Random environments: directional escape and positive speed

Source and verification

This article explains family 220 of the October 6, 2026 OpenAI collection. Research conclusions are attributed to its manuscripts. The review covers statements, introductions, and selected proof mechanisms; it is not an independent verification of the complete proofs.

Overview

A random walk in a random environment returns to the same transition probabilities whenever it revisits a site. Family 220 claims directional zero–one laws and, under stronger assumptions, positive speed. The essential distinction is between control of spatial progress and control of the time spent making that progress.

🏷️ Preliminaries: a persistent random medium

At each $x\in {\mathbb{Z}}^d$, an environment assigns a transition row $\omega (x,e)$ for the $2d$ coordinate steps, with nonnegative entries summing to one. Given $\omega$, the walk is a Markov chain with these fixed transitions. Its law is quenched; averaging over environments gives the annealed law $P_0$.

Strict ellipticity means every allowed transition probability is positive almost surely. Uniform ellipticity means there is a deterministic $\kappa >0$ with $\omega (x,e)\ge \kappa$ everywhere almost surely. Strict ellipticity permits exceptionally small probabilities and gives no common lower bound.

For a nonzero real direction $\ell$, define

$$A_{\ell }=\{X_n\cdot \ell ⟶+\infty \}.$$

Directional transience means $P_0(A_{\ell })=1$. Ballisticity requires a deterministic limiting velocity with positive projection:

$$X_n/n⟶v,v\cdot \ell >0.$$

A path can escape while moving sublinearly, so these are different conclusions.

🏷️ Three claims with different assumptions

The iid strict-ellipticity companion claims $P_0(A_{\ell })\in \{0,1\}$ for every fixed nonzero real direction when $d\ge 3$. No common ellipticity constant or additional moment bound is imposed. (OpenAI, 2026)

The finite-range companion claims the same zero–one conclusion for stationary ergodic uniformly elliptic environments in $d\ge 3$. Finite-range dependence applies to the full transition rows: row collections separated beyond a fixed distance are independent, including collections on infinite deterministic sets. (OpenAI, 2026a)

The ballisticity companion assumes iid uniformly elliptic rows in $d\ge 2$. It claims that almost-sure directional transience implies a deterministic velocity with positive projection. In $d\ge 3$, its combination with the iid zero–one theorem allows positive escape probability to supply almost-sure directional transience. (OpenAI, 2026c)

The ballisticity assertion is not extended here to merely strictly elliptic or finite-range-dependent environments.

🏷️ Regeneration and the missing time moment

A regeneration time is a new directional record whose height the future walk never undercuts. In an iid environment, the unexplored region beyond such a record can supply independent future pieces under the appropriate nonbacktracking conditioning. These times depend on the future, so they are not ordinary stopping times; factorization must be justified rather than obtained by a direct use of the Markov property.

Suppose successive regeneration displacements and durations are iid pairs $(D_k,{\tau }_k)$, with integrable displacement and finite mean duration. The law of large numbers then suggests, and the regenerative argument makes precise,

$$v=\frac{\mathbb{E}{D}_1}{\mathbb{E}{\tau }_1}.$$

Positive mean displacement gives positive speed only after $\mathbb{E}{\tau }_1<\infty$ has been established. Finite mean spatial radius alone does not exclude long traps. This is the central additional obligation in the ballisticity paper.

🏷️ The zero–one contradiction

The iid proof assumes escape with both signs has positive probability. Fresh-row constructions first force the two signs to exhaust the possibilities. Large transverse excursions then permit insertion of a fresh oppositely escaping continuation. Repeating this would contradict the tail of a finite directional maximum, yielding finite mean spatial radius for regeneration pieces under the coexistence assumption.

The manuscript places positive and negative pieces in opposed arrangements and counts intervals of block indices containing a contact. Its two main estimates for the same expected count $m(J)$ are

$$\begin{array}{c}m(J)=O\left(\frac{J}{\log J}\right),\\ m(J)\ge c\frac{J}{\log \log J}.\end{array}$$

Their ratio is incompatible for large $J$. Entropy comparisons produce the upper bound; endpoint alignment and posterior martingale estimates supply the lower bound.

With finite-range dependence, disjoint paths need not use independent rows. The extension inserts marked steps of fixed positive probability, carrying the walk beyond the dependence range without revealing intervening row data. Independence is then applied beyond that buffer. Contacts are enlarged accordingly. Uniform ellipticity pays for these marked connectors.

🏷️ From escape to speed

After obtaining spatial control, the ballisticity paper studies the quenched slab-crossing probability $a_N=P_{0,\omega }(T_N<T_{-1})$. Its decisive target is that $\mathbb{P}(a_N<N^{-b})$ decays faster than every power of $N$, for a fixed $0<b<1$. This controls the duration of successful forward excursions.

Two-copy walk comparisons, layered endpoint distributions, and entropy bounds are used to rule out environments with persistently poor forward crossing. The exposure order keeps each adaptive test below its next top layer. That measurability restriction is what allows fresh-environment reasoning to survive the multiscale construction.

🏷️ Key steps and proof outlines

A useful organizing principle is to record exactly which transition rows a path has used. Arrival at a site does not use its row; departure does. The three companions retain this distinction when building independent pieces, conditioning on bridges, and exposing the upper environment.

A buffer can be crossed without revealing its rows

For the finite-range model, let $\kappa$ be the ellipticity constant and choose $2\lambda <\kappa$. Split each transition into three marked alternatives:

$$\begin{array}{rl}{P}_{\omega }(e,0)& =\lambda ,P_{\omega }(e,1)=\lambda ,\\ P_{\omega }(e,\ast )& =\omega (x,e)-2\lambda .\end{array}$$

Summing over marks recovers $\omega (x,e)$. Only departures marked $\ast$ use environmental information. For a prescribed marked word $\gamma$, its probability is a product depending solely on these active rows. It therefore factors from any row event more than the dependence range $R$ away from that deterministic active set.

Normalize a real direction $v$ by $\Vert v{\Vert }_{\infty }=1$ and choose a coordinate step raising height $v\cdot x$ by one. After a record, prescribe $M>dR+1$ such steps, the first marked zero and the rest marked one. Their probability is the constant ${\lambda }^M$. Earlier active rows have height below the record, while a continuation constrained never to fall below its height at the end of the script uses rows at least $M$ higher. Since $|v\cdot (x-y)|\le d\Vert x-y{\Vert }_{\infty }$, these row sets are separated by more than $R$. The two marks prevent overlapping occurrences of the prescribed script.

This proves the required factorization after first fixing finite words and only then summing them. It does not assert independence of arbitrary disjoint random path ranges (OpenAI, 2026a).

Coexistence forces a spatial first moment

In the iid zero–one argument, suppose both directional signs have positive escape probability. Regeneration slabs already have finite mean projected width; their full radius $R$ still needs control. If its mean were infinite, the truncated mean $I(t)={\sum }_{\pm }{\mathbb{E}}_{\pm }\min (R,t)$ would diverge while satisfying $I(t)=o(t)$.

Choose $N=N(a)$ minimally so $I(aN)\ge \eta a$. Concavity then gives $\eta a\le I(aN)<2\eta a$ for large $a$. A stationary maximal inequality controls all initial averages of truncated radii:

$$P\left(\underset{k\le n}{\max }\frac{1}{k}\sum _{j=1}^k{Z}_j>a\right)\le \frac{\mathbb{E}{Z}_1}{a}.$$

Its proof selects disjoint witness intervals covering all bad starting indices, bounds their total length by total mass divided by $a$, and averages by stationarity. Together with a union bound for the truncation tail, it makes most prefixes spatially controlled. The divergent truncated mean nevertheless forces a large excursion after such a prefix. A slight tilt of the directional functional turns that excursion into a fresh record. An oppositely escaping continuation can then be inserted beyond the old rows with a probability bounded below. This would give a fixed positive probability that the final directional maximum is finite but arbitrarily large, contradicting its tail tending to zero. Thus the radius mean is finite; no duration moment has been used (OpenAI, 2026b).

Entropy and contacts give incompatible scale counts

Arrange positive and negative slab sequences facing each other. Group them into common blocks and let $m(J)$ count, in expectation, how many of the first $J$ dyadic block-index intervals contain a departure-site contact.

For the entropy upper bound, surround a middle chunk by two outer chunks and an independent buffer. Randomize each chunk’s number of blocks over a window of order its scale. Compare the actual arrangement with independent exact-height bridges while retaining the actual joint law of heights and endpoint alignment. This preserves the contact event being tested. The comparison cost is charged to increments in the entropy of sums of iid lateral displacements. Finite first moment bounds that entropy by $C+(d-1)\log (k+1)$ after $k$ blocks. Rare contacts under the reference bridge law and the entropy comparison then yield $m(J)=O(J/\log J)$.

For the lower bound, condition a positive bridge to end at depth $N=2^{2J}$. If a run of depth scales contains no contact, every regeneration cut in a suitable interval must have its first subsequent contact very near the bridge’s top. The source’s posterior first-contact estimate makes that unlikely. Choose a run length

$$G=\lceil 4{\log }_2\log (2+J)\rceil +3.$$

The factor $2^{-G}=O((\log J{)}^{-4})$ absorbs logarithmic losses without requiring a quantitative regeneration-tail rate. Averaging over the many cuts shows that all but $o(J)$ tested scale labels lie within $G$ of a contact label. Each contact covers at most $G+1$ labels, giving $\gg J/(G+1)$ contacts. Finite mean common-block width converts depth scales to block-index scales with bounded shifts. Thus $m(J)\gtrsim J/\log \log J$, contradicting the entropy bound.

Slow crossing would produce a stationary loss process

The ballisticity paper has a separate final contradiction. Suppose the desired superpolynomial bound for $P(a_N<N^{-b})$ fails. Conditioning on these bad environments costs only $O(\log N)$ relative entropy. Adapted layer episodes then accumulate a positive mean logarithmic loss $c_i$ of surviving forward mass, with heights $h_i$ satisfying

$$\mathbb{E}[c_i+\log (1+h_i)]<\infty .$$

Average the law over episode starts. Shifting that average changes only its two endpoints, an error tending to zero, so a subsequential limit is stationary. The state records endpoint profiles, sampled-anchor offsets, and upper environments; it retains diffuse mass and separated components until finite-window entropy estimates exclude those possibilities on the positive-loss event. A surviving component is selected uniformly among the finitely many persistent components, which preserves stationarity.

For its full-support subprobability profile $w$, define

$$T_R(w)=\underset{j}{\max }\min \{w(z_2\le j),w(z_2\ge j+R)\}.$$

This measures two tails separated by $R$. The kernel update and outward-order estimate imply, over $m$ episodes,

$$\begin{array}{r}\log T_R(w_m)-\log T_R(w_0)\ge C_m-A\log (1+H_m)-Z_R-\log 2,\end{array}$$

where $C_m={\sum }_{i<m}{c}_i$, $H_m={\sum }_{i<m}{h}_i$, and $Z_R$ is a controlled environmental remainder. For fixed $m$, send $R\to \infty$: finite-entropy truncation localizes the biased density near the origin, while the two sampled tails leave that neighborhood. Independence under the reference iid field then bounds the remainder by a constant independent of $m$.

Stationarity makes the expected left side zero, justified by clipping the logarithms before removing the cutoff. Hence $m\mathbb{E}{c}_0\le A\mathbb{E}\log (1+H_m)+O(1)$. The logarithmic moment makes the right side $o(m)$: bound it by $\log m+{\max }_{i<m}\log (1+h_i)$ and truncate the maximum. This contradicts positive mean loss. The resulting crossing-tail estimate supplies the finite mean regeneration duration, and the regenerative law of large numbers gives the claimed positive velocity (OpenAI, 2026c).

See Also

• Family 212 studies geodesics chosen by optimization in an iid medium.
• Family 211 highlights a related distinction between convergence of geometry and control of a walk’s elapsed time.

🏷️ Manuscripts and verification scope

• A directional zero-one law for finite-range-dependent random environments — pinned manuscript and companion source.
• A directional zero–one law under strict ellipticity — pinned manuscript and companion source.
• Directional transience implies ballisticity — pinned manuscript and companion source.

The family scope page describes selected formalization coverage. The scope page selects the iid strict-ellipticity zero–one law and the iid uniformly elliptic ballisticity claim. It does not include the finite-range-dependent extension. No Lean build, axiom audit, or manuscript-to-formal-statement equivalence check was performed for these notes.

OpenAI collection provides prerequisites and the reading sequence; coverage register records completed and pending families.

🐻  OpenAI 2026a. A directional zero–one law for finite-range-dependent random environments.

🐻  OpenAI 2026b. A directional zero–one law under strict ellipticity.

🐻  OpenAI 2026c. Directional transience implies ballisticity.