The honeycomb lattice H is the graph whose vertices are the centers of the
triangles of a unit triangular tessellation, with two vertices joined when the
triangles share a side. Three edges meet at every vertex. Euclidean distance,
diameter, and all geometric quantities refer to this embedding. Each triangle
of the tessellation has a center, and every walk below traverses triangle
centers. A port is the midpoint of a tessellation side on the boundary of a
region. A first-exit port-to-port walk in a region D is a self-avoiding walk
inside D with both endpoints on ports of D and no other vertex on a port of D.
A walk traversing N vertices has weight x^N where x = 1/sqrt(2+sqrt(2)) =
0.5411961... is the reciprocal of the honeycomb connective constant. Any
positive weight below is a product of such powers.
The connective constant of the honeycomb lattice equals sqrt(2+sqrt(2)),
established by
H. Duminil-Copin and S. Smirnov, The connective constant of the
honeycomb lattice equals sqrt(2+sqrt(2)), Annals of Mathematics 175
(2012), 1653--1665.
./signed_cylinder_pdfs/duminil-copin-smirnov-2012-1202.5523-honeycomb-connective.pdf
A polygon is a simple closed unoriented cycle on H; its length |lambda| is its
number of edges. Two polygons are incompatible when they share a vertex; each
is incompatible with itself. Polygon collections are summed with product
weights x^{sum |lambda|}.
Equilibrium has the following physical interpretation. Nienhuis's Coulomb-gas
analysis of the O(n) loop model predicted the two-dimensional self-avoiding
walk exponents, including nu = 3/4. The prediction is recorded in
B. Nienhuis, Exact critical point and critical exponents of O(n) models
in two dimensions, Physical Review Letters 49 (1982), 1062--1065,
and in the review
B. Nienhuis, Critical and multicritical O(n) models, Physica A 163
(1990), 152--157.
The exponent nu = 3/4 corresponds to reciprocal 1/nu = 4/3, which appears
below as the mean-first-exit-chord exponent and as the leading decay of the
marked two-bond sum. The empirical vanishing of critical bridges was proved
by Beaton, Bousquet-Melou, de Gier, Duminil-Copin and Guttmann in
N. R. Beaton, M. Bousquet-Melou, J. de Gier, H. Duminil-Copin and A. J.
Guttmann, The critical fugacity for surface adsorption of self-avoiding
walks on the honeycomb lattice is 1+sqrt(2), Communications in
Mathematical Physics 326 (2014), 727--754.
./signed_cylinder_pdfs/beaton-bousquet-melou-de-gier-duminil-copin-guttmann-2014-1406.3163.pdf
For honeycomb bridges Glazman and Manolescu obtained a logarithmic
subsequence bound, and for columnwise rhombic half-plane tilings they proved
invariance of boundary two-point functions; see
A. Glazman and I. Manolescu, Self-avoiding walk on Z^2 with Yang-Baxter
weights, Annales de l'Institut Henri Poincare B 56 (2020), 2281--2300.
Krachun and Panagiotis obtained a polynomial upper bound on the critical bridge
partition and used it to prove quantitative sub-ballisticity for uniform
honeycomb walks; see
D. Krachun and C. Panagiotis, Quantitative sub-ballisticity of
self-avoiding walk on the hexagonal lattice, Annals of Probability 54
(2026), 1109--1125.
The local mechanism on which the present argument depends is the parafermionic
cancellation of Duminil-Copin and Smirnov, extended to rhombic Yang-Baxter
weights by Glazman and Manolescu and by Ikhlef and Cardy:
Y. Ikhlef and J. Cardy, Discretely holomorphic parafermions and integrable
loop models, Journal of Physics A 42 (2009), 102001.
Three observables are treated. The first is a cylinder partition with two
marked points. Fix an even integer M = 2m, identify horizontal tessellation
sites modulo M, and mark the face centers O = (0,0) and O' = (m,0) on the cut.
The cut sites are split into consecutive lists A (from O to O') and B (from
O' back to O), each of size m. The cylinder is infinite vertically and is
defined by the increasing height limit of finite cylinders with both rims
closed (every port on either rim unused). A separating cycle is a simple
closed cycle of H whose complement places O and O' in different components.
Let G_m(g) be the sum over families of pairwise disjoint separating cycles of
prod_{lambda in family} g x^{|lambda|},
the empty family contributing 1. Because a separating cycle crosses the cut
between the marked points, G_m(g) has degree at most 2m in g.
The second observable is the chord law inside a regular tessellation hexagon
D_R of side R. Let Z(R) be the sum over collections of disjoint polygons
inside D_R that surround the central face, with weight 2 x^{|lambda|} per
polygon and empty collection contributing 1. Let C_R be the set of first-exit
chords of D_R of diameter at least R/10, where a chord is a first-exit walk
whose endpoints are boundary ports of D_R, the sum runs over every starting
port and every possible first-exit port, and the direction from start to exit
is retained. Define
Z_R^chord = sum_{gamma in C_R} x^{N(gamma)},
P_R^chord(gamma) = x^{N(gamma)} / Z_R^chord.
The chord length N(gamma) is the number of vertices of gamma.
The third observable is the two-bond sum for planar polygons. For H > 0, put
S(H) = sum_{diam lambda <= H} x^{|lambda|} |lambda|^2,
summed over polygons modulo lattice translations. The squared length factor
|lambda|^2 counts ordered pairs of bonds inside lambda; the sum over such
pairs decomposes S(H) into a sum over two-bond configurations, whose control
is the content of the third theorem.
Through all statements, F(R) = R^{a+o(1)} means that log F(R) / log R
converges to a, an upper bound F(R) <= R^{a+o(1)} means: for every eps > 0
there are C_eps and R_eps such that F(R) <= C_eps R^{a+eps} for all
R >= R_eps, and a lower bound is analogous. A bound is uniform in some data
when the constants C_eps and thresholds R_eps do not depend on that data.
Theorem (cylinder pressure and planar nesting). As m -> infinity through even
positive integers, the balanced cylinder partition with two marks satisfies
G_m(2) = m^{1/6+o(1)},
-(d^2/dg^2) log G_m(g) at g = 0 is O(log m),
the derivative being the total weight of ordered pairs of incompatible
separating cycles (pairs sharing at least one vertex, including each cycle
with itself), and locally uniformly for complex Theta with |Re Theta| < pi,
|G_m(2 cos Theta)| <= m^{Re(1/6 - 2 Theta^2/(3 pi^2)) + o(1)}.
The planar nesting partition satisfies Z(R) = R^{1/12+o(1)} as R -> infinity
through integers.
Theorem (finite chords and marked polygons). For the chord law of D_R,
Z_R^chord = R^{3/4+o(1)},
E_R^chord N = R^{4/3+o(1)},
and uniformly for honeycomb vertices v at distance at most R/10 from the
central face,
P_R^chord(v lies on gamma) = R^{-2/3+o(1)}.
Moreover,
S(H) <= H^{2/3+o(1)}.
Theorem (unequal marked intervals). On a physical tilted cylinder whose
unmarked lists have m+1 and n+1 sites, with n >= m and M = m+n+2 even, with
unmarked logarithmic heights -f and 1-f where 0 <= f <= 0.6, and with site
counts differing from m(1-f), mf and n(1-f), nf by at most a fixed constant,
let H^poly be the total critical weight of a single unoriented polygon
crossing two horizontal-type marked bonds. Then uniformly over the lists and
orders, for every eps > 0 there are C_eps and m_eps such that
H^poly <= C_eps n^{eps} m^{-4/3} (n/m)^{1/3}
for all n >= m >= m_eps.
The remainder of the paper proves these statements in an interleaving that
mirrors their mechanism. The boundary balance and local transfer identities
are established first, since they seed the local cancellation and the finitude
of vacuum limits. The half-plane arch estimate then controls boundary mass.
The cylinder theorems proceed through an oriented transfer, Yang-Baxter and
fusion identities, a finite spin sum, contour and bosonization rewritings, an
absolute stability estimate on the resulting signed gas, and a compact
propagator that isolates the physical power. The chord and visit corollaries
follow by combining the arch estimate, a boundary visit identity, and the
planar nesting exponent. The two-bond estimate is proved independently, with
an analogous signed-transfer pipeline but with two marked bonds instead of one
unmarked cycle.
Throughout, s = 3/8 and alpha = pi/8, so 2 x cos(s pi / 3) = 1. The seed is
the parafermionic turn balance of Duminil-Copin and Smirnov.
Lemma (convex balance). Let D be any finite convex tessellation polygon, and
let a be a port on the boundary of D. The signed sum over first-exit port-to-
port walks gamma of D from a, weighted by x^{N(gamma)} e^{i s T(gamma)}, where
T(gamma) is the total turning of gamma in radians and orientation is counted
with sign, equals one.
Proof. Start with the nonbacktracking tree of walk continuations from a,
with an outgoing factor x e^{i s pi/3} or x e^{-i s pi/3} at each vertex
depending on whether the walk turns left or right. The two outgoing factors
sum to 2 x cos(s pi / 3) = 1 at every interior vertex. The tree stops the
first time an already-visited vertex is about to be entered. Every leaf of
this tree is one of the following: a boundary port (that is, a first-exit
walk, which is the quantity being summed), or a first self-collision loop,
closed by the incoming stem.
A first self-collision closes a simple loop. In the flat annular region
formed by the loop together with the stem, the stem approaches the loop from
outside. The loop's turn contribution is either +5pi/3 or -5pi/3 (the loop
itself has turn +-2pi, but one turn of pi/3 at the closure corner has been
dropped because only two of the three branches at that interior vertex are
used), and the stem's turn onto the loop is -+pi/3. The two orientations of
the loop are summed with weight e^{i s(+-5pi/3 -+ pi/3)}. The total turn per
orientation is +-4pi/3, so the combined phase contribution at the self-
collision is 2 cos(4 pi s / 3) = 2 cos(pi / 2) = 0. Every self-collision
therefore cancels. Only first-exit walks remain, and the total signed sum
equals the same factor that the interior vertices put at each turn, which is
one.
For a first-exit walk to a port b, the total turning T(gamma) equals Delta -
pi, where Delta is the counterclockwise boundary tangent turn accumulated
from a to b along the boundary, together with the two extra quarter turns
needed to close the polygon. Each boundary traversal below contributes
cos(s T) >= sin(alpha) > 0 because Delta is a boundary turn plus a fixed
compensation. Taking real parts of the balance and using the positivity gives
the finiteness of the total positive weight of first-exit walks. QED
Let H(d) be the total positive weight x^N of first-exit walks on the half-
plane above a horizontal tessellation line, from a fixed boundary port to the
port d steps to the right. Reflection in the vertical axis gives the same
total to the left. Taking real parts of the convex balance in the half-plane
and using the two equal sides, gives
2 sin(alpha) sum_{d>0} H(d) <= 1,
with strict equality in the limit and no contribution from d = 0 (a first-exit
arch has positive length). In particular sum_{d>0} H(d) <= 1/(2 sin alpha) <
3/2, so the sum converges.
The following transfer-matrix framework realizes the balance identically as a
matrix-element equation. Consider a cylinder obtained by identifying modulo M
steps horizontally, with rows tiled by M rhombi of basis 1 and e^{i theta},
with theta potentially varying by row. Inside each rhombus, connections
(possibly empty) pair up bond midpoints on distinct sides, using each midpoint
at most once, noncrossingly. The empty weight is 1, a straight connection
between opposite sides weighs v, single turns of magnitudes theta and
pi - theta weigh u and d respectively (where port directions are normal to
sides), and two simultaneous turns weigh a for the pairing left-down and
up-right, or b for left-up and right-down. Write phi = 3 theta / 8 and let
s_k = sin(phi + k alpha), h = sin(2 alpha). The trigonometric SAW weights are
(u, d, v, a, b) = (h s_5, h s_0, s_0 s_5, s_5 s_6, -s_0 s_7) / (s_2 s_3).
At phi = alpha these are (x, x^2, x^2, x^2, 0), exactly the honeycomb weights
after splitting each rhombus into the two equilateral triangles.
For a finite stack, close both rims. The partition Z of one unoriented
noncontractible polygon is the linear coefficient in a fugacity n of the full
loop sum assigning weight zero to contractible loops and weight n per
noncontractible loop. For real angles 0 < theta < pi, orient each loop with a
factor exp(-i T / 4) and twist p to the signed number of seam crossings (east
positive); summing over the two orientations produces the loop weights n =
p + p^{-1}. This is because an embedded loop in the flat cylinder is either
contractible with turn +-2 pi, or has winding +-1 and turn zero. The turning
identities then extend to the rhombic model algebraically.
The transfer matrix. On each horizontal-flow and descending-flow bond put a
spin in {0, +1, -1}. The oriented matrix acts with row indices (left, up) and
column indices (right, down), with positive signs meaning left-to-right or
up-to-down on the corresponding bond. After removing the telescoping phase
e^{i theta (left - right) / 4}, the resulting R(phi) is block-diagonal with
respect to the total sum s_left + s_up = s_right + s_down = J. Writing
lambda = e^{-2 i alpha}, the blocks are
(s', s'), s' = +-1 : [a]
(s', 0), (0, s') : [[v, u], [u, v]]
(1, -1), (0, 0), (-1, 1) : a 3 x 3 block
[[b, d lambda, b lambda^2 + a],
[d lambda^{-1}, 1, d lambda],
[b lambda^{-2} + a, d lambda^{-1}, b]].
The turns connecting left to down have magnitude theta (weights u, d, a, b as
above); those connecting left to up have magnitude pi - theta (weights v, u,
d, b, a with the substitutions s_j -> s_{j+4} of the trig functions).
Exchange. Let V = C^3 with basis |0>, |+>, |->, and let subscripts index
ordered tensor factors. On a copy V_i let s_i act by s_i |k> = k |k>. Write
P_{ij} for interchange of factors i, j, and put Delta_R(psi) = sin(psi + 2
alpha) sin(psi + 3 alpha).
Lemma (local exchange and fusion). The entries of R(psi) are meromorphic,
with possible poles at Delta_R(psi) = 0, and the following are identities of
meromorphic matrix functions:
R_{12}(t) R_{13}(t + phi) R_{23}(phi) = R_{23}(phi) R_{13}(t + phi) R_{12}(t),
R_{12}(phi) R_{21}(-phi) = I.
In particular R_{12}(0) = P_{12}.
Proof. The unitarity R_{12}(phi) R_{21}(-phi) = I follows by block multipli-
cation, using lambda^4 = -1 and the elementary relation aa' = 1, ab' + ba' +
dd' = 0, ad' + d = a'd + d' = 0, vv' + uu' = 1, vu' + uv' = 0, where primes
denote the same quantity at -phi. Each displayed scalar identity reduces to
sin_j' = sin_{8-j} and product-to-sum.
To prove the Yang-Baxter exchange at the three special values t = 0, 2 alpha,
3 alpha, it suffices to check the factorizations R_{12}(0) = P_{12},
R_{12}(2 alpha) = J_2 J_2^*, R_{12}(3 alpha) = J_3 J_3^*, together with the
three-space identity in each block; the block calculations are routine. For
general t the exchange identity is proved by noting that, upon multiplying
the exchange difference by the product of the three Delta_R factors appearing
(t, t + phi, phi), every entry becomes a polynomial in e^{2 i t} of degree at
most 4. The six values 0, 2 alpha, 3 alpha, -phi, 2 alpha - phi, 3 alpha - phi
are distinct modulo pi for generic real phi, so this polynomial of degree at
most 4 vanishes if it vanishes at these six values. Each of the six reduces
by the swap symmetry R(psi + pi) = Z_1 R(psi) Z_1 where Z |k> = (-1)^{k^2}
|k>, to one of the factorizations at 0, 2 alpha, 3 alpha. Continuation in the
two parameters proves the meromorphic identity. QED
The exchange identity has a consequence about the family of rows: rows of any
quantity whose entries are R-matrices with a common twist have a simultaneous
eigenbasis, and their eigenvalues depend only on the row's rapidity. Fusion
extends this to admitting a row to be collapsed.
Fusion. Two adjacent rows can collapse when their rapidities differ by a
quantized amount. Set E_2 = V and E_3 = C, and define injective maps
J_delta : E_delta -> V tensor V by
delta = 2:
J_2 |0> = |00> + x sum_{k = +-1} lambda^k |k, -k>
J_2 |k> = x (|k, 0> + |0, k>) for k = +-1
delta = 3:
J_3 (1) = sum_{k = 0, +-1} lambda^k |k, -k>.
On E_2 put s_f |k> = k|k>, and on E_3 put s_f = 0.
Lemma (fusion factorization). For delta = 2, 3 the identity
R_{12}(delta alpha) = J_delta J_delta^*
holds in V tensor V, and for any extra factor V_k the meromorphic site
intertwining holds:
R_{23}(phi) R_{13}(phi + delta alpha) (J_delta tensor I)
= (J_delta tensor I) R_f(phi)
R_{a1}(psi) R_{a2}(psi + delta alpha) (I_a tensor J_delta)
= (I_a tensor J_delta) R_f(psi + [alpha for delta = 2 else 0])
with R_f(phi) = R(phi + alpha) for delta = 2 and R_f = I for delta = 3.
Proof. At the three values psi = 0, 2 alpha, 3 alpha the five weights are,
respectively,
psi = 0 : (1, 0, 0, 1, 0)
psi = 2 alpha : (x^2, x, x^2, 0, x^2)
psi = 3 alpha : (0, 1, 0, 0, 1).
The R(psi) block decompositions at these values produce the depicted images
directly. For instance, the charge-zero column of J_2 is (x lambda, 1, x
lambda^{-1})^T, and the charge +-1 columns are x (1, 1)^T; substituting into
J_2 J_2^* reproduces R_{12}(2 alpha). The factorizations are not isometries:
J_2^* J_2 = diag(3 - sqrt 2, 2 - sqrt 2, 2 - sqrt 2) in the standard basis,
and J_3^* J_3 = 3.
The three-space intertwining is verified by explicit block multiplication in
each charge sector. The scalar identities needed reduce, after cancelling
common factors, to a finite list of sine product-to-sum identities together
with lambda^2 + lambda^{-2} = 0; the block-by-block check is direct. The
row-space formula j:site-fusion is the transpose-in-spaces image of the
two-space version, obtained by conjugating the row intertwining and exchanging
factors 1 and 2, using the invariance of J_delta under complex conjugation in
real rapidities, and then continuing meromorphically. QED
Lemma (restriction to fusion image, and quotient at fusion). Fix delta in
{2, 3} and put W_delta = im J_delta subset V_1 tensor V_2. For an extra factor
V_k set F_{delta, k}(psi) = R_{2k}(psi) R_{1k}(psi + delta alpha). At regular
parameters the subspace W_delta tensor V_k is invariant; transported through
J_delta tensor I the restriction reproduces R_f(psi). At psi = 0 the full
image of F_{delta, k}(0) lies in W_delta tensor V_k, so the induced operator
on ((V_1 tensor V_2) / W_delta) tensor V_k is zero at psi = 0 and holomorphic
near psi = 0 (being a polynomial in the local parameter divided by Delta_R
factors that do not vanish there). The twist restricts by
p^{s_1 + s_2} J_2 = J_2 p^{s_f} and p^{s_1 + s_2} J_3 = J_3 for every p != 0.
Proof. The invariance and the action on the image are the site intertwining
of the previous lemma, transported through the identification E_delta -> W_delta
supplied by J_delta. At the fusion point the factorization supplies the
stronger inclusion
F_{delta, k}(0) = P_{2k} R_{1k}(delta alpha)
= P_{2k} J_{delta, 1k} J_{delta, 1k}^*,
im F_{delta, k}(0) subset W_delta tensor V_k,
where J_{delta, 1k} is the embedding applied on factors 1 and k. Indeed
P_{2k} carries im J_{delta, 1k} onto im J_{delta, 12} = W_delta. The quotient
operator is holomorphic at psi = 0 because the local parameters 0, delta alpha
are regular; its vanishing at psi = 0 then factors as (z - 1) times a
holomorphic function in the local coordinate z = e^{2 i psi}. The twist
identities follow because every term of J_2 |s_f> has total spin s_f, and
every term of J_3 has total spin zero. QED
Fourier kernel. The finite-vacuum normalization and the boundary moment
problem both use the same elementary transform.
Lemma (Fourier kernel). For 0 < beta < pi and |Im y| < pi - beta,
int_R e^{i k y} (sinh(beta k) / sinh(pi k)) dk
= sin(beta) / (cosh y + cos beta).
The integral and all parameter derivatives converge locally uniformly on
the indicated strip.
Proof. For real y > 0 close the contour by rectangles in the upper half-plane
with upper edges at half-integer heights. The ratio of hyperbolic sines decays
exponentially on the vertical edges, and the factor e^{i k y} controls the
upper edge as its height tends to infinity. The residues at k = i n (n >= 1)
give
2 sum_{n >= 1} (-1)^{n + 1} sin(n beta) e^{-n y}
= sin(beta) / (cosh y + cos beta).
The geometric-series evaluation proves the last equality. Evenness extends it
to negative y, and dominated convergence covers y = 0. Exponential decay in
|k| proves local uniform convergence and parameter differentiation on smaller
strips. Analytic continuation in y gives the asserted domain. QED
The transfer formalism, the convex balance, and the fusion identities seed
everything that follows. Their combination produces a signed-transfer
representation of both the cylinder partition and the two-bond sum. The
remaining work is to make the representation's individual signed terms
summable and to read off the exponents from the absolute sum.
The half-plane arch estimate. Fix a port on the boundary of a horizontal
tessellation half-plane. Let W_R be the total critical weight x^N of first-
exit same-boundary arches reaching Euclidean distance at least R from that
port, and for d > 0 let H(d) be the total weight of first-exit arches from
the fixed port to the port d boundary steps to its right.
Proposition (boundary arch mass). W_R = R^{-1/4+o(1)}, and
2 sin(pi/8) sum_{d>0} H(d) = 1,
sum_{d >= n} H(d) <= n^{-1/4+o(1)}.
Proof.
The fused cylinder increment. On a finite cylinder of circumference M with
closed rims, let Z be the partition of exactly one unoriented noncontractible
polygon. Write T_p(phi) for the oriented row transfer matrix with row
parameter phi, and put alpha = pi/8. Row matrices with the same twist commute:
multiplying the local exchange identity across the sites and tracing over the
two auxiliary spaces, the twist p^{s_1 + s_2} commutes with the exchange
matrix. Then they commute wherever that matrix is invertible, and identically
by continuation.
Set U = im J subset V_1 tensor V_2 for the embedding at 2 alpha, and at site
j put X_j(phi) = R_{2j}(phi) R_{1j}(phi + 2 alpha). Locally this preserves
U tensor V_j and induces R(phi + alpha) through the isomorphism J; at phi = 0,
X_j(0) = P_{2j} R_{1j}(2 alpha), im X_j(0) subset U tensor V_j,
because R_{1j}(2 alpha) = J_{1j} J_{1j}^* and the swap carries embedded
spaces 1, j to 1, 2. The operator induced on the quotient by U tensor V_j is
O(phi), so its product over M sites is O(phi^M). Taking the auxiliary trace
gives
T_p(phi) T_p(phi + 2 alpha) = T_p(phi + alpha) + O(phi^M), (*)
locally holomorphic in (phi, p) near (0, i). Extract the coefficient of n =
p + p^{-1}: at p = i, dn/dp = 2, which preserves the order of vanishing.
Now add one new outer row on L honeycomb rows with all parameters alpha and
subtract the partition without the new row. Let H_M(d) be the weight of half-
plane arches from a fixed port to the port d steps right, disjoint from all
their nontrivial horizontal translates by M. For fixed M these converge as
L -> infinity, since dropping mutual disjointness bounds each by products of
finitely many H(d) and the coarse summability from the convex balance applies.
The tile weights satisfy u d = c_alpha v (1 - v) with c_alpha = 2 sin alpha.
A k-interval term has at most M - 2k straight tiles, so the limiting increment
is a polynomial F(v) of degree at most M, and
F(v) = v^M + M c_alpha v (1 - v) sum_{d = 1}^{M-1} v^{M - d - 1} H_M(d)
+ (1 - v)^2 Q_M(v),
F(1) = 1,
F'_v(1) = M (1 - c_alpha sum_{d = 1}^{M-1} H_M(d)),
for a polynomial Q_M. Inclusion--exclusion for two added rows, and domination
by positive honeycomb weights, pass (*) to the limit. With y = i(2 phi - 3
alpha) and v = (cosh y - sin alpha) / (cosh y + cos alpha), the resulting
identity near y = -3 i alpha is
f(y) + f(y + 4 i alpha) - f(y + 2 i alpha) = 1 + O((y + 3 i alpha)^M),
where f(y) = 1 - F(v(y)).
Positive moment representation. Index columns j = 0, ..., M - 1, and set
A_j(k) = partial_alpha^j sinh(alpha k),
B_j(k) = partial_alpha^j cosh(alpha k),
w(k) = 2(1 + 4 sinh^2(alpha k)) / sinh(pi k),
w_0(k) = k w(k).
Since F(1) = 1, the function f(y) is a linear combination of (cosh y + cos
alpha)^{-r}, 1 <= r <= M. The first M parameter derivatives of sin alpha /
(cosh y + cos alpha) = int_R e^{i k y} sinh(alpha k) / sinh(pi k) dk span
these powers: the derivative of order j has a nonzero coefficient of the
denominator power j + 1. The Fourier kernel lemma gives a vector c with
f(y) = int_R e^{i k y} A(k)^T c / sinh(pi k) dk,
and the integral and required derivatives converge locally uniformly in
|Im y| < pi - alpha. At y = -3 i alpha, splitting positive and negative k gives
the moment matrix equation
D c = B(0), D = int_0^infty B(k) A(k)^T w(k) dk.
D is nonsingular. The ordered determinant expansion of det D is an integral
over 0 < k_1 < ... < k_M of a product of two determinants times positive
weights. Both factors are strictly positive: expanding sinh or cosh in
positive power series gives each via a positive generalized Vandermonde, and
the sign of the second follows by separating the positive k_i in scale. Then
det D = int over configurations of (positive factor) times Delta(k)^2, hence
> 0.
Boundary residue. Set K(z) = A(z)^T D^{-1} B(0). For real y -> +infinity,
the residue at k = i of the Fourier formula for f gives -2 i K(i) e^{-y} as
its first term. On the polynomial side,
1 - v = 2(cos alpha + sin alpha) e^{-y} + O(e^{-2 y}),
so
F'_v(1) = -i K(i) / (cos alpha + sin alpha).
Orbital form. The positive density
det[A(k_1), ..., A(k_M)] det[B(k_1), ..., B(k_M)] prod_i w(k_i)
on 0 < k_1 < ... < k_M is realized by the Harish-Chandra--Itzykson--Zuber
integral: for D_k = diag(k_1, ..., k_M, -k_1, ..., -k_M) and Q = U diag(alpha
I, -alpha I) U^* with U Haar on U(2M), put I(k) = E exp(tr(D_k Q)). Then the
density is proportional to
Delta(k^2)^2 prod_i w_0(k_i) I(k),
which is a positive probability density. The derivation from the HCIZ formula
follows by compressing U D_u U^* to a (2M - 1)-dimensional minor; the ordered
eigenvalues interlace, the last resolvent residue gives uniform simplex
weights, and the resulting Jacobian is Delta(l) / Delta(u). Induction with
test parameters z_i - z_d and column subtraction produce the paired
determinantal formula. Specializing to paired lists (k_1, ..., k_M, -k_1,
..., -k_M) and (u_1, ..., u_M, -u_1, ..., -u_M), using the paired Vandermonde
|Delta(+-k)| = 2^M (prod_i k_i) Delta(k^2)^2,
and splitting the exponential determinant into sinh and cosh blocks gives the
form. The orbital density extends continuously when one point approaches
zero, and it is integrable:
w_0(k) ~ 4 k e^{-6 alpha k} (k -> infinity), w_0(0) = 2 / pi.
Pinning. Let rho(0) be the one-point intensity at zero of the orbital law.
Pinning one point there leaves N = M - 1 positive points with density pro-
portional to Delta(k^2)^2 prod_{j = 1}^N k_j^4 w_0(k_j) I(k_1, ..., k_N, 0),
where the orbital integral still uses the 2M-dimensional orbit with one zero
pair in D_k. For its expectation E_pin,
sin alpha / alpha * rho(0)/w_0(0) * E_pin prod_j (1 + k_j^{-2})
<= -i K(i)
<= rho(0)/w_0(0) * E_pin prod_j (1 + k_j^{-2}).
The pinning proof compares determinants with one evaluated at the missing
zero, using a Bessel function identity
(sinh(k_j t)/f_z(t))' = ((k_j^2 - z^2) / f_z(t)^2)
int_0^t f_z(v) sinh(k_j v) dv
for f_z(u) = sinh(z u)/z, so that ratios of determinants reduce to positive
weighted averages. The averaging weights satisfy the elementary bounds
g(u_M) <= R <= g(u_1)/g(u_M) for g(t) = sin t / t, giving the displayed
two-sided pinning bound.
Orbital derivatives and stochastic comparison. For either orbital factor,
there are q_j >= 0 with partial_{k_j} log I = 2 q_j and
0 <= q_j <= min{alpha, alpha^2 k_j, alpha^2 / sum_{ell} (k_j + k_ell)^{-1}}.
Derivation: set q_j = <Q_{+j, +j}> under the tilting exp(tr(D_k Q)), exchange
all paired positions and send Q to -Q to get <Q_{-j, -j}> = -q_j. Convexity in
k_j from the log-moment-generating structure gives q_j >= 0, and ||Q|| = alpha
gives q_j <= alpha. Infinitesimal conjugation gives
<Q_{aa} - Q_{bb}> = (D_k{aa} - D_k{bb}) <|Q_{ab}|^2>; for a = +j, b = -ell
this becomes <|Q_{+j, -ell}|^2> = (q_j + q_ell)/(k_j + k_ell). Summing over
all ell and using sum_b <|Q_{+j, b}|^2> = alpha^2 (from Q^2 = alpha^2 I) and
dropping q_ell >= 0 gives the third bound.
The orbital density's log-derivative relative to the pure density with weight
k^e w_0(k) e^{2 alpha k} is 2 q_i - 2 alpha <= 0; positive association of the
underlying pure law (Vandermonde determinant squared) then gives
O_N^{(e)} <=_st P_N[k^e w_0(k) e^{2 alpha k}], e = 0, 4,
and the first j such points are stochastically below P_j[k^e w_0(k) e^{2 alpha
k}]. The logarithm of w_0 has derivative D_0(k) satisfying
-6 alpha < D_0(k) < 0,
D_0(k) -> -6 alpha (k -> infinity),
D_0(k) = -(40/3) alpha^2 k + O(k^3) (k -> 0).
Consequently, with m = floor(N^{1 - epsilon}), on an event E of probability
1 - O(e^{-c m}) the small-index coordinates satisfy
sup_{i <= m, k in E} q_i <= C / (epsilon log N) = o(1).
Gamma comparison and evaluation. For ell, beta > 0 let W_{ell, beta}(k) =
|Gamma(ell + i beta k)|^2 and b = pi beta. For every fixed b_+ > 6 alpha there
is ell_+ > 0 such that with beta_+ = b_+/pi,
P_N[k^e W_{ell_+, beta_+}] <=_st O_N^{(e)} for all N, e = 0, 4.
For fixed 0 < b_- < 6 alpha and 0 < epsilon < 1 there is ell_- > 0 such that,
with beta_- = b_-/pi and for sufficiently large N,
L_{O_N^{(e)}}(k_{<= m} | E, k_{> m}) <=_st P_m[k^e W_{ell_-, beta_-}]
almost everywhere on later configurations compatible with E. The Gamma
derivative
G_{ell, beta}(k) = -2 beta^2 k sum_{n >= 0} 1/((n + ell)^2 + beta^2 k^2)
satisfies G < -b + 2/k for 0 < ell <= 1 and G >= -b with |G|/k bounded by
2 beta^2 (ell^{-2} + ell^{-1}) for ell >= 1. Comparing G with D_0 gives the
two stochastic inequalities on bounded intervals and tails.
Evaluate the pure statistic on the Meixner--Pollaczek law. Normalize
|Gamma(ell + i u)|^2 du to a probability on R; its Fourier transform is
cosh(t/2)^{-2 ell}. The generating function (1 - i z)^{-ell + i u} (1 + i z)^{-
ell - i u} = sum_n P_n(u) z^n gives real polynomials P_n of degree n with
leading coefficient 2^n/n!, parity n, and squared norms c_n = (2 ell)_n/n!
~ n^{2 ell - 1}/Gamma(2 ell). Evaluations:
(-1)^j P_{2j}(0) ~ j^{ell - 1}/Gamma(ell),
(-1)^j P_{2j}(i beta) ~ 2^{2 beta - 1} j^{ell + beta - 1}/
Gamma(ell + beta).
The one-point intensity at zero and the pinned statistic give
rho_{Gamma, L}(0) ~ 1 + log L,
E_{Gamma, L}^pin prod_j (1 + k_j^{-2}) ~ L^beta / (1 + log L).
For the unpinned orbital law, the two-sided pin bound and the Gamma
comparisons give
c <= rho(0) <= C (1 + log M),
with constants depending on the comparison parameters. For the pinned law,
the product is decreasing, so the first Gamma comparison gives a lower bound
of the pure statistic; conditioning on E and discarding later factors at
least one gives the opposite, and P(E) = 1 - O(e^{-c m}) closes the loop.
Order of limits. Fix b_+ > 6 alpha for the upper, b_- < 6 alpha and epsilon >
0 for the lower, choose their ell parameters, and only then let M -> infinity.
Put a_M = log F'_v(1) / log M. The pinning and residue bounds give
(1 - epsilon) b_-/pi <= liminf a_M <= limsup a_M <= b_+/pi.
Letting b_-, b_+ -> 6 alpha and epsilon -> 0 gives
F'_v(1) = M^{3/4 + o(1)}. (**)
Deficit and lower exponent. Define the positive deficit
delta_M = 1/c_alpha - sum_{d = 1}^{M-1} H_M(d) = F'_v(1) / (c_alpha M)
= M^{-1/4 + o(1)}.
The coarse bound c_alpha sum_{d > 0} H(d) <= 1 from the convex balance, the
inequality H_M(d) <= H(d), and delta_M -> 0 give saturation sum_{d > 0} H(d) =
1/c_alpha. With A(n) = sum_{d >= n} H(d), this yields A(M) <= delta_M. Every
arch contained within distance strictly less than M/2 of its source is
disjoint from its nontrivial translates by M, so arches counted in the deficit
reach distance at least M/2; hence delta_M <= W_{M/2}. Choosing M just above
2 R gives W_R >= R^{-1/4+o(1)} and A(n) <= n^{-1/4+o(1)}.
Wedge comparison. Intersect the upper half-plane with the wedge between rays
of angles 0 and theta = pi/3. A source on its horizontal ray has horizontal
distance a in {1/2, 3/2, ...} from the vertex. The signed balance holds on
the infinite wedge: exhaust the half-plane by convex tessellation polygons
with extending horizontal bases, use the convex-balance positivity for same-
side arches, and pass to the limit using the absolute convergence of the fixed-
ray exit sums.
Put s = 3/8. For same-side walks the phase is e^{i s pi} to the left and
e^{-i s pi} to the right, and exits to the other ray have phase e^{i s theta}.
Let L_<(a) and L_>(a) be the lost left-side and right-side same-side arch
weights on restriction from the half-plane to the wedge, including endpoints
removed by the restriction. Subtracting the balances and projecting onto
e^{i s theta} gives
sin(s(pi - theta)) L_<(a) - sin(s(pi + theta)) L_>(a) = 0,
so L_<(a) = r L_>(a) with r = sqrt 2. For positive half-integer ports a, b,
let D(a, b) be the lost same-side weight with those fixed endpoints. Reversal
gives D(a, b) = D(b, a). The missing negative ports to the left contribute
A(ceiling a), so
L_<(a) = A(ceiling a) + sum_{0 < b < a} D(a, b),
L_>(a) = sum_{b > a} D(a, b).
Summing L_< = r L_> over 0 < a <= T gives
(r - 1) sum_{0 < a < b <= T} D(a, b) + r sum_{0 < a <= T < b} D(a, b)
= sum_{0 < a <= T} A(ceiling a)
<= T^{3/4 + o(1)}.
Since r > 1, this bounds sum_{0 < a <= T} (L_<(a) + L_>(a)) by the same order.
After translating the source to a fixed port, the wedges increase with a, so
their total loss decreases and
L_<(a) + L_>(a) <= a^{-1/4 + o(1)}.
Two reflected wedges whose vertices are horizontal distance a on opposite
sides of the source intersect in a triangle of diameter O(a). An arch reaching
distance C a for a sufficiently large fixed C must leave at least one wedge,
so its weight is at most the sum of the two losses, giving W_R <=
R^{-1/4+o(1)}. This completes the boundary arch mass proposition. QED
Extremal segments. Lemma. Fix a honeycomb vertex and two distinct height
levels, with the fixed vertex on one level. The total critical weight of SAW
segments from that vertex to any vertex on the other level, whose endpoints
are opposite height extrema of the segment, is bounded by an absolute constant
independent of the levels.
Proof. Extend each endpoint to a port on the adjacent horizontal line beyond
its extremal level. This requires at most one additional triangle at each
end, strictly beyond the segment's range of heights: between two successive
horizontal lines the lower triangles have a port on the bottom and the higher
triangles have a port on the top; a triangle of either type has a neighbor of
the other type. Choose an extension by a fixed local rule. Trimming recovers
the segment, and possible choices have bounded multiplicity. The extended walk
is a first-exit crossing of the slab between the two chosen lines. The convex
boundary balance, exhausted horizontally, bounds its total weight for each of
the finitely many possible starting ports. Removing the at most two extra
vertices costs at most x^{-2}. QED
For later use, split an n-edge walk at a highest vertex and treat the two
halves. Within either half, go to the last occurrence of its lowest remaining
level, then to the last occurrence of its highest remaining level, and continue
alternately. These are segments of the lemma. Their positive integer spans
decrease strictly after a possible equality of the first two, since each
chosen extremum is the last occurrence of its level. The spans sum to at most
2 n, so there are O(sqrt n) segments and at most (2 n + 1)^{O(sqrt n)} span
lists. For a prescribed list, successive use of the lemma bounds the endpoint-
summed product of segment weights by a constant to the number of segments,
and recovering the weight of the joined walk costs x^{-1} per shared endpoint.
These subexponential concatenation bounds will be used below in the marked
argument, and they also recover the reciprocal exponential growth rate of
full-plane SAW by summing over highest-vertex splits.
Chords on the boundary. Let K be a finite convex tessellation polygon D
sharing a starting port with the same inward direction as a containing convex
cap D_0. Let L be the raw weight of first-exit chords of D_0 other than those
confined to K and ending at a boundary port shared by K and D_0, and let E be
the raw weight of first-exit chords of K ending on ports newly created by the
restriction. Then
sin(alpha) L <= E <= L / sin(alpha).
Proof. The common chords, confined to K and ending on a shared true boundary
port, have the same intrinsic turning phase in both balances and cancel. The
two remaining complex sums are equal. Every real coefficient in either convex
domain lies between sin(alpha) and 1. Taking real parts proves both
inequalities. The case with D_0 the saturated half-plane and K a finite convex
cap follows by the same convex exhaustion used in the wedge argument. QED
Visit identity. Let D be a finite convex tessellation polygon and v a
honeycomb vertex not adjacent to a boundary port. Sum over all ordered first-
exit boundary chords of D that visit v and set
S_D(v) = sum_{gamma ni v} x^{N(gamma)} cos(3 T(gamma)/8).
Let Z_v be the partition of disjoint unoriented cycles in D strictly enclosing
v and avoiding v, with weight 2 x^{|lambda|} per cycle. Let Z_out(lambda) count
disjoint cycles in D surrounding and disjoint from lambda with the same
weights. At a vertex of lambda at v, call the corner convex or reflex when its
interior angle is 2 pi/3 or 4 pi/3, and put
c_lambda = sqrt 2 when v is a convex corner of lambda,
= 2 + sqrt 2 when v is a reflex corner.
Proposition. The identity
S_D(v) = 3 Z_v + sum_{lambda ni v} c_lambda x^{|lambda|} Z_out(lambda)
holds. Moreover, letting Z_D(f) denote the analogous partition around a face
center f for each of the three incident face centers,
sum_{f ~ v} Z_D(f) = 3 Z_v + sum_{lambda ni v} 2 k_lambda x^{|lambda|}
Z_out(lambda),
with k_lambda = 1 at a convex corner and 2 at a reflex corner, and therefore
(1/sqrt 2) sum_{f ~ v} Z_D(f) <= S_D(v) <= sum_{f ~ v} Z_D(f),
S_D(v) <= sum_{gamma ni v} x^{N(gamma)} <= S_D(v) / sin(alpha).
Proof. Remove the triangle centered at v, creating three inner ports. For
each outer starting port the signed balance still holds in the punctured
domain: its source is outside every possible first-collision loop, including
a loop surrounding the removed triangle, so the same collision cancellation
applies. Subtracting the punctured balance from the original one cancels
outer chords avoiding v, leaving the outer chords through v equal to the paths
ending at an inner port. Take real parts, sum over the outer sources, and
reverse the paths; reversal changes T to -T and preserves the real part. Thus
S_D(v) is the real sum of paths from the three inner ports to the outer ports.
For each inner source attach a background counted by Z_v and grow the weighted
path tree in the punctured domain. Stop at an exit, a first self-collision, or
first contact with a background cycle. The initial total strength of the three
trees is 3 Z_v. A self-collision loop not enclosing v is approached from
outside and cancels with its reverse exactly as in the convex balance. A loop
enclosing v is approached from inside, since the inner source is in the same
component as v. Its corner at the collision vertex is reflex: the other loop
vertices contribute +-7 pi/3 of turning, and departure from the incoming stem
contributes +-pi/3. The two orientations therefore contribute
2 x^{|lambda|} cos((3/8)(8 pi/3)) = -2 x^{|lambda|},
which cancels first contact with the same loop added to the background. The
inverse operation, removing the first contacted (necessarily innermost)
background cycle and following it from the contact in either orientation, is
defined by first-contact without earlier stem intersection; every background-
contact term is canceled exactly once.
An outer exit can be reached with no background cycle at present, since the
source starts inside every such cycle; its real sum is the S_D(v) identified
above. An exit to another inner port, closed through v, gives a simple cycle
lambda through v, and every remaining background cycle surrounds it. If
L = |lambda|, opening the cycle at v leaves two directed inner-return paths
of weight x^{L-1} each. At a convex corner their turns are +-5 pi/3, at a
reflex corner +-7 pi/3: the full loop turn is +-2 pi with the turn at v
removed. Write t_lambda = 5 pi/3 or 7 pi/3. The remaining real tree identity
is
3 Z_v = S_D(v) + sum_{lambda ni v} 2 x^{|lambda| - 1}
cos((3/8) t_lambda) Z_out(lambda),
and moving the inner-return terms to the left supplies
-(2/x) cos(5 pi/8) = sqrt 2,
-(2/x) cos(7 pi/8) = 2 + sqrt 2,
which is the first displayed identity.
A cycle avoiding v places all three incident face centers in the same
component as v; a cycle through v encloses exactly one incident face center
at a convex corner and two at a reflex corner, and in a surrounding
configuration must be innermost (an inner cycle avoiding v would also enclose
v, lying on the outer cycle). This partitions the Z_D(f) configurations into
the three cases and proves the second displayed identity. Comparing coefficients
gives the (1/sqrt 2) two-sided visit comparison, and convex boundary positivity
gives the ordinary visit mass bound. QED
Flat-hexagon conclusions. Let Z(R) be the +2 nesting partition around the
central face in D_R and, for honeycomb v in D_R, define V_R(v) as the raw
chord mass through v. Then E_R^chord N = Z_R^chord^{-1} sum_v V_R(v). For
v at distance at most R/10 from the center, every first-exit boundary chord
through v is macroscopic for large R, so translating a hexagon of side
floor(R/2) centered at each incident face inside D_R gives
V_R(v) >= c Z(floor(R/2)),
uniformly for these bulk vertices. For v not adjacent to a boundary port,
a translated hexagon of side 3 R contains D_R, giving V_R(v) <= C Z(3 R)
uniformly. There are order R^2 bulk vertices and order R boundary-adjacent
vertices, each of which contributes at most one. Hence
c R^2 Z(floor(R/2)) / Z_R^chord <= E_R^chord N
<= C R^2 Z(3 R) / Z_R^chord + C R.
The chord partition Z_R^chord and the +2 nesting partition Z(R) will be
identified later: Z_R^chord = R^{3/4 + o(1)} and Z(R) = R^{1/12 + o(1)}.
Combining gives the mean E_R^chord N = R^{4/3 + o(1)} and bulk visit
probability R^{-2/3 + o(1)}. The proofs of the two identities occupy the
bulk of the paper.