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

Critical Center Magnetization in the Planar XY Model

Assuming the stated height, renormalization, and spin-field inputs from [Ope26a, Ope26b, Ope26c] in Section 2, there is a constant AXY∈(0,∞)A_{\mathrm{XY}} \in (0, \infty) such that the critical center magnetization an=En,bc0cos⁡θ0a_n = \mathbb{E}^0_{n, b_c} \cos \theta_0 satisfies an=AXYn−1/8(log⁡n)1/16(1+o(1))a_n = A_{\mathrm{XY}} n^{-1/8} (\log n)^{1/16} (1 + o(1)) as n→∞n \to \infty.

For every ϵ>0\epsilon > 0, one can choose a fixed dyadic qq, an arbitrarily large dyadic RR, and then s0s_0 such that for every compatible dyadic s≥s0s \ge s_0 with P=R5P = R^5 and m=qRsm = qRs, averaging the observation noise and reference covariance below mm gives ordinary local activities KmK_m with ∥Km∥m,A,V≤ϵ\|K_m\|_{m, \mathfrak{A}, V} \le \epsilon, and in a pierced free square gives Zm(1)(Kml,Im;ψ)\mathcal{Z}_m^{(1)}(K_m^l, I_m; \psi) with ∥Im∥m,A,V≤CRsCR\|I_m\|_{m, \mathfrak{A}, V} \le C_R s^{C_R} and exact shift-good and pierced-plane copying properties.

On a compatible torus or free square of side MM, the absolute change in the partition sum caused by the product of the face-difference cutoffs χT(df)\chi_T(d_f) with T=sT = \sqrt{s}, divided by the centered partition sum without cutoffs, is at most C(M/s)2e−c1sC(M/s)^2 e^{-c_1 s}, including on a free square with the integral connection ll and on a torus with bounded, polynomial, or exponential smear tests.

On each open preparation graph GG, the affine shape energy decomposes as EGl(y)=LG(d)+∑f,g∈F(G)NG(f,g)dfdgE_G^l(y) = L_G(d) + \sum_{f,g \in \mathcal{F}(G)} N_G(f, g) d_f d_g, where LG(d)L_G(d) has a fixed-radius stencil satisfying 0≤LG(d)≤C∑fdf2+Clog⁡(2s)∑zκz(d)20 \le L_G(d) \le C \sum_f d_f^2 + C \log(2s) \sum_z \kappa_z(d)^2, and NG(f,g)N_G(f, g) has uniform exponential decay in dist(f,g)\mathrm{dist}(f, g) and exponentially localized graph dependence.

For the fixed block factor LL and bulk map parameters, for every sufficiently large dyadic JJ, uniformly in the JJ-history at integer indices J/2≤j≤3JJ/2 \le j \le 3J, the unbalanced gradient coordinate tjt_j, fundamental coordinate zjz_j, and plane remainder Rj\mathcal{R}_j satisfy tj=12jlog⁡L+O(log⁡jj2)t_j = \frac{1}{2j \log L} + O\left(\frac{\log j}{j^2}\right), ∣zj∣=O(j−1)|z_j| = O(j^{-1}), and ∥Rj∥=O(j−2)\|\mathcal{R}_j\| = O(j^{-2}).

For every 0<c<80 < c < 8, uniformly on cJ≤k≤8JcJ \le k \le 8J, the periodic variance readout Uk=Var(X~Nk)/gU_k = \mathrm{Var}(\widetilde{X}_{N_k})/g and mean-phase readout Yk=EeihˉNkY_k = \mathbb{E} e^{i\bar{h}_{N_k}} satisfy Uk=a1−uk(J)+O((Zk(J))2)+O(e−c′J)U_k = \frac{a}{1 - u_k^{(J)}} + O((Z_k^{(J)})^2) + O(e^{-c'J}) and Yk=MkZk(J)+O((∣uk(J)∣+(Zk(J))2)Zk(J))+O(e−c′J)Y_k = M_k Z_k^{(J)} + O((|u_k^{(J)}| + (Z_k^{(J)})^2)Z_k^{(J)}) + O(e^{-c'J}), where MkM_k is bounded above and below by positive constants and common to all histories.

For all sufficiently large jj, the exact maps from entry scale ii to stop hh in the two concentric squares Q∈{M,LM}Q \in \{M, LM\} satisfy max⁡Q∈{M,LM}(∥Kh,Q0∥+∥Kh,Ql∥)≤C/j\max_{Q \in \{M, LM\}}(\|K_{h,Q}^0\| + \|K_{h,Q}^l\|) \le C/j, exact ratio equality PMl/PM0=PLMl/PLM0=:ρjP_M^l/P_M^0 = P_{LM}^l/P_{LM}^0 =: \rho_j with ∣log⁡ρj∣≤ϵ2j|\log \rho_j| \le \epsilon_2 j, and the exact field identity Ei:hIi(Ki,Ql,Ii,Q;ϕ)=PQl{σQZh(Kh,Ql;ϕ)+Ih(Kh,Ql,Ih,Q;ϕ)}\mathbb{E}_{i:h} \mathscr{I}_i(K_{i,Q}^l, I_{i,Q}; \phi) = P_Q^l \{\sigma_Q \mathscr{Z}_h(K_{h,Q}^l; \phi) + \mathscr{I}_h(K_{h,Q}^l, I_{h,Q}; \phi)\} with common scalar σM=σLM=:σj\sigma_M = \sigma_{LM} =: \sigma_j and ∥Ih,Q∥≤CRsCRL−2(h−i)/3\|I_{h,Q}\| \le C_R s^{C_R} L^{-2(h-i)/3}.

For nj=Lj/2n_j = L^j/2, there exists ϵ>0\epsilon > 0 such that log⁡anj+1−log⁡anj=−log⁡L8+116j+OL(j−1−ϵ)\log a_{n_{j+1}} - \log a_{n_j} = -\frac{\log L}{8} + \frac{1}{16j} + O_L(j^{-1-\epsilon}).

With the running gradient coordinate tht_h of Proposition 6.3, the first-order magnetic double difference Δj=(FLMl−FLM0)−(FMl−FM0)\Delta_j = (F_{LM}^l - F_{LM}^0) - (F_M^l - F_M^0) satisfies Δj=th2α2(ELM−EM)+OL(D2j2+1Dj)\Delta_j = \frac{t_h}{2\alpha^2}(E_{LM} - E_M) + O_L\left(\frac{D^2}{j^2} + \frac{1}{Dj}\right).

For the connections in the dual squares of sides M=2nM = 2n and LMLM, there is c4>0c_4 > 0 such that ELM−EM=2πlog⁡L+OL(n−c4)E_{LM} - E_M = 2\pi \log L + O_L(n^{-c_4}).

Methods

Overlapping finite critical histories with cell-averaged observations y=Bsh+ey = B_s h + e at scale sJ≍LJ3s_J \asymp_L J^3 and smooth entire face-difference cutoffs χT(df)\chi_T(d_f) with T=sT = \sqrt{s}.

Source excerpts

“For a dyadic integer s, partition the microscopic graph into square observation cells of side s. Write Bs for cell averaging, and introduce independent standard Gaussian variables ec, one per cell. The observations are y = Bsh + e. (3.2)” · Show in paper

“On every observation face retained in the graph insert χT (df ) = Z T −T 1 √ 2π e−(df −x)2/2 dx, df = yc′ − yc + lf , T = √ s. (3.4)” · Show in paper

“For a large integer J, choose a dyadic observation side s = sJ such that J3 ≤ sJ < CLJ3, qRsJ = Li(J). (5.1)” · Show in paper

Exact covariant isolation expansion with affine shape energies EGl(y)E_G^l(y) and telescoping of nonlocal matrix entries into local links with covering paths.

Source excerpts

“We now apply the finite covariant isolation construction of [Ope26a, Subsection 9.2]. We describe both the scalar normalization and the support labels, because exact copying depends on them.” · Show in paper

“Write the exact coefficient as the first clipped calculation plus successive differences ∆i(f, g) = ∆[D0]i (f, g) + X D≥D0 ∆[D+1]i (f, g) − ∆[D]i (f, g) . (3.30)” · Show in paper

Periodic physical diagnostics (variance of the sampled first horizontal cosine and mean-phase expectation) used to compare coordinates of different finite histories at a common absolute scale.

Source excerpts

“Let f be the first real horizontal cosine on the unit torus, and put g = (f,(−∆T2 )−1f) > 0, XN = (y, fN ), χN = eiy¯. (5.5)” · Show in paper

“Let XeNk be the ordinary microscopic sampled cosine, without observation averaging or noise, and define Uk = Var(XeNk ) g , Yk = Eeih¯Nk . (5.19)” · Show in paper

Degree-one partition algebra over R[ε]/(ε2)\mathbb{R}[\varepsilon]/(\varepsilon^2) that keeps the extracted magnetic scalar σ\sigma outside subsequent renormalization maps.

Source excerpts

“Extend the same exact splitting algebra to coefficients in R[ε]/(ε2). Its degree-zero terms use the background localization; its degree-one terms use the marked localization and allocation specified below.” · Show in paper

“If an earlier step has already produced a scalar σ, linearity applied to σZk(K) + Ik(K, I) gives σ+ = σ + βk, I+ = Tk(K)I. (7.17)” · Show in paper

Simultaneous comparison of two concentric dual squares of sides MM and LMLM stopped at a common intermediate block side v=Lhv = L^h with D=M/v→∞D = M/v \to \infty to cancel the magnetic core scalar exactly.

Source excerpts

“Fix a large integer j, put M = Lj, and consider the free dual squares of sides Q = M and Q = LM, with their common center at the mark. Choose the J-history of Proposition 6.3, where J is a power of two and J ≤ j < 2J.” · Show in paper

“Both ρj and σj are exactly the same for Q = M and Q = LM. The equality of σj is algebraic; its sign has not been assumed.” · Show in paper