editors: nobody
Fix an integer n >= 3 and a real beta with 0 < beta < infinity. A finite
graph G = (V,E) is a subgraph of the nearest-neighbour square lattice
Z^2: the vertex set V is a finite subset of Z^2 and E joins pairs of
lattice points at Euclidean distance one. To each edge e = {x,y} assign
a strength b_e with 0 <= b_e <= beta. A spin sigma_x is a point of the
unit sphere S^{n-1} of R^n. The ferromagnetic Gibbs law at inverse
temperature encoded by b is
d mu_{G,b}^{(n)}(sigma)
= (1 / Z_{G,b}^{(n)}) exp( sum_{{x,y} in E} b_{xy}
sigma_x . sigma_y ) prod_{x in V} d omega_{n-1}(sigma_x),
where omega_{n-1} is the uniform probability measure on S^{n-1},
sigma_x . sigma_y is the Euclidean inner product, and Z_{G,b}^{(n)}
normalizes the total mass to one. No boundary spin is fixed and no
external field is present. Expectation is written < . >_{G,b}^{(n)}.
Main result. For every n >= 3 and every beta < infinity there are
constants A = A(n,beta) < infinity and m = m(n,beta) > 0 such that, for
every finite nearest-neighbour subgraph G of Z^2, every strength array
b in [0,beta]^E, and all x,y in V,
0 <= < sigma_x . sigma_y >_{G,b}^{(n)}
<= A exp( - m |x - y| ),
where |x - y| is the Euclidean distance. The constants do not depend on
G, on b, or on the volume. The correlation is nonnegative and its
magnitude is exponentially small in the distance: symmetry forbids
long-range order, and at any positive inverse temperature it forbids it
at an exponential rate once n >= 3.
The same constants govern the free-boundary laws. On the boxes
Lambda_L = [-L,L]^2 intersected with Z^2, with all nearest-neighbour
edges, free boundary conditions and constant strength beta,
limsup_{L -> infinity} | < sigma_0 . sigma_x >_{Lambda_L,beta}^{(n)} |
<= A exp( - m |x| ),
and every subsequential local weak limit of these box laws obeys the
same two-point bound. The spin susceptibility sum_{x in Z^2}
< sigma_0 . sigma_x > is therefore finite in each such limit. Only the
spin two-point function is controlled: a full transfer-matrix gap,
covariances of rotation-invariant bond energies, the low-temperature
correlation-length asymptotics, and uniqueness of infinite-volume Gibbs
states are separate questions, and the constants may deteriorate as
beta tends to infinity.
Continuous symmetry forces the two-point function of the classical
O(n) model to decay in two dimensions, but the rate is the whole
question. Mermin and Wagner proved the absence of long-range order for
quantum Heisenberg systems, and Mermin gave a direct classical
argument; together the two results rule out magnetization at every positive
temperature for n >= 2. McBryan and Spencer sharpened this to
quantitative algebraic bounds: complex spin rotations gave two-point
correlations bounded by a negative power of the distance. Gagnebin and
Velenik extended the complex-rotation method to continuous interactions
of much weaker regularity and to suitable long-range couplings. Bounds
of this kind quantify the loss of order but permit a zero exponential
rate.
pdf/gagnebin-velenik-1309.2432-upper-bound-decay-correlations.pdf
For two-component spins the distinction is sharp. The Berezinskii-
Kosterlitz-Thouless picture has a low-temperature phase with only
power-law decay, and Frohlich and Spencer proved a polynomial lower
bound on the ordinary plane rotator at sufficiently low temperature.
An exponential upper bound therefore fails at n = 2 for low
temperature, so the restriction n >= 3 in the main result is necessary.
pdf/berezinskii-1971-destruction-long-range-order-jetp.pdf
For three or more components, Polyakov's non-Abelian renormalization
argument predicts mass generation at every positive temperature.
Rigorous results cover parts of this picture. Aizenman and Simon
derived exponential decay from local Ward identities at high
temperature, continuing the Ward-identity estimates of Driessler,
Landau and Perez, and Kupiainen established an asymptotic 1/n
expansion with a mass gap in the large-component regime. These leave
open a fixed small dimension and an arbitrarily low positive
temperature. The all-temperature lattice statement was recorded as an
open conjecture by Aru, Garban and Sepulveda:
pdf/aru-garban-sepulveda-2212.06767-percolation-2d-heisenberg.pdf
"In this case, it has been predicted since Polyakov (1975) that
this model should exhibit exponential decay of correlations at
all inverse temperatures beta > 0. This remains unproven and it
is considered to be one of the major conjectures in statistical
physics."
Geometric objections raise the stakes. Patrascioiu and Seiler studied
percolation of strongly correlated spin regions as a possible route to
a massless phase, arguing against the renormalization prediction for a
constrained model. Aru, Garban and Sepulveda made part of those
arguments rigorous and built ergodic environments of high conductance
in which the projected plane-rotator two-point function still decays
exponentially after averaging; percolation of the strongly coupled
region therefore does not by itself force slow decay. The estimate
holds uniformly over every strength array in [0,beta]^E,
and this uniformity is what permits the final projection without any
assumption on the distribution of the projected couplings.
pdf/patrascioiu-seiler-hep-th-0011199-percolation-soft-phase.pdf
The sign-cluster representation has several ancestors. The
Fortuin-Kasteleyn random-cluster representation and the Edwards-Sokal
coupling tie ferromagnetic correlations to bond connectivity; Wolff's
reflection construction embeds an Ising sign system in a
continuous-spin model; the Fortuin-Kasteleyn-Ginibre and Ginibre
association inequalities supply the monotonicity. For three
components, Campbell and Chayes proved amplitude association, bond
association and boundary comparison, and expressed the spin
correlation as an amplitude-weighted connectivity.
pdf/campbell-chayes-1998-isotropic-o3-wolff.pdf
The estimate holds on a finite graph with arbitrary fixed
boundary spins. Let G = (V,E) be a finite graph, oriented edge by edge
as e = (x,y), with strengths 0 <= b_e <= beta. Fix a set P of pinned
vertices and arbitrary spins s_x in S^2 for x in P; at the remaining
vertices integrate against uniform probability on S^2 with density
proportional to exp( sum_e b_e s_x^T s_y ). Expectation is < . >. A bar
over a symbol denotes complex conjugation; Id is the identity matrix.
Let a, b, c in so(3) act by cross product with the coordinate vectors
e1, e2, e3, so that [a,b] = c, and each has operator norm one. For an
edge field A = (A_e) of skew matrices put
H(A) = sum_{e=(x,y)} b_e s_x^T ( exp(A_e) - Id ) s_y,
W(A) = < exp H(A) >.
W is the ratio of the twisted partition function to the untwisted one.
Derivatives are taken at A = 0; set X = DH and H_k = D^k H. Because the
finite spin space is compact, differentiation passes under the spin
integral. Multilinear derivatives extend complex-linearly in their
arguments, while a change of spin variables always uses a real
rotation.
For a scalar function on vertices, set df(e) = f_y - f_x on the oriented
edge e = (x,y). Scalar norms use counting measure on the indicated
vertices or edges; for an edge-indexed matrix T,
||T||_HS^2 = sum_{e,l in E} |T_{el}|^2.
For g1,...,gk each drawn from {a,b,c}, the local coefficients
J_e(g1,...,gk) = b_e s_x^T ( (1/k!) sum_{pi in S_k}
g_{pi(1)} ... g_{pi(k)} ) s_y
obey |J_e| <= beta, since every matrix product in the average has
operator norm at most one. Consequently
H_k( v1 g1, ..., vk gk )
= sum_e J_e(g1,...,gk) prod_{r=1}^k v_r(e).
The coefficients depend on the spins and the generators, but not on the
scalar test functions. The same formula with k = 1 gives X.
Define the complex bilinear response form
R(h,k) = - D^2 W(hc, kc)
on scalar edge forms. It is symmetric, and it is not assumed to be
positive semidefinite.
Gradient lemma. If f : V -> C vanishes on P, then for g in {a,b,c}
< |X((df) g)|^2 > <= beta ||df||_2^2.
For every complex edge form h,
R(h, hbar) = - < H_2(hc, hbar c) + |X(hc)|^2 >
<= beta ||h||_2^2,
where hbar denotes the conjugate edge form, and adding df in either
argument of R does not change its value whenever f vanishes on P.
Proof. For real f, change variables by s_x^{old} = exp(-t f_x g)
s_x^{new} at each integrated vertex. Fixed spins are untouched because
f vanishes on P, and rotations about a single axis commute, so
W(t (df) g) = 1 identically. Its second derivative at zero is
0 = < H_2((df)g, (df)g) + X((df)g)^2 >.
The coefficient bound |J_e| <= beta gives < |X((df)g)|^2 > <=
beta ||df||_2^2 for real f. For complex f = f1 + i f2 with real f1,f2,
the squared modulus |X((df)g)|^2 equals the sum of the squares for f1
and f2, and adding the two bounds gives the complex case with the same
constant.
Differentiating W twice gives the displayed equality in R(h,h). The
local term obeys |< H_2(hc, hbar c) >| <= beta ||h||_2^2, because
|J_e| <= beta and |h_e hbar_e| = |h_e|^2. The current term |X(hc)|^2 is
nonnegative before its minus sign, so the inequality R(h,h) <=
beta ||h||_2^2 follows. Finally, the same change of variables gives
W((h + df) c) = W(h c) for real h and f; differentiating in independent
scalar multiples of h and df, then extending complex-linearly, gives
R(h, df) = R(df, h) = 0. This proves the last assertion.
For vertex functions f, g define
S(f,g)_e = (1/2)( f_x g_y - g_x f_y ), e = (x,y).
The form S is bilinear and antisymmetric.
Family inequality. Let (f_i)_{i in Idx} be a finite family of complex
vertex functions vanishing on P, and put
u_i = df_i, T_{el} = sum_{i in Idx} u_i(e) ubar_i(l),
Q = sum_{i in Idx} S(f_i, fbar_i),
where ubar_i(l) and fbar_i are the conjugates of u_i(l) and of f_i,
and Qbar is the conjugate of Q. Then
R(Q, Qbar)
<= K_rot(beta) ||T||_HS^2
+ beta sum_{i,j in Idx} min{ ||f_i||_inf^2 ||df_j||_2^2,
||f_j||_inf^2 ||df_i||_2^2 },
with
K_rot(beta) = 6 beta^2 + 4 beta^{3/2} + beta.
The constants are independent of the graph, of the family size, and of
the values of the fixed spins.
Proof of the family inequality. The route is an exact identity that
relates the response R to a fourth derivative of W and then sums it
over the family. Gradient invariance controls the resulting cross
terms, and a nonnegative square supplies the fourth-derivative lower
bound.
The noncommuting gauge identity. Let F_x be a real matrix potential in
the span of a and b, vanishing on P. The substitution
s_x^{old} = exp(-F_x) s_x^{new} gives
W(dF) = W(B), B_e = log( exp(F_x) exp(F_y - F_x) exp(-F_y) )
for F small, with dF(e) = F_y - F_x. The product in the logarithm is
orthogonal and near the identity, so its local logarithm is skew, and
the logarithm commutes with conjugation. Multiplying the exponential
series to second order and taking the logarithm gives
B_e = (1/2)[F_x, F_y] + O(F^3), DB_e(0) = 0.
For directions f a and g b, the mixed second derivative of B_e is
therefore S(f,g)_e c, and it is zero for two directions that use the
same generator.
A fourth-order cancellation is needed as well. Take four independent
parameters whose matrix directions have generator pattern (a,b,a,b),
with arbitrary spatial potentials. The mixed fourth derivative of B_e
vanishes. Conjugation by the half-turn about the first coordinate axis
fixes a and negates b, so the coefficient of all four parameters is
fixed; its two b-directions contribute two minus signs. Conjugation by
the half-turn about the second axis likewise fixes the coefficient, its
two a-directions contributing two minus signs. On so(3) the fixed
spaces of these two conjugations are span{a} and span{b}, whose
intersection is zero. Hence
D_{1234} B_e = 0 for the generator pattern (a,b,a,b).
The quadratic and fourth-order statements extend complex multilinearly;
no complex change of spin variables is involved in the extension.
Fix i and j and take the four edge directions
A1 = u_i a, A2 = ubar_i b, A3 = ubar_j a, A4 = u_j b.
The fourth derivative of the composition in the gauge identity has no
term containing DB, because DB_e(0) = 0, and its term DW[D_{1234}B]
vanishes by the cancellation. Only pairings of two second derivatives
remain. The pairing 13|24 is zero, the two second factors using the
same generator. For the other pairings, set S_i = S(f_i, fbar_i) and
S_{ij} = S(f_i, f_j). Their arguments are
12|34 : ( S_i c, S_j c ),
14|23 : ( S_{ij} c, -overline{S_{ij}} c ),
using S(fbar_j, f_j) = S_j and
S(fbar_j, fbar_i) = -overline{S_{ij}}. Each labeled pairing occurs once
in the composition rule, so
D^4 W(A1,A2,A3,A4) = -R(S_i, S_j)
+ R(S_{ij}, overline{S_{ij}}).
This holds also when some spatial functions coincide, the derivative
being multilinear in the four labeled directions.
Summing over i and j and rearranging gives the identity to be
estimated,
R(Q, Qbar) = sum_{i,j} R(S_{ij}, overline{S_{ij}})
- sum_{i,j} D^4 W(u_i a, ubar_i b, ubar_j a, u_j b).
The first sum is bounded from above using gradient invariance; the
second is bounded from below by keeping a nonnegative square.
The cross terms. For functions f, g vanishing on P, the product fg
vanishes there too, and the edge identities
S(f,g) + (1/2) d(fg) = (1/2)(f_x + f_y) dg,
S(f,g) - (1/2) d(fg) = -(1/2)(g_x + g_y) df
hold on the edge e = (x,y). Applying gradient invariance in both
arguments of R, followed by R(h,h) <= beta ||h||_2^2, gives
R(S(f,g), S(f,g))
<= beta min{ ||f||_inf^2 ||dg||_2^2, ||g||_inf^2 ||df||_2^2 }.
Summing over i,j bounds the first sum in the balance identity by
beta sum_{i,j} min{ ||f_i||_inf^2 ||df_j||_2^2,
||f_j||_inf^2 ||df_i||_2^2 }.
The fourth-derivative lower bound. The claim is
sum_{i,j} D^4 W(u_i a, ubar_i b, ubar_j a, u_j b)
>= - K_rot(beta) ||T||_HS^2.
Set X_r = X(A_r) and H_{r_1...r_k} = H_k(A_{r_1},...,A_{r_k}). Direct
differentiation of exp H gives all fifteen set partitions of four
indices:
D^4 W(A1,A2,A3,A4) = <
X1 X2 X3 X4
+ H12 X3 X4 + H13 X2 X4 + H14 X2 X3
+ H23 X1 X4 + H24 X1 X3 + H34 X1 X2
+ H12 H34 + H13 H24 + H14 H23
+ H123 X4 + H124 X3 + H134 X2 + H234 X1
+ H1234 >.
The four terms (X1 X2 + H12)(X3 X4 + H34), after summation over i,j,
have expectation
< | sum_i [ X(u_i a) X(ubar_i b) + H_2(u_i a, ubar_i b) ] |^2 >
>= 0.
The other eleven terms are bounded in absolute value.
For an edge e write t_e(l) = T_{el} and rho_e = ( sum_l |T_{el}|^2 )^{1/2},
and abbreviate X_g(v) = X(v g). The row-gradient identity is
t_e = d( sum_i u_i(e) fbar_i ).
The potential on the right, and its conjugate, vanish on P, so the
gradient lemma gives
< |X_g(t_e)|^2 > <= beta rho_e^2,
< |X_g(t_e)|^2 > <= beta rho_e^2.
The remaining contractions come from expanding each H_k through the
local coefficients, summing over i,j, and multiplying the contracted
factors. Summation over e follows, except in the two H_2 H_2 rows
where one sums over e and l. In the table the middle column is the
local coefficient and the last column is the contracted expression.
term local coefficient contracted expression
------------------------------------------------------------------
H13 X2 X4 J_e(a,a) X_b(t_e) X_b(t_e)
H14 X2 X3 J_e(a,b) X_b(t_e) X_a(t_e)
H23 X1 X4 J_e(b,a) X_a(t_e) X_b(t_e)
H24 X1 X3 J_e(b,b) X_a(t_e) X_a(t_e)
H13 H24 J_e(a,a) J_l(b,b) T_{el} overline{T_{el}}
H14 H23 J_e(a,b) J_l(b,a) T_{el}^2
H123 X4 J_e(a,b,a) T_{ee} X_b(t_e)
H124 X3 J_e(a,b,b) T_{ee} X_a(t_e)
H134 X2 J_e(a,a,b) T_{ee} X_b(t_e)
H234 X1 J_e(b,a,b) T_{ee} X_a(t_e)
H1234 J_e(a,b,a,b) T_{ee}^2
For the first row, for instance,
sum_{i,j} u_i(e) ubar_j(e) X_b(u_i) X_b(u_j)
= X_b( sum_i u_i(e) ubar_i ) X_b( sum_j ubar_j(e) u_j )
= X_b(t_e) X_b(t_e),
the two factors being the conjugate contractions of the definitions of
t_e. The other rows follow by the same finite summation, with the
conjugated and unconjugated contractions distinguished.
Each of the first four rows has, at edge e, pointwise coefficient
bounded by beta and, by Cauchy-Schwarz followed by the two row bounds,
absolute expectation at most beta^2 rho_e^2; the coefficient may depend
on the spins, and its pointwise modulus is bounded before
Cauchy-Schwarz is applied. These four rows contribute at most
4 beta^2 ||T||_HS^2. The two H_2 H_2 rows contribute at most
2 beta^2 ||T||_HS^2, both contractions having absolute value |T_{el}|^2.
For each of the four H_3 X rows the row bounds give total absolute
expectation at most
beta^{3/2} sum_e T_{ee} rho_e
<= beta^{3/2} ( sum_e T_{ee}^2 )^{1/2} ( sum_e rho_e^2 )^{1/2}
<= beta^{3/2} ||T||_HS^2,
because T_{ee} = sum_i |u_i(e)|^2 >= 0, the squared diagonal entries are
part of the Hilbert-Schmidt sum, and the rho_e have the same l^2 sum.
The last row is bounded by beta sum_e T_{ee}^2 <= beta ||T||_HS^2.
The sum on the left of the lower bound is real: complex conjugation
interchanges i and j and permutes the arguments of the symmetric fourth
derivative. Keeping the nonnegative square and subtracting the bounds
for the remaining terms proves
sum_{i,j} D^4 W(u_i a, ubar_i b, ubar_j a, u_j b)
>= - K_rot(beta) ||T||_HS^2
with K_rot(beta) = 6 beta^2 + 4 beta^{3/2} + beta.
Combining the cross-term bound and this lower bound in the balance
identity proves the family inequality. Only the upper bound and the
gradient invariance of R enter; no positivity property of R is used.
The family inequality is tested on an annulus whose two
boundary layers carry prescribed directions. For an integer N >= 2 set
D_N = { x in Z^2 : N <= ||x||_inf <= 2N },
where ||x||_inf is the coordinate supremum norm. Let G = (V,E) be any
subgraph of the nearest-neighbour graph on D_N, orient each edge in a
positive coordinate direction, and give the edges strengths
0 <= b_e <= beta. Put s_* = (1,0,0) and prescribe
s_x = exp(t c) s_* for ||x||_inf = N,
s_x = s_* for ||x||_inf = 2N (x in V).
All other spins are integrated against uniform probability on S^2. Let
Z(t) be the resulting partition integral, interactions between any
prescribed spins included. Only vertices of V are pinned, so a missing
vertex on either boundary needs no convention.
Annular proposition. For every beta < infinity there is a finite
C_beta such that, for all N >= 2 and all graphs and strengths just
described,
min_{t in R} Z(t) / max_{t in R} Z(t) >= 1 - C_beta / log N,
and in particular Z(pi)/Z(0) >= 1 - C_beta / log N. The constant is
independent of the subgraph, of the strength array, and of the twist.
Boundary gauge and profiles. Fix once and for all a real smooth Theta
equal to one in a neighbourhood of [-1,1]^2 and compactly supported in
(-2,2)^2. Define the real edge form
h_e = Theta(y/N) - Theta(x/N), e = (x,y).
Carry a subscript t on W, X, H_k, R for the quantities of the rotation
estimate evaluated with these boundary values, and write < . >_t for the
corresponding expectation. Although Theta does not vanish on the inner
pinned layer, its gauge action is explicit:
W_t(s h c) = Z(t+s) / Z(t), R_t(h,h) = - Z''(t) / Z(t).
Indeed, substituting s_x^{old} = exp(-s Theta(x/N) c) s_x^{new} cancels
every edge matrix; it changes the inner boundary value to
exp((t+s)c) s_* and leaves the outer value fixed. The gauge changes the
prescribed boundary values rather than acting on integrated variables.
An upper bound of order 1/log N is needed for R_t(h,h). Functions are
chosen whose collective form
sum_i S(f_i, f_i) equals -i lambda_N htilde, with
lambda_N >= (1/2) log N and htilde close to h. The two family costs in
the family inequality then total O(log N), while bilinearity multiplies
R_t(htilde,htilde) by lambda_N^2. This produces the logarithmic gain.
After the family is built, the bound passes from htilde to h without
assuming that R_t is positive.
For m = 1,2 put g_m = partial_m Theta. Let chi_m be a real smooth
cutoff compactly supported in the open annulus
A = { z in R^2 : 1 < ||z||_inf < 2 } and equal to one on the support of
g_m; such a cutoff exists because g_m vanishes near both boundary
squares. The real profiles
phi_{m,+} = (chi_m + g_m)/2, phi_{m,-} = (chi_m - g_m)/2
satisfy phi_{m,+}^2 - phi_{m,-}^2 = g_m. No square root of a signed
smooth function is taken, and all four profiles have compact support in
A.
For p in Z^2 let r_p = ||p||_inf. Introduce four groups of vertex
functions, restricted to V:
f_{m,+-p}(x) = r_p^{-3/2} phi_{m,+-}(x/N)
exp( 2 pi i p . x / (16 N) ),
1 <= r_p <= N, +- p_m > 0.
Every such function vanishes at each pinned vertex. The power r_p^{-3/2}
produces a harmonic sum: the shell face p_m = +- r contains 2r+1 modes,
each with squared normalization r^{-3} and coordinate-m phase increment
of order r/N, so its summed frequency factor in the commutator is of
order 1/(Nr). Opposite frequency signs give opposite coefficients in
the adjacent-point products of phi_{m,+} and phi_{m,-}. Index this
finite family by i, write u_i = df_i, and set
T_{el} = sum_i u_i(e) u_i(l),
S(f,g)_e = (1/2)( f_x g_y - g_x f_y ).
For an edge e = (x, x+e_m) define
htilde_e = (1/N)[ phi_{m,+}(x/N) phi_{m,+}((x+e_m)/N)
- phi_{m,-}(x/N) phi_{m,-}((x+e_m)/N) ].
Mode lemma. There is a constant C, depending only on the fixed
profiles, with the following properties for every N >= 2 and every such
subgraph. First,
sum_i S(f_i, f_i) = -i lambda_N htilde,
lambda_N = N sum_{1 <= r_p <= N, p_1 > 0} r_p^{-3} sin(2 pi p_1 / 16N)
>= (1/2) sum_{r=1}^N 1/r.
Second,
||h||_2 <= C, ||h + htilde||_inf <= C/N,
||h - htilde||_1 <= C/N.
Finally,
||T||_HS^2 + sum_{i,j} min{ ||f_i||_inf^2 ||df_j||_2^2,
||f_j||_inf^2 ||df_i||_2^2 }
<= C (1 + log N).
All edge norms use counting measure on E.
Proof of the mode lemma. For a real profile phi and a single mode on
shell r_p, direct substitution on a positive coordinate-j edge gives
S(f_p, f_p)_e
= -i r_p^{-3} phi(x/N) phi((x+e_j)/N) sin(2 pi p_j / 16N).
When the group is indexed by m different from j, its frequency set is
invariant under p_j -> -p_j, and the sine changes sign while the two
adjacent profile factors do not, so the sum over the group vanishes.
For m = j the negative half-plane reproduces the negative of the
positive-half-plane sine sum, and interchanging coordinates identifies
the positive sums for j = 1 and j = 2. This gives the commutator
identity with the stated lambda_N and the stated htilde.
For the lower bound on lambda_N, on shell r retain the 2r+1 frequencies
with p_1 = r and |p_2| <= r. Since sin( pi r / 8N ) >= r / (4N) for
1 <= r <= N, their contribution to lambda_N is at least
(2r+1)/(4 r^2) >= 1/(2r), so lambda_N >= (1/2) sum_{r=1}^N 1/r.
For an edge in direction m let z = (x+y)/(2N). A centered Taylor
expansion, with uniform bounds on the fixed smooth functions, gives
h_e = N^{-1} partial_m Theta(z) + O(N^{-3}),
phi(x/N) phi(y/N) = phi(z)^2 + O(N^{-2}).
Using phi_{m,+}^2 - phi_{m,-}^2 = g_m = partial_m Theta yields
|h_e - htilde_e| <= C N^{-3}. There are O(N^2) annulus edges and
|h_e| + |htilde_e| <= C/N, so ||h||_2 <= C, ||h + htilde||_inf <= C/N
and ||h - htilde||_1 <= C/N, also after restriction to E.
For a mode on shell r the product rule and |exp(iv) - 1| <= |v| give
||f_i||_inf <= C r^{-3/2}, ||df_i||_2 <= C' r^{-1/2}.
There are O(r) modes on shell r, counting all four groups. For shells
s <= r bound the minimum by its first entry, using the supremum norm of
the shell-r mode; a pair then costs at most C r^{-3} s^{-1}, and all
O(rs) pairs cost at most C r^{-2}. Summing over both shell orderings,
sum_{i,j} min{ ||f_i||_inf^2 ||df_j||_2^2, ||f_j||_inf^2 ||df_i||_2^2 }
<= C sum_{r=1}^N sum_{s=1}^r r^{-2} <= C'(1 + log N).
It remains to bound T, for which mode-by-mode use of the triangle
inequality would give only ||T||_HS <= sum_i ||df_i||_2^2 = O(N). The
sharper bound uses Fourier orthogonality. Work first in one profile
group and let U be the matrix whose p-th column is r_p^{1/2} df_p.
Embed the annulus in the discrete torus Q_N = (Z/(16N Z))^2, whose
representatives may be taken in [-8N, 8N-1]^2; every profile vanishes
near the boundary of this square and so defines a torus function. Let F
have columns
F_{x,p} = exp( 2 pi i p . x / (16 N) ), x in Q_N,
using the frequencies of the group. These frequencies are distinct
modulo 16N, and orthogonality of the unnormalized waves gives
F^* F = (16N)^2 Id, ||F||_op = 16N.
For direction j let P_j restrict functions on Q_N to the origins of the
coordinate-j edges in E. Let M_v be multiplication by v, and set
Delta_j phi(x) = phi((x+e_j)/N) - phi(x/N), phi^{(j)}(x) = phi((x+e_j)/N).
The directional block of U factors exactly as
U_j = P_j M_{Delta_j phi} F diag(r_p^{-1})
+ P_j M_{phi^{(j)}} F
diag( ( exp(2 pi i p_j / 16N) - 1 ) / r_p ).
The operator norms ||P_j||_op <= 1 and ||Delta_j phi||_inf <= C/N hold, and
the last diagonal factor has operator norm at most C/N, since |p_j| <= r_p.
With ||F||_op = 16N this gives ||U_j||_op <= C, hence ||U||_op <= C after
the two directions are combined.
The contribution of the group is T_group = U diag(r_p^{-1}) U^*. The
finite-matrix inequality ||A D B||_HS <= ||A||_op ||D||_HS ||B||_op gives
||T_group||_HS^2 <= C sum_p r_p^{-2}
<= C' sum_{r=1}^N r r^{-2} <= C''(1 + log N).
The triangle inequality over the four groups gives the same order for
||T||_HS^2, and together with the minimum sum this proves the cost
bound.
Proof of the annular proposition. Apply the family inequality in the
law with twist t, using the family of modes and the pinned set P of
boundary vertices, on which every mode vanishes. The left side is
R_t(-i lambda_N htilde, i lambda_N htilde)
= lambda_N^2 R_t(htilde, htilde),
because R_t is bilinear and htilde is real. By the mode lemma,
uniformly in t,
R_t(htilde, htilde) <= C_beta (1 + log N) / (log N)^2
<= C'_beta / log N.
The bound must pass to h without treating R_t as a positive form. If
R_t(h,h) <= 0, the desired positive upper bound already holds.
Otherwise the second-derivative identity and the pointwise bound on
H_{2,t} give
< X_t(hc)^2 >_t < - < H_{2,t}(hc,hc) >_t <= beta ||h||_2^2 <= C beta.
Put d = h - htilde. The current is a real random variable on real edge
forms, and |X_t(dc)| <= beta ||d||_1 <= C beta / N. Ordinary
Cauchy-Schwarz for random variables, followed by the current bound,
gives
| < X_t(hc)^2 - X_t(htilde c)^2 >_t |
<= 2 < X_t(hc)^2 >_t^{1/2} ||X_t(dc)||_inf + ||X_t(dc)||_inf^2
<= C_beta / N.
The local quadratic term has the stronger pointwise bound
| H_{2,t}(hc,hc) - H_{2,t}(htilde c, htilde c) |
<= beta sum_{e in E} |d_e| |h_e + htilde_e| <= C beta / N^2.
The formula for R_t, the two displayed differences, and the preceding
mean bound imply in both cases that
- Z''(t) / Z(t) = R_t(h,h)
<= C_beta / log N + C_beta / N <= delta_N,
delta_N = C'_beta / log N.
All constants are independent of the twist, of the retained vertices
and edges, and of the admissible strengths.
Finally, Z is positive, smooth in t, and 2 pi-periodic. Set
M = max_t Z(t) and choose t_0 with Z(t_0) = M. Then Z'(t_0) = 0, and
the curvature bound gives Z''(t) >= -delta_N Z(t) >= -delta_N M. For any
t choose eps in {-1,1} and 0 <= l <= pi with t = t_0 + eps l modulo
2 pi. Twice integrating the second-derivative bound for
v -> Z(t_0 + eps v) gives
Z(t) >= M - delta_N M l^2 / 2 >= M ( 1 - pi^2 delta_N / 2 ).
Taking the minimum and enlarging C_beta proves the minimum-to-maximum
bound; since Z(pi)/Z(0) >= min Z / max Z, it also proves
Z(pi)/Z(0) >= 1 - C_beta / log N.
The uniform annular estimate turns into a bound on connections through
sign clusters of the type studied for continuous-spin systems by
Campbell and Chayes. The finite-volume comparison statements needed for
that step are proved in full, including the comparison with pinned
spins.
Return to the free spin law on an arbitrary finite graph G = (V,E) with
strengths b_e in [0,beta]. Until the final projection, n = 3 and the
spins are written s_x.
Disintegration. Write
s_x = ( r_x tau_x, q_x cos theta_x, q_x sin theta_x ),
q_x = (1 - r_x^2)^{1/2}.
Under the reference measure the variables r_x, tau_x, theta_x are
independent and uniform on [0,1], {-1,1}, and R/(2 pi Z). The first
coordinate of a uniform point on S^2 is uniform on [-1,1], independently
of the azimuthal angle. Conditional on all amplitudes r = (r_x), the
signs and the angles form independent Gibbs systems with couplings
K_{xy} = b_{xy} r_x r_y, L_{xy} = b_{xy} q_x q_y,
and edge energies tau_x tau_y and cos(theta_x - theta_y). Let Z_I(K) and
Z_P(L) be their partition functions with uniform probability reference
measures; the amplitude density is proportional to
rho(r) = Z_I(K(r)) Z_P(L(r)).
Bonds. Use the embedded Ising representation underlying the reflection
clusters of Wolff, coupled to bonds as in Edwards and Sokal. Given the
spins, sample bonds eta_e in {0,1} independently; for e = {x,y} set
p_e = 1 - exp(-2 K_e), P(eta_e = 1 | s) = p_e 1{tau_x = tau_y}.
The identity exp(K_e tau_x tau_y) = exp(K_e) [ (1 - p_e) + p_e
1{tau_x = tau_y} ] shows that, conditional on r and eta, the signs are
independent fair signs on each open connected component. In particular
the weighted connection identity of Campbell and Chayes holds,
< s_x^1 s_y^1 > = E[ r_x r_y 1{ x <-> y } ],
where x <-> y means that x and y lie in one open component of G, using
only open edges.
To bound these connections it suffices to show that fixing any set of
spins at (1,0,0) cannot decrease the probability of an increasing bond
event; the comparison is applied to simultaneous annulus crossings. Its
proof uses association jointly in amplitudes and bonds, since the
pinning constrains amplitudes as well as cluster signs.
Association. The finite-product form of the
Fortuin-Kasteleyn-Ginibre association criterion is used, with both
discrete and continuous coordinates, and its proof is given. A law on a
coordinatewise ordered product is associated if the covariance of any
two bounded increasing functions is nonnegative. A positive function w
is log-supermodular if
w(join(u,v)) w(meet(u,v)) >= w(u) w(v),
with join and meet taken coordinatewise.
Association lemma. Let w be a positive log-supermodular density on a
finite product of compact intervals or finite chains, with respect to
the product of Lebesgue or counting measures; in the interval case let
w be continuous. The law proportional to w is associated, and its
conditional law on all but one coordinate increases stochastically as
the remaining coordinate increases.
Proof. For one coordinate, association follows from independent copies
X, X' and
2 Cov(F(X),H(X)) = E[ (F(X)-F(X'))(H(X)-H(X')) ] >= 0.
Proceed by induction on the number of coordinates. Write a point as
(u,t) with t the last coordinate. Each conditional density w(u,t) is
log-supermodular in u and its law is associated by induction. For
t' >= t, log-supermodularity makes w(u,t')/w(u,t) increasing in u: apply
the defining inequality to (u,t') and (v,t) when u <= v. Tilting the
conditional law at t by this increasing likelihood ratio, and using its
association, shows that the conditional law at t' stochastically
dominates it; in the continuous case the ratios are bounded by
positivity and compactness.
If F(u,t) and H(u,t) are increasing, their conditional expectations are
increasing functions of t. The decomposition
Cov(F,H) = E[ Cov(F,H | t) ] + Cov( E[F|t], E[H|t] )
has nonnegative terms by induction and by the one-coordinate case. This
proves both assertions.
The Ising and plane-rotator cases of the method of Ginibre supply the
required inequalities. Their restricted scope matters: no such inequality
for the full vector-spin law is assumed.
Edge-correlation lemma. In either the free Ising or the free planar
system on a finite graph with nonnegative couplings, every edge energy
has nonnegative mean, and any two edge energies have nonnegative
covariance.
Proof. The mean inequality follows by expanding the Gibbs exponential.
For Ising spins every reference moment of a sign monomial is zero or
one; for planar spins each cosine of an angle difference has
nonnegative Fourier coefficients, so every product of such cosines has
a nonnegative constant Fourier coefficient.
For covariances take two independent copies of the Gibbs system. If T_e
and T_f are edge energies and primes mark the second copy, then
2 Cov(T_e, T_f) = E[ (T_e - T'_e)(T_f - T'_f) ].
In the planar reference integral make the substitution theta = phi + psi,
theta' = phi - psi modulo 2 pi. Independent uniform phi, psi produce
independent uniform theta, theta': the map is a surjective homomorphism
of compact tori and preserves Haar probability. For an oriented edge g
write phi_g = phi_x - phi_y and similarly psi_g. The doubled density
numerator becomes exp(2 sum_g L_g cos phi_g cos psi_g) and T_e - T'_e =
-2 sin phi_e sin psi_e. Hence the unnormalized numerator on the right is
4 sum_{a in N_0^E} ( prod_{g in E} (2 L_g)^{a_g} / a_g! )
( int sin phi_e sin phi_f prod_{g in E} (cos phi_g)^{a_g} dphi )^2
>= 0,
the integral against product Haar probability and each factor real; the
possible multiplicity of the torus map introduces no factor because the
measures are probabilities.
For Ising reference signs d, d' set u = (d + d')/2 and v = (d - d')/2
at each site. The sitewise values are (1,0), (-1,0), (0,1), (0,-1) with
equal probabilities, and every mixed monomial in u and v has
nonnegative expectation: a monomial with positive powers of both
vanishes, an odd pure moment vanishes, and an even pure moment is
nonnegative. The doubled density is exp(2 sum_g K_g (u_x u_y + v_x v_y))
and an edge-energy difference is 2 (u_x v_y + v_x u_y). Expanding the
exponential and the product of two differences gives only nonnegative
coefficients multiplying nonnegative reference moments, and all
expansions are absolutely integrable on the finite products.
The amplitude law is associated. This is the finite-graph statement of
Campbell and Chayes; its proof is included because both its sign
structure and its boundary-free hypothesis are used.
Amplitude-association lemma. The amplitude law with density rho(r) =
Z_I(K(r)) Z_P(L(r)) is associated.
Proof. For either partition function, with couplings J_e(r) and edge
energies T_e, differentiation in distinct site variables r_x, r_y gives
partial_x partial_y log Z(J(r))
= sum_e < T_e > partial_x partial_y J_e
+ sum_{e,f} Cov(T_e, T_f) (partial_x J_e)(partial_y J_f).
For J = K the first derivatives are nonnegative; for J = L the first
derivatives are nonpositive, so the products in the second sum are
again nonnegative.
The distinct-site mixed derivatives of individual couplings vanish
unless e = {x,y}, in which case
partial_x partial_y K_{xy} = b_{xy},
partial_x partial_y L_{xy} = b_{xy} r_x r_y / (q_x q_y) >= 0
in the open amplitude cube. The edge-correlation lemma therefore shows
that every distinct-site mixed derivative of log rho is nonnegative
there. An increment of log rho in one coordinate is thus nondecreasing
in every other coordinate. For two points u, v compare the successive
increments from meet(u,v) to u with the same coordinate increments from
v to join(u,v): their sum gives the log-supermodular inequality in the
open cube, and continuity extends it to the closed cube. A possible
singularity of derivatives at r_x = 1 is immaterial. The density is
positive and continuous, so the association lemma applies.
The joint law of amplitudes and bonds is also associated. The
conditioning argument used for bond association by Campbell and Chayes
gives the following statement.
Joint-association lemma. The joint free law of r and eta is associated,
with amplitudes and bond indicators ordered coordinatewise.
Proof. Conditional on r, summing the signs in the bond expansion gives
a bond weight proportional to
2^{k(eta)} prod_{e in E} p_e^{eta_e} (1 - p_e)^{1 - eta_e},
where k(eta) counts all open components, singletons included. The
function k is supermodular: the change on opening an edge is -1 when
its endpoints were disconnected and 0 otherwise, and this change
increases as other bonds are opened. Hence the weight is
log-supermodular and its law is associated by the association lemma.
Increasing any nondegenerate parameter p_e tilts the weight by an
increasing likelihood ratio, so association implies stochastic increase
of the conditional bond law in each p_e and consequently in r. A zero
parameter forces its edge closed; restriction to the remaining edges, or
continuity, handles these degenerate cases.
For bounded increasing functions F(r,eta) and H(r,eta), conditional
association makes E[ Cov(F,H | r) ] nonnegative. Their conditional means
are increasing in r, by stochastic monotonicity of the bond law and
their direct monotonicity in r. Their covariance is nonnegative by the
amplitude-association lemma. The total covariance identity proves the
claim.
Pinning and crossings. Let s_* = (1,0,0). For a set B of vertices, the
pinned spin law means that s_x = s_* for x in B, while all other spins
keep uniform reference measures and the same edge interactions. Bonds
are sampled by the same rule p_e 1{tau_x = tau_y}.
Pinning-comparison proposition. For the three-component free spin law on
any finite graph with nonnegative strengths, and bonds sampled by the same
rule, pinning any set of spins to s_* cannot decrease the probability of
any increasing bond event A.
Proof. For 0 < eps < 1 condition the free system on r_x >= 1 - eps and
tau_x = 1 for every x in B. Conditional on r and eta, the probability of
the sign requirement is 2^{-k_B(eta)}, where k_B is the number of open
components meeting B. The induced law of r and eta is thus tilted by
T_eps(r,eta) = 1{ r_x >= 1 - eps for all x in B } 2^{-k_B(eta)}.
Opening an edge cannot increase k_B, including when it merges a
component meeting B with one not meeting B, so T_eps is increasing. The
joint-association lemma gives
E[ 1_A T_eps ] / E[ T_eps ] >= P(A).
As eps decreases to zero the conditioned reference measures on the
positive caps converge to point masses at s_*, and the finite-volume
Gibbs weight is bounded, positive and continuous. The spin-to-bond
kernel is also continuous, its edge-opening probability being
1 - exp(-2 b_{xy} max{s_x^1 s_y^1, 0}), with no discontinuity at a zero
first coordinate. Integrating any event of the finitely many bonds and
passing to the limit gives the comparison for exact pinning.
Specialize now to finite subgraphs of the nearest-neighbour graph on
Z^2, with strengths in [0,beta]. For z in Z^2 and an integer N >= 2 set
D_N(z) = { v in Z^2 : N <= ||v - z||_inf <= 2N }.
An annulus crossing is an open path in the restriction of G to D_N(z)
joining its layers of sup radius N and 2N; edges or vertices outside the
annulus are excluded from the path.
Crossing-suppression proposition. For each beta < infinity there are
numbers p_N = p_N(beta) > 0 tending to zero as N tends to infinity with
the following property. For every finite lattice subgraph, every
strength array in [0,beta], and every list of centers whose closed boxes
of sup radius 2N are pairwise disjoint, the free probability that all
the corresponding annuli are crossed is at most p_N to the number of
annuli. The boxes need not lie inside the finite graph.
Proof. First work in one restricted annulus graph. Pin both layers to
s_* and let Z(0) be its partition function; let Z(pi) be the partition
function with the inner layer changed to -s_* and the outer layer left at
s_*. Then
P_pinned(crossing) = 1 - Z(pi)/Z(0).
Indeed, in the bond expansion every open component missing the pinned
layers has two sign choices, and a component meeting a pinned layer has
one sign choice provided all prescribed signs on it agree. Changing the
inner signs removes exactly the bond configurations joining the two
layers and changes the weight of no remaining configuration: the
boundary amplitudes are r = 1, q = 0, and the reference normalization
for the unpinned signs is unchanged. This holds at every interior
amplitude configuration and hence after integration. If a layer has no
vertices, both sides vanish.
By the annular proposition the right side is bounded uniformly by a
nonnegative sequence tending to zero; enlarge it slightly to a strictly
positive sequence p_N with the same limit.
For several annuli pin all their present boundary-layer sites to s_*
simultaneously. The intersection of the crossing events is increasing,
so the pinning-comparison proposition bounds its free probability by its
pinned probability. In the pinned system each unpinned annulus region is
separated from every vertex outside that annulus by the pinned layers,
since a nearest-neighbour path cannot jump across either layer. The
unpinned annulus spins are then independent with the required fixed
boundary values, and each resulting annulus marginal is evaluated by the
crossing ratio applied to the partition functions of that restricted
annulus graph. Boundary-only interactions contribute constants, and
independent sampling of edge bonds preserves independence of the
crossing events, so their probabilities multiply.
This separation uses only edges present in G; missing vertices or edges,
including those beyond the finite volume, cannot create an unseparated
nearest-neighbour interaction. Pinning the present layer sites and
applying the annular proposition to the restricted subgraph gives the
bound also for truncated boxes.
Connection-decay proposition. For every beta < infinity there are
constants A_0 < infinity and m_0 > 0, depending only on beta, such that
the three-component free bond law on every finite square-lattice
subgraph with strengths in [0,beta] satisfies
P(x <-> y) <= A_0 exp( - m_0 |x - y| ), x,y in V.
Proof. Translate the lattice so that x = 0. Choose a finite integer
N >= 2 so large that, for the sequence of the crossing-suppression
proposition,
0 < p := p_N < 1, alpha := 4 p^{1/25} < 1.
Partition Z^2 into boxes z + {0,...,N-1}^2 with coarse vertices z in
N Z^2. Suppose 0 and y are joined by an open path and d := ||y||_inf >
4N. Recording the visited boxes and erasing loops produces a simple
nearest-neighbour coarse path from the box of 0 to the box of y; if its
length is k then
k >= d/N - 1.
Each vertex z of this simple coarse path is a box visited by the
original open path, at some vertex within sup distance N of z. At least
one original endpoint is farther than 2N from z, for otherwise their
mutual distance would be at most 4N. A segment of the original path from
the visited vertex to that endpoint therefore contains a crossing of
D_N(z): its first hit on the outer layer and its last preceding hit on
the inner layer are joined inside the annulus. It does not matter
whether that segment survived coarse loop erasure; its edges are still
open.
Among the k+1 distinct coarse vertices, one of the 25 residue classes
modulo 5N in the two coordinates contains at least (k+1)/25 vertices.
Their closed outer boxes of sup radius 2N are disjoint, because distinct
centers in the class differ by at least 5N in some coordinate. Choose
one maximizing class deterministically for each candidate coarse path.
The crossing-suppression proposition bounds the probability that this
path has all its required crossings by p^{(k+1)/25}. There are at most
4^k coarse paths of length k from the initial box, so the union bound
and the length inequality give
P(0 <-> y) <= p^{1/25} sum_{k >= ceil(d/N - 1)} alpha^k
<= ( p^{1/25} / ( alpha (1 - alpha) ) )
exp( - ( - log alpha / N ) d ).
Since d >= |y| / sqrt(2), take m_0 = ( - log alpha ) / ( sqrt(2) N ) > 0
and increase the finite prefactor to cover d <= 4N. Every estimate was
uniform in the finite subgraph and its strengths.
Conditioning to three components. The uniform connection estimate is
already available in three dimensions. Passing to spheres of higher
dimension needs only a statement about the conditional reference measure.
Three-component conditioning lemma. Let n > 3 and let sigma_x in S^{n-1}
have the free Gibbs law on a finite graph. Write sigma_x = (rho_x s_x,
v_x) with v_x in R^{n-3}, rho_x = (1 - |v_x|^2)^{1/2}, and s_x in S^2
whenever rho_x > 0. Conditional on (v_x), the directions (s_x) have
exactly the free three-component Gibbs law with strengths
b_{xy} rho_x rho_y in [0,beta].
Proof. Represent a uniform sphere point as (U,V)/(|U|^2 + |V|^2)^{1/2},
where U in R^3 and V in R^{n-3} have independent standard Gaussian
coordinates. The direction U/|U| is uniform on S^2 and independent of
(|U|, V). Since v = V/(|U|^2 + |V|^2)^{1/2} depends only on (|U|, V),
the three-component direction is independent of v under the reference
law. Applying this representation independently at every site shows
that, conditional on all v_x, the reference directions remain
independent and uniform on S^2.
The conditioned interaction is
sum_{{x,y} in E} b_{xy} v_x . v_y
+ sum_{{x,y} in E} b_{xy} rho_x rho_y s_x . s_y.
The first sum is constant under the conditional law and cancels from the
normalization; the second gives the asserted couplings. The event
rho_x = 0 has reference probability zero and Gibbs probability zero, the
finite Gibbs density being positive and bounded; alternatively, choosing
any independent reference direction at such a site decouples it without
changing any spin product.
Completion of the main result. For n = 3, the connection identity and the
connection-decay proposition give
0 <= < s_x^1 s_y^1 >_{G,b}^{(3)} <= A_0 exp(- m_0 |x - y|).
For n > 3 apply the conditioning lemma. Conditional on all v_x, the
first-coordinate correlation equals rho_x rho_y times the corresponding
three-component first-coordinate correlation, which is nonnegative and
obeys the same bound uniformly in the conditional strength array. Since
0 <= rho_x rho_y <= 1, averaging gives, for every n >= 3,
0 <= < sigma_x^1 sigma_y^1 >_{G,b}^{(n)} <= A_0 exp(- m_0 |x - y|).
The free spin law is invariant under coordinate permutations, so
< sigma_x . sigma_y >_{G,b}^{(n)}
= n < sigma_x^1 sigma_y^1 >_{G,b}^{(n)} >= 0.
Taking A = n A_0 and m = m_0 gives 0 <= < sigma_x . sigma_y >_{G,b}^{(n)}
<= A exp(- m |x - y|) for every n >= 3, uniformly in the finite subgraph
and in the strengths.
The constants are independent of the volume, so the estimate passes
unchanged to the free-boundary limsup. If a subsequence of the box laws
converges locally weakly, the expectation of the bounded continuous
local function sigma -> sigma_0 . sigma_x converges along it, and its
limit obeys the same bound. This gives the free-boundary two-point
estimate and the bound on subsequential limits without any uniqueness
assertion.