By OpenAISourceGitHub

Manuscripts from openai/math

Fork
Clone
git clone https://openversions.com/git/index/openai-math.git
Cite
Version link
https://openversions.com/index/openai-math/commit/dfd8c46870b48511e697cbb43a3c8c91d375fa2f
Version identifier (commit hash)
dfd8c46870b48511e697cbb43a3c8c91d375fa2f

Fork to create your own copy with credit to the source. Clone to get a Git copy on your computer.

All papers

Claims

Findings

Hard-sphere fluctuations on the regular Boltzmann lifespan

Under Assumption 1.1, the centered Gaussian generalized field ζ\zeta solving dζt(ϕ)=ζt(Atϕ) dt+dWt(ϕ)\mathrm{d}\zeta_t(\phi) = \zeta_t(A_t\phi)\,\mathrm{d}t + \mathrm{d}W_t(\phi) is well defined by its backward mild form, and for every m≥1m \ge 1, fixed times t1,…,tm∈[0,T]t_1, \dots, t_m \in [0, T], and real tests ϕ1,…,ϕm∈Cc∞(R6)\phi_1, \dots, \phi_m \in C_c^\infty(\mathbb{R}^6), the exact-centered hard-sphere fluctuation vector (ζt1ϵ(ϕ1),…,ζtmϵ(ϕm))(\zeta_{t_1}^\epsilon(\phi_1), \dots, \zeta_{t_m}^\epsilon(\phi_m)) converges in distribution to (ζt1(ϕ1),…,ζtm(ϕm))(\zeta_{t_1}(\phi_1), \dots, \zeta_{t_m}(\phi_m)) as ϵ↓0\epsilon \downarrow 0.

The fluctuation equation in Theorem 1.2 has a unique centered Gaussian solution on the test scales Ea0E_a^0, 0≤a<β0 \le a < \beta, with covariance E[ζt(ϕ)ζs(ψ)]=∫f0(z)U(0,t)ϕ(z)U(0,s)ψ(z) dz+∫0min⁡(s,t)Cr(U(r,t)ϕ,U(r,s)ψ) dr\mathbb{E}[\zeta_t(\phi)\zeta_s(\psi)] = \int f_0(z)U(0,t)\phi(z)U(0,s)\psi(z)\,\mathrm{d}z + \int_0^{\min(s,t)} C_r(U(r,t)\phi, U(r,s)\psi)\,\mathrm{d}r.

For the initial grand-canonical ensemble, the exact-centered initial fluctuations converge jointly to a centered Gaussian vector with covariance ∫f0ϕψ\int f_0\phi\psi, while lim⁡ϵ→0μ(Eπ0ϵ(ϕ)−∫f0ϕ)=−4π3∫R6f0(x,v)ϕ(x,v)ϱ0(x) dx dv\lim_{\epsilon\to 0}\sqrt{\mu}(\mathbb{E}\pi_0^\epsilon(\phi) - \int f_0\phi) = -\frac{4\pi}{3}\int_{\mathbb{R}^6} f_0(x,v)\phi(x,v)\varrho_0(x)\,\mathrm{d}x\,\mathrm{d}v.

For Γ\Gamma and LL chosen independently of the moment order and observation list, the normalized marked factorial cumulant chϵc_h^\epsilon of the pasted dynamics equals the sum of retained globally connected diagrams plus a total variation error of at most Che−cΛLC_h e^{-c\Lambda_L}, and diagrams with complexity ρ>K\rho > K have summed absolute variation at most ϵP\epsilon^P.

For the pasted dynamics with Γ,L\Gamma, L chosen as in Theorem 4.9 (and LL increased if needed depending only on ff and TT), the normalized marked factorial cumulants satisfy ∥chϵ∥TV≤Chμ1−h\|c_h^\epsilon\|_{\mathrm{TV}} \le C_h \mu^{1-h} for each fixed hh, and for h=1,2h=1,2 the measures μh−1chϵ\mu^{h-1}c_h^\epsilon converge against bounded continuous record tests to the signed sums of retained global trees with trivial initial cumulants.

On the single pasted dynamics, for every fixed observation list, (∑iyi(T)−E∑iyi(T))/μ⟹∑jζtj(ϕj)(\sum_i y_i(T) - \mathbb{E}\sum_i y_i(T))/\sqrt{\mu} \Longrightarrow \sum_j \zeta_{t_j}(\phi_j), with convergence of all centered moments.

For sufficiently large Γ\Gamma and LL, the true hard-sphere dynamics and the pasted dynamics coupled using the same initial configuration satisfy P{the two trajectories differ at some time in [0,T]}≤Cϵ1+η\mathbb{P}\{\text{the two trajectories differ at some time in } [0, T]\} \le C\epsilon^{1+\eta} for all sufficiently small ϵ\epsilon, with η=1/20\eta = 1/20.

Under the cutoff estimate of Proposition 7.1, the expectation of any fixed bounded empirical observation list differs between the true and pasted dynamics by o(μ−1/2)o(\mu^{-1/2}), so the corresponding exact-centered fluctuation vectors have the same weak limits.

Methods

Restarted truncated ('pasted') hard-sphere dynamics on LL time slabs with particle-number caps Λi\Lambda_i and collision-excess cap Γ\Gamma, combined with layered C/O collision histories.

Source excerpts

“The restarted truncated dynamics is defined on all configurations, with overlaps permitted. Within slab i, start the collision prehistory anew. At each incoming contact, accept the elastic collision precisely when adding that contact to the existing collision graph leaves its component with at most \Lambda_i particle labels and excess at most \Gamma.” · Show in paper

Penrose tree-graph partition scheme for connected-graph expansions of the initial grand-canonical ensemble.

Source excerpts

“The Penrose partition scheme [9], in the form proved in [4, Proposition 2.3.3], gives the following tree bound. |\mathcal{U}_n(Z_n)| <= \sum_{\mathcal{T} \text{ tree on } [n]} \prod_{\{i,j\}\in\mathcal{T}} \chi_{ij}.” · Show in paper

Marked componentwise expansion with particle records and one-root resummation for factorial cumulants of multitime observations.

Source excerpts

“For a particle label a, a record is the sum of its requested test values at pulses already passed. Only labels whose records are required carry this extra coordinate... The retained hierarchy admits an exact finite expansion \widehat{F}_{i,S}^R = \sum_{H: R\subseteq H\subseteq S} Ei,HE_{i,H}^R \otimes g_i⊗(S∖H)i^{\otimes(S\setminus H)}.” · Show in paper

Chronological merger coordinates and joint energy-time-order integration for sharp connected cumulant bounds.

Source excerpts

“Away from nontransversal contacts and simultaneous events, its history integral has coordinates consisting of the birth velocities, one absolute translation for each global component, a vector in a ball of radius RK,DR_{K,D}\epsilon for each physical initial merger, and (t, \omega) \in \mathbb{R} \times S^2 for each dynamic merger.” · Show in paper

Amplified pure-collision cutting sequences, local adjacent-pair estimates, and literal free-parameter extraction with common velocity containing sets for bounding the discrepancy probability.

Source excerpts

“Either it contains a freely cut pair estimated by (7.9) and no other activity loss, or its retained improved pieces and free singletons satisfy \delta g - 2s_4 >= 6/5.” · Show in paper

“Remove the factors a_j from the ordinary measures, obtaining a positive measure \nu. If E is a set of complete histories and \mathcal{D} is a measurable set containing its projection onto the literal variables \vartheta = (\vartheta_1, ..., \vartheta_q), then \nu(E) <= (\prod_\ell H_\ell) \int_{\mathcal{D}} \prod_{j=1}^q dt_j d\omega_j dw_j.” · Show in paper