on self-avoiding walks and the honeycomb connective constant

Overview

Duminil-Copin and Smirnov prove that the connective constant of the hexagonal, or honeycomb, lattice is exactly (Duminil-Copin & Smirnov, 2012)

$${\mu }_{\mathrm{hex}}=\sqrt{2+\sqrt{2}}.$$

Equivalently, if $c_n$ is the number of $n$-step self-avoiding walks from a fixed vertex, then $c_n$ grows exponentially like ${\mu }_{\mathrm{hex}}^n$ up to subexponential factors.

The proof identifies the critical fugacity

$$x_c=\frac{1}{\sqrt{2+\sqrt{2}}}$$

by constructing a parafermionic observable

$$F(z)=\sum _{\gamma :a\to z}{e}^{-i(5/8)W_{\gamma }(a,z)}{x}_c^{\ell (\gamma )}.$$

At this one value of $x$, local cancellations around each honeycomb vertex give a discrete contour identity. Summed in a strip, that identity controls the partition functions of walks reaching the left and right sides of the strip. A Hammersley-Welsh bridge decomposition then converts those strip estimates into the two radius-of-convergence inequalities that force ${\mu }_{\mathrm{hex}}=x_c^{-1}$.

🏷️ Self-Avoiding Walks

Let $G$ be an infinite locally finite lattice graph. A self-avoiding walk of length $n$ is a nearest-neighbor path

$$\gamma =({\gamma }_0,{\gamma }_1,\dots ,{\gamma }_n)$$

whose vertices are all distinct. On a transitive lattice, the number

$$c_n=\#\{\text{n-step self-avoiding walks from a fixed origin}\}$$

does not depend on the origin.

The basic exponential growth rate exists by submultiplicativity. Cutting an $(n+m)$-step walk after $n$ steps and translating the remaining tail back to the origin gives

$$c_{n+m}\le c_n{c}_m.$$

Thus $a_n=\log c_n$ is subadditive, and Fekete’s lemma gives

$$\underset{n\to \infty }{\lim }\frac{1}{n}\log c_n=\underset{n\ge 1}{\inf }\frac{1}{n}\log c_n.$$

The connective constant is therefore

$$\mu =\underset{n\to \infty }{\lim }{c}_n^{1/n}.$$

For the honeycomb lattice one has elementary bounds such as

$$(\sqrt{2}{)}^n\le c_n\le 3\cdot 2^{n-1},$$

so the connective constant is finite and nonzero. These bounds are not sharp; they only show that there is a genuine exponential scale.

Fugacity and radius of convergence

It is often cleaner to encode all lengths at once in a generating function

$$C(x)=\sum _{n\ge 0}{c}_n{x}^n.$$

The radius of convergence is $1/\mu$. Therefore proving $\mu =x_c^{-1}$ is equivalent to proving that $C(x)$ is finite for $x<x_c$ and divergent for $x>x_c$. Duminil-Copin and Smirnov work with a mid-edge partition function whose length convention differs by a bounded shift, but it has the same critical radius.

The paper proves the value predicted by Nienhuis from the Coulomb-gas analysis of the two-dimensional $O(n)$ model at $n=0$ (Nienhuis, 1982). It does not prove the sharper conjectural asymptotic

$$c_n\sim A{n}^{\gamma -1}{\mu }^n,\gamma =\frac{43}{32},$$

nor the predicted end-to-end exponent $\nu =3/4$. The theorem identifies the lattice-dependent exponential constant, not the universal critical exponents.

🏷️ Honeycomb Mid-Edges

The proof uses the honeycomb lattice embedded in $\mathbb{C}$. Vertices have degree three, and it is convenient to put the endpoints of walks at mid-edges rather than vertices. Let $H$ be the set of midpoints of honeycomb edges. A walk

$$\gamma :a\to b,a,b\in H,$$

starts at a mid-edge, visits honeycomb vertices without repetition, and ends at a mid-edge. Its length $\ell (\gamma )$ is the number of honeycomb vertices visited.

This mid-edge convention has two advantages. First, the three local possibilities around a vertex are perfectly symmetric. Second, the winding from one mid-edge to another has increments that are multiples of $\pi /3$.

For a self-avoiding walk $\gamma$ between mid-edges $a$ and $b$, define

$$W_{\gamma }(a,b)$$

to be the total signed rotation, in radians, of the tangent direction as $\gamma$ is traversed from $a$ to $b$. A left turn contributes $+\pi /3$ and a right turn contributes $-\pi /3$ in this convention.

A finite honeycomb domain $\Omega$ is described by a set $V(\Omega )$ of vertices and all mid-edges emanating from them. A mid-edge belongs to the boundary $\partial \Omega$ if exactly one endpoint of its original honeycomb edge lies in $V(\Omega )$. The paper assumes simply connected domains, so windings to boundary arcs are determined by the geometry of the boundary rather than by hidden holes.

🏷️ The Parafermionic Observable

Fix a boundary mid-edge $a\in \partial \Omega$. For a mid-edge $z\in \Omega$, a fugacity $x>0$, and a spin parameter $\sigma$, define

$$F_{\Omega }(a,z;x,\sigma )=\sum _{\gamma \subset \Omega :a\to z}{e}^{-i\sigma W_{\gamma }(a,z)}{x}^{\ell (\gamma )}.$$

This is a weighted generating function for self-avoiding walks ending at $z$, with a complex phase recording winding.

The special values are

$$x=x_c=\frac{1}{\sqrt{2+\sqrt{2}}}=\frac{1}{2\cos (\pi /8)},\sigma =\frac{5}{8}.$$

Write

$$\lambda =\exp \left(-i\frac{5\pi }{24}\right),j=e^{2\pi i/3}.$$

Then each left or right turn contributes a factor $\lambda$ or $\bar{\lambda }$ to the winding phase.

Local parafermionic cancellation

Let $p,q,r$ be the three mid-edges adjacent to a vertex $v$, listed counterclockwise. At $x=x_c$ and $\sigma =5/8$,

$$(p-v)F(p)+(q-v)F(q)+(r-v)F(r)=0.$$

Here $p-v$, $q-v$, and $r-v$ are complex edge vectors. The identity says that the discrete contour integral of $F$ around the elementary dual face vanishes.

Proof idea: local pairings

Expand the left side as a sum over walks ending at one of $p,q,r$. The walks split into two types.

If a walk visits all three adjacent mid-edges, then near $v$ it contains a self-avoiding loop segment. Pair it with the walk using the same edges but traversing that loop segment in the opposite direction. The two walks have the same length and their windings differ by $8\pi /3$ across the paired endpoints. Their contribution is proportional to

$$j{\bar{\lambda }}^4+\bar{j}{\lambda }^4,$$

which is zero for $\lambda =e^{-5\pi i/24}$.

If a walk visits only one adjacent mid-edge, pair it with the two walks obtained by extending it through $v$ to the two other mid-edges. This gives a triplet. The two extended walks have one extra visited vertex, and the winding changes by $\pm \pi /3$. Their contribution is proportional to

$$1+x_{c}j\bar{\lambda }+x_c\bar{j}\lambda ,$$

which is zero exactly because $x_c^{-1}=2\cos (\pi /8)$.

Summing these zero contributions over all pairs and triplets proves the local identity.

The identity is the engine of the proof. It is not full discrete holomorphicity in the square-lattice sense: there is one relation per honeycomb vertex but one value of $F$ per mid-edge, so there are not enough equations to reconstruct the observable from boundary data. Nevertheless, it is strong enough after summation over a strip.

🏷️ Strip Partition Functions

Let $S_T$ be a vertical honeycomb strip of width $T$, and let $S_{T,L}$ be a finite truncation with top and bottom cutoffs. The boundary of $S_{T,L}$ is divided into four pieces:

$$\alpha =\text{left side},\beta =\text{right side},\epsilon =\text{top},\bar{\epsilon }=\text{bottom}.$$

The starting mid-edge $a$ lies on $\alpha$.

Define positive partition functions

$$A_{T,L}^x=\sum _{\gamma \subset S_{T,L}:a\to \alpha \setminus \{a\}}{x}^{\ell (\gamma )},$$ $$B_{T,L}^x=\sum _{\gamma \subset S_{T,L}:a\to \beta }{x}^{\ell (\gamma )},$$

and

$$E_{T,L}^x=\sum _{\gamma \subset S_{T,L}:a\to \epsilon \cup \bar{\epsilon }}{x}^{\ell (\gamma )}.$$

They count walks exiting through the left side, the right side, or the top and bottom.

Summing the local cancellation over all vertices in $S_{T,L}$ makes interior mid-edge terms cancel. Only boundary terms remain:

$$0=-\sum _{z\in \alpha }F(z)+\sum _{z\in \beta }F(z)+j\sum _{z\in \epsilon }F(z)+\bar{j}\sum _{z\in \bar{\epsilon }}F(z).$$

The strip symmetry and the fixed boundary windings remove the complex phases. Walks ending on $\beta$ have winding $0$; walks ending on the two parts of $\alpha \setminus \{a\}$ have winding $\pm \pi$; walks ending on the top and bottom have winding $\pm 2\pi /3$. The trivial walk from $a$ to $a$ contributes $F(a)=1$.

Consequently, at $x=x_c$,

$$1=\cos \left(\frac{3\pi }{8}\right)A_{T,L}^{x_c}+B_{T,L}^{x_c}+\cos \left(\frac{\pi }{4}\right)E_{T,L}^{x_c}.$$

Letting $L\to \infty$ gives

$$1=c_{\alpha }{A}_T^{x_c}+B_T^{x_c}+c_{\epsilon }{E}_T^{x_c},c_{\alpha }=\cos \left(\frac{3\pi }{8}\right),c_{\epsilon }=\cos \left(\frac{\pi }{4}\right).$$

This identity has two important consequences. First,

$$B_T^{x_c}\le 1$$

for every strip width $T$. Second, the identity prevents all right-crossing partition functions from becoming too small at once.

🏷️ Divergence at the Critical Point

Let

$$Z(x)=\sum _{\gamma :a\to H}{x}^{\ell (\gamma )}$$

be the full-plane mid-edge partition function. The first half of the theorem is

$$Z(x_c)=+\infty .$$

This implies $\mu \ge x_c^{-1}$.

There are two cases. If $E_T^{x_c}>0$ for some strip width $T$, then the weights of walks exiting through the top or bottom of taller and taller truncations force

$$Z(x_c)\ge \sum _{L>0}{E}_{T,L}^{x_c}=+\infty .$$

Otherwise $E_T^{x_c}=0$ for every $T$. The strip identity becomes

$$1=c_{\alpha }{A}_T^{x_c}+B_T^{x_c}.$$

A walk counted by $A_{T+1}^{x_c}$ but not by $A_T^{x_c}$ must reach the new right edge region of $S_{T+1}$ before returning to the left boundary. Cutting at the first such visit decomposes it into two strip-crossing bridges, with one extra vertex. Therefore

$$A_{T+1}^{x_c}-A_T^{x_c}\le x_c{\left(B_{T+1}^{x_c}\right)}^2.$$

Subtracting the strip identities for $T$ and $T+1$ gives

$$B_T^{x_c}\le B_{T+1}^{x_c}+c_{\alpha }{x}_c{\left(B_{T+1}^{x_c}\right)}^2.$$

This recurrence implies

$$B_T^{x_c}\ge \frac{\min \{B_1^{x_c},(c_{\alpha }{x}_c{)}^{-1}\}}{T}.$$

Indeed, if $b_T=B_T^{x_c}$ and $K=c_{\alpha }{x}_c$, then $b_T\le b_{T+1}+K{b}_{T+1}^2$ cannot drive a positive initial value below the harmonic profile $m/T$ when $m\le \min \{b_1,K^{-1}\}$.

Thus

$$Z(x_c)\ge \sum _{T\ge 1}{B}_T^{x_c}=+\infty .$$

This proves the lower bound

$$\mu \ge \sqrt{2+\sqrt{2}}.$$

🏷️ Bridge Decomposition

The opposite inequality uses the Hammersley-Welsh decomposition of self-avoiding walks into bridges (Hammersley & Welsh, 1962; Madras & Slade, 1993).

A bridge of width $T$ is a self-avoiding walk crossing a strip of width $T$ from one side to the other, considered up to vertical translation. Its partition function is $B_T^x$. Since every such bridge has length at least $T$ and $B_T^{x_c}\le 1$,

$$B_T^x\le {\left(\frac{x}{x_c}\right)}^T{B}_T^{x_c}\le {\left(\frac{x}{x_c}\right)}^T$$

for every $x<x_c$.

Hence

$$\sum _{T\ge 1}{B}_T^x<\infty ,\prod _{T\ge 1}\left(1+B_T^x\right)<\infty .$$

Now take any full-plane self-avoiding walk. Cut it at the first vertex with maximal real part. The part before the cut is a reverse half-plane walk; the part after the cut is a half-plane walk. Each half-plane walk decomposes canonically into bridges of strictly decreasing widths. Equivalently, the original walk decomposes into a sequence of bridge widths

$$T_{-i}<\cdots <T_{-1},T_0>\cdots >T_j.$$

Once the starting mid-edge and the first visited vertex are fixed, this bridge list uniquely reconstructs the walk.

Therefore, for $x<x_c$,

$$Z(x)\le 2\sum _{\begin{array}{c}{T}_{-i}<\cdots <T_{-1}\\ T_j<\cdots <T_0\end{array}}\prod _{k=-i}^j{B}_{T_k}^x=2\prod _{T\ge 1}{\left(1+B_T^x\right)}^2<\infty .$$

The factor $2$ accounts for the two possible first vertices from the starting mid-edge.

Thus the partition function is finite for every $x<x_c$, and hence

$$\mu \le x_c^{-1}=\sqrt{2+\sqrt{2}}.$$

Combined with critical divergence, this proves the theorem.

Honeycomb connective constant

For the hexagonal lattice,

$${\mu }_{\mathrm{hex}}=\sqrt{2+\sqrt{2}}.$$

🏷️ Why the Argument Is Rigid

The value of $x_c$ is not guessed during the proof. It is forced by the local triplet cancellation

$$1+xj\bar{\lambda }+x\bar{j}\lambda =0$$

with spin $\sigma =5/8$. Solving this trigonometric identity gives

$$x^{-1}=2\cos \left(\frac{\pi }{8}\right)=\sqrt{2+\sqrt{2}}.$$

This is also why the proof is special to the honeycomb lattice. The degree-three geometry gives exactly the local pair and triplet classification needed for the parafermionic cancellation. The square lattice has a different connective constant, and no analogous exact evaluation is known.

The observable also explains the connection to conformal invariance. If the missing half of the discrete Cauchy-Riemann equations could be recovered in the scaling limit, the normalized observable would solve a Riemann boundary value problem with covariance exponent $5/8$. Lawler, Schramm, and Werner identify the expected scaling limit as chordal ${\mathrm{SLE}}_{8/3}$ under conformal-invariance assumptions (Lawler, Schramm & Werner, 2004).

The Duminil-Copin and Smirnov theorem therefore proves the exact critical point while leaving the deeper scaling-limit conjecture open.

📊 Numerical Verification

The script

python3 'content/codes/2026 Summer/saw_honeycomb_verification.py'

performs two finite checks. First, it exactly enumerates honeycomb self-avoiding walks from one vertex. The finite ratios are still above the limiting constant, but they move toward

$$x_c^{-1}=(0.541196100146{)}^{-1}=\mathrm{1.847759065023.}$$

Representative values from the default run are:

$n$	$c_n$	$c_n^{1/n}$	$c_n/c_{n-1}$
12	4416	2.012576615	1.896907216
18	199350	1.969804496	1.877118644
24	8710212	1.946112148	1.871367358

Second, it enumerates all walks ending on the boundary of small finite strips and forms the parafermionic observable at $x_c$ and $\sigma =5/8$. Instead of hard-coding the scalar strip identity, it checks the vector boundary identity obtained by summing the local cancellation with the actual boundary edge vectors. The residual is therefore

$$\left|\sum _{v\in V(\Omega )}\sum _{z\sim v,z\in \partial \Omega }(z-v)F(z)\right|,$$

which should vanish up to floating-point roundoff.

$T$	$L$	vertices	boundary-ending walks	residual
1	1	6	6	$5.004\cdot 10^{-17}$
3	2	28	236	$1.602\cdot 10^{-16}$
4	3	54	42376	$2.352\cdot 10^{-15}$
5	3	66	575449	$2.540\cdot 10^{-14}$

This is not a proof of the limiting theorem. It checks the two computationally accessible pieces: the finite-length growth trend and the finite-domain parafermionic cancellation that drives the proof.

🏷️ Links

• on Ising models: The parafermionic observable is part of the same discrete-complex-analysis philosophy that makes two-dimensional Ising observables tractable.
• on cover times for two-dimensional random walks: Both notes use two-dimensional recurrence, boundary hitting, and scale-sensitive path decompositions, though the SAW proof is combinatorial rather than Markovian.
• on high-dimensional cover times and random interlacements: A contrast point where the geometry is dominated by high-dimensional transience instead of planar criticality.
• on sum-product in finite fields via entropy: The proof has a similar flavor of extracting a sharp global threshold from a local algebraic identity.

📚 References

🐻  Duminil-Copin, H. & Smirnov, S. 2012. The connective constant of the honeycomb lattice equals \sqrt2+\sqrt 2. Annals of Mathematics 175(3), 1653–1665.

🐻  Hammersley, J.M. & Welsh, D.J.A. 1962. Further results on the rate of convergence to the connective constant of the hypercubical lattice. Quarterly Journal of Mathematics 13(1), 108–110.

🐻  Lawler, G.F., Schramm, O. & Werner, W. 2004. On the scaling limit of planar self-avoiding walk. In Fractal Geometry and Applications: A Jubilee of Benoit Mandelbrot, Part 2, pp. 339–364. Proceedings of Symposia in Pure Mathematics, American Mathematical Society.

🐻  Madras, N. & Slade, G. 1993. The Self-Avoiding Walk, Birkhäuser Boston,p.

🐻  Nienhuis, B. 1982. Exact critical point and critical exponents of O(n) models in two dimensions. Physical Review Letters 49(15), 1062–1065.