorem as though the particles were independent at each restart. Earlier collisions have already created correlations. The manuscripts retain their histories while dividing the macroscopic interval into layers.

For potentials, entire interacting components are kept intact so that potential energy is not lost from the estimates. Stability then gives one Gaussian energy budget across the full history. With $L$ layers, the source organizes the history bounds around a small factor of the form $CT/\sqrt{L}$. Centering removes independent contributions; a causal exposure argument estimates extra contacts using fresh scattering parameters. Both outputs of an as-yet-unexposed genuine collision must remain hidden for the independence argument to work.

Fluctuations need finer information than ordinary chaos. Connected cumulants of $h$ histories must be controlled at the scale ${\mu }^{1-h}$. At central-limit normalization, pair connections survive while higher cumulants vanish. The surviving collision trees determine the noise covariance through the change

$$\begin{array}{c}\Delta \varphi \\ =\varphi (v^{\prime })+\varphi (v_{\ast }^{\prime })-\varphi (v)-\varphi (v_{\ast }).\end{array}$$

Collision invariants have $\Delta \varphi =0$, which explains why the collision noise respects their conservation laws. Establishing the sharp cumulant bounds, including summability over long histories, is the central new estimate.

🏷️ Key steps and proof outlines

Center independent roots without breaking interacting components

The potential problem cannot restart each time layer from a product distribution: correlations created before the cut remain present. The source uses an exact inclusion–exclusion transform. For a contact pattern $\pi$, let $S$ be its singleton labels and $B$ all labels in nontrivial components. Write $f_D={\prod }_{i\in D}f(z_i)$ and ${\tilde{F}}_D^{\epsilon }$ for the normalized microscopic factorial marginal. Define

$$E_{\pi }^{\epsilon }={\chi }_{\pi }^{\epsilon }\sum _{A\subset S}(-1{)}^{|S\setminus A|}{f}_{S\setminus A}{\tilde{F}}_{B\cup A}^{\epsilon }.$$

The full-pattern indicator ${\chi }_{\pi }^{\epsilon }$ remains in every summand. Deleting a singleton from a marginal therefore does not erase its separation conditions from the other displayed particles. Nonsingleton components are never dismantled by this transform.

To invert, multiply the smaller centered terms by the missing independent factors and sum. The coefficient of a proper subset $D⊊S$ is

$$\sum _{D\subset A\subset S}(-1{)}^{|A\setminus D|}=(1-1{)}^{|S\setminus D|}=0.$$

Only the full marginal survives. Because exactly one contact pattern occurs at a given array, summing patterns does not introduce a Bell-number cost. The norm loss is at most $2^{|R|}$ from subset choices. This exact cancellation is the mechanism for retaining dependence across layers while still extracting small centered errors.

Collision trees carry their own factorial cancellation

When a new particle is inserted at a collision, its relative position has the form

$$\epsilon \omega +(c-s)(w-v).$$

Differentiation in collision time $s$ and the two angular directions gives the spatial Jacobian ${\epsilon }^2((w-v)\cdot \omega {)}_{+}$. Under the three-dimensional Boltzmann–Grad normalization $\mu ={\epsilon }^{-2}$, the activity cancels the factor ${\epsilon }^2$. The remaining measure is the familiar collision flux times ordered time and velocity measures.

For $n$ births the ordered time simplex supplies $1/n!$. Increasing labelings of a forest contribute

$$\frac{n!}{{\prod }_{v}s(v)},$$

where $s(v)$ is the number of internal vertices below $v$, including $v$. This formula follows by recursively distributing later labels among child subtrees. Consequently, many vertices with large descendant subtrees incur a product of large denominators. After excluding those rare histories, many defects force many disjoint small terminal trees. Fresh translation or angular coordinates on those trees then provide repeated small geometric factors. This is why a single recollision estimate is insufficient but an extensive defect count can beat the combinatorial cost. The source carefully performs ordinary parent counting before changing to the global exposure coordinates. (OpenAI, 2026b s “Centering at the layer cuts” and “Terminal trees and exposure of first defects”)

Records make multiple observation times part of one hierarchy

For hard-sphere fluctuations, attach a scalar record $y$ to each particle. At observation time $t_j$, update it by $y\mapsto y+{\varphi }_j(x,v)$; collisions leave each particle’s record attached to that particle. Thus the final sum of records equals the chosen linear combination of all observations exactly.

In the limiting expansion, $F_t$ is the one-root marked law and $J_t$ the two-root connected law. Deleting the last collision of a tree splits it into two components. This identifies their evolution with the Boltzmann equation and its linearization. The full covariance also contains the same-particle contribution:

$$G_t(\Phi ,\Psi )=J_t(\Phi \otimes \Psi )+F_t(\Phi \Psi ).$$

Restoring this diagonal term and using pre/postcollision symmetry produces the covariance source with products of collision increments $\Delta \varphi ={\varphi }^{\prime }+{\varphi }_{\ast }^{\prime }-\varphi -{\varphi }_{\ast }$. Higher connected cumulants vanish at the fluctuation normalization; the surviving first two cumulants determine the Gaussian limit. These conclusions require the source’s sharp tree bounds and spatial disintegration estimates, which justify multiplying component measures at a common position.

Exact centering requires more than a likely coupling

Let ${\mathcal{E}}_{\epsilon }$ be the event on which the true and restarted trajectories differ. The manuscript proves $\mathbb{P}({\mathcal{E}}_{\epsilon })=O({\epsilon }^{1+\eta })$. For a bounded record sum $S$, the difference is at most $2M_{\mathrm{obs}}N{1}_{{\mathcal{E}}_{\epsilon }}$. Hölder’s inequality and bounded normalized particle-number moments give

$$\frac{|\mathbb{E}{S}^{\mathrm{true}}-\mathbb{E}{S}^{\mathrm{restart}}|}{\sqrt{\mu }}\le C_p{\epsilon }^{-1+(1+\eta )(1-1/p)}.$$

Choose $p>(1+\eta )/\eta$ so the exponent is positive. The expectation difference is then negligible at the fluctuation scale. Agreement with probability tending to one would suffice for uncentered differences in probability, but not for this expectation comparison. This final estimate transfers the Gaussian law with its exact microscopic centering. (OpenAI, 2026a s “Identification of the covariance and the Gaussian limit” and “Exact centering and the hard-sphere fluctuation limit”)

See Also

• Family 363 explains why regularity assumptions on the kinetic solution matter.
• Family 362 introduces phase-space transport in a collisionless model.

🏷️ Manuscripts and verification scope

• The Boltzmann–Grad limit for stable radial potentials on regular kinetic intervals — pinned manuscript and companion source.
• Hard-sphere fluctuations on the regular Boltzmann lifespan — pinned manuscript and companion source.

No family-specific scope page was available at lean/docs/364.md in this snapshot. This records the catalogue lookup, not a claim that no related Lean code exists. No Lean build or axiom audit was performed.

OpenAI collection provides prerequisites and the reading sequence; coverage register records completed and pending families.

🐻  OpenAI 2026a. Hard-sphere fluctuations on the regular Boltzmann lifespan.

🐻  OpenAI 2026b. The Boltzmann–Grad limit for stable radial potentials on regular kinetic intervals.