Honeycomb walks: strip estimates and the three-quarter exponent

Source and verification

This article explains family 237 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

The thirteen companions in family 237 build toward the three-quarter diameter exponent for honeycomb self-avoiding walks. They move through strip crossings, marked polygons, cylinder estimates, and renewal comparisons. The probability law must be specified at each stage: fixed length, critical all-length weighting, and polygon conditioning are different ensembles.

🏷️ Preliminaries: walks, critical weights, and bridges

A self-avoiding walk visits no vertex twice. Let ${\mathcal{W}}_n$ be the length-$n$ honeycomb walks starting at a fixed vertex, with endpoint free, and give them uniform probability. For a path $\gamma$, write $D(\gamma )$ for its Euclidean diameter.

The honeycomb connective constant is $\mu =\sqrt{2+\sqrt{2}}$, so the critical activity is $\rho ={\mu }^{-1}$. In an all-length ensemble, a path of length $\ell$ receives unnormalized weight ${\rho }^{\ell }$. A mass is a sum of these weights before division by a partition sum.

A strip bridge starts at one boundary and ends at the other without otherwise touching either boundary. A returning path is an arch. For port-to-port paths, ports are edge midpoints and the source counts visited vertices; this differs from the usual edge length by a fixed endpoint convention.

A polygon is a simple closed path, often counted unrooted and modulo translations. Multiplying polygon weights by their length changes the law.

🏷️ The fixed-length claim

Diameter, local mass, and covering exponents

The fixed-length companion claims that, for every positive exponent slack $\delta$ and every polynomial failure target $n^{-k}$, all sufficiently large integers $n$ satisfy, with probability at least $1-C_{\delta ,k}{n}^{-k}$,

$$n^{3/4-\delta }\le D(\gamma )\le n^{3/4+\delta }.$$

It also claims simultaneous local-mass and covering bounds with exponent $4/3$, over all centers on the path and all radii $1\le s\le n$. (OpenAI, 2026)

Writing $M_{\gamma }(z,s)$ for the number of visited vertices in a radius-$s$ ball, the local quantity is bounded above and below by $n^{\pm \delta }\min \{n,s^{4/3}\}$. The minimal number of such balls covering the trace is bounded by $n^{\pm \delta }(1+n{s}^{-4/3})$.

These statements describe power exponents with arbitrary slack. They do not claim an SLE₈/₃ scaling limit, a limiting diameter distribution, or a pointwise asymptotic with a fixed amplitude.

🏷️ Strip mass and length normalization

The strip-mass companion claims total bridge mass $B_h\asymp h^{-1/4}$, with the source port fixed and the terminal port summed. It also identifies an arch displacement moment of order $h^{3/4}$. (OpenAI, 2026a)

The sharp finite-moment companion gives

$$\begin{array}{c}{B}_h\asymp h^{-1/4},\\ \sum _{\gamma \text{ bridge}}\ell (\gamma ){\rho }^{\ell (\gamma )}\le C{h}^{13/12},\end{array}$$

and a fixed fraction of bridge mass on lengths at least $c{h}^{4/3}$. Dividing the first-length mass by $B_h$ gives the upper mean exponent $13/12+1/4=4/3$; the mass-fraction statement gives the matching lower order. (OpenAI, 2026c)

The polynomial-vacuum and disk-transfer companions establish all-length bridge mean laws $h^{4/3+o(1)}$, central-visit estimates, and versions confined to a fixed multiple of the strip width. Confinement is extra information: a strip-crossing estimate alone does not keep a bridge inside a corridor where another path could be modified safely. (OpenAI, 2026c) (OpenAI, 2026b)

🏷️ Chords, cylinders, and marked polygons

Several companions supply the geometric inputs used in changing ensembles:

• The prescribed-endpoint paper treats macroscopically separated boundary ports, half-plane arches, and specified cuts. Some additional strip endpoint statements use a density-one set of heights or a weighted mixture of heights, rather than every fixed height. (OpenAI, 2026a)
• Cylinder loop weights identify a $1/6$ partition exponent at loop fugacity two and a $1/12$ planar nesting exponent. The signed-propagation companion also treats free-endpoint hexagonal chords and a truncated polygon second moment. (OpenAI, 2026c) (OpenAI, 2026g)
• The cap-selected paper obtains triangle chord partition mass comparable to $R^{3/4}$, mean length $R^{4/3+o(1)}$, and a cylinder amplitude selected by its boundary caps. Both chord endpoints are summed. (OpenAI, 2026a)
• The marked-correlation companion controls squared-length polygon mass through two marks and a one-arc estimate uniform for unequal marked intervals. The logarithmic-window companion retains bridge mass within a specified range around the $h^{4/3}$ scale, with logarithmic losses. (OpenAI, 2026f) (OpenAI, 2026d)
• The radial-transfer companion claims polygon diameter-tail exponents $-2$ and $-2/3$ under unweighted and length-weighted critical polygon measures. Its diameter conclusion conditions on length at least $n$, rather than exactly $n$. (OpenAI, 2026j)

These are distinct inputs with different rooting and endpoint conventions. Their exponents cannot be moved between laws solely by matching a scaling heuristic.

🏷️ Renewal and the passage to every length

The renewal companion constructs infinite irreducible-bridge and thermally discounted laws, derives spatial exponent $3/4$, and treats all sufficiently large compatible even bridge lengths. For unrestricted walks, one endpoint lower law is stated only on a common natural-density-one set of lengths. It also claims a small-force free-energy exponent $4/3$ and a near-critical correlation scale. (OpenAI, 2026k)

The later fixed-length paper has a stronger all-integer diameter conclusion. It imports precise positive estimates from the analytic companions and supplies a geometric transfer. Smaller bridges are sewn into localized corridors with recoverable sewing lines, so the multiplicity of the operation is controlled. First and second moments give a positive chance of a long confined bridge; renewal trials amplify that chance.

Ordered tests on neighboring renewal paths then suppress excessively fast travel and excessive local accumulation of length. The proof must transfer these controls to arbitrary subpaths of an unrestricted fixed-length walk and retain polynomially small exceptional probabilities. This is the missing work in the informal inference “length scales as $R^{4/3}$, therefore diameter scales as $n^{3/4}$.”

🏷️ Verification boundary

The documented formalization selections cover supporting summability of bridge measures, existence of a force-dependent limiting free energy, and strip mass and balance estimates. They do not document the complete all-length or fixed-length three-quarter spatial theorem. The thirteen manuscripts’ claims remain source-attributed throughout this article.

🏷️ Key steps and proof outlines

Positive integrals make the strip exponent accessible

The strip-crossing paper first obtains exact polynomial and Pfaffian expressions. A quotient of signed expressions is not yet an asymptotic estimate: its denominator could vanish when parameters coincide. Its positive-integral step proves nonsingularity of the homogeneous problem and represents the arch moment using ordered points $0<y_1<\cdots <y_N<1$ with density proportional to

$$\prod _i\omega (y_i)\prod _{i<j}\frac{(y_j-y_i{)}^2}{y_i+y_j},\omega (x)=\frac{1}{2}\sqrt{\sqrt{\frac{1+x}{2x}}-1}.$$

The normalizer is finite and positive: $\omega (x)=O(x^{-1/4})$ at zero and the pair factors are bounded on the ordered region. In the integral formula, the relevant factor is $1-R(x)$, where $R(x)={\prod }_i(y_i-x)/(y_i+x)$. Each individual ratio is in $[-1,1]$, so

$$0\le 1-R(x)\le 2\min \left\{1,x\sum _i{y}_i^{-1}\right\}.$$

On $y_1/3<x<y_1/2$, all factors are nonnegative and the first is at most $1/2$, giving a lower bound $1-R(x)\ge 1/2$. Integrating these two bounds against the kernel $\omega (x)/x$ yields

$$\mathbb{E}{y}_1^{-1/4}\lesssim m_N\lesssim \mathbb{E}{\left(\sum _i{y}_i^{-1}\right)}^{1/4}.$$

Thus the arch moment is controlled by particles nearest zero. The source compares this positive measure with a Laguerre-type reference law; its inverse moments and rescaling yield $m_N\asymp N^{3/4}$. A key comparison check is

$${\partial }_x{\partial }_y\log \frac{(y-x{)}^2}{x+y}=\frac{2}{(y-x{)}^2}+\frac{1}{(x+y{)}^2}>0.$$

Conditioning successively on the largest coordinate proves positive association. A monotone likelihood ratio therefore moves increasing observables in the corresponding direction. This supplies a measure-level comparison rather than taking absolute values of a signed Pfaffian. The strip balance identities then connect the arch moment to $B_h\asymp h^{-1/4}$. (OpenAI, 2026c)

Recover a physical lower bound from a complex upper estimate

The cylinder-loop paper’s contour analysis supplies a locally uniform upper exponent for $Z_N(2\cos \eta )$. Its normalization matters: the contributions are

$$\frac{7}{8}-\frac{1}{2}-\frac{5}{24}=\frac{1}{6}.$$

These come respectively from a deformation determinant, two reference masses and the scalar normalization. The resulting upper function is harmonic in the complex parameter:

$$\frac{\log |Z_N(2\cos \eta )|}{\log N}\le \mathcal{H}(\eta )+o(1),\mathcal{H}(\eta )=\frac{1}{6}-\frac{2}{3{\pi }^2}\mathrm{Re}({\eta }^2).$$

At $\eta =\pi /2$, zero loop fugacity leaves only the empty configuration, so both sides equal zero. The nonnegative superharmonic deficit propagates smallness along a finite chain of disks by the area mean inequality. This yields some nearly saturating point in each target disk, not a pointwise lower bound throughout it. Nonnegative coefficients finally give $|Z_N(z)|\le Z_N(|z|)$, transferring those points to a lower bound at positive real fugacity. Letting the disks approach the desired parameter matches the upper exponent. (OpenAI, 2026e)

The sharp finite-moment companion strengthens normalization control using a positive Gram determinant at the zero-loop parameter. If ${\Pi }_m$ is its polynomial projection and the multiplier $R\ge 1$, Jensen’s inequality on an eigenbasis gives

$$\log \det ({\Pi }_{m}R{\Pi }_m{|}_{\mathrm{ran}{\Pi }_m})\ge \mathrm{Tr}({\Pi }_m\log R{\Pi }_m).$$

A nonnegative integral and Fatou’s lemma identify the lower logarithmic correction $5/24$. Subharmonic propagation and positivity then identify the other correction $3/8$, whose difference is again $1/6$. These normalization calculations are inputs to its later bounded-factor nesting and finite bridge estimates. (OpenAI, 2026m)

Signed transfer must retain the boundary coefficient

A transfer eigenvalue alone does not determine the physical partition amplitude. The signed-propagation paper isolates an abstract coefficient argument. A compact exterior operator has finitely many eigenvalues $\zeta$ of modulus one; on each generalized eigenspace,

$$E^n{P}_{\zeta }={\zeta }^n\sum _{j=0}^{q_{\zeta }-1}\left(\genfrac{}{}{0ex}{}{n}{j}\right){\zeta }^{-j}{J}_{\zeta }^j{P}_{\zeta }.$$

Here $P_{\zeta }$ is the spectral projection and $J_{\zeta }$ its nilpotent part. Exponentially summable boundary perturbation words permit termwise coefficient extraction, leaving a finite sum of unit-modulus exponentials times polynomials in $n$, plus a vanishing remainder. If that sum tends to zero, divide by its highest polynomial power, multiply by a conjugate frequency, and take Cesàro averages. Distinct frequencies average to zero; repeating by descending degree kills every coefficient. No diagonalizability assumption is required. (OpenAI, 2026l)

The cap-selected paper instead compares poles of a finite-size generating function. A generic ordinary trace term has a nonzero isolated pole. The cap estimate allows only one pole in the relevant disk, so both select the same eigenvalue and residue. Normalize at the vacuum, where the cap coefficient is one, then continue analytically to the homogeneous physical parameters. Its spectral gaps are established for each fixed circumference; uniformity in circumference belongs to the subsequent asymptotic estimate. This separation prevents a boundary amplitude from being silently replaced by an eigenvalue estimate. (OpenAI, 2026a)

Turn nesting into actual visits and length moments

The disk-transfer paper proves a finite recursion, so no infinite-volume path law is needed. Give each nonbacktracking continuation the critical complex weights $\rho e^{\pm i\pi /8}$, whose sum is one. Stop exploration at an exit or first repeated vertex. A repeated loop attached from the exterior has two orientations whose phases sum to zero. A loop enclosing the marked starting vertex instead contributes a negative factor; moving it to the other side produces a positive recursion for the signed excursion mass through that vertex.

The remaining through-vertex coefficients are $-\cos (5\pi /8)$ and $-\cos (7\pi /8)$, both positive. Unrolling the recursion through strictly smaller interiors expresses the visit observable as a positive sum over nested polygon chains. Comparing the final local factors gives comparability with the sum of nesting partition functions around the three incident faces. On convex domains, the winding factors are also bounded between fixed positive constants, converting this signed observable to unsigned visit mass. Filling the domain is essential to the exterior-loop cancellation. (OpenAI, 2026f)

The polynomial-vacuum companion then distinguishes arches from bridges using a marked tripod and translated-bridge intersection comparison. Its central-strip bridge visit mass is $h^{1/12+o(1)}$. There are order $h$ horizontal-translation classes of vertices; translating each bridge source to one fixed port makes the sum of these visit masses exactly its first-length mass. Thus

$$M_h=h^{13/12+o(1)},M_h/B_h=h^{4/3+o(1)}.$$

This division changes an unnormalized moment into a mean under the height-conditioned law. (OpenAI, 2026i)

For triangle chords with both endpoints summed, the cap-selected paper instead has order $R^2$ central contacts. Each has visit mass $R^{1/12+o(1)}$, giving length numerator $R^{25/12+o(1)}$. Its denominator is $R^{3/4}$, so their ratio again has exponent $4/3$. The matching exponent arises from explicitly different numerator and denominator counts.

Sewing needs recoverable seams and controlled multiplicity

To insert or close a path, one must keep the added pieces disjoint and bound how many inputs yield the same output. The fixed-length paper first localizes bridge mass: averaging a boundary-flux comparison over order $M$ sources, reversing side exits, and using $B_M\asymp M^{-1/4}$ bounds lateral loss by $Ch{M}^{-5/4}$. Choosing $M$ a sufficiently large multiple of $h$ preserves a fixed fraction of mass in a usable corridor.

A seam crossed once splits a bridge into two bridges; a twice-crossed seam cuts off one arch excursion. Ordered seam moments bound reconstruction multiplicity. The logarithmic-window paper illustrates the count sharply. If $M_g(P)$ counts eligible double cuts of a polygon at gap scale $g$, selecting $r$ distinct cuts contributes: $O(g)$ choices at the first cut, two exterior arch factors $O(g^{-5/4})$, and $r-1$ intermediate sums ${\sum }_{j\le Cg}{B}_j^2=O(g^{1/2})$. Hence

$$\sum _P{\rho }^{|P|}{M}_g(P{)}^k\le C_k{g}^{-2+k/2}$$

for fixed $k$, after grouping repeated cut choices. Hölder combines this with a squared-length polygon moment. This controls the price of recovering the original arch or bridge and retains paths in a logarithmic window around $h^{4/3}$. It does not infer a typical length from a first moment alone. (OpenAI, 2026d)

The marked-correlation companion supplies the squared-length input through two marks and a separate one-arc estimate. In the latter, oriented loops disjoint from the marked arc cancel; the arc’s winding cancels the imposed twist, leaving one constant phase times its unsigned mass. The contour identity is established with generic site offsets before coalescence. Simultaneous exponential decay in the particles and the two probe variables justifies closing the contours. Its positive real comparison treats unequal marked intervals without an imaginary displacement of the reference particles. (OpenAI, 2026g)

Renewal identities explain the spatial exponent and the conditioning cost

Let $p(\gamma )={\rho }^{L(\gamma )}$ on irreducible bridges, with joint height and length $(H,L)$. Unique cutting at once-crossed levels gives

$$b(s,u)=\frac{1}{1-{\mathbb{E}}_p{e}^{-sH-uL}}.$$

Since $b(s,0)\asymp s^{-3/4}$ diverges as $s\downarrow 0$, the irreducible weights sum to one. Marking one irreducible piece gives the exact convolution identity $B_f=B\ast p_f\ast B$ for any nonnegative additive reward $f$. For length, summing outer bridge heights up to $x$ yields

$${\left(\sum _{h\le x}{B}_h\right)}^2{\mathbb{E}}_p[L;H\le x]\le \sum _{h\le 3x}{M}_h,$$

so the truncated reward is at most $C{x}^{7/12}$. Splitting $\{L>n\}$ at height $x=n^{3/4}$ gives a length tail $C{n}^{-9/16}$. The positive mass of long finite bridges supplies the matching Laplace-deficit lower bound. Independent irreducible sums then have length exponent $16/9$ and height exponent $4/3$; eliminating the number of pieces gives $(4/3)/(16/9)=3/4$. Height, length and displacement of one piece remain dependent throughout. (OpenAI, 2026k)

For finite bridges, confined moments give $m_1(h)\ge h^{13/12-o(1)}$, $m_2(h)\le h^{29/12+o(1)}$, and $m_0(h)\le C{h}^{-1/4}$. A second-moment argument yields only a seed probability $h^{-o(1)}$ of length $h^{4/3-\xi }$. Independent trials restart at renewal times to amplify it. Conditioning on a renewal at height $H$ must still be paid for: if a stopped prefix ends at height $U_{\tau }$, then

$$\mathbb{P}(E\mid H\text{ is a renewal})=\frac{\mathbb{E}[1_E{B}_{H-U_{\tau }}]}{B_H}.$$

The ratio is bounded when $U_{\tau }\le H/2$; overshoots require a separate renewal-tail estimate. (OpenAI, 2026h)

Prescribed boundary endpoints impose further costs. The endpoint companion allocates disjoint groups of irreducible pieces to top-order survival, bottom-order survival, a marked length reward, and an exact two-coordinate endpoint match. Only those disjoint groups are treated independently. Dyadic summation bounds the length numerator, and division by the prescribed-gap mass gives the mean law. A terminal-summed mixture and a single prescribed gap have different normalizers. (OpenAI, 2026b)

The final transfer to every fixed length is quantitative

The radial companion’s polygon tails arise from projection losses. A polygon lost on a cylinder of period $N$ has diameter at least order $N$, giving a lower tail bound. Conversely, a large polygon has large span in one of three lattice directions. At a slightly smaller period it is either lost or survives as an exceptionally tall contractible cylinder polygon; the latter mass is negligible by the cylinder-height estimate. This yields unrooted mass $R^{-2+o(1)}$ and length-weighted mass $R^{-2/3+o(1)}$. Their ratio is a diameter-conditioned mean, not an exact-length theorem. (OpenAI, 2026j)

For unrestricted fixed-length walks, the later companion proves $Z_n=c_n{\rho }^n\ge 1$ for every integer $n$. Indeed $c_{m+n}\le c_m{c}_n$; if some $Z_{n_0}<1$, iteration would force exponential decay of all $Z_n$, hence of bridge mass, contradicting its polynomial lower bound. Therefore an unnormalized bad mass $O(H^{-A})$ in a box of size $H\asymp n$ remains $O(n^{-A})$ after restricting to length $n$ and dividing by $Z_n$.

Its fast/slow-path arguments provide simultaneous subpath bounds

$$\begin{array}{rl}j-i& \le C{n}^{\eta }(1+\mathrm{diam}\gamma [i,j]{)}^{4/3},\\ \mathrm{diam}\gamma [i,j]& \le C{n}^{\eta }(1+j-i{)}^{3/4}.\end{array}$$

Applying them to the whole walk proves both diameter bounds. For a local-mass lower bound, take a contiguous block of about $n^{-\delta /2}\min (n,s^{4/3})$ edges next to the chosen vertex. The second inequality puts that entire block in its radius-$s$ ball. For covering, split time into such blocks to obtain an upper bound; for the lower bound, recenter each covering ball at a visited vertex and use the local-mass upper bound at radius $2s$. These final deductions are deterministic on one simultaneous good event. The substantive full-box fast/slow estimates, contour estimates and geometric sewing constructions remain manuscript inputs rather than independently verified proofs in this post.

See Also

• The honeycomb connective-constant article explains the exact critical activity; knowing that activity alone does not determine spatial exponents.
• Family 230 distinguishes a dimension exponent from an exact geometric measure.

🏷️ Manuscripts and verification scope

• Radial transfer estimates and polygon length laws for honeycomb walks — pinned manuscript and companion source.
• Critical honeycomb chords with prescribed boundary endpoints — pinned manuscript and companion source.
• Cylinder loop weights and planar nesting — pinned manuscript and companion source.
• Mass and covering exponents for fixed-length honeycomb walks — pinned manuscript and companion source.
• Signed cylinder propagation and marked polygons on the honeycomb lattice — pinned manuscript and companion source.
• Cylinder amplitudes and logarithmic bridge-length windows on the honeycomb lattice — pinned manuscript and companion source.
• Marked polygon correlations and one-arc bounds — pinned manuscript and companion source.
• Disk transfer representations and confined bridge mass — pinned manuscript and companion source.
• Polynomial vacuum representations and bridge mass for honeycomb walks — pinned manuscript and companion source.
• Renewal and changes of law for critical honeycomb walks — pinned manuscript and companion source.
• Critical strip-crossing mass on the honeycomb lattice — pinned manuscript and companion source.
• Uniform marked-polygon estimates and sharp finite bridge moments — pinned manuscript and companion source.
• Cap-selected amplitudes and triangle chords for honeycomb walks — pinned manuscript and companion source.

The family scope page describes selected formalization coverage. The selected statements concern bridge summability, existence of a force-dependent limiting free energy, and sharp strip masses and balance estimates. They do not document the complete three-quarter spatial laws or the all-integer fixed-length mass and covering theorem. 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. Cap-selected amplitudes and triangle chords for honeycomb walks.

🐻  OpenAI 2026b. Critical honeycomb chords with prescribed boundary endpoints.

🐻  OpenAI 2026c. Critical strip-crossing mass on the honeycomb lattice.

🐻  OpenAI 2026d. Cylinder amplitudes and logarithmic bridge-length windows on the honeycomb lattice.

🐻  OpenAI 2026e. Cylinder loop weights and planar nesting.

🐻  OpenAI 2026f. Disk transfer representations and confined bridge mass.

🐻  OpenAI 2026g. Marked polygon correlations and one-arc bounds.

🐻  OpenAI 2026h. Mass and covering exponents for fixed-length honeycomb walks.

🐻  OpenAI 2026i. Polynomial vacuum representations and bridge mass for honeycomb walks.

🐻  OpenAI 2026j. Radial transfer estimates and polygon length laws for honeycomb walks.

🐻  OpenAI 2026k. Renewal and changes of law for critical honeycomb walks.

🐻  OpenAI 2026l. Signed cylinder propagation and marked polygons on the honeycomb lattice.

🐻  OpenAI 2026m. Uniform marked-polygon estimates and sharp finite bridge moments.