# Renewal and changes of law for critical honeycomb walks
*An expository walkthrough, written to be readable with high-school mathematics.*
---
## 0. Prologue: the paper in one sentence
Draw the honeycomb (hexagonal) grid and consider a path on it that never
visits the same vertex twice (a *self-avoiding walk*, SAW). A long SAW does
not spread out like a straight line and does not stay put: it fills space
with a very specific power-law profile. After `n` steps one is typically
about `n^(3/4)` units from the start, and the walk visits about `R^(4/3)`
vertices inside a ball of radius `R`.
The number `3/4` is the *spatial exponent* (traditionally written `nu`). It
had been predicted for decades from physics (Nienhuis 1982, an `O(n)` model at
`n -> 0`) but proving it rigorously for a single fixed-length walk, or for a
single exact height, or for a thermal (length-penalized) ensemble is very
hard. This paper proves it, simultaneously, in several natural random laws
on the honeycomb grid, and gives the exact power with which each law agrees.
In one sentence: *on the honeycomb lattice the critical self-avoiding walk
scales in space with exponent exactly `3/4`, and this remains true across a
whole family of length and height conditioning rules; the paper builds one
renewal "machine" for all of them.*
Everything below is the map of how the proof works. The paper is long
(about 6,000 lines of LaTeX) and this exposition is much shorter; the deepest
technical steps are summarized by their exact statement and the reason they
are needed, rather than by their twenty-page constant-chasing.
---
## 1. The cast of characters
### 1.1 The honeycomb graph and the triangular tiling
Put a regular triangular tiling of the plane with unit side length. Each
little triangle has a center. The centers are the vertices of the honeycomb
graph: two centers are joined when their triangles share a side. A honeycomb
vertex therefore has three neighbours, and the graph looks like a honeycomb
of hexagons.
The vertical direction matters a lot. Choose one family of horizontal
tiling lines. Successive horizontal lines are `d0 = sqrt(3)/2` units apart;
we measure integer *height* in units of that gap, so the `h`-th horizontal
line is at height exactly `h` (an integer). Later we sometimes use the
*unit-edge* embedding of the same graph (where a height unit has length
`3/2`); the exponents never change, only the unit convention.
### 1.2 Self-avoiding walks
A *self-avoiding walk* (SAW) is a sequence of distinct neighbouring vertices
`gamma_0, gamma_1, ..., gamma_n`. `n` is the *length* (we mostly count
visited vertices `L = n + 1`, and switching between vertex count and edge
count changes only a bounded factor). We care about two spatial quantities:
A(gamma) = |gamma_n - gamma_0| (endpoint distance)
D(gamma) = diam{gamma_0, ..., gamma_n} (diameter of the trace)
Together with `L`, these measure how "stretched out" a walk is.
### 1.3 Ports, strips, bridges
A *port* is a little fixed point on the boundary of the region we are
walking in: technically the midpoint of a tiling edge where the walk is
allowed to enter or leave a horizontal strip. Ports let us chop a walk into
pieces without losing bookkeeping.
A *strict bridge of height h* (for an integer `h >= 1`) is a walk with:
- one endpoint port on a horizontal line,
- the other endpoint port on the line `h` rows above it,
- every visited vertex strictly between those two lines.
The terminal port and its horizontal position are *free* unless we fix
them. The bridge is the basic building block. The whole paper is written
so that pieces concatenate cleanly: the top port of one bridge becomes the
bottom port of the next, and the pieces sit in disjoint horizontal slabs.
We also need a critical weight. Each *visited vertex* gets a factor
rho = (2 + sqrt(2))^(-1/2) (= 1 / sqrt(2 + sqrt(2)))
and ports get weight `1`. A walk with `L` vertices then has weight `rho^L`.
The number `rho` is the famous critical value of the honeycomb SAW (proved by
Duminil-Copin and Smirnov in 2012); it is the exact "density" at which the
number of walks stops growing or shrinking. We always work at *criticality*
`rho`, never above or below, unless we explicitly discount.
### 1.4 Why a critical weight?
Fix a start vertex and let `c_n` count self-avoiding walks of `n` steps from
it. By submultiplicativity one has `c_{n+m} <= c_n c_m`, so `c_n^(1/n)`
converges to a limit `mu >= 1`, the *connective constant* (the "growth rate"
of SAWs). At criticality, `rho` is chosen so that `rho * mu = 1`. Then the
total SAW mass grows polynomially instead of exponentially; this is where
power laws live. The whole paper lives exactly on this boundary.
---
## 2. The three laws (what the paper compares)
A single walk can be picked in three very different ways. The paper proves
the `3/4` exponent in essentially all of them, and explains why the three
arguments cannot simply be copied from one another.
### 2.1 Law 1: a uniform walk of a fixed length
Fix `n` and pick uniformly at random among all SAWs of exactly `n` steps
(or among all strict bridges with exactly `n` vertices, or exactly `n`
half-plane vertices, etc.). The normalisation divides by the exact count;
the count is a single coefficient of a generating function, so controlling
it is *local in length*.
### 2.2 Law 2: the infinite irreducible-bridge concatenation
Chop any strict bridge into *irreducible bridges* by cutting at every
horizontal line that the walk crosses exactly once. An irreducible bridge is
a bridge that does this no more: once you have entered it and left it, you
never come back across either boundary. The single key fact is:
> The total critical mass of all irreducible bridges is exactly `1`.
So they can be normalised to a probability law, call it `p`. Sample an
*infinite* sequence of independent irreducibles from `p` and concatenate
them. Successive pieces sit in disjoint slabs, so no self-avoidance
correction is ever needed. This is the *infinite law*; it is the most
"canonical" of the three.
### 2.3 Law 3: the thermal law
Fix a "temperature-like" parameter `N > 0`. Give every finite walk of `L`
vertices weight
rho^L * exp(-L / N).
This is a *Boltzmann* / thermal law: lengths much bigger than `N` are
exponentially suppressed, lengths much smaller than `N` are fine. `N` sets the
discount scale. As `N -> infinity` the discount disappears and we approach the
infinite critical law; for finite `N` we have a genuine probability law with
`L` of order `N`.
### 2.4 Why the three are not the same problem
- Law 1 fixes one coefficient of the generating function.
- Law 2 sums *all* lengths at criticality (coefficient sum with no discount).
- Law 3 replaces the sum by a discount `e^{-L/N}` and sums.
Showing that all three have the same *spatial* exponent `3/4` needs a
completely different argument for each. This is the technical core of the
paper, and the reason it is long. Once we have the machine in Section 4
below, each change of law becomes a specific estimate.
---
## 3. The main results
All theorems are of the form `X_t = t^(a+o(1)) in probability`. To unpack:
for every fixed `epsilon > 0`, the event
t^(a - epsilon) <= X_t <= t^(a + epsilon)
has probability tending to 1 as `t -> infinity`. Saying `a` is the exact
exponent means we cannot push this to `a - delta` or `a + delta` for a fixed
positive `delta`. Equivalently `log X_t / log t -> a` in probability. The
phrase `almost surely` means the same claim holds with probability one
under the underlying random model (each probability-one event may depend on
the model, and the statement is not just in probability over a large `t`).
The main theorems (numbered as in the paper, reproduced as claims):
A. Infinite irreducible-bridge law.
Under `p`, for the concatenated infinite walk with vertices gamma_n:
|gamma_n - gamma_1| = n^(3/4 + o(1)) almost surely,
#{n : |gamma_n - gamma_1| <= R} = R^(4/3 + o(1)) almost surely.
(Theorem `R:renew:infinite`.)
B. Uniform strict bridge at exact even length.
Pick a strict bridge uniformly with `n` visited vertices (`n` even,
sufficiently large), start port fixed, terminal free.
n^(3/4 - eps) <= A <= D <= n^(3/4 + eps) with prob -> 1.
(Theorem `R:exact:bridge`.)
C. Uniform strict bridge at exact height `h`.
Pick a strict bridge uniformly with exact height `h`.
L = h^(4/3 + o(1)) and D = h^(1 + o(1)), in probability;
and the same in first mean.
(Theorem `R:exact:height`.)
D. Uniform endpoint-free half-plane and full-plane walks.
There is one set of lengths of *natural density one* on which the
uniform walk (endpoint free) has
A = D = n^(3/4 + o(1)) in probability;
the upper spatial bound holds for *every* sufficiently large `n`,
to arbitrarily high polynomial accuracy; and all positive moments
`E A^p`, `E D^p` have exponents `3p/4` on that same density-one set.
(Theorems `R:cal:halfplane-laws`, `R:cal:plane-laws`.)
E. Thermal law.
Under `rho^L e^{-L/N}` on half- or full-plane walks, for every `xi > 0`:
N^(1 - xi) <= L <= N^(1 + xi),
N^(3/4 - xi) <= A <= D <= N^(3/4 + xi) with prob -> 1.
For every fixed `p > 0`:
E L^p = N^(p + o(1)), E A^p, E D^p = N^(3p/4 + o(1)).
(Theorem `R:thermal:law`.)
F. Small-force free energy (near-critical response).
Let `f_e(s)` be the fixed-length free energy in a fixed direction
`e` at force `s`. As `s -> 0`, uniformly in `e`:
f_e(s) = s^(4/3 + o(1)),
f'_{e,-}(s) = f'_{e,+}(s) = s^(1/3 + o(1)).
(Theorem `R:resp:force`, Corollary `R:resp:force-derivatives`.
The `1/3` matches Pincus's force-extension scaling `1/nu - 1 = 1/3`.)
G. Prescribed terminal vertex.
Let `b_N` with `|b_N - o| = N^(3/4 + o(1))`. The thermal law from
`o` conditioned to end at `b_N` has
L = N^(1 + o(1)), D = N^(3/4 + o(1)) in probability,
and all positive moments with corresponding exponents.
(Theorem `R:resp:prescribed`.)
The two exponents `nu = 3/4` (spatial scale of distance in `n`) and
`d* = 4/3` (fractal dimension: count in a ball `R^(4/3)`) are reciprocal:
`d* * nu = 1`. This is automatic once you know the distance grows like the
inverse of the number of steps needed to reach it.
There is one important technical honesty in the paper: statement D (the
uniform fixed-length *endpoint-free* law) is proven only along a
density-one set of lengths, not for every length, because that law needs
"mass at nearby lengths" and no single exact-length denominator cancels.
Statement B (uniform *strict bridge*) is stronger, exact at every even
length; and statement E (thermal) holds at every discount `N`. This
precision matters: the proofs of B, C, E all first prove or assume enough
mass at a single length or a single discount, which D does not get for
free.
---
## 4. The renewal machine
The whole paper turns on one structure: cutting infinite walks into
irreducible pieces. The finite estimates behind it are shipped in from
companion preprints; the paper itself uses only the renewal identities and
a small set of exponent facts.
### 4.1 The irreducible probability law `p`
Fix a starting port. Let `B_h` be the total critical mass of strict
bridges from that port at exact height `h`, with terminal port free. The
finite input is:
B_0 = 1,
c * h^(-1/4) <= B_h <= C * h^(-1/4) (all large h),
M_h := sum over bridges of height h of L * rho^L <= C * h^(13/12),
B_h { L >= c * h^(4/3) } >= c * h^(-1/4),
B_h { D > x } <= C * x^(-1/4) (all x >= 1).
Those constants `c, C` depend only on `h`, `x` ranges and are not small or
large in any dangerous way. Cutting a bridge at every horizontal line it
crosses *exactly once* splits it uniquely into irreducible pieces, and the
total critical mass of irreducibles is exactly `1`, giving the probability
law `p`.
### 4.2 The generating-function identity
The single identity the rest of the paper is built on:
sum_{h >= 0} B_h e^{-s * h} = 1 / (1 - E_p[e^{-s * H}]),
where `H` is the height of one `p`-distributed irreducible. From the finite
input `B_h ~ h^(-1/4)`, the left side grows like `s^(-3/4)`. So
E_p[e^{-s * H}] = 1 - s^(3/4) * (lower-order), as s -> 0.
The `3/4` on the right is the *same* `3/4` as the spatial exponent. It is
the seed of everything.
### 4.3 The two exponents `alpha = 9/16` and `beta = 3/4`
Write `H`, `L`, `D` for the height, vertex length and diameter of a single
`p`-distributed irreducible. The finite input plus one extra moment gives:
(height) 1 - E_p[e^{-s H}] ~ s^(3/4) (i.e. exponent beta = 3/4)
P_p[H > x] <= C x^(-3/4),
(diameter) P_p[D > x] <= C x^(-3/4),
(length) 1 - E_p[e^{-u L}] ~ u^(9/16) (i.e. exponent alpha = 9/16)
P_p[L > n] <= C n^(-9/16),
E_p[L ; H <= x] <= C x^(7/12).
The length tail exponent `alpha = 9/16` is *smaller* than `beta = 3/4`:
irreducibles can be taller than long in the tails. The height and diameter
have the same `3/4` tail; the length is spikier.
### 4.4 Why the spatial exponent is `alpha / beta = 3/4`
Concatenate `k` independent irreducibles. Their height sums grow as
`k^(1/beta) = k^(4/3)`; their lengths grow as `k^(1/alpha) = k^(16/9)`.
The nth visited vertex sits in piece `k` where `k` is the first integer with
total length at least `n`, i.e. `k = n^(9/16 + o(1))`. Then:
distance from start <= total diameter <= sum D_i <= k^(4/3 + o(1))
= (n^(9/16))^(4/3 + o(1)) = n^(3/4 + o(1)),
and similarly the sum of heights, which lower-bounds the vertical distance,
grows as `k^(4/3) = n^(3/4)`. The two bounds meet at `3/4`; that is `nu`.
The trick that makes this rigorous is an elementary conversion lemma:
if `Y` is nonnegative with tail `P[Y > t] <= t^(-a+o(1))`, then
`sum_{j<=k} Y_j <= k^(1/a + o(1))` almost surely; and if additionally
`1 - E[e^{-uY}] = u^(a+o(1))`, then the same sum is `k^(1/a + o(1))`
almost surely (both bounds). This is exactly Borel-Cantelli plus Chebyshev
at dyadic scales `k = 2^j`.
### 4.5 The half-plane cancellation
One important cancellation that keeps the thermal half-plane denominator
from blowing up: in a half-plane walk, write the walk as a list of complete
irreducibles followed by one final "remainder" that has no further renewal
boundary. If `U_N` is the total `z`-weight of those remainders (with
`z = rho * e^{-1/N}`), then the total half-plane weight is
H_N = U_N / q_N, q_N = 1 - E_p[e^{-L/N}] = N^(-9/16 + o(1)).
So the half-plane normaliser factors as a denominator `q_N` and a
remainder mass `U_N`. When we ask for the *probability* of a fixed initial
tuple of irreducibles, `U_N` cancels identically out of numerator and
denominator. That is why fixed half-plane prefix limits converge, and it is
the reason the thermal law in the half-plane is tractable. In the full
plane one *cannot* use the same trick: one needs an explicit lower bound on
the full-plane denominator instead. The paper supplies that lower bound by
sandwiching a bridge between two no-renewal ends (Section 6).
---
## 5. Guided tour: what each of the paper's 12 sections actually does
The paper is organised as 12 numbered sections. Here is what each one builds
and why. The numbering matches `main.tex`.
### Section 1: Introduction
Sets up the three laws (uniform fixed length, infinite irreducible
concatenation, thermal) and states the principal spatial conclusions A-E
above. Also lists the companion preprints whose finite estimates are used
as black boxes and explains which companion supplies which input.
### Section 2: One probability space for the different laws (`renewal`)
The foundation. Defines ports, strict bridges, `B_h`, `M_h`; states the
finite bridge input; proves the renewal factorization `B(s) = 1/(1 - I(s))`
and the reward identity `B_f = B * p_f * B` (mark one piece and cut before
and after). Proves the joint increment estimates of Section 4.3 above.
Proves the infinite spatial law (statement A) directly: the sum of heights
grows like `k^(4/3)`, the sum of lengths like `k^(16/9)`, the sum of
diameters is `<= k^(4/3)`. Then proves the ratio limit
`a_{n+2}/a_n -> rho^(-2)` for uniform half-plane paths by hexagon flips and
a slot-counting argument, giving statement A's prefix limit.
The flips-vs-nonflips argument: a path's turn word has runs of at most four
equal turns (five would close a hexagon and revisit a vertex). The
proportion of "length-3 runs" (of which there are many) that are flippable
tends to a fixed ratio; this is what makes the ratio `a_{n+2} / a_n` exist.
### Section 3: Conditioning on one exact length (`exact`)
Proves statements B and C. Two ingredients:
- a *local renewal lower bound*: if `X` has span-one support and tail
`P[X > x] <= C x^(-a)`, and Laplace deficit `1 - E e^{-uX}` is comparable
to `u^a`, then `V_m := sum_{k} P(S_k = m)` satisfies
`V_m >= c m^(a-1)` for all large `m`. (Theorem `R:exact:local`.)
- a *uniform-bridge decomposition*: at even length `n`, masses at exact
length are of size `V_{n/2} ~ n^(-7/16)`. Then split the piece count `k`
into two halves, one providing displacement and one giving exact
residual length. Only one scalar equation (the length) is imposed on
the fresh half; Fourier concentration in Section 2 makes this cheap.
The Fourier proof of the local renewal bound is: tilt the law by
`e^{-uX}`, use the tilted characteristic function's Gaussian in a window of
width `u` and its algebraic tail `e^{-c |theta|^a}` outside. Fourier
inversion in the window yields `P_u(S_k = m) >= c / m` uniformly. Remove
the tilt and sum over a window of `c m^a` compatible `k`. This is the
standard "Garsia-Lamperti" style local renewal theorem, redone from
exponent bounds rather than a regular-variation limit.
### Section 4: Conditioning on a logarithmic window (`calibrated`)
Alternative route to statements B and C from a *weaker* finite input: only
a window of bridge lengths of width logarithmic in the height (that is,
`L` lying between `h^(4/3) / log^(1/8)` and `h^(4/3) * log^(1/2)`). From
this and Fourier differences, a long even interval of lengths gets a
matching lower bound for the local renewal mass. Then a *landing* argument:
from a current length gap `r`, expose a fixed block whose success chance
is only `(log r)^(-C)` per stage; each successful stage shrinks the gap by
a factor at least `(log r)^(-3/4)`. Only `log log r` factors accumulate,
so the total landing cost is `n^{-o(1)}`. Conclusion:
`u_n >= n^(alpha-1-o(1))` and `u_{n+2}/u_n -> 1`, therefore local
convergence.
### Section 5: Geometric estimates before a change of law (`geometry`)
The technical section. Proves the key avoidance estimate for *two* bridges
from adjacent ports: probability they stay disjoint through height `r` is
`<= r^(-3/4+eps)`. The proof uses the diameter tail and a recursive
"fence" argument (Lemma `R:geo:fence`). Then:
- the *sewing* lemma (close two bridges to a polygon) gives mass bounds
for the seam between adjacent bridges;
- the *marked-pair* estimate, uniform in the mark function, controls how
much mass can survive at a given height when a subset of one bridge is
forced to miss the other;
- the *turning-chain* and *absolute count* bounds "consume" these to give
exponential control on paths with lots of turning.
This section supplies the absolute spatial tails used everywhere downstream.
### Section 6: The unrestricted thermal law (`thermal`)
Proves statement E. Splits the plane Boltzmann sum into: bridge factor
`(q_N^{-1} - 1) * U_N^2` as denominator (an explicit *lower bound* built by
sandwiching a nonempty bridge between two exterior no-renewal ends). The
same construction is a bijection (recover the two single-crossing lines from
the output). Then bounds the numerator: paths whose endpoints have small
*vertical* separation `V <= r` have mass `<= C N^eps (1+r)^(3/4) U_N^2`,
which after division by `q_N^{-1} U_N^2` gives `N^(eps - 3 delta /4)`.
Taking `r = N^(3/4 - delta)` and `delta -> 0` gives the thermal spatial
lower bound. The long-length tail is controlled by earlier absolute
estimates. All this is done at every`N` ; there is no density-one
qualification on the thermal statement.
### Section 7: Censored arms and absolute spatial estimates (`censored`)
Supplies an independent proof of the *rapid-travel* theorem: the total
critical mass of plane walks that traverse a diameter `~ R` in length
`<= R^(4/3 - xi)` is bounded by `exp(-R^c)`. This is what rules out walks
that "run across a long span without growing", the main obstruction to a
spatial lower bound. The two-arm survival lemma (probability that two
censored independent arms stay disjoint up to height `g`) is
`<= C g^(-3/4+eta)`. The censored reference measure is transferred to
weighted bridge counts by a "kernel" identity which sums the crossing jumps.
Finally a volume bound: walks confined to a box of side `R` have mass
`exp(C (1+R)^nu)` for every `nu > 0`.
### Section 8: Uniform endpoint-free laws and moments (`uniform`)
Proves statement D. Two very different insertion arguments:
- Half-plane: adjoin a long initial block of `k` independent irreducibles
before comparing nearby lengths. This is the "good tuple" device; the
Jensen/logarithm telescoping over a dyad `n in [N, 2N)` leaves boundary
errors only, and the exceptional fraction of lengths in each dyad tends
to zero.
- Plane: choose an *internal* global height minimum, reflect the incoming
arm below it, insert a tall bridge of controlled multiplicity, and
translate the outgoing arm to glue. Then bound the inverse multiplicity
using the censored survival estimate: an output can be produced by at
most `s^a (r/s)^vartheta N^(2 eta)` distinct insertions, which suffices
after Jensen over `n` on a dyad.
The resulting endpoint lower bound holds for a single set of lengths of
natural density one. The upper spatial bound holds for every large length
at super-polynomial accuracy. The moments follow from these plus the
deterministic `D <= L`.
### Section 9: Forcing, spatial rates, prescribed endpoints (`responses`)
The forcing free energy (statement F). Gives the *directional
susceptibility* threshold: `m_e(t) := lim inf (-log G_t(b)) / |b-o|` is
`t^(3/4+o(1))` uniformly in `e`. The free-energy exponent
`f_e(s) = s^(4/3+o(1))` follows by Fekete subadditivity and convexity. The
corollary identifies the force-response exponent as `1/3`, matching
Pincus.
The prescribed-endpoint statement G. First proves a uniform two-dimensional
lattice lower bound (Proposition `R:resp:point-local`): for any compatible
count `k` near `H^(3/4)` and any target `y = (H, x)` with `|x| <= CH`, the
`p`-sum of `k` irreducibles has mass at least `H^(-5/4)` in a box. Then
chains these segments end-to-end in a corridor, giving a lower bound on
`G_{1/N}(b_N)`. Combined with the earlier absolute estimate on exceptional
paths, this identifies `L` and `D` at prescribed endpoint.
### Section 10: Confined renewal rewards (`alternatives`)
A self-contained weaker-input route to the infinite spatial law and the
completed-prefix expectation in a fixed-proportion ball. Uses only the
polynomial-vacuum estimates (`b_h = h^(-1/4+o(1))`, `m_h = h^(13/12+o(1))`)
and lateral confinement. Proves: infinite height lower bound
`n^(3/4 - o(1))`, distance upper bound `n^(17/18 + o(1))` (note: `17/18`,
not `3/4`; the weaker route does not reach the sharp spatial exponent), and
expected completed-prefix mass inside a ball `R^(4/3 + o(1))`.
### Section 11: Irreducible length tail from signed cylinder estimates (`signed`)
A second route to the length-tail exponent `alpha = 9/16`. Starts from
the *signed cylinder* boundary balance identity
`sum_b exp(3 i T(a,b)/8) K(a,b) = 1` (complex coefficients with positive
real parts under convexity). Derives strip masses, a slit bound, the
two-source positive kernel, then closes chords into polygons (via the
marked-polygon second-moment estimate). Converts mid-visiting chord mass
into bridge- or arch-mass by recovering cut levels. Yields
`P_p[L >= n] = n^(-9/16 + o(1))`, which is exactly `alpha - 1` applied to
the irreducible length tail, and also proves `p(H > r) = r^(-3/4 + o(1))`
and `p(D > r) <= r^(-3/4 + o(1))` again. Then packages into the fixed-height
and length-tilted conclusions.
### Section 12: Geometric transfers to infinite and thermal (`geometric-transfers`)
A third route, using insertion at a minimum again but with a different
multiplicity analysis and different finite inputs. Proves the infinite
spatial law (distance `n^(3/4)`, count `R^(4/3)`) by averaging over future
renewals: if `Gamma_h` is the stopped concatenation just after height `h`,
then `L(Gamma_h) ~ h^(4/3)` and `diam(Gamma_h) <= C h^(1+eps)`. Then
proves the thermal law by *tall-bridge insertion*: pick a tall bridge of
height `H ~ T^(3/4)`, place it between the two arms of an input, and bound
inverse multiplicities by the same avoidance machinery. Conclusion is
statements A and E again, from independent finite inputs.
---
## 6. Why 3/4 in the first place (informal)
You can read the paper as one long technical proof that a quantity equals
`3/4`. Where does `3/4` come from?
### 6.1 Not 1/2
An ordinary random walk (each step chosen uniformly among the neighbours,
no self-avoidance) after `n` steps is typically at distance about
`sqrt(n) = n^(1/2)` from the start. That is the classical *central limit*
picture. But a self-avoiding walk is *not* a random walk: it cannot revisit
a vertex, and that constraint pushes it out. In space it grows like
`n^(nu)` for some `nu > 1/2`. Finding the exact `nu` is a famous problem.
### 6.2 The Nienhuis guess
In 1982 Nienhuis analysed the `O(n)` spin model on the plane (a continuum
of `n`-component spins), which has a combinatorial expansion as a gas of
closed loops on the honeycomb. In the limit `n -> 0`, the loops become an
ensemble of self-avoiding walks. Nienhuis computed the critical point of
this `O(n)` model, and found the corresponding `nu` obeys
nu = 3/4 (exactly; at n -> 0, d = 2).
So `3/4` is a number predicted by a physical field theory in 1982, not a
trivial fact. Making it a theorem about actual combinatorial SAWs on the
honeycomb is what the paper does, in several different ensembles.
### 6.3 The critical density is known
The critical weight `rho = 1 / sqrt(2 + sqrt(2))` was proven by
Duminil-Copin and Smirnov in 2012 using a *parafermionic observable* on the
honeycomb: a complex-valued function on mid-edges of the lattice whose
discrete contour integral gives exact relations on `rho`. This gives one
boundary of the picture. The paper takes `rho` as an input and asks what
spatial law critical SAW obeys.
### 6.4 The conjectural scaling limit
It is conjectured (and strongly supported) that the scaling limit of
critical planar SAW is `SLE(8/3)`, a Schramm-Loewner evolution curve. SLE
curves have a deterministically-known *fractal dimension*; `SLE(8/3)` has
dimension `4/3`. A curve of fractal dimension `d_f` fills space so that a
ball of radius `R` contains about `R^(d_f)` length; the relation between
`d_f` and `nu` is `d_f = 1/nu`. With `nu = 3/4` you get `d_f = 4/3`; with
`d_f = 4/3` you get `nu = 3/4`. The two are the same fact.
### 6.5 Why it is hard
Information-theoretically, deciding `nu` from first principles is hard
because the count of SAWs is a single integer that jumps erratically
between lengths; the "local" behaviour at one exact `n` is not a smooth
function. That is why sections 3 and 4 of the paper spend most of their
effort getting *exact* single-length denominators rather than an averaged
one. In the plane, the further obstruction is that a uniform random point
can be reached by many distinct configurations; counting how many distinct
configurations produce the same spatial pattern (the "inverse
multiplicity") is one of the recurring technical difficulties.
### 6.6 Analogy to Brownian motion
If you replace the self-avoidance constraint by a weak repulsion (as in the
weakly self-avoiding walk), the walk is diffusive and `nu = 1/2`. The
critical `3/4` of SAW is a *critical exponent* tied to a phase transition;
the paper shows that at the point where the walk's generating function has
a singularity, the way the singularity approaches gives `3/4`.
---
## 7. Technical honesty: what is proven, what is qualified
The paper is careful about the scope of every theorem. Here is a checklist
of the qualifications, so the high-school reader knows not to overclaim:
1. **"3/4 in probability"** means for every fixed `epsilon > 0` the event
`n^(3/4 - epsilon) <= X <= n^(3/4 + epsilon)` has probability tending to
one. It does not say `X = n^(3/4)` exactly, nor that the ratio
`X / n^(3/4)` converges to a constant.
2. **The uniform fixed-length statement (D) is on a density-one set of
lengths, not all lengths.** The reason is structural: proving the
endpoint lower bound at a single length needs mass around that length,
and mass at precise lengths is only known to be large for certain
lengths (not all). The paper's alternative proofs (sections 3-4) do give
exact-length lower bounds for *strict bridges*, but the endpoint-free
fixed-length case is genuinely harder.
3. **Even length.** In the port convention, bridges have even length
because the two ends live on ports of opposite triangle types. Odd
lengths require bounded local modifications of the port convention, which
the paper does supply (see stated corollaries) but at bounded cost in
exponents.
4. **Bounded exponents in `log log` are ignored.** All conclusions have
`o(1)` error in the exponent, i.e. `n^(3/4 + o(1))`, not
`n^(3/4 + c)` for a specific `c`. This is the same precision as
conformal field theory predictions; it is enough to identify `nu` but it
is not a sharp asymptotic expansion.
5. **Black boxes from companion preprints.** The paper proves the *changes
of law* itself. The *finite estimates* (bridge masses, polygon moments,
the parafermionic boundary balance) are imported from companion
preprints cited in the bibliography. Each section states which companion
supplies which estimate, so the dependency graph is explicit.
6. **Some results are proved twice for robustness.** In particular the
irreducible length tail `alpha = 9/16` is reached both from the local
renewal theorem (section 3) and from the signed cylinder boundary balance
(section 11). The infinite spatial law `nu = 3/4` is reached both by
renewal growth (section 2) and by geometric transfer (sections 10 and 12).
Each route uses its own finite input; if one were repaired by different
companion estimates the other would still run.
7. **The thermal statement is unconditional in `N`.** Once the half-plane
remainder cancellation is available (Section 4.5) and the full-plane
denominator is bounded below (Section 6), the thermal spatial laws hold for
every large discount `N`, not just along a subsequence or a density-one
set. This is one of the cleanest conclusions of the paper and the reason
the thermal framework is emphasised.
8. **Not a proof of the full conformal invariance.** The paper identifies
the exact power `nu = 3/4` under several conditioning rules. It does not
prove that the full scaling limit is `SLE(8/3)`; that is a stronger,
separate conjecture. The paper only pins the exponent.
9. **Compatibility condition on the target length.** In some statements
(e.g. statement B) the length must be a "compatible" length, i.e. of
the parity and vertex type that allows a bridge. For all large even `n`
this is automatic. This is a parity artifact of the port construction,
not a real restriction on the exponent.
10. **Prescribed-endpoint conditioning works only at scale `|b - o| ~
N^(3/4)`.** The theorem G assumes the target vertex is at exactly the
typical distance. Conditioning on an *atypically close* or *far*
endpoint is a different problem, and the paper explicitly does not
address it; the "spatial rate" theorem treats that regime only
asymptotically in `N`.
All of these qualifications together mean: this is a *complete* proof of
statements A-G as literally stated (with the density-one and even-length
qualifiers), but not a proof of every conceivable statement one might
wish for critical honeycomb SAW.
---
## 8. Glossary of symbols and quantities
Here is every symbol that appears, in one place.
Geometry
`honeycomb graph` the graph of triangle centers in the unit triangular tiling
`d0` Euclidean height of one integer row; d0 = sqrt(3)/2
`port` boundary point where a walk enters or leaves a strip
`strict bridge` walk between two horizontal lines, vertices strictly inside
`height h` integer number of rows between bottom and top lines
`L` number of vertices visited (vertex length)
`H` integer height of an irreducible
`X` signed horizontal displacement (transverse coordinate)
`A` endpoint distance |last vertex - first port|
`D` Euclidean diameter of the visited trace (ports included)
`S` horizontal span (range of transverse coordinates)
Critical weights
`rho` critical density, rho = (2 + sqrt(2))^(-1/2)
`weight` rho^L for a walk visiting L vertices; ports weigh 1
`mu` connective constant: c_n^(1/n) -> mu, rho * mu = 1
Mass and probability
`B_h` total critical mass of strict bridges of height h
`M_h` sum over height-h bridges of L * rho^L (first length moment)
`B(s)` generating function sum_h B_h e^(-s h) = 1 / (1 - E_p[e^{-sH}])
`p` probability law on irreducible bridges (sum of rho^L is 1)
`E_p[f]` expectation of f under p
`q_L(t)` 1 - E_p[e^{-tL}], the length Laplace deficit
`q_N` = q_L(1/N) = N^(-9/16 + o(1))
Exponents (the ones that matter)
`nu` spatial exponent, nu = 3/4
`d*` fractal dimension, d* = 4/3 = 1/nu
`alpha` irreducible-length tail exponent, alpha = 9/16
`beta` irreducible-height (and diameter) tail exponent, beta = 3/4
`theta` rate rho^2 / (1 + rho^2) used in the half-plane ratio
`vartheta` 1 - beta = 1/4 in the calibrated / censored / uniform sections (boundary exponent); in the signed-cylinder section 11 the symbol instead denotes the angle pi/8, so watch it across sections
Laws
`p` the irreducible probability law (as above)
`rho^L` fixed-length critical mass
`rho^L e^{-L/N}` thermal (Boltzmann) weight at discount scale N
`U_N` total z-weight of "no further renewal" half-plane ends
`H_N` total half-plane z-mass = U_N / q_N
`Z_N` total full-plane z-mass (>= c / q_N)
Paths and events
`S_k` sum of k independent copies of a random variable
`Gamma_h` stopped concatenation of irreducibles just above height h
`g` height threshold in the censored two-arm survival Lemma
`E_k` event that first k irreducibles of two adjacent arms stay disjoint
Control quantities
`epsilon`, `xi`, `delta`, `eta`, `zeta` arbitrary fixed positive numbers; every conclusion has an `o(1)` (or `o_k(1)`) slack and never depends on the choice of these constants after the proof
Theory names
`O(n) model` loop gas of closed paths with "n" loop weight; SAW at n=0
`SLE(kappa)` Schramm-Loewner evolution; kappa = 8/3 for SAW scaling
`parafermion` the Duminil-Copin - Smirnov holomorphic observable
`signed cylinder` the representation with complex boundary weights exp(3 i T / 8) used in section 11
`marked polygon` a polygon with a distinguished face, used for chiral / signed estimates
---
## 9. Reading guide
If you are a high-school reader who wants the headlines, read in this order:
1. Sections 0, 1.1-1.4 (cast of characters).
2. Section 3 A, B, C, D, E (main theorems; you can skip the notation caveats).
3. Section 6 (why 3/4).
4. Section 7 (what is proven; the honesty checklist).
If you want a bit of the mechanism, also read:
5. Section 2 (the three laws and why they are different).
6. Section 4.1-4.4 (the renewal machine and `alpha/beta = 3/4`).
If you are a graduate student or professional who wants to know what is
actually proven here, the full mathematical content is the 12 paper
sections. Nothing in Sections 0-8 above goes beyond the paper in strength
or generality.
The prerequisite for a serious reader is:
- basic probability (expectation, variance, Borel-Cantelli, characteristic function, Fourier inversion);
- generating functions (formal power series, coefficient extraction, submultiplicativity);
- the definition of self-avoiding walk and the honeycomb lattice;
- the classical local renewal theorem in one dimension (Garsia-Lamperti / Caravenna-Doney).
Everything else (the parafermionic observable, the marked-polygon machinery,
the signed cylinder representation) is developed in the paper itself or in
its companion preprints, and this exposition makes no attempt to duplicate
the analytic content of those.
---
## 10. References
Classical references
Nienhuis 1982. Exact critical point and critical exponents of O(n) models
in two dimensions. Phys. Rev. Lett. 49, 1062-1065.
(Predicts nu = 3/4 for SAW.)
Duminil-Copin, Smirnov 2012. The connective constant of the honeycomb
lattice equals sqrt(2 + sqrt(2)). Ann. of Math. 175, 1653-1665.
(Proves the critical density rho.)
Kesten 1963. On the number of self-avoiding walks.
J. Math. Phys. 4, 960-969.
Hammersley, Welsh 1962. Further results on the rate of convergence to
the connective constant of the hypercubical lattice.
Quart. J. Math. 13, 108-110.
Madras, Slade 1993. The Self-Avoiding Walk. Birkhauser.
(Standard book with the irreducible-bridge decomposition.)
Garsia, Lamperti 1962. A discrete renewal theorem with infinite mean.
Comment. Math. Helv. 37, 221-234.
Caravenna, Doney 2019. Local large deviations and the strong renewal
theorem. Electron. J. Probab. 24, paper 72, 1-48.
On the half-space and strip points of view
Lawler, Schramm, Werner 2004. On the scaling limit of planar
self-avoiding walk. Proc. Symp. Pure Math. 72, 339-364.
Dyhr, Gilbert, Kennedy, Lawler, Passon 2011. The self-avoiding walk
spanning a strip. J. Stat. Phys. 144, 1-22.
(Explicit finite-prefix description used in section 2.)
Beaton, Bousquet-Melou, de Gier, Duminil-Copin, Guttmann 2014.
The critical fugacity for surface adsorption of self-avoiding walks on
the honeycomb lattice is 1 + sqrt(2). Comm. Math. Phys. 326, 727-754.
On sub-ballisticity (the counterpart result)
Krachun, Panagiotis 2026. Quantitative sub-ballisticity of self-avoiding
walk on the hexagonal lattice. Ann. Probab. 54 (no. 3), 1109-1125.
On forcing and Pincus scaling
Pincus 1976. Excluded volume effects and stretched polymer chains.
Macromolecules 9, 386-388.
Ioffe, Velenik 2008, 2010. Ballistic phase of self-interacting random
walks; The statistical mechanics of stretched polymers.
Schilling 2016. An introduction to Levy and Feller processes.
(Background on infinitely divisible limits.)
On the general probabilistic infrastructure
Ioffe 1998. Ornstein-Zernike behaviour and analyticity of shapes for
self-avoiding walks on Z^d. Markov Process. Related Fields 4, 323-350.
Borgs, Chayes, King, Madras 2000. Anisotropic self-avoiding walks.
J. Math. Phys. 41, 1321-1337.
Companion preprints (finite estimates used as black boxes)
Radial transfer estimates and polygon length laws for honeycomb walks.
Uniform marked-polygon estimates and sharp finite bridge moments.
Cylinder amplitudes and logarithmic bridge-length windows on the honeycomb
lattice.
Mass and covering exponents for fixed-length honeycomb walks.
Cylinder loop weights and planar nesting.
Disk transfer representations and confined bridge mass.
Signed cylinder propagation and marked polygons on the honeycomb lattice.
Polynomial vacuum representations and bridge mass for honeycomb walks.
Critical strip-crossing mass on the honeycomb lattice.
Each section of the main paper cites exactly which of these supplies its
input, so the reader who wants to see the finite estimates proved in full
can follow the citations mechanically.
---
## 11. Closing note for the reader
The single sentence of this paper is: *on the critical honeycomb lattice,
self-avoiding walk scales with the exactly determined spatial exponent
`nu = 3/4` — equivalently fractal dimension `4/3` — and this remains true
across uniform fixed-length, infinite irreducible, thermal, and
prescribed-endpoint ensembles, each proven from the corresponding renewal /
geometric identity.*
What makes it hard is that the exponent has to survive a change of
measure. Counting walks by length, by height, by area, by discount, and by
endpoint produces five different denominators; the paper shows that they
all produce the same spatial exponent. What makes it tractable is the
irreducible-bridge factorization: at criticality the irreducibles form a
probability law, and every other law becomes a *change of that law*. What
makes it plausible is the physical conformal picture, which this paper
does not prove but which motivates the exact numbers `9/16`, `3/4` and
`4/3` that appear throughout.
That is the whole story. Every technical line in the original paper exists
to make one of the ten honesty checks of Section 7 come out right.