Gaussian Moat Conjecture: How OpenAI’s Proof Solves the 64-Year Walk-to-Infinity Mystery

Imagine stepping from prime to prime across the complex plane, always keeping the next jump below the same fixed limit. OpenAI’s September 26, 2026 manuscript answers that challenge negatively: the Gaussian moat conjecture admits no infinite walk through distinct Gaussian primes with bounded steps. Its stronger conclusion caps every connected component, wherever the walk starts. The surprise is how it gets there. Geometry supplies differences, arithmetic separates them, and information theory shows that avoiding composite points would demand more information than bounded movements can carry.

1. Gaussian Moat Conjecture Solved: What Exactly Did OpenAI Prove?

Bounded-Step Walks on Gaussian Primes states a proof of the uniform Gaussian moat conjecture. For each finite real step bound $D$, one finite number $B_D$ bounds all components of the Gaussian prime graph. The manuscript belongs to OpenAI’s October mathematics release, attributed to an internal frontier model.

1.1. Gaussian Moat Conjecture: Key Facts

“Fixed” matters. You may choose an enormous jump allowance before starting, but cannot keep increasing it indefinitely. The conclusion concerns distinct visited primes. Repeatedly shuttling between two nearby primes is no counterexample. The theorem is stronger than saying that each particular starting point eventually traps you.

2. What Are Gaussian Primes, and How Can You Walk Between Them?

Gaussian integers form $\mathbb Z[i]=\{a+bi:a,b\in\mathbb Z\}$, with $i^2=-1$. Their norm is $N(a+bi)=a^2+b^2$, and Euclidean length is $|z|=\sqrt{N(z)}$. A Gaussian prime is irreducible: any factorization has a unit factor, one of $1,-1,i,-i$.

For example, $2+i$ is prime because its norm is the ordinary prime $5$. The integer $3$ is also Gaussian prime, despite having norm $9$. By contrast, $5=(2+i)(2-i)$ is reducible. Ordinary primes congruent to $1$ modulo $4$ split into conjugate factors. Those congruent to $3$ remain irreducible, while $2$ ramifies.

A legal short walk is $2+i,3,3+2i$. Its jumps have lengths $\sqrt2$ and $2$, so $D=2$ permits both. All three vertices are prime and distinct. Distance, rather than the difference between their norms, determines legality.

3. The Original 1962 Gaussian Moat Problem: Can a Prime Walk Continue Forever?

The manuscript traces the question to Basil Gordon at the 1962 International Congress of Mathematicians. Erdős later credited Gordon and Motzkin, recalling a conversation with Motzkin in 1963. That history explains the 64-year framing.

Define $G_D$ by joining distinct Gaussian primes whose distance is at most $D$. A self-avoiding walk never revisits a vertex. The question asks whether some fixed $D$ permits an infinite such walk.

On the ordinary prime number line, a sufficiently long composite interval blocks progress outright. In two dimensions, sideways escape remains possible. The connectivity language resembles percolation theory, but this graph is determined by arithmetic. Its vertices are not independently retained at random.

4. Why Previous Attempts Could Not Prove the Gaussian Moat Conjecture

Earlier work constructed arbitrarily large prime-free disks and arbitrarily isolated real Gaussian primes. These are genuine obstructions, but an empty disk is a lake, not necessarily a coastline extending across every possible route.

Tsuchimura’s computation bounded the region reachable with jumps of length at most six from the origin, added as an initial vertex. That controls one component, not every starting prime for every bound.

Gethner and Stark established periodic obstructions for step bounds $\sqrt2$ and $2$, while Vardi studied periodic coprimality graphs and uniform component bounds. The missing ingredient was a finite collection of congruence restrictions that works for every fixed $D$. Searching farther from one origin could not supply that universal quantifier.

5. Theorem 1.1: Why Every Gaussian Prime Component Has a Uniform Finite Bound

In mathematical form, the main assertion is:

Here $\#C$ counts vertices. The same $B_D$ works for every starting prime, including axis primes and all associates, meaning multiplication by a unit. Consequently, every injective bounded-step walk has at most $B_D$ terms, hence at most $B_D-1$ steps.

This uniformity is substantial. A graph can have finite components of sizes one, two, three, and so on without any infinite component. The theorem excludes that possibility here. For $D<1$, distinct lattice points cannot be adjacent, so $B_D=1$ suffices. For larger bounds, the argument establishes existence without providing a numerical estimate.

6. The Finite Sieve Obstruction: How OpenAI Restricts Every Possible Walk

Choose finitely many ordinary primes $p\equiv1\pmod4$, each splitting as $p=\pi_p\overline{\pi_p}$. Reduction modulo either Gaussian factor gives a field with $p$ elements. These residue fields are related to the algebra discussed in our prime-field factorization explainer.

For a finite selection $P$, define the avoiding set:

Theorem 1.2 supplies one finite set $P_D$, depending only on $D$, whose avoiding set contains no infinite self-avoiding bounded-step lattice walk. Importantly, the surviving points need not themselves be prime. Blocking this larger set blocks prime walks too, apart from finitely many exceptions.

That finite sieve obstruction is the real engine. The proof must establish that suitable primes exist uniformly over all hypothetical walks. Choosing different restrictions after seeing each walk would not establish the theorem.

7. Why Periodicity Converts the Sieve Into a Uniform Component Bound

Put $Q=\prod_{p\in P}p$. Translating by either coordinate multiple of $Q$ preserves the avoiding set. This is the repeating-cell idea familiar from periodic lattice tilings.

Each vertex has finitely many bounded-length neighbors. An infinite component would therefore contain an infinite simple path, by the finitely branching tree of finite paths. The sieve theorem rules that out.

Now suppose two points in one finite component agree modulo $Q$ in both coordinates. Translating by their nonzero difference maps that component onto itself. Repeating the translation would produce infinitely many points, a contradiction. Each component therefore injects into the $Q^2$ residue cells.

Let $E$ contain the associates of the selected Gaussian factors, and let $K_D$ count lattice vectors of length at most $D$, including zero. Deleting $E$ puts every remaining Gaussian prime inside the avoiding graph. Restoring those exceptions gives Proposition 2.1:

Only finitely many pieces can attach to the exceptional primes. This transfers the sieve’s bound to the entire prime graph.

8. Torus Geometry: How Long Walks Generate Enough Distinct Differences

To prove the sieve exists, fix any deterministic infinite self-avoiding lattice walk. For a segment of $n$ steps, temporarily use $E$ for its visited points. Let $R$ be its diameter and $W$ its greatest distance from a diameter line, increased to at least one. Set $A=RW$.

Lemma 4.1 establishes:

Here $E-E$ contains distinct displacement vectors, and the positive constants have the dependencies indicated by their subscripts. Packing disjoint small disks around visited lattice points proves the first lower bound.

The difference bound needs topology. When $W>1$, diameter endpoints and a farthest point generate a lattice of area $A$. Subtract two interpolated walk paths modulo that lattice. The resulting torus map has degree minus one, so it must cover the target torus. Every image lies within $2D$ of a vertex difference. Disks around those differences must therefore cover area $A$.

The sampling rule chooses a distinct difference uniformly, selects a representative endpoint pair, then flips a coin between its endpoints. Uniformly sampling pairs instead would overweight frequently repeated differences.

9. Random Signed Prime Factors and the Growth of Residue Entropy

For each selected split prime, choose one conjugate factor with a fair coin. “Signed” refers to that choice. The walk remains deterministic. Randomness belongs to the proof’s measurements.

Lemma 4.2 controls collisions after reduction modulo products of chosen factors. If nearby sign choices have witnesses in a sufficiently small rectangle, common divisors force their determinant to be divisible by a number larger than its possible nonzero magnitude. The witnesses must be collinear.

More precisely, write $P=\prod_{j=1}^r p_j$ for distinct split primes in $[T,2T]$, with $T\ge5$. For a rectangle with half-lengths $R_1\ge W_1\ge1$ satisfying $R_1W_1\le P^{1-d}$, where $0<d<1/4$, the probability of a nonzero multiple of the signed product inside it is at most $Ce^{-cd^3r}$, provided $d^3r$ is sufficiently large. That turns geometric separation into a quantitative arithmetic estimate.

For a fixed direction, Hoeffding concentration makes such a witness unlikely. Expansion in the Boolean cube of sign choices then controls all directions together: distinct direction classes have disjoint enlarged neighborhoods. This avoids an uncontrolled count of possible lines.

Define $f_Z(s)$ as the average joint residue entropy of a sampled position $Z$ over signed subsets of size $s$. All logarithms are natural. With $L_*$ the batch’s mean logarithmic prime size, suitable input and output sizes $q,b$ satisfy Lemma 4.3:

Repeated enrichment halves the entropy deficit up to controlled error. Entropy is a rigorous uncertainty measure, as in the Courtade–Kumar information problem, rather than a synonym for visual disorder.

The sizes are $q=\lfloor U/L_*\rfloor$ and $b=\lfloor g^2U/L_*\rfloor$, with segment length $\lfloor e^{gU}\rfloor$. Large scales and small $g$ make the errors controllable while matching consecutive sampling operations.

For probabilities $p_x$, entropy is $H(X)=-\sum_xp_x\log p_x$, with zero-probability terms contributing zero. A uniform variable with $p$ possible values has entropy $\log p$. Conditional mutual information, $I(X;Y\mid Z)$, measures how much learning $Y$ reduces uncertainty about $X$ after $Z$ is already known.

10. From High Entropy to Residue Coverage: The Proof’s Critical Bridge

High joint entropy does not ensure that each residue gets useful probability. A distribution can have nearly maximal entropy while entirely omitting some symbols. The needed coverage statement is:

Thus almost every residue has probability at least $\tau/p$, with an explicitly controlled exceptional set. The proof obtains this by transferring coverage backward through checkpoints, although every sampled walk time moves forward.

Fix an exact later checkpoint. Run independent continuations from it, recording their displacements together in one data vector $H$. If terminal coverage fails, candidate starting residues can be tested against the deficient set. The true residue lands on a short candidate list with appreciable probability. Coverage from that exact checkpoint makes most false candidates unlikely to survive repeated trials.

The short-list inequality then forces a conditional entropy deficit. But high joint entropy and the small entropy of $H$ cannot support such deficits across many coordinates. Crucially, the shared displacement vector is charged once, not separately for every residue.

The quantitative principle is simple. If an alphabet has size $a$ and data identify a list of size at most $b$ containing the true value with probability $\rho$, then:

Here $h$ is binary entropy and $Y$ is the observed data. A short successful list reveals information even when success depends on the hidden value. Independence is required for repeated continuations only after fixing their exact checkpoint, not before that conditioning.

10.1. What the Sampling Schedule Supplies

The schedule uses decreasing geometric scales in widely separated windows. Later movements have small displacement entropy compared with information accumulated earlier. An independent uniform time offset makes the common terminal law nearly invariant under the short shifts used in block comparisons. No stationarity assumption is imposed on the original walk.

At the initial checkpoint, at least 99% of signed factors receive terminal coverage with probability floor $1/(4p)$ outside fewer than $p^{199/200}$ residues. These are auxiliary sampling guarantees, not measured properties of randomly generated prime walks.

11. The Final Contradiction: Why Zero Avoidance Costs More Information Than a Walk Can Carry

Assume the walk avoids every selected zero class. A short increment word tests candidate starting residues: a candidate is impossible if its translated word hits zero. Self-avoidance makes the hit opportunities disjoint when $DL<\sqrt p$, since a nonzero multiple of a factor of norm $p$ has length at least $\sqrt p$.

Coverage therefore makes repeated tests eliminate most candidates. The manuscript uses dyadic lengths $L_T$ comparable to $T^{2/5}$ for primes near $T$. After accounting for displacement entropy, Proposition 7.1 charges conditional mutual information of at least $c_1/\log T$ per increment to each batch.

Let $X’_T$ be that batch’s terminal residue vector and $O’_T$ contain earlier batches. Write $J_T$ for the conditional mutual information $I(X’_T;W_{L_T}(t_*)\mid O’_T)$, averaged over auxiliary choices. Nested residue vectors, dyadic block subadditivity, and smoothing yield:

Here $W_{L_T}$ is the increment word, $C$ records auxiliary prime choices, and $o(1)$ tends to zero as the schedule grows. The paper connects this budgeting method to Tao’s entropy-decrement argument, also relevant background to the two-point Chowla story.

Each increment has only $K_D$ possible values. Yet the selected prime windows force:

Each window contributes a fixed positive amount, and the number of windows grows without bound. The right side stays finite. Uniformity over walks lets the proof choose one sufficiently large finite schedule, giving $P_D$, then Theorem 1.1 through periodicity.

Why does a window contribute enough? Its logarithmic width grows in proportion to its location. Geometrically spaced batches therefore supply proportionally more terms, each proportionally smaller. Their aggregate stays bounded below rather than disappearing as the primes become enormous.

The bound remains nonexplicit. This is not a practical maximum-walk calculator. Developers can illustrate finite components, but a large simulation cannot replace the uniform argument.

The official Lean scope reports formalization of the no-infinite-walk result and uniform component bound, including axis primes and associates. That concerns the encoded theorem, not separate verification of every paragraph of the 29-page exposition. We did not rebuild the Lean project. The distinction between discovery and formalization also guides our Fermat proof explainer.

The Gaussian moat conjecture’s resolution turns a local jump restriction into a global information contradiction. For more explanations that follow the actual mathematical mechanism, explore Binary Verse AI’s AI mathematical proofs coverage and follow the research series.

12. Gaussian Moat Conjecture: Frequently Asked Questions