From Newtonian particles to Boltzmann evolution and Gaussian fluctuations

← All digests

Family 364 · Partial differential equations
Importance 5/5 · Provisional editorial rating
Sources consulted 2026-10-26

1. The question

A gas can be described by tracking every particle through Newton’s equations, or by evolving one density through the Boltzmann equation. The second description is far simpler, but why should it correctly summarize the first?

The Boltzmann–Grad limit makes this question precise. Particle interaction ranges shrink while particle numbers grow so that collisions remain frequent enough to matter. At leading order, one expects a deterministic kinetic density. At the next order, random initial sampling produces fluctuations around the mean, even though subsequent particle motion is deterministic.

Family 364 makes two claims on these two levels. M1 derives the nonlinear Boltzmann equation for a broad class of stable radial finite-range interaction potentials, allowing attractive wells. M2 derives Gaussian fluctuations for hard spheres away from equilibrium.

Both extend through any prescribed finite interval on which the corresponding classical Boltzmann solution exists with the stated Gaussian velocity bound. That is a conditional lifespan statement. Neither manuscript proves global regularity of the Boltzmann equation for arbitrary data.

2. Background and essential definitions

The dilute scaling and initial ensemble

Work in the whole position space $\mathbb R^3$, with velocities in $\mathbb R^3$. Let $\varepsilon$ be the interaction range or hard-sphere diameter. The activity parameter is

$$\mu_\varepsilon=\varepsilon^{-2}.$$

The collision cross-sectional area is of order $\varepsilon^2$, so particle number of order $\varepsilon^{-2}$ keeps a typical collision frequency of order one. Meanwhile, occupied volume is of order $\mu_\varepsilon\varepsilon^3=\varepsilon$, which tends to zero.

The papers use a grand-canonical law: the number $N$ is random. Given a probability density $f_0(z)$, with $z=(x,v)$, the initial configuration law is proportional to

$$\frac{\mu_\varepsilon^N}{N!} \prod_{i=1}^N f_0(z_i)\, \mathbf1_{\{|x_i-x_j|>\varepsilon\ \text{for }i\ne j\}}.$$

An exact normalizing constant makes this a probability measure, including the empty configuration. The exclusion prevents particles from starting within interaction range. For M1 it is an initial-law restriction only: the later Newtonian evolution is not artificially forbidden from creating close interacting groups.

Once initial data are sampled, the dynamics are deterministic. There is no stochastic collision rule or repeated resampling.

Stable potentials and the kinetic equation

M1 uses $\Phi_\varepsilon(x)=\Phi(x/\varepsilon)$, where $\Phi$ is radial and zero outside the unit ball. Thermodynamic stability means that for some fixed $B\ge0$, every finite configuration satisfies

$$\sum_{i<j}\Phi(q_i-q_j)\ge-BN.$$

Attraction is allowed, but total potential energy cannot decrease faster than linearly with particle number. The potential is $C^2$ everywhere, or $C^2$ away from a repulsive singularity with $\Phi(q)\to+\infty$ at the origin.

The limiting density solves

$$\partial_tf+v\cdot\nabla_xf=Q_\Phi(f,f),$$

where the collision operator is determined by actual two-body scattering for $\Phi$. M2 instead uses instantaneous elastic hard-sphere collisions.

Independence and fluctuations are different scales

The rescaled factorial marginal $F_s^\varepsilon$ describes ordered selections of $s$ distinct particles, divided by $\mu_\varepsilon^s$. Convergence to $f^{\otimes s}$ means asymptotic independence of any fixed number of sampled particles, often called propagation of chaos.

For a smooth compactly supported test function $a$, the empirical observable is

$$\pi_t^\varepsilon(a)=\mu_\varepsilon^{-1}\sum_{i=1}^N a(z_i(t)).$$

Its leading value is $\int af_t$. Central-limit fluctuations use the larger magnification

$$\zeta_t^\varepsilon(a)=\sqrt{\mu_\varepsilon} \bigl(\pi_t^\varepsilon(a)-\mathbb E\pi_t^\varepsilon(a)\bigr).$$

An error that disappears at the first scale may survive at the second. This is why the exact microscopic expectation appears in the definition.

3. Development history and previous partial results

Grad introduced the dilute kinetic scaling, and Lanford proved the hard-sphere Boltzmann limit for a short time interval. King treated positive finite-range potentials. Gallagher, Saint-Raymond and Texier developed detailed geometric control of recollisions for hard spheres and repulsive interactions.

Pulvirenti, Saffirio and Simonella obtained a short-time theorem for the stable radial class relevant to M1, including attractive parts and scattering maps that need not have a globally single-valued differential cross-section. Global-time results for sufficiently rare dispersive gases were also obtained by Illner and Pulvirenti.

A major recent advance is Deng, Hani and Ma’s hard-sphere derivation throughout every regular finite kinetic interval. Their layered expansion retains correlations from earlier times rather than restarting a short-time theorem as if particles had become independent again. M1 extends this regular-lifespan objective to finite-duration interactions governed by stable potentials.

Fluctuation theory has a parallel history. Van Beijeren, Lanford, Lebowitz and Spohn obtained short-time equilibrium covariance results; Spohn treated nonequilibrium second moments. A Gaussian limit additionally requires control of higher correlations, not just the covariance.

Bodineau, Gallagher, Saint-Raymond and Simonella developed short-time nonequilibrium fluctuation theory, including process convergence, and later proved a long-time equilibrium Gaussian limit. Their fluctuation setting is periodic. M2 works in the whole space with integrable initial data, extends the time interval away from equilibrium, and claims finite-dimensional laws rather than convergence of entire random paths.

These distinctions locate the contributions accurately: neither long-time hard-sphere kinetic limits nor Gaussian Boltzmann fluctuations originate with this family. The new claims concern the stated interaction class, nonequilibrium lifespan and sharper control required at fluctuation scale.

4. What the new manuscripts claim

Common regular-interval assumptions

The initial probability density has Gaussian velocity decay that is summable over spatial unit cells, as does its spatial gradient. Explicitly, for some $\beta>0$, the relevant sums include

$$\sum_{k\in\mathbb Z^3}\ sup_{|x-k|\le1,\ v\in\mathbb R^3} e^{2\beta|v|^2}\bigl(|f_0(x,v)|+|\nabla_xf_0(x,v)|\bigr)<\infty.$$

M1 assumes $f_0\in C^1$; M2 assumes smoothness. Both take as input a nonnegative classical probability-density solution on $[0,T]$ with

$$\sup_{0\le t\le T}\sup_{x,v}e^{2\beta|v|^2}f(t,x,v)<\infty.$$

No near-equilibrium or small-data condition replaces this lifespan hypothesis.

M1: all fixed marginals converge in full phase-space norm

For every fixed positive integer $s$, M1 claims

$$\sup_{0\le t\le T} \|F_s^\varepsilon(t)-f(t)^{\otimes s}\|_{L^1((\mathbb R^6)^s)} \longrightarrow0.$$

The potential and $T$ are fixed independently of $\varepsilon$. Attractive wells, long scattering delays, noninjective scattering maps, simultaneous interactions, recollisions and dynamically formed clusters are allowed in the actual process.

For each fixed observable, empirical averages consequently converge in probability to their kinetic values, uniformly over deterministic observation times. This means $\sup_t\mathbb P(|\text{error at }t|>\eta)\to0$, not a probability bound on the supremum of the error along an entire sample path.

M2: Gaussian hard-sphere fluctuations at finitely many times

For hard spheres, M2 claims that every finite vector

$$\bigl(\zeta_{t_1}^\varepsilon(a_1),\ldots, \zeta_{t_m}^\varepsilon(a_m)\bigr)$$

converges in distribution to the corresponding centered Gaussian vector from the linear fluctuating Boltzmann equation. The times and smooth compactly supported real tests are fixed in $[0,T]$.

The limiting field has weak evolution

$$d\zeta_t(a)=\zeta_t(A_ta)\,dt+dW_t(a),$$

where $A_t$ is the dual linearization of transport and collisions about $f_t$, and $W$ is Gaussian noise with independent increments. This noise describes the limiting statistics of deterministic collisions with random initial data.

For one elastic collision, define

$$\Delta a=a(x,v')+a(x,v_*')-a(x,v)-a(x,v_*).$$

With $B=[(v-v_*)\cdot\omega]_+$, the covariance rate of the noise is

$$\mathcal C_t(a,b)=\frac12\int B f_t(x,v)f_t(x,v_*)\Delta a\,\Delta b\,dx\,dv\,dv_*\,d\omega.$$

This incoming-flux normalization fixes the collision-rate constant. The initial covariance is

$$\mathbb E[\zeta_0(a)\zeta_0(b)]=\int f_0ab.$$

There is no subtraction of $(\int f_0a)(\int f_0b)$, because particle number is random. Initial fluctuations are independent of the collision noise.

5. Why the advance matters

Importance: 5/5, conditional on the claims. The family would strengthen the mathematical connection between microscopic mechanics and kinetic statistics at both the law-of-large-numbers and central-limit levels.

M1 would cover stable interactions with attraction through the full assumed regular kinetic interval, rather than only short times or purely repulsive settings. It treats actual interacting clusters instead of removing them from the microscopic model by definition.

M2 would explain nonequilibrium fluctuations over the same kind of regular lifespan. Identifying the Gaussian covariance and eliminating higher cumulants gives information that leading-order independence cannot provide.

The importance is conditional on kinetic regularity. Deriving a PDE from particle mechanics on its smooth lifespan and proving that the PDE remains smooth forever are different problems.

6. The main idea

Retain correlations across time layers

A collision-tree expansion tracks which particles have influenced an observed particle. Over longer times, those histories proliferate and create correlations. One cannot simply declare independence again at the start of each short interval.

Both manuscripts use coarse time layers, with finer subdivisions inside them, and center parts of the expansion against the known Boltzmann solution. Correlated remainders retain their earlier histories. Exact finite identities are established before limits and cutoffs are removed.

In M1, interactions last for positive time and several particles may interact together. The expansion therefore follows whole connected interaction components, rather than breaking every event into isolated instantaneous binary collisions. Keeping components intact preserves their potential-energy bookkeeping.

Thermodynamic stability provides one energy budget across the history. Each particle input is charged once rather than repeatedly at every layer. Combined with the time factors, this yields a summable expansion when the fixed number of coarse layers is chosen sufficiently large.

Expose exceptional encounters in causal order

Recollisions, crowded cells and long interactions are treated through geometric estimates. Fresh scattering parameters give small sets for an extra encounter with a trajectory independent of those parameters.

To multiply such small factors, the exposure order matters. For a genuine scattering, both not-yet-exposed outgoing branches must remain hidden until the fresh parameters are integrated. A virtual overlap between independently evolved components requires different bookkeeping. Exposing a branch too early could destroy the independence on which the small-factor estimate relies.

The manuscript combines these causal estimates with finite-history counting and chronological cutoff removal. Close configurations and any clusters in the real dynamics are finally controlled through their small nonnegative marginal mass; they were not excluded from the evolution.

Fluctuations require sharp connected estimates

A cumulant records the part of a correlation not explained by products of lower-order correlations. For Gaussian variables, cumulants above order two vanish.

M2 obtains connected factorial-cumulant bounds at scale $\mu^{1-h}$ for $h$ observed particle histories. Geometric mergers contribute the powers of $\varepsilon^2$ needed for that scale. After central-limit normalization, higher ordinary cumulants vanish, while the surviving two-particle terms and the same-particle diagonal contribution yield the stated covariance.

Observation marks attached to particle histories encode several fixed times in one calculation. This gives joint Gaussian laws for any finite list without asserting tightness in a space of whole trajectories.

Exact centering cannot be replaced by the kinetic mean

The initial exclusion already changes the empirical expectation at fluctuation scale. Writing $\rho_0(x)=\int f_0(x,v)\,dv$, M2 states

$$\sqrt\mu\left(\mathbb E\pi_0^\varepsilon(a)-\int f_0a\right) \longrightarrow-\frac{4\pi}{3}\int f_0(x,v)a(x,v)\rho_0(x)\,dx\,dv.$$

Thus centering by $\int af_t$ alone generally introduces a deterministic shift already at time zero. Leading-order convergence of the mean is not accurate enough.

The proof first obtains fluctuation limits for a modified, truncated dynamics. It then bounds the probability that this dynamics ever differs from the true one before $T$ by $O(\varepsilon^{1+\eta})$ for some $\eta>0$. Together with moments of $N/\mu$, this stronger estimate transfers expectations at order $o(\mu^{-1/2})$, and hence transfers exact centering. A discrepancy probability merely tending to zero would not suffice for that step.

7. Scope, limitations, and open questions

M1 treats fixed stable finite-range radial potentials in its stated smooth or repulsively singular class. It does not include arbitrary long-range forces, unstable attractions or nonspherical interactions. M2’s fluctuation theorem is for hard spheres; it is not automatically a fluctuation theorem for every potential in M1.

Both rely on a classical kinetic solution with uniform Gaussian velocity decay over the chosen finite interval. Neither extends the conclusion past a kinetic singularity or chooses among rough nonunique Boltzmann solutions.

M1 controls each fixed marginal order, not an order growing with particle number. Its empirical corollary is uniform over deterministic times, while M2 gives finite-dimensional distributions rather than process convergence or a path-space tightness theorem.

Extending fluctuations to stable attractive potentials, establishing full path-space convergence on the regular nonequilibrium lifespan, and weakening the kinetic envelope assumptions are separate questions. The expansions and convergence claims are reported as stated without independent proof verification.

8. Sources and further reading

M1 states the stable-potential theorem and deterministic-time empirical corollary. M2 specifies the exact microscopic centering, grand-canonical covariance and finite-dimensional fluctuation limit.

Primary predecessors include Pulvirenti–Saffirio–Simonella on stable short-range potentials, Deng–Hani–Ma on the regular-lifespan hard-sphere limit, BGSS on short-time statistical dynamics, and their equilibrium fluctuation theorem.

Source manuscripts

• M1. OpenAI, The Boltzmann–Grad limit for stable radial potentials on regular kinetic intervals (September 23, 2026).
• M2. OpenAI, Hard-sphere fluctuations on the regular Boltzmann lifespan (September 23, 2026).

Sources consulted on 26 October 2026 at repository revision 1581d4c. Download the manuscript citations.