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.

Latest version

Checks complete

Sync OpenAI Math manuscripts from 4eaee87548250b2682affda7d05d947c05554bae

Committed

All papers

Claims

Findings

Directional transience implies ballisticity

For d≥2d \ge 2 and P=ν⊗ZdP = \nu^{\otimes \mathbb{Z}^d} an iid law of nearest-neighbor transition rows satisfying uniform ellipticity, if ℓ∈Sd−1\ell \in S^{d-1} and P0(Aℓ)=1P_0(A_\ell) = 1, then there exists a deterministic vector v=v(ν)∈Rdv = v(\nu) \in \mathbb{R}^d such that v⋅ℓ>0v \cdot \ell > 0 and P0(lim⁡n→∞Xn/n=v)=1P_0(\lim_{n \to \infty} X_n / n = v) = 1.

For d≥3d \ge 3 and P=ν⊗ZdP = \nu^{\otimes \mathbb{Z}^d} an iid law of nearest-neighbor transition rows satisfying uniform ellipticity, if P0(Aℓ)>0P_0(A_\ell) > 0 for ℓ∈Rd∖{0}\ell \in \mathbb{R}^d \setminus \{0\}, then there is a deterministic nonzero vector v=v(ν)v = v(\nu) such that v⋅ℓ>0v \cdot \ell > 0 and P0(lim⁡n→∞Xn/n=v)=1P_0(\lim_{n \to \infty} X_n / n = v) = 1, and the direction sets satisfy {u∈Sd−1:P0(Au)=1}={u∈Sd−1:P0(Au)>0}={u∈Sd−1:v⋅u>0}\{u \in S^{d-1} : P_0(A_u) = 1\} = \{u \in S^{d-1} : P_0(A_u) > 0\} = \{u \in S^{d-1} : v \cdot u > 0\}.

Under the conditioned word law P0(⋅∣Dh)P_0(\cdot \mid D_h), the spatial radius RR of a regeneration word has finite mean, ER<∞\mathbb{E}R < \infty, with no moment of the word duration assumed.

There is a deterministic unit vector rr with h(r)>0h(r) > 0 such that Xn/∥Xn∥→rX_n / \|X_n\| \to r almost surely, and after a deterministic signed permutation of the coordinate axes, r1>0r_1 > 0 and all preceding word conclusions hold with height x1x_1.

For every site xx, aH(x)↓q(x)>0a_H(x) \downarrow q(x) > 0 for almost every environment, and for some λ>0\lambda > 0, Eq(0)−λ<∞E q(0)^{-\lambda} < \infty.

If the crossing probability estimate P(aN<N−β)<N−D0P(a_N < N^{-\beta}) < N^{-D_0} holds eventually for each fixed finite D0>0D_0 > 0 (for one fixed 0<β<10 < \beta < 1), then Eτs<∞\mathbb{E}\tau_s < \infty, and the walk has an almost-sure deterministic velocity with strictly positive projection on the original direction ℓ\ell.

Along every Gaussian sequence, under P0,0++P_{0,0}^{++}, for each fixed finite TT, lim⁡ρ↓0lim sup⁡r→∞1n(r)E0,0++∑0≤h≤Tn(r)1{∣Zh∣≤ρr}=0\lim_{\rho \downarrow 0} \limsup_{r \to \infty} \frac{1}{n(r)} E_{0,0}^{++} \sum_{0 \le h \le T n(r)} \mathbf{1}_{\{|Z_h| \le \rho r\}} = 0.

Under P0,0++P_{0,0}^{++}, along any Gaussian sequence, the pair of normalized first-hit processes Bir(h/n(r))=(Yh(i)−hθr)/rB_i^r(h/n(r)) = (Y_h^{(i)} - h\theta_r)/r converges functionally to two independent Brownian motions, each with variance parameter σ∗2=1/(2Lˉ)\sigma_*^2 = 1/(2\bar{L}).

Along Gaussian sequences, the quenched path under DD-conditioning from a fixed start has the Brownian limit of (16) in PP-probability, and for every ρ>0\rho > 0 and b∗(r)→∞b_*(r) \to \infty with H=⌊n(r)⌋H = \lfloor n(r) \rfloor, P(KH(0,{WH≤b∗(r)r,∣YH−bH∣≤ρr})/aH≥cρ)→1P(K_H(0, \{W_H \le b_*(r)r, |Y_H - b_H| \le \rho r\}) / a_H \ge c_\rho) \to 1.

For any fixed 0<T<∞0 < T < \infty, η>0\eta > 0, perr>0p_{\text{err}} > 0, there is d0>0d_0 > 0 such that for every sufficiently large rr, some deterministic slope srs_r satisfies P(KHtot(x,{max⁡j≤Htot∣Yj−jsr∣≤ηr})≥d0)>1−perrP(K_{H_{\text{tot}}}(x, \{\max_{j \le H_{\text{tot}}} |Y_j - j s_r| \le \eta r\}) \ge d_0) > 1 - p_{\text{err}} at every prescribed starting site xx, with Htot=⌈Tn(r)⌉H_{\text{tot}} = \lceil T n(r) \rceil.

There are fixed A→,c>0A_{\to}, c > 0 and a nonnegative random function Z(x,y)\mathcal{Z}(x,y) of the upper environment such that, simultaneously in integer H≥1H \ge 1, −log⁡pH→(x,y)≤A→log⁡(1+H)+Z(x,y)-\log p_H^{\to}(x,y) \le A_{\to} \log(1+H) + \mathcal{Z}(x,y) and sup⁡x,yEexp⁡(cZ(x,y))<∞\sup_{x,y} E \exp(c \mathcal{Z}(x,y)) < \infty, with every cap Z∧B′\mathcal{Z} \wedge B' approximable in PP-probability uniformly in x,yx,y by bounded functions of fixed-radius upper row neighborhoods.

Under the bad-environment law QN=P(⋅∣aN<N−β)Q_N = P(\cdot \mid a_N < N^{-\beta}), the episode construction satisfies clog⁡N≤EQNMc \log N \le E_{Q_N} M, EQN(M+J)≤Clog⁡NE_{Q_N}(M+J) \le C \log N, and EQN∑i<Mci1{Ji≤L0}≥clog⁡NE_{Q_N} \sum_{i < M} c_i \mathbf{1}_{\{J_i \le L_0\}} \ge c \log N for fixed constants c,C>0c, C > 0 and L0<∞L_0 < \infty.

On the positive-probability shift-invariant event E={lim⁡m→∞C0,m/m>ζ}\mathcal{E} = \{\lim_{m \to \infty} C_{0,m}/m > \zeta\} under the stationary episode limit L\mathbb{L}, there is no dust, and the proper classes form the same finite nonempty set at every time, with each class profile having full support on horizontal space.

For any stationary marked profile law with full-support subprobability profiles, nonnegative marks cic_i, positive integer marks hih_i satisfying E[c0+log⁡(1+h0)]<∞E[c_0 + \log(1+h_0)] < \infty, coordinatewise update (43), finite entropy (45) for every fixed mm, and outward-order estimate (33) with uniform local approximation, we have Ec0=0Ec_0 = 0.

Methods

True-record regeneration words in arbitrary real projected directions h(x)=(ℓ/∥ℓ∥∞)⋅xh(x) = (\ell / \|\ell\|_\infty) \cdot x using ordinary strict record indices to obtain finite mean projected width without assuming a lattice structure for projected heights.

Source excerpts

“Put h(x) = (\ell/\|\ell\|_\infty) \cdot x, so that a lattice step has height increment of absolute value at most one, and some step has increment exactly one.” · Show in paper

“Here we count ordinary record indices so that the projected heights themselves need not form a lattice.” · Show in paper

Truncated-variance scale n(r)=r2/E(S2∧r2)n(r) = r^2 / E(S^2 \wedge r^2) and deterministic median centering bjb_j for two-walk comparisons at common true cuts without assuming a finite second moment.

Source excerpts

“Choose a deterministic integer median b_j of Y_j under P_0^+, with b_0 = 0, and write W_H = \max_{0\le j\le H} |Y_j - b_j|.” · Show in paper

“For r > 0, define m(r) = \mathbb{E}(S^2 \wedge r^2), n(r) = r^2/m(r). ... The scale n(r) uses only the finite first moment of S; a finite second moment is not assumed.” · Show in paper

Gaussian-sequence two-copy Brownian limit and second-moment quenched concentration under the q(0)2q(0)^2-tilted environment law.

Source excerpts

“A sequence r -> \infty is Gaussian if n(r)\Pr(|S| > ar) -> 0 for every fixed a > 0.” · Show in paper

“Give the environment the probability law \hat{P}(d\omega) = \frac{q(0)^2}{E[q(0)^2]} P(d\omega). Under this weighting, two conditionally independent quenched conditioned paths have exactly the annealed law P0,0P_{0,0}^{++}.” · Show in paper

Clipping paths on arbitrary scales by concatenating short blocks with auxiliary dummy jumps and transferring truncated moments from iid common-word sums.

Source excerpts

“We will concatenate a fixed number of short blocks. In each block we choose a probability law on a restricted set of successful prefixes, except on a rare failure event.” · Show in paper

“The dummy jumps keep the auxiliary process defined after a failure; only trajectories without a failure will later contribute genuine quenched path mass.” · Show in paper

Outward-order kernel construction using an iterative buffer-growth schedule to bound −log⁡pH→(x,y)-\log p_H^{\to}(x,y) by A→log⁡(1+H)+Z(x,y)A_{\to} \log(1+H) + \mathcal{Z}(x,y) with an exponentially integrable local remainder.

Source excerpts

“A stage starts with an extra signed gap, or buffer, between the walks. It usually increases this buffer by a fixed factor; on failure it reduces the buffer by a smaller factor.” · Show in paper

Adaptive episode construction with stopping-layer tests and KL-entropy transfer from the fresh product law PP to the bad-environment law QNQ_N.

Source excerpts

“Suppose (5) fails along integers N -> \infty for some fixed D_0, and let Q_N be the law of the environment conditioned on a_N < N−βN^{-\beta} along this sequence. Then KL(Q_N | P) <= D_0 \log N.” · Show in paper

“We now use the stage estimate to construct the actual episode boundaries. Only the kernel tests in (35) determine those boundaries; the longer cell certificates in its proof do not enter the stopping rule.” · Show in paper

Stationary marked profile array with sampled anchors, finite-window relative entropy bounds, and a translation-invariant two-tail statistic TR(w)T_R(w).

Source excerpts

“For a full-support subprobability w on horizontal space, define T_R(w) = \max_{j \in \mathbb{Z}} \min\{w(z_2 <= j), w(z_2 >= j + R)\}, R \in \mathbb{Z}_{>= 1}.” · Show in paper

“Conditional on the entire environment, sample anchors Ui,aU_{i,a} ~ \mu_i^* independently for indices i \in \mathbb{Z} and a >= 1.” · Show in paper