The Collision Integral and Molecular Chaos

A dilute gas of $N$ molecules is described microscopically by the $N$ -particle Liouville equation, but tracking $6N$ coordinates is both hopeless and unnecessary. Boltzmann's strategy is to derive a closed equation for the single-particle distribution $f(\mathbf r,\mathbf p,t)$ , the expected number of molecules per unit phase-space volume. Free streaming under external forces is handled exactly; the entire difficulty sits in the collisions, which couple $f$ to the two-particle distribution $f^{(2)}$ . The collision integral is the term that encodes this coupling, and closing it requires an assumption about how the velocities of two molecules about to collide are correlated. The Stosszahlansatz — literally "collision-number hypothesis," usually called molecular chaos — asserts that two molecules about to collide arrive with independently distributed velocities, so $f^{(2)}(\mathbf r,\mathbf p_1,\mathbf p_2)=f(\mathbf r,\mathbf p_1)f(\mathbf r,\mathbf p_2)$ at the instant before impact. This single closure truncates the exact hierarchy of reduced distributions at first order and produces a nonlinear, bilinear collision term. The bilinearity reflects that a collision needs two partners; the factorization is the precise place where information about correlations is thrown away. The price and the prize of the ansatz are the same: irreversibility. Microscopic mechanics is time-reversal invariant, yet the Boltzmann equation has a monotone Lyapunov functional, the H-function, that can only decrease. The resolution is that molecular chaos is imposed on incoming pairs but not outgoing ones — after a collision the partners are correlated, so applying the ansatz again on the next collision quietly discards that correlation. This asymmetry between past and future selects a thermodynamic arrow of time and is the conceptual heart of how Boltzmann reconciled mechanics with the second law. The collision integral also structures the rest of kinetic theory: its zeros are the local Maxwellians, its linearization governs transport coefficients, and its weak-coupling limit connects to the Fokker-Planck and master equations of stochastic dynamics.

Key takeaways

The Boltzmann collision integral $C[f]$ is a bilinear functional of the single-particle distribution that counts the rate at which binary elastic collisions scatter molecules into and out of a phase-space cell.
Its central, non-rigorous input is the Stosszahlansatz (molecular chaos): the pre-collision two-particle distribution factorizes, $f^{(2)}(\mathbf p_1,\mathbf p_2)\approx f(\mathbf p_1)f(\mathbf p_2)$ , treating incoming velocities as uncorrelated.
This factorization breaks time-reversal symmetry and is exactly what makes the equation irreversible, underwriting the H-theorem and the second law from reversible microdynamics.
The gain-loss structure uses microscopic reversibility and the collisional invariants $1,\mathbf p,p^2$ , so $C[f]$ vanishes precisely for the local Maxwell-Boltzmann distribution.
Molecular chaos is an assumption about the state, not the dynamics: correlations build up over a collision, so the ansatz holds for dilute gases on times longer than the collision duration but shorter than the mean free time.
Key idea: closing the hierarchy by assuming uncorrelated incoming velocities ( $f^{(2)}\to ff$ ) turns the exact two-body collision term into a bilinear, irreversible operator whose null space is exactly the Maxwell-Boltzmann distribution.

The core insight

Closing the BBGKY hierarchy by assuming uncorrelated incoming velocities ( $f^{(2)}\to ff$ ) turns the exact two-body collision term into a bilinear, irreversible operator whose null space is exactly the Maxwell-Boltzmann distribution.

Where this lives today

The collision integral, and the molecular-chaos closure that defines it, remains the conceptual and computational core of rarefied-gas dynamics. Direct Simulation Monte-Carlo — the standard tool for spacecraft re-entry, vacuum systems, and micro-electro-mechanical flows — literally samples the gain-loss collision integral stochastically, pairing simulation particles and scattering them with the correct cross section, and thereby reproduces slip flow, thermal creep, and shock structure that hydrodynamics cannot capture. Molecular chaos enters as the assumption that colliding simulation particles are drawn independently from the local distribution. In plasma and astrophysical kinetics the grazing-collision limit — the Landau collision integral with its Coulomb logarithm — governs the relaxation of fusion plasmas and the two-body relaxation of star clusters, and underlies transport coefficients like Spitzer resistivity. And the same operator's weak-coupling, small-momentum-transfer limit is the microscopic origin of Brownian motion: a heavy particle in a light gas experiences many small kicks that coarse-grain to a Fokker-Planck (Langevin) equation with friction and noise obeying the Einstein relation, which is how kinetic theory ultimately connects to the whole stochastic side of statistical mechanics.

The physical picture

Picture a single tagged molecule drifting through the gas. Most of the time it free-streams; occasionally it strikes another molecule and is kicked to a new velocity. The collision integral is just the bookkeeping of these kicks: a "loss" term removes molecules from velocity $\mathbf p_1$ when they scatter away, and a "gain" term adds molecules that scatter into $\mathbf p_1$ from some other pair $(\mathbf p_1',\mathbf p_2')$ via the inverse collision. The rate of either process is proportional to how many suitable partners are available — hence two factors of $f$ , the bilinearity. The molecular-chaos closure is the leap of faith that the partner you are about to hit is a "fresh" molecule, statistically independent of you. Right after you collide, you and your partner fly apart correlated: your velocities are linked by the conservation laws. Molecular chaos says that by the time either of you collides again, you have each met so many other molecules that this correlation is washed out and can be ignored for the next incoming pair. It is a statement about the state of colliding molecules, not about the reversible dynamics. Why does this break time symmetry? Run the film backwards. The time-reversed motion would require outgoing pairs to be uncorrelated and incoming ones correlated — the opposite of what the ansatz assumes. The Boltzmann equation is therefore not the time-reversal of itself; the arrow of time is injected by always factorizing the past, never the future. The dedicated resolution of the Loschmidt and Zermelo objections that this raises lives in the irreversibility article; here the factorization is the source of the arrow. Equilibrium is the state where every collision is exactly balanced by its reverse — detailed balance in velocity space. The integrand of $C[f]$ then vanishes pointwise, which forces $\ln f$ to be a linear combination of the collisional invariants $1,\mathbf p,p^2$ , i.e. a Maxwellian. The collision integral thus both relaxes the gas and defines what equilibrium is.

Figure

The gain-loss collision pair on the momentum sphere, with molecular chaos as the T-breaking input

Figure

Punchline: the H-function slides monotonically to its Maxwellian floor

The core equations

The Boltzmann transport equation splits the evolution of $f$ into streaming and collisions; the collision integral $C[f]$ is the entire right-hand side. The streaming part is the collisionless Liouville derivative, established elsewhere; here everything of interest is in $C[f]$ .

\left(\frac{\partial}{\partial t}+\frac{\mathbf p}{m}\cdot\nabla_{\mathbf r}+\mathbf F\cdot\nabla_{\mathbf p}\right)f(\mathbf r,\mathbf p,t)=C[f] \qquad (1)

where: $f(\mathbf r,\mathbf p,t)$ is the single-particle distribution; $\mathbf p/m$ the velocity; $\mathbf F$ the external force; $C[f]$ the net rate of change from binary collisions. The three streaming terms advect $f$ along free trajectories (Liouville); $C[f]$ is the only source of relaxation. The collision integral is gain minus loss, each a rate built from the flux factor, the cross section, and — through molecular chaos — a product of one-particle distributions.

C[f]=\int\mathrm d^3p_2\!\int\mathrm d\Omega\,|\mathbf v_1-\mathbf v_2|\,\frac{\mathrm d\sigma}{\mathrm d\Omega}\left(f_1'f_2'-f_1f_2\right) \qquad (2)

where: $f_1,f_2$ are the distributions at incoming momenta $\mathbf p_1,\mathbf p_2$ ; $f_1',f_2'$ at the outgoing momenta of the inverse collision; $|\mathbf v_1-\mathbf v_2|$ is the relative speed (flux factor); $\mathrm d\sigma/\mathrm d\Omega$ the differential cross section. The $+f_1'f_2'$ gain and $-f_1f_2$ loss appear with the same kernel because the inverse collision has identical relative speed, cross section, and a unit Jacobian $\mathrm d^3p_1'\mathrm d^3p_2'=\mathrm d^3p_1\mathrm d^3p_2$ . Each binary collision conserves momentum and kinetic energy, which fixes the post-collision momenta given the pre-collision ones and the scattering angle.

\mathbf p_1+\mathbf p_2=\mathbf p_1'+\mathbf p_2',\qquad p_1^2+p_2^2=p_1'^2+p_2'^2 \qquad (3)

where: $\mathbf p_i,\mathbf p_i'$ are pre- and post-collision momenta (equal masses). These four constraints leave the scattering direction as the only free parameter, integrated over by $\mathrm d\Omega$ . Any quantity conserved in a collision has a vanishing moment of $C[f]$ — the source of the hydrodynamic conservation laws.

\chi(\mathbf p_1)\in\{1,\ \mathbf p_1,\ p_1^2\}\ \Rightarrow\ \int\mathrm d^3p_1\,\chi(\mathbf p_1)\,C[f]=0 \qquad (4)

where: $\chi$ is any collisional invariant — number, momentum, kinetic energy. The derivation below symmetrizes over the four momenta to make this vanishing manifest. The H-function is the Lyapunov functional; its manifestly non-positive rate of change is the H-theorem.

H(t)=\int\mathrm d^3r\,\mathrm d^3p\;f\ln f,\qquad \frac{\mathrm dH}{\mathrm dt}=-\frac14\!\int\mathrm d\Gamma\,(f_1'f_2'-f_1f_2)\ln\frac{f_1'f_2'}{f_1f_2}\le0 \qquad (5)

where: $\mathrm d\Gamma=\mathrm d^3p_1\mathrm d^3p_2\mathrm d\Omega\,|\mathbf v_1-\mathbf v_2|\,\mathrm d\sigma/\mathrm d\Omega$ is the collision measure; $-k_BH$ is the Boltzmann entropy. The non-positivity follows from the elementary inequality $(x-y)\ln(x/y)\ge0$ ; the dedicated H-theorem article completes the argument, which this equation summarizes. The time-scale hierarchy is what justifies molecular chaos physically: collisions are brief compared to the mean free time, and correlations decay on the collision scale.

\tau_{\text{coll}}\ll\tau_{\text{mfp}}\sim\frac{1}{n\sigma\bar v}\ll\tau_{\text{hydro}} \qquad (6)

where: $\tau_{\text{coll}}$ is the collision duration, $\tau_{\text{mfp}}$ the mean free time, $n\sigma$ the density times cross section. Diluteness $n\sigma^{3/2}\ll1$ makes the correlations built up in a collision short-lived compared to the time until the next one.

Derivation AFrom the hierarchy to the collision integral via molecular chaos

The goal is to obtain Boltzmann's bilinear collision term from the exact first hierarchy equation by imposing the diluteness limit and the molecular-chaos factorization, and to show why that factorization is the irreversible step. The strategy is to write the exact equation for the one-particle distribution — which couples to the two-particle distribution through the interparticle force — then argue that in a dilute gas the coupling is local in space and needs $f^{(2)}$ only at the moment of impact, at which point the Stosszahlansatz replaces it by a product. Watch that the factorization is applied to the incoming pair; this is the single place where correlations are discarded and time-reversal symmetry is lost.

Step 1. Start from the exact first hierarchy equation, in which the streaming of $f^{(1)}$ is driven by the interparticle force averaged over the second particle:

\left(\partial_t+\tfrac{\mathbf p_1}{m}\!\cdot\!\nabla_{\mathbf r_1}+\mathbf F\!\cdot\!\nabla_{\mathbf p_1}\right)f^{(1)}=\int\mathrm d^3r_2\,\mathrm d^3p_2\,(\nabla_{\mathbf r_1}V_{12})\!\cdot\!\nabla_{\mathbf p_1}f^{(2)}.

This is not closed — it needs $f^{(2)}$ .

Step 2. Invoke diluteness and locality. For short-range $V_{12}$ and a dilute gas, the interaction matters only when $|\mathbf r_1-\mathbf r_2|\lesssim$ range, so the coupling collapses to an integral over the collision cylinder and $f^{(2)}$ is needed only at contact.

Step 3. Impose the Stosszahlansatz on the incoming pair at the same point:

f^{(2)}(\mathbf r_1,\mathbf p_1,\mathbf r_1,\mathbf p_2;t)\to f^{(1)}(\mathbf r_1,\mathbf p_1,t)\,f^{(1)}(\mathbf r_1,\mathbf p_2,t).

The correlation built up in previous collisions is dropped — the irreversible step.

Step 4. Convert the force integral to a gain-loss form. Parameterize collisions by impact parameter, use $b\,\mathrm db\,\mathrm d\phi=(\mathrm d\sigma/\mathrm d\Omega)\mathrm d\Omega$ , and split trajectories entering versus leaving the collision cylinder into loss ( $-f_1f_2$ ) and gain ( $+f_1'f_2'$ ). Microscopic reversibility — identical $|\mathbf v_{\text{rel}}|\,\mathrm d\sigma$ and unit Jacobian for the inverse collision — lets the gain share the same kernel, giving Eq. (2). Why irreversibility appears here and nowhere else: every other step is an exact reversible manipulation; only the incoming-pair factorization treats past and future asymmetrically. *Verified: the loss/gain split, the Jacobian invariance, and the Maxwellian-as-null-space are checked symbolically — verify_collision-integral-molecular-chaos.py. Matches Pathria & Beale ch. 13 and Kardar ch. 3.*

Go deeper · Derivation B · 1 minThe Maxwellian is the unique zero of the collision integral

Derivation BThe Maxwellian is the unique zero of the collision integral

The goal is to prove that $C[f]=0$ pointwise requires $\ln f$ to be a summational (collisional) invariant, which forces $f$ to be a local Maxwellian — establishing that the collision integral not only relaxes the gas but defines equilibrium. The strategy is to note that a vanishing collision integral demands the integrand vanish for every collision, translate that into an additive functional equation for $\ln f$ , and solve it using the fact that the only additive collisional invariants are $1,\mathbf p,p^2$ . Watch that this is stronger than the H-theorem stationarity: it pins the functional form of $f$ , not just $\mathrm dH/\mathrm dt=0$ .

Step 1. Require the integrand of $C[f]$ to vanish for all collisions: $f_1f_2=f_1'f_2'$ for every $(\mathbf p_1,\mathbf p_2)\to(\mathbf p_1',\mathbf p_2')$ . Vanishing of the integral is not enough — the H-theorem shows the integrand is one-signed, so it must vanish identically.

Step 2. Take the logarithm: $\ln f_1+\ln f_2=\ln f_1'+\ln f_2'$ . So $\ln f$ is conserved in every collision — it is a summational invariant.

Step 3. Solve the functional equation. The most general additive quantity conserved in a binary collision is a linear combination of the independent invariants: $\ln f=a+\mathbf b\cdot\mathbf p+c\,p^2$ .

Step 4. Complete the square and identify coefficients with the hydrodynamic fields:

f^{(0)}=n\left(\frac{1}{2\pi mk_BT}\right)^{3/2}e^{-(\mathbf p-m\mathbf u)^2/2mk_BT},

the local Maxwellian, with $a,\mathbf b,c$ fixed by $n,\mathbf u,T$ . This is the unique null distribution. Why uniqueness holds: the strict convexity used in the H-theorem makes $f_1f_2=f_1'f_2'$ both necessary and sufficient for stationarity, and the invariant-counting leaves no other functional form. *Verified: $\ln f=a+c\,p^2$ is checked to be conserved under energy conservation $p_1^2+p_2^2=p_1'^2+p_2'^2$ symbolically — verify_collision-integral-molecular-chaos.py. Matches Reif §14, Pathria §13.4.*

Limiting cases

The collision integral simplifies in physically distinct regimes:

Case	Result	Meaning
Equilibrium (null space)	$C[f]=0\iff f=f_0=n\left(\frac{m}{2\pi k_BT}\right)^{3/2}e^{-m(\mathbf v-\mathbf u)^2/2k_BT}$	the local Maxwellian is the unique zero, fixed by the five collisional invariants
Linearized (near equilibrium)	$f=f_0(1+h)$ , $C[f]\approx\hat Lh$ , $\hat L(1,\mathbf p,p^2)=0$	Chapman-Enskog regime: transport coefficients are matrix elements of the self-adjoint, negative-semidefinite $\hat L$
Grazing / Coulomb	$C[f]\to-\partial_{v_i}\!\int\mathrm d^3v'\,U_{ij}(f'\partial_{v_j}f-f\partial_{v_j'}f')$	small-angle scattering (plasmas) gives the Landau collision integral — a Fokker-Planck operator with Coulomb-logarithm cutoff
BGK relaxation	$C[f]\approx-(f-f_0)/\tau$	single-time relaxation to local equilibrium; keeps invariants and the H-theorem, sacrifices the correct Prandtl number

Worked examples

Example 1 — Mean free path of air (the real number). For air at standard temperature and pressure ( $T=273.15\,$ K, $P=101325\,$ Pa) modeled as hard spheres of kinetic diameter $d=3.64\times10^{-10}\,$ m, find the mean free path $\ell=1/(\sqrt2\,n\pi d^2)$ and the collision frequency.

Step 1. Number density from the ideal-gas law: $n=P/(k_BT)=101325/(1.381\times10^{-23}\times273.15)=2.69\times10^{25}\,$ m $^{-3}$ .

Step 2. The collision cross section is $\pi d^2=\pi(3.64\times10^{-10})^2=4.16\times10^{-19}\,$ m $^2$ . The $\sqrt2$ comes from averaging the relative speed of two moving molecules, $\langle|\mathbf v_1-\mathbf v_2|\rangle=\sqrt2\,\bar v$ , not the speed of one against a fixed target.

Step 3. $\ell=\dfrac{1}{\sqrt2\,(2.69\times10^{25})(4.16\times10^{-19})}=\mathbf{6.3\times10^{-8}\ m}\approx63\ \mathrm{nm}$ , the textbook air value (quoted as $\sim68\,$ nm for $d\approx3.7\,$ Å). With mean speed $\bar v\approx460\,$ m/s the collision frequency is $\bar v/\ell\approx7\times10^{9}\,$ s $^{-1}$ .

Example 2 — The H-theorem inequality on a two-cell relaxation. A spatially homogeneous gas is coarse-grained into two velocity classes with populations $f_1,f_2$ that a symmetric collision exchanges toward equality. Show $H=f_1\ln f_1+f_2\ln f_2$ decreases when the populations relax from $(f_1,f_2)=(0.8,0.2)$ toward $(0.5,0.5)$ .

Step 1. Initial: $H_i=0.8\ln0.8+0.2\ln0.2=0.8(-0.2231)+0.2(-1.6094)=-0.1785-0.3219=-0.5004$ .

Step 2. After one symmetric relaxation step to $(0.65,0.35)$ : $H=0.65\ln0.65+0.35\ln0.35=0.65(-0.4308)+0.35(-1.0498)=-0.2800-0.3674=-0.6474$ .

Step 3. $\Delta H=-0.6474-(-0.5004)=\mathbf{-0.147<0}$ , and continuing to the equal-population fixed point gives the minimum $H=\ln0.5=-0.693$ . The H-function decreases monotonically to its floor at the "Maxwellian" (equal populations), exactly as $(x-y)\ln(x/y)\ge0$ demands.

Origins · 1 minmolecular chaos and how this was discovered

molecular chaos

the Stosszahlansatz, introduced by Ludwig Boltzmann (1872) and named "Molekularunordnung" by Boltzmann in 1895; hierarchy foundations by Bogoliubov, Born, Green, Kirkwood, Yvon · Molecular chaos — the assumption that two colliding molecules have statistically independent pre-collision velocities — is the closure Boltzmann introduced in 1872 (the "Stosszahlansatz," collision-number hypothesis) and later named "Molekularunordnung," molecular disorder, in his 1895 reply to critics. It is the ingredient that converts the exact, reversible many-body dynamics into the closed, irreversible Boltzmann equation, and it was at the center of the Loschmidt reversibility and Zermelo recurrence objections. Its status was clarified only decades later by the systematic derivation of the collision integral from the exact hierarchy of reduced distributions (the BBGKY chain of Bogoliubov, Born, Green, Kirkwood, and Yvon), which located molecular chaos as a controlled dilute-gas limit rather than an arbitrary assumption.

History

1872

— Boltzmann introduces the collision integral and the Stosszahlansatz

1876

— Loschmidt raises the reversibility objection to the H-theorem

1895

— Boltzmann names the assumption "Molekularunordnung" (molecular disorder) in reply to critics

1896

— Zermelo raises the Poincaré-recurrence objection

1936

— Landau derives the small-angle collision integral for Coulomb interactions

1946

— Bogoliubov, Born and Green, and Kirkwood set the collision integral within the exact hierarchy

Going further

References. Pathria & Beale, Statistical Mechanics, 3rd ed., ch. 13 (The Boltzmann transport equation) · Kardar, Statistical Physics of Particles, ch. 3 (Kinetic theory) · Reif, Fundamentals of Statistical and Thermal Physics, ch. 12-14 · Sethna, Statistical Mechanics: Entropy, Order Parameters, and Complexity, ch. 5 · Landau & Lifshitz, Physical Kinetics (vol. 10), ch. 1-3 · Cercignani, The Boltzmann Equation and Its Applications.

Provenance. All equations are standard kinetic theory. A verifier should: (1) reproduce the gain-loss collision integral and its derivation from the hierarchy via the Stosszahlansatz against Pathria & Beale ch. 13 and Kardar ch. 3; (2) confirm the four-fold symmetrization and the inequality $(x-y)\ln(x/y)\ge0$ giving $\mathrm dH/\mathrm dt\le0$ (Reif ch. 12-13, Cercignani); (3) check the collisional-invariant identity and that the Maxwellian is the unique null distribution; (4) verify the Landau/Fokker-Planck reduction and Coulomb logarithm against Landau-Lifshitz Physical Kinetics §41-46; (5) recompute the hard-sphere mean free path $\ell=1/(\sqrt2 n\pi d^2)$ for air. Numerical constants are standard textbook values; none fabricated.

Builds on: The Master Equation · The Fokker-Planck Equation · The Kramers-Moyal Expansion. Leads to: The Fokker-Planck Equation · Brownian Motion · The Fluctuation-Dissipation Theorem.

What is the Stosszahlansatz in plain words?

It is the assumption — literally "collision-number hypothesis" — that the two molecules about to collide have statistically independent velocities, so the probability of finding them together is the product of the two single-particle probabilities: $f^{(2)}=f\cdot f$ . Boltzmann needed it to close his equation: without it, the one-particle distribution depends on the two-particle distribution, which depends on the three-particle one, and so on forever. The ansatz cuts this chain by declaring incoming pairs uncorrelated, which is well justified in a dilute gas where each molecule meets many fresh partners between correlated encounters.

Why does factorizing the incoming pair make the equation irreversible?

Because it treats the past and the future asymmetrically. Before a collision the two molecules may plausibly be uncorrelated; after the collision they are definitely correlated, their velocities linked by conservation laws. Boltzmann factorizes only the incoming pair. If you time-reverse the motion, the roles swap: the (now incoming) formerly-outgoing pair is correlated, violating the ansatz. So the Boltzmann equation is not equal to its own time-reversal — the factorization of the past but not the future is exactly the mathematical origin of the arrow of time, even though the underlying Newtonian dynamics is perfectly reversible.

How can the H-theorem give irreversibility from reversible mechanics without contradiction?

There is no contradiction because the H-theorem is a theorem about the closed kinetic equation, not about the exact $N$ -body trajectory. The exact trajectory is reversible and recurrent. The Boltzmann equation is obtained from it by imposing molecular chaos, which breaks time-reversal symmetry. So $\mathrm dH/\mathrm dt\le0$ is a genuine property of the coarse-grained description, while the microscopic mechanics remains reversible. Loschmidt's objection (reverse all velocities and $H$ should increase) and Zermelo's (Poincaré recurrence must return $H$ to its start) are both answered by noting that molecular chaos fails on the reversed or recurrent trajectory — these are resolved in detail in the dedicated irreversibility-paradox article.

Why are there exactly the invariants

1

\mathbf p

, and

p^2

Because those are the only quantities conserved in a binary elastic collision: particle number ( $1$ ), the three components of momentum ( $\mathbf p$ ), and kinetic energy ( $p^2$ ). Any function of momentum that is conserved in every collision must be a linear combination of these — this is a theorem about the collision kinematics. Their importance is twofold: the vanishing moments of $C[f]$ against them give the five hydrodynamic conservation laws, and the requirement that $\ln f$ be built from them at equilibrium is exactly what forces the Maxwellian form. So the invariants control both the approach to equilibrium and equilibrium itself.

What is the role of microscopic reversibility if the equation is irreversible?

Microscopic reversibility is what lets the gain term (inverse collisions) share the same cross section and Jacobian as the loss term (direct collisions), giving the clean $f_1'f_2'-f_1f_2$ structure. It is a property of the underlying two-body scattering and is fully time-symmetric. The irreversibility is a separate ingredient — molecular chaos — layered on top. So the collision integral uses reversibility to build its kernel and molecular chaos to break time symmetry; both are needed, and they act at different levels (the two-body dynamics versus the statistical closure).

When exactly does molecular chaos break down?

It breaks down when correlations between colliding molecules cannot be neglected: at high density, where a molecule re-encounters its former partner before decorrelating (recollisions and ring correlations), so the diluteness condition $n\sigma^{3/2}\ll1$ fails; in the presence of long-lived hydrodynamic correlations (the source of the algebraic long-time tails); and in strongly coupled systems where three-body and higher collisions matter. In these regimes one must go beyond the Boltzmann closure — to the ring-kinetic or full hierarchy treatments — or use correlation-function (Green-Kubo) methods that make no factorization assumption.