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 directional zero–one law under strict ellipticity

For any integer d≥3d \ge 3, any independent and identically distributed law P\mathbb{P} of nearest-neighbor transition rows on Zd\mathbb{Z}^d satisfying strict ellipticity μ({p∈PU:p(e)>0 for every e∈U})=1\mu(\{p \in \mathcal{P}_U : p(e) > 0 \text{ for every } e \in U\}) = 1, and any nonzero real direction ℓ∈Rd∖{0}\ell \in \mathbb{R}^d \setminus \{0\}, the annealed probability of escape in direction ℓ\ell satisfies P0(Aℓ)∈{0,1}P_0(A_\ell) \in \{0, 1\}.

Under iid strictly positive transition rows, if P0(Av)>0P_0(A_v) > 0 and P0(A−v)>0P_0(A_{-v}) > 0 for some v∈Rd∖{0}v \in \mathbb{R}^d \setminus \{0\}, then the regeneration slab laws νv\nu_v and ν−v\nu_{-v} satisfy ∫R dνv+∫R dν−v<∞\int R \, d\nu_v + \int R \, d\nu_{-v} < \infty, where RR is the maximum Euclidean distance from a slab's initial site.

If the annealed walk satisfies 0<P0(Av)<10 < P_0(A_v) < 1 for some v∈Rd∖{0}v \in \mathbb{R}^d \setminus \{0\}, then there exists a coordinate i∈{1,…,d}i \in \{1, \dots, d\} for which P0(Aei)>0P_0(A_{e_i}) > 0 and P0(A−ei)>0P_0(A_{-e_i}) > 0.

When both coordinate signs have positive directional-transience probability, the expected number m(J)m(J) of dyadic common-block index intervals (2j,2j+1](2^j, 2^{j+1}], 1≤j≤J1 \le j \le J, containing a departure-site contact in the opposed arrangement of Section 4 satisfies m(J)=O(J/log⁡J)m(J) = O(J / \log J) as J→∞J \to \infty.

When both coordinate signs have positive directional-transience probability, the expected number m(J)m(J) of dyadic common-block index intervals (2j,2j+1](2^j, 2^{j+1}], 1≤j≤J1 \le j \le J, containing a departure-site contact in the opposed arrangement of Section 4 satisfies m(J)≥cJlog⁡log⁡(3+J)m(J) \ge \frac{c J}{\log\log(3+J)} for all J≥J0J \ge J_0, for some constants c>0c > 0 and J0<∞J_0 < \infty.

In the random-endpoint bridge experiment of Section 6.2 with 0<k<r<N0 < k < r < N, the first contact time τ=inf⁡{i≥Tk:i<TN, Yi∈{βj:j≥0}}\tau = \inf\{i \ge T_k : i < T_N, \, Y_i \in \{\beta_j : j \ge 0\}\} satisfies P(Tr≤τ<TN, min⁡Tk≤i≤τ(Yi)1≥k, ∥E∥≤N2)≤Clog⁡(2N)Δlog⁡2(e/Δ)P\left(T_r \le \tau < T_N, \, \min_{T_k \le i \le \tau} (Y_i)_1 \ge k, \, \|E\| \le N^2\right) \le C \log(2N) \Delta \log^2(e/\Delta), where Δ=ur−k+−uN−k+\Delta = u_{r-k}^+ - u_{N-k}^+ and the right side is zero when Δ=0\Delta = 0.

For nonnegative martingales (wi(e))i≥0(w_i(e))_{i \ge 0}, 1≤e≤m1 \le e \le m, with ∑e=1mwi(e)≤1\sum_{e=1}^m w_i(e) \le 1 almost surely, the sum of running maxima A∞=∑e=1msup⁡i≥0wi(e)A_\infty = \sum_{e=1}^m \sup_{i \ge 0} w_i(e) satisfies Eexp⁡(cA∞log⁡(m+1))≤C\mathbb{E} \exp\left(\frac{c A_\infty}{\log(m+1)}\right) \le C and E[A∞1V]≤Clog⁡(m+1)plog⁡(e/p)\mathbb{E}[A_\infty \mathbf{1}_V] \le C \log(m+1) p \log(e/p) for any event VV of probability pp on a joint space.

Methods

Transfer of finite nearest-neighbor word weights through unused departure rows to factor out nonnegative measurable functions of the outside environment.

Source excerpts

“Lemma 2.1 (Transfer through unused departure rows). Let γ = (x0, . . . , xm) be a finite nearest-neighbor word and let f(ω) be a nonnegative measurable function of the rows outside Dep(γ). Then E[∏_{j=0}^{m-1} ω(xj, xj+1 − xj) f(ω)] = W(γ) E[f].” · Show in paper

Repeated-trial column-partitioning argument for finite height supremum under strict ellipticity without uniform lower bounds.

Source excerpts

“Partition the lattice into columns, the cosets of Za. Every column has only finitely many sites with height in [−k, b]: its successive heights differ by exactly one, regardless of the other coordinates of v.” · Show in paper

Opposed regeneration-tape arrangement from a common top with a one-level offset, dividing both tapes into iid common blocks.

Source excerpts

“Set the positive top endpoint at zero. Place A1 with its terminal there, then place A2 with its terminal at the start of A1, and continue downwards. ... Place the negative tape chronologically starting from −e1, so the placed copy of Bi is translated by −e1 + ∑_{r<i} d(Br). Only translations are used; the steps inside each word retain their chronological orientation.” · Show in paper

Randomized chunk-count comparison between the opposed arrangement and independent exact-cut bridges, controlling lateral alignment cost via telescoping displacement entropy.

Source excerpts

“We randomize the number of blocks in each portion as well as in the buffer. Choosing a chunk count uniformly from n possibilities will offset the cost of specifying its random height. Randomizing the buffer count will let us compare its entropy with that of the whole experiment without paying an additional log n.” · Show in paper

Environment-separated first-contact prefix transfer combined with stopped annealed endpoint-posterior martingales and quenched exit-probability binning.

Source excerpts

“We first express the event in (37) by raw path weights, then transfer those weights to a shared environment, and only afterward introduce quenched exit probabilities.” · Show in paper