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

A radial continuum phase transition with algebraic decay

There exists a bounded, continuous, stable radial pair potential ϕ:[0,∞)→R\phi : [0, \infty) \to \mathbb{R} in three dimensions with ∣ϕ(r)∣≤Cr−3−1/32|\phi(r)| \le C r^{-3-1/32} for r≥1r \ge 1 and ∑i<jϕ(∣xi−xj∣)≥−BN\sum_{i<j} \phi(|x_i - x_j|) \ge -BN, such that throughout an open density interval I⊂(0,∞)I \subset (0, \infty) centered at 5p/35p/3 and at a common inverse temperature βc∈[7/8,9/8]\beta_c \in [7/8, 9/8], the canonical free energy f(β,ρ)=−lim⁡L→∞1βL3log⁡ZL,⌊ρL3⌋ϕ(β)f(\beta, \rho) = -\lim_{L \to \infty} \frac{1}{\beta L^3} \log Z_{L, \lfloor \rho L^3 \rfloor}^\phi(\beta) exists, is finite for every β>0\beta > 0, and satisfies f−′(βc,ρ)>f+′(βc,ρ)f'_-(\beta_c, \rho) > f'_+(\beta_c, \rho).

Every member of the potential family (2.2) is stable with the same constant B0=t2/2B_0 = t^2/2, and for tt sufficiently large satisfying D0≥8MD_0 \ge 8M and ct>4(A+1)ct > 4(A+1) with c=1/3456c = 1/3456, any interaction vv of the form (2.4) in an axis-parallel cube of side L≥1L \ge 1 and volume V=L3V = L^3 satisfies Uv≥ct2N2/V−(t2/2)NU^v \ge c t^2 N^2 / V - (t^2/2)N.

For every member φ\varphi of (2.2), including countably infinite sums, the canonical limit Hφ(β,d)H_\varphi(\beta, d) and grand-canonical pressure Qφ(β,h)Q_\varphi(\beta, h) exist for all β>0\beta > 0 and d≥0d \ge 0, are finite, satisfy convexity/concavity properties, and obey the Fenchel duality relations Qφ(β,h)=sup⁡d≥0{hd−Hφ(β,d)}Q_\varphi(\beta, h) = \sup_{d \ge 0} \{hd - H_\varphi(\beta, d)\} and Hφ(β,d)=sup⁡h∈R{hd−Qφ(β,h)}H_\varphi(\beta, d) = \sup_{h \in \mathbb{R}} \{hd - Q_\varphi(\beta, h)\} for d>0d > 0.

For fixed tt satisfying (2.7) and a compact D⊂Ω\mathcal{D} \subset \Omega, there exist constants Cexp⁡,C1<∞C_{\exp}, C_1 < \infty depending only on tt and D\mathcal{D} such that every grand canonical measure for an interaction (2.4) satisfies sup⁡iEexp⁡(βD0ni2/4)≤Cexp⁡\sup_i \mathbb{E}\exp(\beta D_0 n_i^2 / 4) \le C_{\exp} and sup⁡i,jE[ninj]≤C1\sup_{i,j} \mathbb{E}[n_i n_j] \le C_1 for (β,h)∈D(\beta, h) \in \mathcal{D}, uniformly over bounded regions and all permitted deletions and interpolations.

For an admissible tt and compact D⊂Ω\mathcal{D} \subset \Omega, there exist constants CD,t,KD,t<∞C_{\mathcal{D},t}, K_{\mathcal{D},t} < \infty independent of R0≥2R_0 \ge 2, the number and values of reference ranges Rj∈[1,R0]R_j \in [1, R_0], reference masses αj≥0\alpha_j \ge 0, and a>0a > 0 such that whenever ∑jαj+a≤A+1\sum_j \alpha_j + a \le A + 1 and R/R0≥KD,tR/R_0 \ge K_{\mathcal{D},t}, sup⁡λ∈D∣Qφ0−tawR(λ)−TaQφ0(λ)∣≤CD,tR0/R\sup_{\lambda \in \mathcal{D}} |Q_{\varphi_0 - taw_R}(\lambda) - T_a Q_{\varphi_0}(\lambda)| \le C_{\mathcal{D},t} \sqrt{R_0/R}.

As t→∞t \to \infty, the bare-core pressure q(t)=Qψtq^{(t)} = Q_{\psi_t} converges locally uniformly on Ω=(0,∞)×R\Omega = (0, \infty) \times \mathbb{R} to q∗(β,h):=max⁡{0,ph,4ph−9p}q_*(\beta, h) := \max\{0, ph, 4ph - 9p\}.

For all sufficiently large finite tt, the size conditions in Propositions 2.3 and 4.2 hold, sup⁡λ∈W1∣P1t(λ)−P∗(λ)∣<γ/64\sup_{\lambda \in W_1} |P_1^t(\lambda) - P_*(\lambda)| < \gamma/64, and at every λ∈W1=B‾(λ∗,1/4)\lambda \in W_1 = \overline{B}(\lambda_*, 1/4), ∂P1t(λ)=conv⁡(∂P1t(λ)∩⋃i=13B‾(vi,τ/2))\partial P_1^t(\lambda) = \operatorname{conv}(\partial P_1^t(\lambda) \cap \bigcup_{i=1}^3 \overline{B}(v_i, \tau/2)).

For every n≥1n \ge 1 in the multiscale sequence (7.6)-(7.7), there exists λn∈int⁡W(1/8)\lambda_n \in \operatorname{int} W(1/8) such that B‾(g,γ)⊂∂Pn(λn)\overline{B}(g, \gamma) \subset \partial P_n(\lambda_n).

Methods

Construct a bounded continuous repulsive core ψt\psi_t with a narrow zero-energy shell It=[lt,(1+Δ)lt]I_t = [l_t, (1+\Delta)l_t] (lt=e−tl_t = e^{-t}) that permits 4-particle tetrahedral clusters while excluding 5-particle cliques via a Gram matrix rank obstruction and Motzkin-Straus extremal graph bound.

Source excerpts

“Define ψt : [0, ∞) → [0, t2] by ψt(r) = t2 on [0, lt] ∪ [(1 + ∆)lt, 1], ψt(r) = 0 on [(1 + ∆/3)lt,(1 + 2∆/3)lt] ∪ [1 + 1/t, ∞), (2.1)” · Show in paper

“Lemma 2.2 (The shell graph). For any finite configuration in R3, join two particles when their distance belongs to It. This graph has no clique of size five. On any m vertices it has at most 3m2/8 edges.” · Show in paper

Establish uniform exponential and second-moment occupation bounds for small cells of side b=1/4b = 1/4 by conditioning on exterior configurations and closing the moment inequality sublinearly using D0≥8MD_0 \ge 8M.

Source excerpts

“To bound this maximum independently of the number of cells, we shall prove Fmax ≤ C2F η0 max, η0 = 4M2/D2 0 ≤ 1/16 < 1,” · Show in paper

Analyze the ideal attraction transform (TaQ)(β,h)=sup⁡u∈R{Q(β,h+βau)−βa2u2}(T_a Q)(\beta, h) = \sup_{u \in \mathbb{R}} \{Q(\beta, h + \beta a u) - \frac{\beta a}{2} u^2\} via its parabolic upper bound at active points and exact convex-hull subdifferential formula.

Source excerpts

“∂(TaQ)(λ) = conv [ u active at λ ( e + a 2 u 2 , u ) : (e, u) ∈ ∂Q(xu) .” · Show in paper

Compare a physical long-range attraction −tawR-t a w_R with the ideal transform TaQφ0T_a Q_{\varphi_0} by partitioning space into intermediate boxes of side s≍R0Rs \asymp \sqrt{R_0 R}, deleting inter-box reference interactions, and using independent reference grand ensembles with square completion.

Source excerpts

“Use boxes of an intermediate side s, with R0 ≤ s ≤ R. Deleting the reference interaction between different boxes costs O(R0/s) in pressure. ... Balancing these two errors explains the choice s ≍ √ R0R.” · Show in paper

Iterate the attraction addition along the multiscale schedule Rn=R12n−1R_n = R_1^{2^{n-1}} and an=Rn−1/16a_n = R_n^{-1/16} to preserve three separated slope groups, obtain a 2D ball B‾(g,γ)\overline{B}(g, \gamma) in ∂QΦ(βc,hc)\partial Q_\Phi(\beta_c, h_c), and apply Fenchel duality (Lemma 7.5) to deduce the canonical free-energy derivative jump.

Source excerpts

“For R1 > 2, define Rn = R 2 n−1 1 , a1 = A, an = R −1/16 n (n ≥ 2). (7.6)” · Show in paper

“Lemma 7.5 (From pressure supports to canonical slopes).” · Show in paper