Audience undergraduate + curious expertReading time 75–110 minInteractive labs 14+
zj = (xj, vj)
𝔼N · εd−1 ≈ 1
fs → f⊗s
Conceptual gas-particle field · not a numerical hard-sphere simulation
One physical gas.
MICROMESOMACRO
Three mathematical languages.
HOW TO READ THIS
A map of a proof—not a replacement for 192 pages
This guide makes every logical dependency visible, proves the elementary mechanics and scaling calculations in full, and gives a step-by-step reconstruction of the long-time argument. The deepest cutting-algorithm estimates are explained structurally and schematically; reproducing their full hypotheses, constants, and case analysis requires the research article.
Figure policy. Every SVG, canvas, graph, and workflow on this site is an original pedagogical schematic—not a reproduced research-paper figure. Source trails beside the key displays point to the authors’ exact statements and constructions.
Intuition Derivation Research-level bridge
PROOF RELAY · FOLLOW THE SAME OBJECT FORWARD
Each step must hand the next step exactly what it needs.
Select a step to see its input, mathematical move, passed-forward output, and unpaid question. The output line is repeated verbatim as the next step’s input.
STEP 1 OF 7 · MICROSCOPIC RULES
Evolve the initial ensemble exactly
INPUT
An initial ensemble WN(0) supported on non-overlapping hard-sphere configurations.
MOVE
Apply the hard-sphere flow HN(t): free flight plus the exact elastic collision map.
WHY THIS MOVE WORKS
The flow is invertible and preserves phase-space volume away from a null singular set, so pushing the initial law through it gives the exact law at time t.
OUTPUT · PASSED FORWARD
The exact ensemble law WN(t), transported by reversible hard-sphere dynamics.
SCOPE · DO NOT OVERCLAIM
Nothing has been averaged, closed, or sent to a limit yet; this is still the full microscopic law in one fixed-N sector.
UNPAID QUESTION
How can we retain only the few particles an observer asks about?
Use the scale switcher. The physical gas is the same, but the mathematical object changes as we deliberately retain less information.
What are all positions and velocities?
State: (x₁,v₁), …, (xₙ,vₙ)
MODEL
Newton + elastic collisions
Follow every hard sphere. Between collisions it travels in a straight line; at collision, momentum and kinetic energy are conserved.
KEEPS
Exact particle identities and trajectories.
FORGETS
Nothing—this is why the state has about 6N numbers in 3D.
What the switcher establishes: changing scale changes which variables we retain. It does not simulate either limiting theorem or claim that a finite particle system literally equals a kinetic or fluid field.
Stage 1: let ε→0 with activity α fixed. For the excluded grand-canonical law, (𝔼N)εd−1=α+O(ε), so the population is asymptotic to αε−(d−1); this produces Boltzmann. Stage 2: after the kinetic equation is obtained, let the Knudsen number Kn→0; in the nondimensional convention used for this bridge, Kn=α−1. This is a separate hydrodynamic limit, and the companion paper specifies the permissible iterated/scaling variants.
CONTEXT INTERLUDE · OUTSIDE THE NUMBERED PROOF CHAIN
The citation names three research programs. Chapters 02–07 now follow the hard-sphere program without a side trip; the wave and random-PDE pillars return only after the particle-to-fluid chain is complete.
2026
01 · WHAT THE MEDAL RECOGNIZES
“For his work in partial differential equations, including the rigorous derivation of the Boltzmann equation from hard-sphere dynamics for rarefied gases, the derivation of wave kinetic equations from nonlinear dispersive systems, and probabilistic approaches to nonlinear Schrödinger dynamics.”
Long-time control of collision histories for a dilute hard-sphere gas, with Zaher Hani and Xiao Ma.
02
Where the method matured
Waves → kinetic theory
Rigorous statistical laws for weakly nonlinear waves at their natural kinetic timescale, with Zaher Hani.
03
A second probability program
Randomness → PDE control
Random tensors, probabilistic scaling, and Gibbs-measure dynamics, with Andrea Nahmod and Haitian Yue.
The hard-sphere Boltzmann theorem and its companion fluid-limit paper are joint work of Yu Deng, Zaher Hani, and Xiao Ma. The wave-kinetic results discussed here are joint work of Deng and Zaher Hani; the highlighted probabilistic NLS results are joint work of Deng, Andrea Nahmod, and Haitian Yue.
02 · START AT THE BOTTOM
One collision, derived completely
Before the probability and combinatorics, the microscopic rule is just geometry plus two conservation laws.
STATE
zj = (xj, vj) ∈ ℝd × ℝd
Particle j has a center xj and velocity vj. In 3D that is six real numbers.
NO OVERLAP
|xi − xj| ≥ ε
Each sphere has diameter ε. Their centers can approach to distance ε, never less.
FREE FLIGHT
ẋj = vj, ṽj = 0
Until a collision, velocity stays constant and position changes linearly.
INTERACTIVE LAB 1
Turn the collision normal
At impact, only the component of relative velocity parallel to the line of centers is exchanged. The perpendicular component is untouched.
Collision replay ready.
Total momentum before(1.20, 0.10)
Relative normal speed+a → −a
Total momentum after(1.20, 0.10)
Kinetic energy error0.000000
Follow the live calculation
1 · normal componenta=(v−v*)·ω
2 · particle 1v′=v−aω
3 · particle 2v′*=v*+aω
particle 1 particle 2 collision normal ω
Exact two-body free-flight and impact kinematics; not a many-particle simulation.
In the center-of-mass frame, a collision is a reflection
Define the center-of-mass velocity and relative velocity
V = (v+v*)/2, g = v−v*.
Momentum conservation keeps V fixed. For identical, frictionless, perfectly elastic spheres, the tangent component of g is unchanged and the normal component reverses:
g = gtan + (g·ω)ω g′ = gtan − (g·ω)ω = g − 2(g·ω)ω.
Since v=V+g/2 and v*=V−g/2, substituting g′ gives the collision rule used above.
Momentum is (2,1) on both sides. The squared-speed sums are 5 before and 2000/625+1125/625=3.2+1.8=5 after; the normal relative velocity changes from +2/5 to −2/5.
Scope. “Exchange the parallel component” is special to equal-mass, smooth, perfectly elastic spheres. Unequal masses or rough bodies obey different formulas.
DERIVATION A
Why the collision formula conserves momentum and energy
1
Isolate the exchanged scalar
Let v,v* be the velocities before impact and v′,v′* the corresponding velocities after impact. Set a = (v − v*) · ω, the signed normal component of relative velocity along the unit collision normal ω.
v′ = v − aω, v′* = v* + aω
The same vector aω is subtracted from one velocity and added to the other.
2
Add the two outgoing velocities
v′ + v′* = (v − aω) + (v* + aω) = v + v*
So total momentum is exactly conserved (equal masses are normalized to 1).
Because |ω| = 1 and a = (v − v*)·ω, the last two terms are 2a² − 2a² = 0.
✓
Conclude
|v′|² + |v′*|² = |v|² + |v*|²
The microscopic map is invertible: reverse every velocity and the spheres retrace their paths.
03 · FROM ONE TRAJECTORY TO A LAW
Randomness enters at time zero
Newton evolves each chosen configuration deterministically. Probability describes which configuration we chose—and what a typical particle does.
UNDERGRADUATE BRIDGE · FLOW → LAW → CORRELATION
One deterministic flow transports the entire probability law
Fix one particle-number sector N and write ZN=(z₁,…,zN). The allowed configuration domain is DNε:={ZN:|xi−xj|≥ε for i≠j}; W0,N:=WN(0) is its initial density, and G below denotes any density on that domain. The same hard-sphere flow that moves one configuration moves every such density.
SOURCE FORMULA · WHERE THE INITIAL RANDOMNESS ACTUALLY COMES FROM
Start with independent weights—then forbid overlaps
The companion paper's general-α normalization makes the law concrete. Write n₀(z) for the normalized one-particle profile and 𝒵ε for the partition function.
Activity chooses the scale. The weight αε−(d−1) makes the expected population of order ε−(d−1), so the collision rate can remain finite.
2
The product is the independence template. Without the indicator, the zj carry independent n₀-weights and the sector weights form a Poisson law after normalization.
3
Exclusion creates the first correlation. 𝟙DNε deletes overlapping configurations; 𝒵ε renormalizes what remains. Thus the finite-ε law is almost—but not exactly—factorized from time zero.
HN(t) follows free flights and applies the elastic velocity rule at each regular collision. It is invertible and preserves phase-space volume away from a null exceptional set.
→2 · TRANSPORTED LAW
SN(t)G=G∘HN(t)−1 WN(t)=SN(t)W0,N=W0,N∘HN(t)−1
To find the density at a present state, run that state backward and read the initial density there. Randomness remains in the initial draw; the subsequent map is deterministic.
Integrating all unobserved states forms a factorial correlation; the prefactor makes it order one when 𝔼N·εd−1≈α. There is no 1/s! because the s displayed coordinates remain ordered. This is a joint intensity, not generally a probability density of total mass one.
WHY THERE IS NO JACOBIAN IN STEP 2
Push a density forward, then use volume preservation
WN(t,Z)=W0,N(HN(t)−1Z) · |det DHN(t)−1(Z)| |det DHN(t)|=1 almost everywhere ⇒ WN(t)=W0,N∘HN(t)−1
Free flight(x,v)↦(x+tv,v) is a shear, so its determinant is 1.
Regular impactThe specular collision map preserves the natural boundary-flux measure. This includes the dependence of collision time and contact normal on the incoming state.
Full billiard flowThe standard billiard-flow argument composes free shears and regular boundary maps to give |det DHN(t)|=1 almost everywhere. Grazing and simultaneous multiple collisions form a null exceptional set.
Standard derivation. This Jacobian calculation explains the paper's transported-law formula; it is not a separate verbatim display from the paper.
DERIVED STANDARD DIFFERENTIAL FORM
∂tWN+∑j=1Nvj·∇xjWN=0 in int(DNε)
Between collisions this Liouville equation is the differential form of WN(t)=W0,N∘HN(t)−1. At a regular contact boundary, the incoming and outgoing traces have equal density after their velocities are related by the elastic collision map: this is the specular trace condition. The displayed PDE and trace wording are a standard derivation from the paper’s flow formulas, not a verbatim equation printed there.
The sequence (WN/N!) gives the sector densities of the random-N grand-canonical law. Each N-sector lives on a 2dN-dimensional phase space.
02
s-particle correlation
fs(t,z₁,…,zs)
A rescaled factorial-moment density built from the grand-canonical sequence (WN). It plays the role of an s-particle joint intensity, but is not generally a normalized tagged-particle marginal.
03
Propagation of chaos
fsε(t) ≈ f(t)⊗s𝟙Ds
At finite ε, exclusion forbids overlap. For fixed s, the appropriately scaled correlation approaches the factorized density on the allowed domain as ε → 0.
WORKED EXAMPLE · WHAT “INTEGRATE OUT” MEANS
Keep one particle; forget the other two
First use a fixed, labelled three-particle system. Let P₃(z₁,z₂,z₃) be a normalized joint probability density, so integrating over all three states gives 1. If we observe only particle 1, every possible state of particles 2 and 3 must be added up:
These tagged marginals are still normalized to 1. They answer probability questions about specified labels, such as “where is particle 1?”
WHY A FACTORIAL CORRELATION IS DIFFERENT
3choices of one particle→3×2=6ordered choices of two distinct particles
∫ρ₁(z₁)dz₁ = 𝔼N ∫ρ₂(z₁,z₂)dz₁dz₂ = 𝔼[N(N−1)]
A factorial-moment density ρs counts ordered distinct s-tuples, rather than assigning total probability 1 to fixed labels. Here ρs is temporary notation for this example: the unsubscripted ρ used later denotes molecule complexity, while ρ(t,x) in the fluid chapter denotes mass density. In the paper, fsε is a rescaled version adapted to Boltzmann–Grad scaling, so its exact normalization follows that convention.
Carry forward. A marginal forgets variables. A factorial correlation also forgets variables, but retains how many distinct tuples are available. Chaos asks whether its rescaled joint intensity approaches a product.
f2(z₁,z₂)≈f(z₁) f(z₂)
THE CRUCIAL WORD: “CHAOS”
It means independence, not disorder.
If knowing particle 1’s state gives asymptotically no information about particle 2’s state, their limiting correlation factorizes. Collisions do create dependence; the theorem controls the aggregate connected contribution of all relevant histories through estimates and cancellations.
Why factorization closes the equation
The exact evolution equation for f₁ depends on f₂. If f₂ ≈ f₁f₁, that unknown two-particle term becomes a quadratic expression in f₁. That quadratic expression is precisely the Boltzmann collision operator.
VISUAL CHECKPOINT · CHAOS
Equal marginals do not imply independence
Two binary variables X and Y always have 50–50 marginals here. Move only their connected correlation c, then find the one setting where knowing X tells you nothing about Y.
Goal: make P(Y=L | X=L) equal the marginal P(Y=L)=0.500.
Joint probability table
X=L, Y=L0.450
X=L, Y=R0.050
X=R, Y=L0.050
X=R, Y=R0.450
P(Y=L | X=L)
0.900
marginal P(Y=L)
0.500
connected cell remainder
c = +0.200
L¹ factorization error
hidden until checked
The four joint probabilities and factorization diagnostics update with the correlation slider.
Licensed inference: equal one-variable marginals do not imply independence; in this binary model, c=0 is exactly the product law. This table does not prove propagation of chaos for hard spheres. There the theorem controls the bulk factorization defect in an integrated norm (with 𝟙D₂), and collision histories separately control its contact-flux contribution.
Grand-canonical and factorial correlation—in plain language
Grand-canonical means the particle count N is itself random, concentrated around a large mean 𝔼N; this simplifies cluster bookkeeping. The function fs counts expected ordered s-tuples near (z₁,…,zs), so it need not integrate to 1. For a fixed-N normalized marginal, the analogous statement is the joint law of s tagged particles. Factorization means f⊗s(z₁,…,zs)=f(z₁)···f(zs).
04 · THE MESOSCOPIC LAW
Read the Boltzmann equation like a balance sheet
Left: particles move through space. Right: collisions move probability into and out of velocity v.
DESTINATION PREVIEW · NOT YET DERIVED
This opening display is the equation we want to reach, not a consequence established yet. Read its symbols once; the visible rewind below then returns to the exact finite-ε gas.
Read the targettransport = gain − loss
Choose the scaleNεd−1 ≍ 1 keeps collisions visible
Recover the exact ladderBBGKY couples f₁ to contact values of f₂
Close only in the limitf₂ε → f⊗f produces Q(f,f)
(∂t + v·∇x) f=Q(f,f)
∫∫
f′ f′*−f f*
[(v − v*)·ω]+
dω dv*
Symbol key. f=f(t,x,v), f*=f(t,x,v*), f′=f(t,x,v′), f′*=f(t,x,v′*). The partner velocity v* runs over ℝd, the impact normal ω over Sd−1, and [a]+=max(a,0). For general Boltzmann–Grad density α, the right side is αQ(f,f). The destination preview immediately above uses the α=1 convention; the later hierarchy and two-limit bridge retain α explicitly.
Transport
Along a free particle path x(t) = x₀ + tv, the operator ∂t + v·∇x is the total rate of change of f.
VISUAL CHECKPOINT · GAIN MINUS LOSS
Does one collision channel add or remove density at v?
For one fixed partner velocity and collision normal, predict the sign of Δpair=f′f′*−ff*. Then reveal the multiplication. (The letter B is reserved below for the nonnegative collision-rate weight.)
Pair-density products
GAIN · f′f′*0.20
LOSS · ff*0.48
NET · gain−losshidden
0.40×0.50 − 0.80×0.60 = −0.28
Bars show the gain-antecedent and current/loss pair products in this operator convention. This is one integrand channel—not the sign of the fully integrated Q(f,f)(v).
DERIVATION B
Why the collision term has “gain minus loss” form
1
Pick the state
Fix position x and velocity v. We ask how collisions change the number of particles near this state.
2
Count arrivals
A pre-collision pair (v′,v′*) can scatter into (v,v*). Under independence its density is f′f′*.
3
Count departures
A present pair (v,v*) can collide and leave velocity v. Under independence its density is ff*.
4
Weight geometry
[(v−v*)·ω]+ is the positive part of the signed normal component of relative velocity. Integrate over partners and impact directions.
INTERACTIVE LAB 2
The Boltzmann–Grad balance
Make each sphere smaller. To keep an order-one collision-frequency scale—and, when tagged and relative speeds are comparable and nonzero, an order-one mean flight length—add enough particles to compensate for the shrinking collision cross-section.
Collision cross-sectionε² = 10⁻⁶
Required particle countN ≈ 10⁶
ProductN ε² ≈ 1
Mean free path order (nondegenerate velocities)ℓ ≍ 1
Occupied volume fractionN ε³ ≈ ε = 10⁻³
Licensed inference: balancing partner count against collision cross-section gives the exponent (𝔼N)εd−1. The dots are capped and subsampled, and this unit-volume heuristic suppresses fixed geometric constants and the partner-velocity average. A tagged mean path also contains the ratio |v|/⟨|v−v*|⟩, so the displayed order assumes comparable, finite, nonzero tagged and relative speeds. It does not prove convergence or compute the theorem’s error. Here ≍ or ≈ means “the same order up to fixed constants.”
Worked 3D numbers with geometric constants
Take ε=10−3 and 𝔼N=ε−2=106 in unit volume. The center-to-center collision cross-section is σ=πε²=π·10−6. If both the tagged speed and the relevant averaged relative speed are normalized to one, the simplified collision frequency is nσ≈π and the tagged mean flight length is ℓ≈1/π≈0.318—still order one.
The collision opportunity stays finite while the volume occupied by matter tends to zero. The theorem absorbs fixed geometric constants into its normalization; this is the physical order-of-magnitude calculation.
DERIVATION C
Why Nεd−1 must stay of order one
1
Sweep a relative-motion collision tube
Fix a tagged velocity v and a partner velocity v*. In time Δt their relative center moves by |v−v*|Δt. A collision occurs when this relative path enters a (d−1)-dimensional cross-section of scale εd−1.
relative swept volume ≍ |v−v*| Δt · εd−1
2
Average over partner velocities and multiply by density
In unit volume the number density is of order 𝔼N. Integrating the relative speed against the partner-velocity law gives a typical relative-speed factor ⟨|v−v*|⟩.
expected collisions ≍ (𝔼N) εd−1 ⟨|v−v*|⟩ Δt
3
Demand a nontrivial prefactor
For velocity laws with a finite, nonzero relative-speed scale, (𝔼N)εd−1→0 suppresses collisions and divergence makes their prefactor singular. Keeping it near α∈(0,∞) preserves an order-one collision-frequency scale. A monokinetic gas is a reminder that α alone does not force collisions: if every relative speed is zero, the actual rate is zero.
𝔼N → ∞, ε → 0, (𝔼N)εd−1 → α
EXACT MECHANISM · DERIVED CONTACT-FLUX FORM · BBGKY
Before taking a limit, one equation never closes
The Bogoliubov–Born–Green–Kirkwood–Yvon hierarchy is the ladder where “independence” must be proved—not inserted by hand. To predict particle 1, we need the chance that a partner is at contact: two-particle information f₂. The f₂ equation similarly needs f₃. The coupling mechanism is exact; the one-line operator display below compresses its boundary geometry.
Derived standard contact-flux form. The differential PDE holds in the interior of the closed non-overlap domain Dsε. Collisions among the s already-observed particles live in specular reflecting trace conditions on ∂Dsε; they are suppressed in this compact line. Cs,s+1ε records contact with one additional partner. Setting α=1 gives the normalized displays used elsewhere.
Integration removes a variable—not the transport we still observe
Suppose a density ρ(t,x,y) satisfies a two-variable conservation law on 0<y<1. We keep x and integrate out y:
∂tρ + ∂xJx + ∂yJy = 0
∂t∫₀¹ρdy + ∂x∫₀¹Jxdy = Jy(t,x,0) − Jy(t,x,1)
KEEP
The x-transport remains because x was not integrated out.
CONVERT
The y-derivative becomes flux through the two y-boundaries by the fundamental theorem of calculus.
TRANSLATE
For hard spheres, z₁ is the retained variable and the relevant boundary is contact, |x₂−x₁|=ε. Its flux depends on the pair density f₂ε.
WORKED BRIDGE · WHY f₁ NEEDS f₂
1
Integrate Liouville’s equation over every variable except z₁=(x₁,v₁). Transport in the integrated-out variables becomes boundary flux; the retained term v₁·∇x₁f₁ε remains on the left.
2
Use the guide's ω=(x₂−x₁)/ε convention. The current/loss trace lies at x₂=x₁+εω. After reflecting velocities and reversing the normal to keep the same positive-part parameterization, the gain-antecedent trace lies at x₂=x₁−εω. For s≥2, collisions among observed particles remain boundary conditions on Dsε.
3
Flux through a surface patch is weighted by B=[(v₁−v₂)·ω]+. With this orientation, B>0 is precisely the incoming/current loss channel.
4
For normalized tagged marginals, contact area εd−1 and the order-𝔼N partner count give the prefactor (𝔼N)εd−1→α. The paper’s rescaled factorial correlations do not remove this parameter: its limiting hierarchy has αCs,s+1 and its kinetic equation has αQ(f,f). Our displayed coefficient 1 is the α=1 convention.
5
Formally, the exact pair density first approaches its finite-ε one-body product, that product approaches f⊗f, and both displaced contact points x₁±εω merge with x₁. The three debts are separated explicitly below. Bulk L¹ convergence alone does not control the first boundary-trace debt; collision-history estimates do.
Exact s=1 contact-flux split in the guide's ω=(x₂−x₁)/ε convention and the paper's rescaled-correlation normalization. Here (v₁′,v₂′) is the gain antecedent obtained by the involutive collision map. The opposite contact shifts are essential. Setting α=1 recovers the normalized displays used elsewhere.
factorization + ε→0 ↓
(∂t+v·∇x)f = Q(f,f)
THE CLOSURE RELAY · THREE DEBTS, NOT ONE SUBSTITUTION
Separate exact finite-ε factorization from the limiting Boltzmann product
Define the finite-ε one-body product and its exact factorization defect by
P2ε:=f1ε⊗f1ε, g2ε:=f2ε−P2ε so exactly f2ε=P2ε+g2ε.
The common factor α is omitted from this ledger. At order s=1, C0(f⊗f)=Q(f,f). Each equality above is algebra; proving that all three terms on the right vanish requires different information.
DEBT 1 · CORRELATION AT CONTACT
Cεg2ε→0
The hard term samples collision-boundary traces. Bulk L¹ chaos does not imply it; recollision and collision-history estimates supply the missing control.
DEBT 2 · ONE-BODY CONVERGENCE
Cε(P2ε−f⊗f)→0
Once f1ε→f in the norms needed by the flux, its tensor product approaches the limiting product. This is not true merely by defining g2ε.
DEBT 3 · CONTACT SHIFT
(Cε−C0)(f⊗f)→0
For smooth f, the mean-value theorem controls each of the two spatial displacements:
|f(x±εω,v*)−f(x,v*)| ≤ ε sup0≤r≤1|∇xf(x±rεω,v*)|.
Theorem/heuristic boundary. The decomposition is exact. The displayed mean-value estimate proves only the smooth spatial-shift mechanism; velocity weights, integration, exclusion, and uniform bounds still belong to the rigorous limiting argument.
f₁needsf₂needsf₃needs…
propagation of chaos
f₂ ≈ f⊗f⇒∂tf + v·∇xf = αQ(f,f)
INTERACTIVE CLOSURE CHECKPOINT
Inspect the first debt: finite-ε correlation at contact
This lab isolates Debt 1 above. Its product is P2ε=f1ε⊗f1ε, not yet the limiting product f⊗f:
f2ε = P2ε + g2ε.
For one gain-antecedent/current-loss channel, normalize the positive contact rate B=[(v−v*)·ω]+ to 1. Change the finite-ε product values and their exact defect below.
The ranges keep f2ε=P2ε+g2ε nonnegative in this toy channel. Negative g2 is allowed: it is a difference from a product, not itself a probability density.
f′2−f2=(P2ε,′−P2ε)+(g′2−g2)
exact gain antecedent f′2=P′+g′2
0.40
exact current f2=P+g2
0.42
Contact-flux ledger, B=1 · bars share a scale and saturate at frame width
finite-ε product channel P2ε,′−P2ε+0.15
factorization-defect term g′2−g2−0.17
exact channel f′2−f2−0.02
(0.45−0.30) + (−0.05−0.12) = −0.02
The finite-ε product contribution is +0.15, the factorization-defect term is −0.17, and their exact sum is −0.02.
INTERACTIVE LOGIC WARNING · BULK ≠ CONTACT TRACE
A function can disappear in the square and explode on its edge
On the unit square, choose the explicit representative
gδ(x,y)=δ−1/2 𝟙{0≤y≤δ}, (x,y)∈[0,1]2.
Its height grows while its support becomes a thinner strip. Use the logarithmic control to compare the two-dimensional bulk mass with the one-dimensional value on the contact edge y=0.
Choose k from 1 through 6 in δ=10−k. The main square uses the true strip thickness; the inset magnifies it vertically so it remains inspectable.
The left strip has its true thickness inside the unit square. The inset enlarges only the vertical direction; all exact values are written below.
strip width δ
10⁻² = 0.01
strip height δ−1/2
10
bulk L¹ norm δ1/2
0.1
contact L¹ value δ−1/2
10
‖gδ‖L¹([0,1]²) =∫₀¹∫₀δδ−1/2dy dx=δ1/2→0
∫₀¹|gδ(x,0)|dx =∫₀¹δ−1/2dx=δ−1/2→∞
At δ=0.01, the bulk L1 norm is 0.1, but the boundary L1 value is 10.
Back to BBGKY. This is why ‖g2ε‖L¹(bulk)→0 alone cannot justify Cεg2ε→0: the collision operator samples g2ε on the gain/loss contact traces x2=x1∓εω.
1
Exact: Liouville’s transport equation for the full N-particle density preserves all microscopic information.
2
Form factorial correlations: integrating out particles makes fs depend on fs+1.
3
Close in three controlled passages: remove the contact defect, pass the finite-ε one-body product to f⊗f, then remove the ε contact shift so C0(f⊗f)=Q(f,f).
05 · WHY 1975 WAS NOT ENOUGH
The short-time proof grows a forest
Lanford expanded a tagged particle’s possible collision histories backward in time. The method was revolutionary—but its absolute series converged only for a small fraction of a mean free time.
THE MISSING ALGEBRAIC BRIDGE
Why iterating BBGKY draws a tree
Cεg2ε(t)=Cε[f2ε(t)−f1ε(t)⊗f1ε(t)]
This is why trees answer the contact-trace question. Set s=2 in the recurrence below to expand f2ε. Expand each f1ε in the product in the same ordered-history language, then subtract. Histories that split into two independent one-root histories form the product part; connected histories remain in g2ε, and applying Cε samples that remainder at root contact. Signed/exclusion and return-contact histories belong to this comparison. Representation and truncation are separate proof remainders introduced when the comparison is organized.
Let Ssε(t) carry s observed hard spheres, including their reflecting collisions, between collision creations, and let Cs,s+1ε attach one new collision partner. Duhamel’s formula turns the differential hierarchy into an integral equation. The limiting kinetic hierarchy instead uses the free-flow operator Ss0; the superscript keeps these two propagators distinct.
two-creation term = α²∫0<t₂<t₁<t Ssε(t−t₁)Cs,s+1ε ×Ss+1ε(t₁−t₂)Cs+1,s+2εfs+2ε(t₂) dt₂dt₁
Read left to right: each Cε adds exactly one partner, while substitution forces the ordered times 0<t₂<t₁<t. Repeating this step produces the parent choices, nodes, and time simplex shown in the tree lab below. Each creation also contributes one factor α; setting α=1 recovers the earlier coefficient-one preview.
Provenance. These displayed semigroup formulas are the standard Duhamel consequence of the derived hierarchy, reconstructed here to expose the algebra. Section 1.3.1 of the paper supplies the microscopic/kinetic expansion comparison and the divergence rationale; it does not print these exact two display lines verbatim.
At each backward collision, one new possible partner is attached to one particle already in the history. Compare that actual structure with a binary branching metaphor.
particles in one history5
possible parent sequences24
ordered-time simplext⁴ / 4!
Licensed inference · tagged case s=1: n creations produce n+1 particles, n! possible parent sequences, and an ordered-time volume tn/n!. The displayed graph is only one selected history; it does not estimate velocity integrals, recollisions, or convergence, and particles do not literally double.
EXACT COMBINATORIAL COUNT · TAGGED CASE s=1
Parent choices and ordered times contain opposite factorials
n creations→1·2···n = n! parent choices×Vol{0<tn<···<t₁<t} = tn/n!→(Ct)n after collision integrals
Where C comes from. In the chosen Gaussian-weighted norm, integrating one newly created partner’s velocity and impact direction against the relative-speed collision kernel costs at most one uniform constant C. Repeating this for n creations costs Cn; the ordered-time factor supplies tn.
starting from order s: s(s+1)···(s+n−1)=(s+n−1)!(s−1)! parent sequences
For the tagged case s=1, the two factorials cancel exactly at the counting-and-time-volume level. For fixed general s, their ratio leaves only a polynomial factor in n. Thus the obstruction is not literal binary population doubling; after the collision integrals are bounded, the schematic absolute series still has only a finite time radius. Time layering makes Cτ<1 locally, while cumulants carry the correlations that a naïve restart would erase.
absolute n-collision bound ≲ (Ct)n
Then ∑n≥0(Ct)n is a geometric series and converges only when Ct<1. Time layering keeps each local width τ in that safe regime; the difficult part is carrying old correlations across the layer boundaries.
1975
Lanford’s expansion
Expand all collision histories from time t back to 0.
Bound the absolute value of every tree contribution.
Sum the resulting series.
Works only while the series stays small
→
2024–25
Deng–Hani–Ma’s iteration
Divide long time into locally manageable layers.
Stop expanding the factorized leading part.
Locally update the leading density; carry multi-layer history only through cumulants.
Reaches every regular Boltzmann time
06 · THE PROOF ENGINE
Keep the signal. Expand the error.
The central innovation is an inductive representation that remembers the entire relevant collision history without expanding the already-understood Boltzmann part.
Model: hard spheres of diameter ε in ℝd, d≥2; the specified grand-canonical initial ensemble, conditioned on non-overlap.
Initial profile: f0 is a normalized nonnegative density, sufficiently regular in x, spatially localized, and bounded by a Gaussian tail in v. The paper packages these requirements in a weighted norm denoted ‖·‖Bol,2β; β and the bound B0 are fixed positive constants.
Effective solution: f solves the hard-sphere Boltzmann equation on [0,T] and retains a uniform Gaussian velocity bound, summarized by a fixed constant A.
Limit: the ensemble has 𝔼N·εd−1≈1 (the α=1 normalization used in this theorem box); first choose T and its solution bounds, then take ε sufficiently small.
fsε is the exact rescaled s-particle factorial correlation of the grand-canonical hard-sphere ensemble. It is a joint intensity, not generally a normalized tagged-particle probability density.
INTERACTIVE LOGIC MAP · THE THEOREM BECOMES FIVE JOBS
Every object introduced next pays one named term.
Do not enter the proof as a list of techniques. Start with the target norm, split it into debts, and ask which representation or estimate removes each debt.
The locally updated signal fA must still approach f.
Layering prevents a divergent global expansion, but it repeatedly updates the leading one-particle profile. Boltzmann stability controls the accumulated difference.
OBJECT · fA→ESTIMATE · tensor telescoping + ‖fA−f‖₁→PAYS · DA
Here [s]={1,…,s}; H is the nonempty subset carrying one defect, and (fA)⊗([s]∖H) means one factor fA for every remaining root.
Schematic decomposition. The exact paper notation is reconciled in “How the layer estimates add up” below. This relay exists to show why fA, EH, Errcumul, ftrunc, and the exclusion indicator all appear before their detailed estimates.
THE INVARIANT CARRIED FROM LAYER TO LAYER
Begin with two roots; generalize only after every term has a job.
Here “invariant” means a representation preserved by induction, not a conserved physical quantity. Superscript A is a label for the chosen leading profile, not a numerical power.
K={1,2}
the set of particle labels, called roots, observed at the current layer boundary
H⊆K
the subset of roots assigned to one represented defect; K∖H stays in the leading product
f̃Kε
the correlation represented by the admissible partial expansion, before the separately truncated rare histories are restored
fA
the locally updated one-particle leading state, chosen because it follows a stable Boltzmann evolution
E{1}, E{2}
singleton or one-root defects: deviations of one represented root from fA, not correlations between two roots
E{1,2}
the genuinely connected two-root term: the part not reconstructed from the product and singleton defects
ErrKcumul
the error in the cumulant representation itself; it is carried and estimated, not silently absorbed into EH
fKtrunc
the separate error from rare histories removed by the truncated dynamics
Why the last line is separate. Errcumul measures accuracy of the representation for the truncated dynamics; ftrunc measures the histories removed when that dynamics was truncated. They require different estimates.
WORKED RETRIEVAL · SOLVE FOR THE CONNECTED MEMORY
The cumulant is a remainder with a precise subtraction rule
At one fixed pair of phase-space points, suppose the represented two-root value is 0.37. Predict E{1,2} before revealing the subtraction.
Givens in reading order: total 0.37; leading +0.30; singleton contributions −0.03 and +0.05; representation error +0.02.
represented total f̃120.37
leading f₁Af₂A0.60×0.50 = +0.30
singleton f₁AE{2}0.60×(−0.05) = −0.03
singleton f₂AE{1}0.50×0.10 = +0.05
representation error+0.02
connected E{1,2}?
E{1,2}=0.37−0.30−(−0.03)−0.05−0.02=0.03
Interpretation. A singleton defect may be signed. “Connected” does not mean “whatever is positive”; it means exactly what remains after the leading, both singleton, and representation-error pieces are removed.
TWO ROOTS→ARBITRARY FINITE ROOT SET K
GENERAL SUBSET RULE
Choose every possible defect set H⊆K exactly once
Roots outside H contribute independent copies of fA; roots in H are carried together inside EH. The empty choice H=∅ is the fully factorized leading term because E∅=1.
Once the representation-error family is fixed, all EH with H⊊K are already known when EK is reached. Subtract the leading product and every proper-subset contribution; the remainder is the unique K-root defect.
EKε=f̃Kε−∑H⊊K(fA)⊗(K∖H)EHε−ErrKcumul K={1,2,3}: subtract the leading triple, three singleton terms, three pair terms, and Err123cumul; what remains is E123.
The method neither forgets correlation nor re-expands everything
At a layer boundary, the full represented state contains the leading product, two singleton defects, one genuinely connected defect, and two separately estimated error families. Which operation preserves this information without rebuilding the finite-radius forest?
Licensed inference and limit. This checkpoint now displays every term in the two-root representation. It tests the architecture of the next operation; it does not execute the molecule estimates, the representation-error estimate, or the separate truncation estimate.
Discard
Erasing E{1}, E{2}, and E{1,2} assumes the factorization conclusion at every boundary; erasing the two error families also drops obligations rather than estimating them.
Carry
Update fA; partially expand each represented EH, including singletons; keep Errcumul and ftrunc in their own estimates.
Re-expand
This is algebraically legal. It is analytically unusable because expanding the already-understood product through all prior layers regenerates the finite-radius forest.
δℓ+1 ≤ (1+Cτ)δℓ + Cεν
δL ≲ eCT(δ₀+LCεν) → 0, L=T/τ fixed before ε→0
This scalar recurrence is a pedagogical stability analogy, not the paper’s estimate. The proof replaces δℓ by weighted cumulants and molecule integrals with detailed support conditions.
WHY EACH TOOL IS FORCED
Every proof device pays one debt—and creates the next precise question.
Original pedagogical matrix. Read each row from left to right; the final cell becomes the obstruction in the following row.
[M]
one collision-diagram type: internal particle names may change, but the root labels and graph structure stay fixed
m=|M|
the number of collision or overlap events in that molecule
ρ
the molecule’s multi-layer complexity: carried lines plus circuit ranks
𝓜m,ρ
the family of molecule types with event count m and complexity ρ
#𝓜
the number of members in that family; # is a counting symbol
INM, ν>0
the history contribution for M, and the positive geometric-gain exponent
Graph proliferationHistory types grow with size and complexity.
Molecules + countingEncode histories and truncate rare large clusters.
Controlled count#𝓜m,ρ≤Cm|log ε|Cρ.
Per-history sizeCounting alone gives no smallness.
Recollision geometryCycles constrain variables on thin sets.
Cutting / Fubini orderExpose the constrained variable before summing.
Geometric gainsup[M]∈𝓜m,ρ‖INM‖≤τmενρ.
Leading-state driftfA must still match Boltzmann f.
Leading-state driftLocal updates accumulate across L layers.
Signal identification + stabilityMatch no-recollision trees and compare nearby Boltzmann data.
Stable leading state‖fA−f‖1→0.
Final summationCombine every nonempty H and both error families.
STEP 1 OF 8intuition
Step 1 of 8: Replace one impossible interval by L small ones. Level: intuition.
Replace one impossible interval by L small ones
Choose a short layer width τ = T/L. Local cluster expansions remain convergent inside each layer, even though a single expansion across [0,T] would diverge.
INPUT
A regular Boltzmann solution on [0,T]
OUTPUT
Short intervals [(ℓ−1)τ, ℓτ]
RISK
Correlations from earlier layers cannot be forgotten
IDENTIFICATION STEP · WHY THE MAIN TERMS ARE BOLTZMANN
No-recollision histories reproduce the kinetic Duhamel series
f(t) = S10(t)f₀ + α∫₀t S10(t−r)Q(f,f)(r) dr S10(t)g(x,v)=g(x−tv,v)
Substituting this integral equation into itself attaches an earlier collision partner and builds a tree, with one factor α per collision just as in the BBGKY series above. A microscopic history with no recollision has the same four limiting ingredients: free flight, the elastic collision map, relative-speed weighting, and a genuinely new particle label. That label is not literally sampled independently inside one deterministic trajectory. Only after averaging over the initially chaotic ensemble and taking ε→0 does its weight asymptotically factorize as a new factor f.
microscopic no-recollision history
term by term →
Boltzmann Duhamel term
+ one return-contact edgegeometry-controlled channel
This extra edge depicts only one error channel: a return contact with a label already in the history. It is controlled by geometry and cutting. Other non-leading terms need different tools—signed cluster/exclusion terms can cancel by an involution, while oversized histories are truncated and estimated separately.
THE COMBINATORIAL HEART
A collision history becomes a “molecule”
Two graph languages appear below. In the teaching graph, vertices are particle labels and edges are encounters. In Deng’s molecule, atoms are collision or overlap events, while bonds and serial colored paths carry particle trajectories. The translator makes the distinction before either graph is used for counting.
INTERACTIVE TRANSLATOR · ONE HISTORY, TWO GRAPH LANGUAGES
First ask: what do the dots mean?
The episode “A meets B, B meets C, then A meets C” can be drawn in two useful ways. Select a convention and watch the same episode be re-encoded. The drawings are related, but their vertices and edges are not the same objects.
A pathB pathC pathreturn / cycle cue
CONVENTION A · PEDAGOGICAL
Dots are particles; lines are encounters.
Here V counts particle labels and E counts recorded pair encounters. This makes “new partner” versus “already present” immediately visible, so it is the right picture for the elementary circuit-rank calculation below.
Do not rename this a molecule. It is an intuition graph created for this guide.
CONVENTION A ONLY
Derive the circuit rank instead of memorizing it
rpg = E − V + C
For a graph with E encounter edges, V particle-label vertices, and C connected components, rpg counts independent circuits. In the connected history below C=1 throughout.
0
Begin with A. E=0, V=1, C=1, so rpg=0−1+1=0.
1
Attach new partner B. ΔV=1, ΔE=1, ΔC=0. Therefore Δrpg=1−1+0=0.
2
Attach new partner C. Again ΔV=1 and ΔE=1, so Δrpg=0. The A–B–C path is still a tree.
3
Meet within the existing connected history. A and C are already labels, so ΔV=0, ΔE=1, ΔC=0 and Δrpg=1−0+0=+1.
The last rule assumes the two old labels already lie in the same connected component. Joining two previously separate components would instead have ΔC=−1 and would not create a circuit.
TRANSLATION, NOT IDENTIFICATION
From one teaching cycle to Deng’s multi-layer complexity
The particle graph suggests why a return encounter removes a “new-label freedom.” Deng’s proof does not insert rpg into its estimates. It first redraws the history as event atoms joined by particle-line bonds.
MAIN PAPER · DEFINITION 7.1
ρ = ∑ℓ′=0ℓ sℓ′ + ∑ℓ′=1ℓ Rℓ′
Within layer ℓ′: Rℓ′ is the circuit rank of that layer’s atom–bond molecule graph: bonds minus atoms plus components.
Across layers: sℓ′=|r(Mℓ′)| counts particle lines crossing into layer ℓ′; s₀ counts particles involved in the initial links.
Therefore: ρ charges both within-layer cycles and memory carried between layers. Although both circuit ranks use E−V+C, their E and V refer to different objects, so rpg and ρ are not numerically identical by redrawing.
FOUR-WAY LEDGER · MATCH EACH HISTORY TO ITS TOOL
“Not leading” does not mean “one extra edge”
The proof first classifies what kind of term it has. Only then does it choose cancellation, geometry, or truncation.
1 · LEADING
New-partner trees
Each backward creation brings a fresh label. After averaging and ε→0, these histories match the Boltzmann Duhamel series term by term.
Destination: identify the signal2 · SIGNED
Exclusion / cluster terms
Alternating signs encode forbidden overlaps or cluster bookkeeping. The paper pairs appropriate terms through an involution before taking absolute values.
Tool: cancellation3 · GEOMETRIC
Unintended or return contacts
A prescribed trajectory must hit another ε-scale target. Support restrictions and the cutting order expose a small geometric set.
Tool: geometry + cutting4 · TOO LARGE
Oversized / highly recollisional histories
Clusters beyond the local size cutoff, or histories beyond the recollision cutoff, are removed from the main expansion and bounded as truncation errors.
Tool: truncation estimates
No one-edge dictionary: a cumulant records dependence, and an error term records a proof remainder. Neither definition says “exactly one recollision edge.” A term may cancel by sign, contain several return contacts, or collect an entire truncated family.
WHY EXCLUSION CREATES SIGNS · SOURCE IDENTITY + TOY CHECK
Pair first; take absolute values afterward
A cluster is compatible precisely when two pieces do not overlap. The elementary indicator identity expands every compatibility condition into a positive “do nothing” term and a negative overlap term:
An O-atom is bookkeeping for an overlap restriction; it is not a physical collision event. Later, for the particular non-leading one-root tree family 𝓜n∖𝓜n*, the paper toggles one C/O atom and proves equality after integrating the toggled top variables, while the O-count changes parity. This cancels that selected signed family—not every history containing an O-atom.
𝓘𝕄twi(Q𝕄twi)=𝓘𝕄(Q𝕄), |𝕄twi|O=|𝕄|O±1
INM=+0.08paired withINιM=−0.08
|0.08+(−0.08)|=0, whereas |0.08|+|−0.08|=0.16
Toy arithmetic, exact lesson. The numbers are pedagogical. The research involution has to prove that paired terms really have matching domains and weights; the identity above explains where their alternating signs originate.
EXACT SOURCE IDENTITY · WHAT “CUT” MEANS ANALYTICALLY
The inner fragment is integrated with its cut ends frozen
𝓘𝕄=𝓘𝕄₁∘𝓘𝕄₂
1
Cut a bond and make one endpoint a free output of the outer fragment 𝕄₁.
2
The matching endpoint is a fixed input of the inner fragment 𝕄₂. Hold that boundary variable frozen while integrating 𝕄₂.
3
The result 𝓘𝕄₂(Q) is still a function of the frozen end. Feed that function into the outer integral 𝓘𝕄₁.
Why the order matters: a useful inner integral exposes an ε-thin permitted set uniformly in the variables that remain frozen. An arbitrary cut may hide that restriction or spend an integration freedom too early.
Original pedagogical reconstruction. It reproduces the free/fixed-end logic, not the artwork of a paper figure.
WORKED BRIDGE · PARTICLE-GRAPH CONVENTION A · NOT LITERALLY A PAPER MOLECULE
A–B, then B–C, then A–C
Here the letters A, B, C are vertices and encounters are edges. Use this picture to understand “new partner” versus “return contact”; use convention B above when reading the paper’s atom-and-particle-line molecule definitions.
1
Backward event A–B introduces B as a new partner of A.
2
Event B–C introduces C. The two edges form a tree, so each event still brings a new particle label.
3
Demanding a later A–C encounter introduces no new label: A and C are already positioned by the earlier history. Their relative trajectory must now hit an ε-sized target. The added edge closes the triangle.
For each fixed x, a good Fubini order exposes only an O(ε)-length interval of allowed y-values. This strip is an analogy, not the hard-sphere estimate itself: real recollisions also involve times, velocities, grazing configurations, and cross-layer degeneracies, so the proof extracts the weaker robust gain ενρ.
ε
GEOMETRIC MODEL · WHERE SMALLNESS COMES FROM
A second prescribed encounter is a thin target
Fix one trajectory during a layer of length τ. Another trajectory hits its ε-neighborhood only if the transverse impact parameter lies in a (d−1)-dimensional disk of radius O(ε):
allowed transverse measure = O(εd−1).
An ordinary first collision uses the freedom already balanced by (𝔼N)εd−1≈1. A further separately prescribed encounter imposes an extra constraint. The complete proof earns a weaker but robust net ενρ, because grazing, nearly parallel, degenerate, and cross-layer configurations must also be covered.
INTERACTIVE LAB 4
Cut a complex molecule
Choose cuts in a legal order. Each cut is a Fubini decision: which variables will be integrated first, uniformly in the variables that remain fixed?
Normal {3}0
Good {33}0
Bad {4}0
Net ε-gainpending
Start at the top to respect time ordering.
Licensed inference: cutting is a choice of integration order that can expose a thin-set restriction. This toy displays one nondegenerate support configuration in which the shown {33} is good; it neither certifies a piece from topology alone nor executes the paper’s full cutting budget.
atom particle line cycle constraint chosen cut
THE RETURN PATH · FROM GRAPH INTEGRALS BACK TO THE THEOREM
A molecule estimate is useful because molecules control EH.
After the paper’s decompositions, cancellations, and support restrictions, the molecule representation gives the pointwise domination shown at right. Representation and truncation remainders are handled separately.
|EHε(ℓτ)| ≤ ∑m,ρ ∑[M]∈𝓜(H;m,ρ) |INM|
‖EHε‖₁ ≤ ∑m,ρ #𝓜(H;m,ρ) · sup[M]‖INM‖₁
Here m is the event count, ρ is molecule complexity, and 𝓜(H;m,ρ) is the family rooted at H with those indices. The brackets [M] mean one diagram type: arbitrary names on internal, non-root particle lines are ignored, while the observed/root labels stay fixed. INM is its collision-history contribution.
CountHow many admissible molecule types can contribute?
CostHow large can the integral attached to one molecule be?
SumDoes geometric ε-smallness beat the graph count and make EH small?
ReturnInsert that EH bound into DE in the target-error ledger.
Exact versus grouped. The first display is the paper’s pointwise molecule bound, regrouped here by event count m and complexity ρ. The second line is the teaching-level count × worst-cost consequence after taking L¹; Equation (2.1), the cancellation steps, and the separate remainder propositions supply the full bookkeeping.
Overview-level normalized estimates. Let m=|M| be one molecule’s event/atom count, 𝓜m,ρ the family of molecules with size m and complexity ρ, and INM the collision-history contribution associated with one M. The constant C is independent of ε once the time horizon and data bounds are fixed. These formulas mirror the paper’s introductory reduction; the detailed proof includes root-normalization factors, refined τ powers, and Proposition 9.7’s full cutting budget before Proposition 6.2 emerges.
COUNT
#𝓜m,ρ ≤ Cm|log ε|Cρ
Each unit of multi-layer complexity incurs a logarithmic counting loss.
×
COST
sup[M]∈𝓜m,ρ‖INM‖ ≤ τm ενρ
The full cutting/root budget combines good thin-set restrictions, compensates bad pieces, and earns a net ενρ. It is not literally one gain per graph cycle.
→
NET
|log ε|Cρ ενρ → 0
Any fixed positive power of ε beats every logarithmic power.
First sum molecule sizes. Choose τ so Cτ<1; then ∑m≥0(Cτ)m converges. What remains per unit of complexity is |log ε|Cεν→0. Here ν>0 and c>0 below are fixed small constants depending on dimension; ν is not a velocity.
Repeated L’Hôpital (or the exponential series) shows yC/eνy → 0. Raising this to ρ preserves the gain.
Research-level precision: two different cutting-budget statements
MOTIVATION · EQ. (2.13)
υ · #good + (d−1)(|H| − #bad) ≥ cρ
This is the introduction’s schematic target: enough good pieces and root normalization should pay for bad pieces and a positive fraction of complexity.
DETAILED RESULT · PROPOSITION 9.7, EQ. (9.51)
υ2 · #good + (d−1)(|H| − #bad) ≥ (C13*)−1ρ
The proved proposition uses υ/2 to absorb logarithmic losses, specifies the operation sequence and subcase count, and adds −C*|H| on the right for the stated error-molecule class.
Do not conflate them. Eq. (2.13) motivates the algorithm; Proposition 9.7 and Eq. (9.51) are the detailed statement with hypotheses and extra bookkeeping. The paper fixes υ=3−d−1; this is distinct from the site’s final schematic gain exponent ν.
THE MISSING PACKAGING ARROW · OVERVIEW GAIN → THE PROVED CUMULANT EXPONENT
ενρ is the engine; εc₀+a|H| is the delivered estimate
PER TYPEcount × cost
(Cτ)m|log ε|Cρενρ
Cutting supplies complexity-dependent smallness.
→SUMall m and ρ
Cτ<1; εν|log ε|C→0
Short layers pay size; ε-powers pay complexity.
→RESTOREroot normalization + detailed budgets
Prop. 9.7 + local integral bounds + molecule counting
The omitted root factors and fixed exponent slack now matter.
→DELIVERProposition 6.2
‖EH‖₁≤εc₀+a|H|
One common gain c₀ plus an additional gain per observed root.
Do not identify the two exponents term by term. Proposition 6.2 follows after velocity-support decomposition, volume and cross-section bounds, root normalization, cutting, Proposition 9.7’s combinatorial budget, and summation over molecule size and rank. The signed leading-tree cancellation belongs instead to the separate fA approximation in Proposition 6.1.
VISUAL CHECKPOINT · ASYMPTOTIC BUDGET
When does the ε-power actually win?
Explore the explicitly pedagogical bound Rρ(ε)=(|log ε|Aεν)ρ, with A=1 and ν=0.15. These are chosen for visibility, not claimed constants from the paper.
At ε=10⁻³ and complexity ρ=3, log₁₀Rρ=+1.17.
Base-10 exponent ledger
smaller bound ←larger bound →
log-count loss+2.52
ε-power gain−1.35
log₁₀ Rρ+1.17
3[log₁₀(3 log 10) − 0.15·3] = +1.17
Positive log₁₀R means this illustrative numerical bound exceeds one; negative means it is small. The theorem is asymptotic and does not claim practical accuracy at moderate ε.
NORMAL
{2}
Two fixed ends determine the remaining collision variables. No power gained or lost.
NORMAL
{3}
One free incoming trajectory supplies exactly the ordinary collision freedom.
GOOD
{33}
A second prescribed encounter constrains a free trajectory to a thin tube: a positive ε-power gain.
BAD
{4}
No end is fixed, so an integration degree is wasted. The algorithms ensure good pieces compensate these losses.
Research-level precision: topology alone does not determine “good”
The four cards are the intuition used in the exposé. In the paper, “good” depends on support restrictions created during splitting, not topology alone: a {3} or {33A} piece may be normal or good; {33B} supplies important good cases; a {4} is counted as bad but can also carry a good restriction; and good {44} pieces handle strongly degenerate configurations.
A
Boltzmann stability keeps fA(ℓτ) close to f(ℓτ).
B
Cutting/root gains make every nonempty defect/cumulant EH, H≠∅, small after the full budget is summed.
C
The two remainder mechanisms stay separate: Proposition 6.2 controls Err1, and recurrence (3.19) propagates Err2, making the representation remainder Err=Err1+Err2 negligible. Proposition 6.3 separately controls the truncated-dynamics remainder fserr.
∴
Induction closes for ℓ = 1,…,L.
THE MISSING LAST DERIVATION
How the layer estimates actually add up to the theorem
The paper proves the hard estimates. This box performs the final bookkeeping: nonempty subsets, drift of the leading tensor, and the two error families.
1
Sum every nonempty defect set
The detailed representation (2.1) contributes one term for each ∅≠H⊆[s]. Proposition 6.2 gives the proved bound ‖EH‖1≤εc₀+a|H|. Group subsets by h=|H|:
What happened: exponentially many subsets were not ignored; the binomial theorem summed all of them at once, and the gain per root defeated their number in the permitted range of s.
s=3 WORKED CHECK · SEE THE BINOMIAL THEOREM DO THE COUNTING
How many nonempty defect sets are there?
3 singletons{1}, {2}, {3}3A²x
3 pairs{1,2}, {1,3}, {2,3}3Ax²
1 triple{1,2,3}x³
A:=1+ηε, x:=εa 3A²x+3Ax²+x³=(A+x)³−A³
The eighth subset is ∅. It gives the leading A³ term and is subtracted, leaving exactly seven nonempty defect terms.
2
Replace the leading profile fA by the Boltzmann solution f
The difference of two s-fold products telescopes one factor at a time:
Multiplication by 𝟙Ds can only decrease an L¹ norm. The factor (1+ηε)s−1 is harmless because |log ε|ηε→0; no false mass-one assumption on fA is needed.
3
Restore what the clean cumulant formula suppressed
Err=Err1+Err2representation remainder in the detailed cumulant formula
fserrseparate remainder from truncating the dynamics
Exclsεhard-core conditioning / excluded-set correction when an intermediate product was not already restricted to Ds
Proposition 6.2 controls Err1; recurrence (3.19) propagates Err2 across the layers, and Section 6 combines them to make Err negligible. Proposition 6.3 separately controls fserr. If every comparison is kept on Ds, the exclusion correction is already encoded by 𝟙Ds; if one passes through an unrestricted product, a union bound gives the schematic Exclsε≲Cs²εd under a bounded spatial-density hypothesis.
Schematic consequence, not a verbatim paper inequality. The displayed sum exposes the final logic. The paper’s weighted estimates, constants, and exponent choices in Section 6 and the later proofs supply the quantitative margins used next.
Set δ=α−1. For the branch-specific well-prepared families, under the regularity and uniform estimates required by the cited hydrodynamic results, the leading-order collision balance is Q(F0,F0)=0. Under the stated positivity and nondegeneracy hypotheses, entropy equality identifies that equilibrium as Maxwellian; solvability then determines which fluid equation its parameters obey.
ORDER-OF-LIMITS CHECKPOINT
Two small parameters, two different approximations
STAGE 1 · PARTICLES → KINETIC
ε→0 with α=(𝔼N)εd−1 fixed
For each fixed finite α, the many-sphere correlations converge to a Boltzmann solution nα satisfying (∂t+v·∇x)nα=αQ(nα,nα).
THEN →STAGE 2 · KINETIC → FLUID
α→∞ ⇔ Kn=α−1→0
Only after the kinetic description is controlled does this second limit begin. It forks: NSF uses the slow clock τ=δt and O(δ) fluctuations around a global Maxwellian, whereas Euler uses t=O(1) and an order-one local Maxwellian.
THE LOGICAL QUANTIFIERS
fsε,δ ⟶ε→0, δ fixed (fδ)⊗s
NSF: 𝒪δNSF(fδ)(τ):=[δ−1(∫fδ(τ/δ)dv−1), δ−1∫vfδ(τ/δ)dv] → (ρ,u), while raw fδ→μ
Euler: fδ(t,x,v)→M[ρ,u,T](t,x,v), so its ordinary Maxwellian moments → (ρ,u,T)
For every fixed δ>0, the inner particle-to-Boltzmann limit is completed first and produces a one-particle distribution. Only then does the outer hydrodynamic limit act through a branch-specific observation map 𝒪δ. A distribution, an s-particle tensor, and a tuple of fields are different object types; the arrows above never equate them.
A simultaneous diagonal choice δ=δ(ε) is a different statement and must satisfy the companion paper's additional joint condition (1.24).
Notation bridge. In this fluid section, δ:=α−1 and fδ:=n1/δ; thus the density has only been renamed after reparametrizing the collision rate. This δ is local to the fluid limit—not the strip width in the earlier trace toy and not a Dirac delta.
Quick check: may we replace both arrows by “ε→0”?
No. ε measures particle diameter in the Newton→Boltzmann limit; Kn=α−1 measures mean-free-path scale in the Boltzmann→fluid limit. For a coordinated limit, condition (1.24) requires max(1,δ−1)max(1,δ−1Tfin)≪(log|log ε|)1/2. The iterated limit needs no silent identification of the parameters.
Optional Stage-1 torus extension · why periodic returns and long bonds need extra geometry
COMPANION-PAPER INNOVATION · ORIGINAL RECONSTRUCTION
On a torus, unfold a return into a lattice displacement
The main long-time theorem is Euclidean. To reach periodic fluid limits, the companion paper must control trajectories that wrap around 𝕋d and may meet a periodic copy of the same ε-target.
Original pedagogical schematic. It reconstructs the unfolding idea, not a paper figure and not a simulation of the full recollision estimate.
1
Lift to ℝd. Replace one periodic cell by all of its lattice translates. A wraparound contact becomes an ordinary contact with a translated target.
2
Repeated-pair obstruction. Two contacts of the same relative trajectory force its velocity near one nonzero lattice direction. That thinness controls a periodic double bond.
3
Unbounded-count obstruction. If too many collisions occur in a fixed interval, some adjacent collision times are separated by order one. The proof records that segment as a long bond and extracts a separate {33A} gain.
nearby collision timesLONG BONDwell-separated adjacent times
For d=2,3, the long-bond {33A} molecule gains ≲εd−1−ϑ for every ϑ>0. The lattice-direction estimate and long-bond argument solve different torus-specific problems; neither is the general Euclidean mechanism of the main theorem.
Collisions act only on the velocity variable at a fixed position. The argument below is therefore local in x; its coefficients may still vary with time and position.
1
Derived conditional balance: fast scale → leading collision equilibrium
If the transport side stays controlled under the branch's prepared-data estimates, a smooth leading profile F0 must satisfy Q(F0,F0)=0. Fast collisions alone do not justify this without those bounds; an initial layer or loss of compactness could otherwise intervene.
2
Zero collision term → zero local entropy production
Here A=F0(v), A*=F0(v*), and primes denote the post-collision velocities. Earlier, the gain formula called the same primed pair an antecedent because it read this involution backward at fixed v. Detailed balance is symmetric between the two readings. Under suitable positivity and regularity hypotheses, Q(F0,F0)=0 gives Dx(F0)=0.
3
Equality → detailed balance
(a−c)log(a/c)≥0, Dx=0 ⇒ F0(v)F0(v*)=F0(v′)F0(v′*)
Taking φ(v)=log F0(v) turns the product identity into an additive one: φ(v)+φ(v*)=φ(v′)+φ(v′*). Thus φ is a collision invariant.
An elastic collision fixes V and |w| while rotating w, so FV(w) depends only on |w|. The Hessian of a radial C² function at the origin commutes with every rotation; therefore Dw²FV(0), and hence D²φ(V), is a scalar multiple λ(V)I. In d≥2, the zero mixed derivatives make each diagonal derivative depend on only one coordinate; equality of all diagonal derivatives then forces λ to be constant. Integrating twice gives
φ(v)=a+b·v+c|v|².
This proof sketch assumes sufficient C² regularity. Measurable classification theorems reach the same collision-invariant family under their own nondegeneracy hypotheses.
5
Integrability + completing the square → Maxwellian
Rotate one elastic collision and test an additive quantity
Fix the center-of-mass velocity V=(1,0) and half-relative velocity w=(½,0). The incoming pair is v=V+w=(3/2,0) and v*=V−w=(1/2,0). Rotate w by θ to obtain the outgoing pair v′=V+Rθw and v′*=V−Rθw.
Why every displayed rotation is an elastic collision. For θ∈(0,180°), choose ω=(w−Rθw)/|w−Rθw|. Then w−2(w·ω)ω=Rθw. At θ=0 the map is the grazing/identity case; at θ=180° it swaps the two velocities.
Licensed inference. A nonzero Δφ is a counterexample and proves that candidate is not invariant. Several zero tests do not prove the full classification; Step 4 above supplies the C² argument.
incoming pair
v=(1.500,0.000), v*=(0.500,0.000)
outgoing pair
at θ=60°: v′=(1.250,0.433), v′*=(0.750,−0.433)
exact numerical test
Δφ=0.000000
Δφ=φ(v)+φ(v*)−φ(v′)−φ(v′*); Δ|v|⁴=2sin²θ
For φ=1, the pair sum is 2 before and after every elastic collision.
One-dimensional Maxwellian on velocity minus eight to eight, centered at 0.7, with temperature 1.3, full-line area 1.0.
mass density—the full-line area under the velocity bell
u
bulk velocity—the bell’s center
T
temperature—the bell’s width
∫M dv=ρ, ρ−1∫vM dv=u, (dρ)−1∫|v−u|²M dv=T
M₁(v)=ρ(2πT)−1/2e−(v−u)²/(2T), v∈ℝ
So the three fields really are moments of one velocity bell. The canvas graphs this exact one-dimensional marginal on the fixed window −8≤v≤8; its full-line area is exactly ρ, while the finite viewport clips only the far tails. Particle mass and Boltzmann’s constant are normalized to 1.
THE SHARED BRIDGE · EQUILIBRIUM GIVES SHAPE; INVARIANTS GIVE EVOLUTION
Multiply, integrate, and watch the collision term disappear for a precise reason.
It is not true that Q(f,f)=0 for every f. What is true is that its three conserved moments vanish for every sufficiently integrable f.
The bracket vanishes pointwise for ψ=1, v, and |v|²/2 because one elastic collision preserves pair mass, momentum, and energy. Therefore the tested moment of Q is zero even when Q(f,f) itself is not.
Source status. This is the standard symmetrized weak form derived from companion Eq. (1.15), with its positive-part kernel convention; it is not a verbatim statement of Propositions 1.6 or 1.8.
These are three parallel tests of the same kinetic equation—not successive operations.
TEST WITH 1
∫f dv = ρ
Counts mass. Its flux is ∫vf dv.
ANDTEST WITH v
∫vf dv = ρu
Counts momentum. Its flux is the stress ∫v⊗vf dv.
ANDTEST WITH |v|²/2
∫½|v|²f dv = E
Counts energy. Its flux is ∫½|v|²vf dv.
ZEROTH-ORDER CLOSURE
Insert fδ≈M[ρ,u,T]. Gaussian moments express every flux using ρ, u, and T, giving compressible Euler.
ORFIRST NON-EQUILIBRIUM CORRECTION
Near the global μ, solve for the next component. Its stress and heat flux produce viscosity and thermal diffusion under the NSF scaling.
The first line is the Taylor tangent to the Maxwellian family at (ρ,u,T)=(1,0,1). The Boussinesq relation ϑ=−r turns its density-and-temperature part into exactly the coefficient displayed in g; the site then writes r as ρ.
RAW VIEWfδ→μ, so its density tends 1 and mean velocity tends 0.
× δ−1
MAGNIFIED VIEWFor the prepared Boltzmann ansatz, δ−1(∫fδdv−1)→ρ and δ−1∫vfδdv→u.
The companion theorem assumes d∈{2,3}, smooth divergence-free u₀ and smooth ρ₀ with zero spatial means, plus the Eq. (1.20) constraints on gR. Its particle-level limits use |v|≤ε−κ; the uncut formulas above explain the moments, not a replacement for the theorem statement.
The fluid equation uses the slow time τ
∂τu+u·∇u−μ₁Δu=−∇p ∂τρ+u·∇ρ−μ₂Δρ=0 ∇·u=0, ρ+ϑ=0
The paper writes the scalar equation in ρ. The temperature fluctuation is ϑ=−ρ: this is the Boussinesq relation that filters the leading acoustic mode.
With Fδ(τ)=fδ(τ/δ), the kinetic equation becomes δ²∂τFδ+δv·∇xFδ=Q(Fδ,Fδ). The first fluctuation lies in μ·span{1,v,|v|²}, the tangent space to the Maxwellian family at μ.
LμF:=Q(μ,F)+Q(F,μ), ker Lμ=μ·span{1,v₁,…,vd,|v|²} P is the orthogonal projection in L²(μ−1dv); Lμ−1 is restricted to (ker Lμ)⊥, and its input must obey the collision-invariant solvability conditions.
The orthogonality conditions are exactly the solvability conditions. This displayed expansion is a formal kinetic derivation explaining the mechanism; the companion theorem supplies the rigorous prepared-data estimate. The non-equilibrium correction has constitutive moments, schematically,
Σdev(H⊥)=−μ₁(∇u+∇u⊤), q(H⊥)=−κth∇ϑ.
Divergence of deviatoric stress yields viscosity; divergence of heat flux yields thermal diffusion. After normalizing the scalar balance, the latter becomes the PDE coefficient μ₂. With ∇·u=0 and ϑ=−ρ, these give the displayed Laplacians.
Open the exact prepared ansatz and remainder from Eqs. (1.21)–(1.23)
The smooth remainder gR is constrained as in Eq. (1.20). Prepared data suppress the fast acoustic and initial-layer behavior that the displayed limit is not designed to resolve.
The two odd centered moments vanish. The term ∫(u·c)cMdv=(∫c⊗cMdv)u=ρTu supplies the extra “+2” in d+2.
The collision invariants annihilate every collision moment. The five Gaussian identities above then close the three conservation laws using only ρ, u, and T. This is the missing calculation from local Maxwellian to compressible Euler.
Why M alone is not enough
Q(M,M)=0, but generally (∂t+v·∇x)M≠0. The Hilbert correctors solve that mismatch; the requirement that each linearized equation be solvable is exactly what constrains ρ, u, and T.
Open the Hilbert recurrence and prepared expansion from Eqs. (1.40)–(1.46)
Centered test-window averages converge in probability to NSF fields
This closes Route A. Euler uses unmagnified order-one observables instead of the δ−1-centered quantities below. Here ψ is fixed and smooth, κ=θ/4, and t is fixed in [0,δ−1Tfin].
VELOCITY OBSERVABLE
uem[ψ](t)=1δ1N∑j=1N𝟙|vj(t)|≤ε−κψ(xj(t))vj(t)
Weight velocities inside the window, discard the technical high-velocity tail, average over particles, and magnify the O(δ) signal.
What “resolves Hilbert’s sixth problem” means here
PROVED
Deng, Hani, and Ma derive the Boltzmann equation from dilute identical hard spheres on ℝd, d≥2, through every fixed interval on which the required regular Boltzmann solution exists. Their companion paper extends the kinetic limit to 𝕋d, d=2,3, and combines it with established Boltzmann-to-fluid limit theorems to derive incompressible Navier–Stokes–Fourier and compressible Euler under the paper’s stated prepared-data, regularity, and iterated/scaling limits.
NOT CLAIMED
A complete axiomatization of physics; dense gases or liquids; arbitrary interactions, shapes, domains, or boundary conditions; every hydrodynamic scaling; one unrestricted simultaneous limit; or global regularity for the Boltzmann or three-dimensional Navier–Stokes equations.
THE ARROW OF TIME
Reversible rules, irreversible statistics
x(t)=x(0)+tv mod box; v↦−v retraces the same path
Exact periodic free flight: reversing every velocity retraces every wrapped path. This is a kinematic analogy, not a hard-sphere correlation simulation.
1
The hard-sphere flow preserves phase-space volume and can be reversed.
2
The initial ensemble selects “chaotic” states: particles are almost independent.
3
The time-reversed state after mixing has extraordinarily organized correlations; it is not a fresh chaotic initial state.
4
Thus an irreversible effective law can emerge for the prescribed forward-chaotic ensemble in the kinetic limit without contradicting reversible microscopic mechanics.
The theorem justifies the Boltzmann description across its regular lifespan. The entropy interpretation additionally uses Boltzmann’s H-theorem and the statistical choice of initial ensemble.
DERIVATION D · THE H-THEOREM
Where the irreversible sign appears
Formal calculation for positive smooth f, with periodic x or sufficient spatial/velocity decay. The displays below use the α=1 convention. For Eq. (1.15), dH/dt=α∫Q(f,f)log f dx dv; because α>0, it changes the rate but not the sign. Rigorous entropy identities or inequalities require the corresponding integrability and trace hypotheses.
1 · DEFINE
H[f] = ∫ f log f dx dv
Up to physical constants and a sign, −H is entropy.
2 · DIFFERENTIATE
dH/dt = ∫ Q(f,f) log f dx dv
The transport term integrates to zero on a periodic domain or with sufficient decay. The “+1” from differentiating f log f vanishes because ∫Q(f,f)dv=0: collisions conserve mass.
3 · SYMMETRIZE
dH/dt = −¼∫ [(v−v*)·ω]+ (A−C) log(A/C) dx dv dv* dω
Here A=f′f′* and C=ff*. The factor ¼ comes from averaging the equivalent expressions obtained by swapping particles and swapping pre/post-collision variables; with the positive-part kernel, the latter change is accompanied by ω↦−ω.
Because log is increasing. Under suitable positivity, regularity, integrability, and collision-kernel nondegeneracy hypotheses, equality forces detailed balance almost everywhere; log f is then an additive collision invariant, and c<0 gives a Maxwellian. Return to the local five-step classification proof ↑
NUMERICAL SIGN CHECK
B=1, A=4, C=1 ⇒ −¼B(A−C)log(A/C)=−¾log4≈−1.040
One symmetrized channel is negative. If A and C are interchanged, both A−C and log(A/C) change sign, so their product stays nonnegative and the leading minus sign still makes entropy production nonpositive.
08 · THE BIGGER PICTURE
One idea migrates across fields
Deng’s programs share a theme: randomness is not noise to discard—it is structure that can survive nonlinear evolution in a controlled form.
WAVES
Nonlinear Schrödinger equation
(i∂t − Δβ)u + λ²|u|²u = 0
Fourier modes interact weakly. Δβ is an anisotropic Laplacian with dispersion ω(k)=|k|β². Picard iterates produce paired trees and Feynman diagrams.
SHARED TEMPLATElayer long timepreserve statistical historyorganize nonlinear diagramsmake smallness beat their count
PARTICLES
Hard-sphere Newtonian flow
ẋj=vj, ṽj=0 between impacts; vj jumps by the elastic rule at impact
Collision histories create clusters and recollisions. The wave strategy inspired the architecture; particle geometry required new algorithms.
Cumulants, collision/overlap molecules, hard-sphere recollision geometry, truncation, and cutting into elementary integrals.
THE SAME META-PROBLEM IN THREE DIALECTS
Propagation of randomness
Nonlinearity creates correlations. Deng’s work finds enough structure inside those correlations to retain a statistical law—or construct a rough random flow—over meaningful times. Both wave pillars start with random Fourier data, but wave kinetics averages over realizations to derive a closed law for a spectrum, whereas probabilistic PDE fixes almost every realization and constructs its canonical rough trajectory. This table compares proof problems; it does not claim the theorems are instances of one formal framework.
Comparison categoryParticlesWeak wavesRough random PDE
Microscopic system
Newtonian hard spheres
random-mode cubic NLS
Wick-ordered NLS flow
Object retained
marginals fs; density f
spectrum n(k)=𝔼|ûk|²
Gibbs-typical solution law
Naïve obstacle
recollisions destroy exact independence
resonant modes build correlations
rough products are not classically defined
Memory object
cumulants & molecules
paired trees & layered gardens
random averaging operators & tensors
Source of gain
thin recollision geometry εν
oscillation, pairings, cancellations
Gaussian contraction & operator norms
Output
quadratic Boltzmann closure
cubic wave-kinetic closure
strong flow + invariant Gibbs measure
VISUAL CHECKPOINT · THE BIGGER PICTURE
Match each source of smallness to its program
The three pillars share a count-versus-structure problem, but the small parameter and output object are different. Choose a program for each mechanism.
Licensed inference: the three proofs obtain smallness from different structures. This analogy does not identify their expansions or theorems term by term.
PILLAR 2 · WAVE TURBULENCE
From nonlinear waves to a wave kinetic equation
random Fourier modes→NLS dynamics→closed kinetic law for n(t,k)
Set αwave=λ²L−d, the effective nonlinear coupling in the paper’s large-box normalization. In their 2023 Inventiones paper, Deng and Hani derived the wave kinetic equation from cubic NLS up to t=δTkin, where δ>0 is sufficiently small but independent of L and Tkin=1/(2αwave²). Their later long-time preprint reaches arbitrary fixed kinetic times strictly inside the wave kinetic equation’s lifespan in its admissible regimes. These diagrammatic ideas inspired—but did not mechanically transfer to—the hard-sphere proof.
They also derived higher statistics: distinct modes remain asymptotically independent, and Gaussianity propagates when the initial modes are Gaussian.
4The result is a canonical global strong flow on a full Gibbs-measure set: canonical truncations converge to a unique cutoff-independent limit, and the flow preserves the Gibbs measure, for every odd defocusing Wick power p=2r+1≥3 on 𝕋².
The general random-tensor theory also gives almost-sure local well-posedness throughout the range subcritical for probabilistic scaling.
Before drawing a tree, fix what is random, what its variance means, and why the kinetic clock is quadratic.
The normalization matters. These formulas use Deng–Hani’s large-box Fourier convention; changing the Fourier normalization changes the expression for αwave, though not the underlying theory.
Input law. The ηk are independent, centered, normalized complex Gaussians, or independent variables uniform on the complex unit circle. Thus nin(k) is not an amplitude: it is the expected spectral density at mode k.
1One vertex costs αwave
A cubic Duhamel correction is first order in the weak effective coupling.
2The first secular transfer is second order
In the spectrum 𝔼|ûk|², the leading diagonal/self-energy effect only changes phase after renormalization: a deterministic nonlinear frequency shift is removed by a phase rotation, which changes phase but not |ûk|². Transfer between modes first survives at order αwave².
3Balance size against time
A rate of size αwave² becomes order one when αwave²t≈1, hence t≈αwave−2. The factor 1/2 is the paper’s convention.
Pass this object forward: the initial variance nin enters the leaves of the Duhamel expansion; two copies of that expansion meet when we form 𝔼|ak|²; their admissible pairings produce the cubic products in the kinetic collision operator.
MISSING MIDDLE STEP · AMPLITUDE → SPECTRUM → KINETIC LAW
The phase mismatch appears in Duhamel; the kinetic resonance emerges after squaring, averaging, and taking the limit.
Two valid formulas live at different clocks. The first is the exact physical-time equation before renormalization. The second is the exact rescaled, Wick-ordered equation used in the paper’s proof. Keeping them separate prevents a coefficient or phase from silently changing conventions.
Let M be the conserved spatially averaged mass, set v=e−2iλ²Mtu, and then ak(σ):=e−δπiL²|k|β²σv̂(δTkinσ,k). On σ∈[0,1],
∂σak=iδ2Ld−1∑k₁−k₂+k₃=kεk₁k₂k₃eδπiL²Ωσak₁ak₂ak₃
The selector εk₁k₂k₃ implements the paper’s Wick/self-energy removal. Here δ is the paper’s small time-window parameter in the wave theorem—not the fluid Knudsen parameter used in the preceding chapter.
The frequency mismatch exists before any expectation is taken. Squaring, averaging, and kinetic scaling decide which terms survive in nk; the paper’s signed trees organize repeated use of its second equation.
TREE → COUPLE · WHAT “PAIR THE INPUTS” ACTUALLY MEANS
One spectrum term contains a positive tree, a negative tree, and a matching of their leaves.
A scale-n signed ternary tree has n branching nodes and 2n+1 leaves. At a node of sign ζ, its three children have signs (ζ,−ζ,ζ). A couple is a + tree, a − tree, and a partition of all their leaves into opposite-sign pairs. This scale-one illustration chooses three cross-tree pairs; the definition also permits opposite-sign pairings within one tree.
Convention guardrail. The displayed identity is Isserlis’s rule for normalized complex Gaussians. Uniform random phases obey index-balance rules; they do not carry identical Gaussian pairing multiplicities.
SIX LEAVES → THREE SPECTRA → FOUR-TERM BRACKET
Pairing explains why the closure is cubic; orientation explains the signs.
For three distinct modes, each opposite-sign pair contributes one variance ni. Six random leaves therefore become three spectral factors. This is the degree-three bridge from a pair of cubic vertices to the WKE.
ALGEBRAIC ROLEMONOMIALSIGN
all three internal slotsn₁n₂n₃+
output n replaces slot 1nn₂n₃−
output n replaces conjugated slot 2nn₁n₃+
output n replaces slot 3nn₁n₂−
Schematic reading guide—not a diagram proof. Isserlis pairing supplies products of variances, but it does not by itself supply these four signs. The signs arise from the oriented expansion after Wick/self-energy renormalization and the summation of the regular diagram families.
Quick check: does “Gaussian pairing” alone prove the four-term bracket?
No. Pairing explains which mode labels can survive and why three n-factors appear. One still has to classify the oriented couples, prove that non-leading diagrams are negligible or cancel, and compute the four signed regular contributions.
1Duhamel iterate
One cubic interaction contributes an oscillatory time integral and three mode amplitudes.
2Form the spectrum
Insert the iterate into 𝔼|ak(t)|²; products of random amplitudes appear.
3Pair random inputs
For the well-prepared independent Gaussian or uniform-phase inputs treated in the derivation, expectation suppresses incompatible terms and selects admissible diagram pairings.
4Use oscillation
Cancellation is effective when |Ω|t≫1; interactions with |Ω|t≲1 can accumulate over that observation time. Here Ω is the energy mismatch defined in the exact Duhamel formula above.
5Take the kinetic limit
Surviving quartets concentrate on the resonant manifold and close an evolution law for n.
On a periodic box, u(x)=∑kûkeik·x. A wavevector k labels a Fourier wave’s direction and wavelength; points here are wavevector tips, not physical positions. Momentum matching determines k=k₁−k₂+k₃. On the square torus, lasting exact resonance also matches ω(k)=|k|².
Goal: find the perpendicular rectangle where Ω=−2(k₁−k₂)·(k₃−k₂)=0.
Why energy mismatch matters: the exact phase above is e2πisΩ, so ∫₀te2πisΩds=(e2πitΩ−1)/(2πiΩ). Its size is at most 1/(π|Ω|) when Ω≠0; at Ω=0 it equals t and accumulates.
Scope. This slider uses continuum Euclidean geometry. Finite-torus Fourier modes lie on a lattice, so an arbitrary angle need not be an allowed discrete quartet. The α=L−1 theorem in the 2023 paper uses generic anisotropic β; the lab explains resonance algebra, not its number-theoretic estimates.
move k₃ to (1,1), so k=(2,1) Ω=1+2−5=−2 ∫₀1/2e2πisΩds=∫₀1/2e−4πisds=0, while resonance gives 1/2
Momentum matching builds a parallelogram. For Euclidean quadratic dispersion, writing u=k₁−k₂ and w=k₃−k₂ gives Ω=−2u·w, so a rectangle is exactly resonant.
INTERACTIVE LIMIT LAB · A DELTA IS A WEAK LIMIT, NOT A TALL FUNCTION
Watch a finite-time resonance window narrow into a Dirac mass
For this teaching plot, set ξ:=2πΩ, the angular phase mismatch in the exact Duhamel exponent. Squaring one oscillatory time integral gives a nonnegative kernel in ξ. This harmless fixed rescaling preserves the resonant set ξ=0⇔Ω=0; with the normalization below, the kernel's total area is always one.
Kt(ξ):=12πt|∫₀teisξds|²=2sin²(tξ/2)πtξ² ⇀ δ(ξ)
central height Kt(0)
1.273
first-zero half-width
0.785
analytic full-line area
1 · unchanged on ℝ
Read the graph: nonzero ξ cancels increasingly strongly as t grows; only a window of width ≍1/t around resonance retains mass. The limit symbol ⇀ means weak convergence: ∫Kt(ξ)φ(ξ)dξ→φ(0), not pointwise convergence at ξ=0.
What remains in the theorem. This one-dimensional kernel explains the frequency delta only. The rigorous argument also takes L→∞ with weak coupling, controls all diagrams and remainders, and turns the lattice momentum constraint into the second continuum Dirac delta.
THE SECOND DELTA · DISCRETE MOMENTUM → CONTINUUM CONSTRAINT
A normalized lattice sum becomes an integral over a hyperplane.
The discrete Kronecker constraint leaves two independent lattice vectors, hence the L−2d Riemann-sum normalization. In the continuum, the Dirac mass restricts the three-vector integral to the same 2d-dimensional momentum hyperplane.
THE WAVE KINETIC COLLISION OPERATOR
Near-resonances concentrate onto a resonant manifold
Write n=n(τ,k) and ni=n(τ,ki), where τ=t/Tkin. At finite L, near-resonant quartets accumulate; in the kinetic limit their contribution concentrates onto the continuum manifold represented by the two Dirac deltas. For d≥3 and the admissible scalings, n satisfies a closed cubic law. The β-subscript covers the anisotropic dispersion; on the square torus β is Euclidean.
The Dirac deltas impose wavevector and frequency matching. The bracket is the wave analogue of gain minus loss—cubic rather than Boltzmann’s quadratic closure.
WHAT THE 2023 THEOREM ACTUALLY DELIVERS
The diagrams close into a uniform, quantitative kinetic approximation.
Let d≥3, choose β outside the paper’s exceptional Lebesgue-null set, take a nonnegative Schwartz profile nin, and use the well-prepared random data above. Under λ=L(d−1)/2, one has αwave=L−1 and Tkin=L²/2.
For every fixed A≥40d and sufficiently small fixed δ depending on A, β, and nin, the WKE has its unique solution n on [0,δ]. For all sufficiently large L depending on δ, NLS has a smooth solution through δTkin with probability at least 1−L−A; C is independent of L and ν=ν(d)>0.
Scope. This endpoint theorem uses a generic anisotropic torus. Following its convention, the quantity inside the expectation is defined as zero on the exceptional event where no smooth solution exists through the stated time. The rectangle slider uses square Euclidean geometry only to teach the resonance identity; it is not the theorem’s arithmetic setting.
The coefficients gk are independent complex Gaussians and ⟨k⟩=(1+|k|²)1/2. Almost surely u0 lies in Hs for every s<0 (written H0−), but not generally in L². Naïve cutoff powers fail to converge as classical products; Wick ordering subtracts divergent self-contractions before the cutoff is removed.
u = ulin + 𝒬(ulin) + Y (schematic)
Wick ordering subtracts divergent self-contractions before the cutoff limit. The random averaging correction 𝒬 isolates rough linear randomness; Y is smoother. Random tensors record how Gaussian inputs pair through nonlinear iterations.
SCALING THRESHOLDS FOR p-POWER NLS
sprob = −1/(p−1), sdet = d/2 − 2/(p−1)
For odd nonlinear degree p≥3 (excluding the exceptional general-theorem case (d,p)=(1,3)), randomness can lower the relevant threshold from sdet toward sprob; the full probabilistically subcritical range s>sprob is covered. A tensor “flattening” groups input/output indices into a matrix; operator norms give stability while Hilbert–Schmidt bounds capture square-root cancellation.
INTERACTIVE LAB 6 · RANDOM-PDE PILLAR
Why H0− converges while L² diverges
Truncate u at |k|≤K in two dimensions. Its expected squared Sobolev norm is ∑⟨k⟩2s−2. A dyadic shell |k|≈2j has ≍22j modes, so—after suppressing fixed lattice constants—its normalized shell contribution is ≍22sj.
Through shell 8, the normalized L2 model sums to 8, while the H minus one quarter model sums to 2.26 and approaches 2.41.
normalized L² shell model
8.00 · grows like log K
normalized H−1/4 shell model
2.26 · approaches 2.41
This plots an asymptotically equivalent, normalized dyadic-shell model—not exact lattice sums and not a random sample. Standard independence/Gaussian arguments upgrade the threshold to the almost-sure statement.
INTELLECTUAL LINEAGE
From a physical guess to a long-time theorem
1872
A statistical equation is born
Ludwig Boltzmann formulates a kinetic equation using “molecular chaos”: particles about to collide are treated as statistically independent.
What it suppliedA correct and extraordinarily successful target equation—but not a theorem deriving it from Newtonian particles.
WAVE-TURBULENCE LINEAGE
Peierls, Hasselmann, and Zakharov formulated kinetic turbulence; Spohn, Lukkarinen–Spohn, Collot–Germain, and Buckmaster–Germain–Hani–Shatah established rigorous partial regimes before the full kinetic-time program.
RANDOM-PDE LINEAGE
Lebowitz–Rose–Speer and Bourgain built Gibbs–NLS foundations; Burq–Tzvetkov developed random-data well-posedness. Parabolic work such as Da Prato–Debussche and Hairer is adjacent, but dispersive NLS lacks parabolic smoothing.
WHAT CAME BEFORE THE LONG-TIME BREAKTHROUGH?
Earlier theorems were profound—but lived in different regimes
No single row “supersedes” all the others. Each controls a different combination of time, perturbation size, and closeness to equilibrium.
1975 · LANFORD
Non-perturbative regular dilute data, very short time
Collision-tree convergence for a small fraction of a mean-free time. This established the microscopic Boltzmann derivation.
1989 · ILLNER–PULVIRENTI
Long time, near vacuum
Global-in-time results in a dispersive low-density regime where particles separate strongly.
2010s–23 · EQUILIBRIUM PROGRAM
Long kinetic times at exact equilibrium
Bodineau, Gallagher, Saint-Raymond, Simonella and collaborators obtained fluctuation and linearized regimes where equilibrium cancellations are available—not nonlinear propagation of chaos for arbitrary near-equilibrium data.
2024–25 · DENG–HANI–MA
Long time, nonlinear off-equilibrium
Every fixed time in the lifespan of the required regular nonlinear Boltzmann solution, without a near-vacuum or equilibrium restriction.
KNOWLEDGE MAP
Prerequisites for the particle theorem
WHERE THE MACHINERY GOES NEXT
Extensions, bridges, and open terrain
These statuses matter: a related use of the machinery is not automatically a consequence of the hard-sphere theorem.
EXTENSION
Forced and inhomogeneous waves
Damped-driven wave turbulence and inhomogeneous kinetic limits bring forcing, dissipation, and space dependence into the picture.
Deng–Ionescu–Pusateri establish deterministic estimates and propagation of randomness. This is progress toward—not yet a full quasilinear wave-kinetic derivation.
Long-time limits for smooth or long-range potentials, boundaries, mixtures, non-spherical particles, off-equilibrium fluctuations, large deviations, and weak Boltzmann solutions.
OPEN · WAVES
Harder resonance worlds
Quasilinear water waves, lower-dimensional degeneracy, condensation or blow-up, and general quantum Boltzmann limits.
OPEN · PROOF THEORY
How universal is “memory + cutting”?
Which many-body systems admit a structured history object whose geometric or oscillatory gain beats combinatorial growth?
NOTATION WITHOUT FEAR
Search the glossary
ε
sphere diameter
The small parameter sent to zero in the dilute-gas limit.
𝔼N
expected particle count
In the grand-canonical ensemble, grows while ε shrinks, balanced by (𝔼N)εd−1 ≈ α.
f(t,x,v)
one-particle density
Expected density near position x and velocity v at time t.
fs
s-particle correlation
Rescaled factorial-moment density of the grand-canonical ensemble; a joint intensity, not generally a normalized marginal.
chaos
asymptotic independence
fs converges to the product f⊗s for fixed s.
EH
defect / cumulant
Defined relative to the leading state fA. For |H|=1 it is a one-root defect; terms with |H|≥2 carry genuine multi-root dependence.
M
molecule
A layered graph encoding collisions, overlaps, particle lines, and time order.
ρ
multi-layer complexity
Built from per-layer roots/crossings and recollision ranks; ordinary circuit rank E−V+C is the one-layer prototype.
τ
layer width
T/L, chosen short enough for local cluster expansion.
α
collision-frequency scale
The limiting geometric prefactor (𝔼N)εd−1. The realized rate also depends on the velocity law; this α is unrelated to weak-coupling notation in some wave papers.
ω
collision normal
Unit vector along the centers of two spheres at impact.
𝟙Ds
exclusion indicator
Zero when any of the s spheres overlap, one otherwise.
Kn
Knudsen number
Mean free path divided by macroscopic length. Kn→0 is the fluid limit after the kinetic equation has been obtained.
ν
geometric gain exponent
A fixed positive number depending on dimension; εν is the smallness earned by a good collision restriction.
fA
leading one-particle state
The locally updated, almost-factorized density carried across time layers. A is a label, not an exponent.
δ
local small parameter
In the trace toy δ is a strip width; in the fluid chapter δ=α−1 and fδ=n1/δ. These local uses are unrelated.
n(τ,k), δ(·)
wave spectrum and Dirac constraint
n is expected Fourier-mode energy; here δ(·) is the Dirac distribution restricting the kinetic integral to a matching manifold, not either small parameter above.
Hs
Sobolev regularity
A scale measuring derivatives in an averaged sense. H0− means membership for every s<0, generally rougher than L².
Wick ordering
renormalized product
Subtract divergent self-contractions before removing a frequency cutoff, producing a meaningful random nonlinear term.
Gibbs measure
energy-weighted random field law
A nonlinear energy weight applied to a Gaussian reference field; invariance means the PDE flow preserves this probability law.
≍
same order
Two quantities bound each other up to fixed positive multiplicative constants.
No glossary term matches that search.
CHECK YOUR MODEL
Fourteen questions in dependency order
Part A follows the particle proof from mechanics to fluids. Part B checks the two related Fields Medal pillars. A wrong choice stays available as a learning step: review the clue, then retry until every question is mastered.
The displayed main theorem is a pedagogical compression, checked against arXiv:2408.07818 v3; consult the paper for the complete hypotheses and normalization. The proof architecture was also compared with the 2026 Bourbaki exposé. “Resolution of Hilbert’s sixth problem” is always qualified by model, limit, domain, regularity, and scope. Interactive simulations are conceptual illustrations, not numerical evidence for the theorem; the site’s figures are original pedagogical diagrams, not reproductions of figures in the cited papers.