Ann. Henri Poincaré 23 (2022), 1061–1139
c 2021 The Author(s)
1424-0637/22/031061-79
published online October 11, 2021
https://doi.org/10.1007/s00023-021-01107-3
Annales Henri Poincaré
Non-integrable Ising Models in Cylindrical
Geometry: Grassmann Representation and
Infinite Volume Limit
Giovanni Antinucci, Alessandro Giuliani and Rafael L. Greenblatt
Abstract. In this paper, meant as a companion to Antinucci et al. (Energy
correlations of non-integrable Ising models: the scaling limit in the cylinder, 2020. arXiv: 1701.05356), we consider a class of non-integrable 2D
Ising models in cylindrical domains, and we discuss two key aspects of
the multiscale construction of their scaling limit. In particular, we provide a detailed derivation of the Grassmann representation of the model,
including a self-contained presentation of the exact solution of the nearest
neighbor model in the cylinder. Moreover, we prove precise asymptotic
estimates of the fermionic Green’s function in the cylinder, required for
the multiscale analysis of the model. We also review the multiscale construction of the effective potentials in the infinite volume limit, in a form
suitable for the generalization to finite cylinders. Compared to previous
works, we introduce a few important simplifications in the localization
procedure and in the iterative bounds on the kernels of the effective
potentials, which are crucial for the adaptation of the construction to
domains with boundaries.
Contents
1.
2.
Introduction
The Nearest-Neighbor Model
2.1. Diagonalization of the Free Action
2.1.1. Introduction to the Grassmann Variables and
Representation
2.1.2. Diagonalization of St1 ,t2
2.2. The Critical Propagator: Multiscale Decomposition and
Decay Bounds
1062
1069
1070
1070
1071
1075
1062
G. Antinucci et al.
2.2.1.
2.2.2.
Ann. Henri Poincaré
Multiscale and Bulk/Edge Decompositions
Decay Bounds and Gram Decomposition: Statement
of the Main Results
2.3. Asymptotic Behavior of the Critical Propagator
2.4. Symmetries of the Propagator
3. Grassmann Representation of the Generating Function
4. The Renormalized Expansion in the Full-Plane Limit
4.1. Effective Potentials and Kernels: Representation and
Equivalence
4.2. Localization and Interpolation
4.3. Trees and Tree Expansions
4.4. Bounds on the Kernels of the Full Plane Effective Potentials
4.5. Beta Function Equation and Choice of the Counterterms
Acknowledgements
A. Diagonalization of the Matrix Ac
B. Proof of Proposition 2.3
C. Proof of Proposition 2.9
D. Non-interacting Correlation Functions in the Scaling Limit
References
1075
1076
1078
1079
1080
1086
1089
1091
1096
1102
1110
1115
1116
1120
1129
1135
1137
1. Introduction
In this article, which is a companion to [4], we consider a class of non-integrable
perturbation of the 2D nearest-neighbor Ising model in cylindrical geometry
and discuss some of the key ingredients required in the multiscale construction
of the scaling limit of the energy correlations in finite domains. The material
presented here generalizes and simplifies the approach proposed by two of the
authors in [18], where a similar problem in the translationally invariant setting
was investigated. As discussed extensively in [4, Section 3.1], which we refer
to for additional motivations and references, the methods of [18], as well as of
several other related works on the Renormalization Group (RG) construction
of the bulk scaling limit of non-integrable lattice models at the critical point,
are insufficient for controlling the effects of the boundaries at the precision
required for the construction of the scaling limit in finite domains. This is a
serious obstacle in the program of proving conformal invariance of the scaling
limit of statistical mechanics models [17]; the goal would be to prove results
comparable to the remarkable ones obtained for the nearest neighbor 2D Ising
model [10,12,29], but for a class of non-integrable models, such as perturbed
Ising [2] or dimer models [22] in two dimensions, via methods that do not
rely on the exact solvability of the microscopic model. In this paper and in its
companion [4], we attack this program by constructing the scaling limit of the
energy correlations of a class of non-integrable perturbations of the standard
2D Ising model in the simplest possible finite domain with boundary, that is,
a finite cylinder.
Vol. 23 (2022) Non-integrable Ising Models in Cylindrical Geometry
1063
Let us define the setting more precisely. For positive integers L and M ,
with L even, we let GΛ be the discrete cylinder with sides L and M in the
horizontal and vertical directions, respectively, with periodic boundary conditions in the horizontal direction and open boundary conditions in the vertical
directions. We consider GΛ as a graph with vertex set Λ = ZL × (Z ∩ [1, M ]),
where ZL = Z/LZ (in the following we shall identify the elements of ZL with
{1, . . . , L}, unless otherwise stated) and edge set BΛ consisting of all pairs
of the form1 {z, z + êj } for z ∈ Λ, j ∈ {1, 2} and ê1 , ê2 the unit vectors in
the two coordinate directions. For x ∈ BΛ , we let j(x) be the j for which
x = {z, z + êj } for some z ∈ Λ, so that j(x) = 1 for a horizontal bond and
j(x) = 2 for a vertical bond. The model is defined by the Hamiltonian
Jj(x) ǫx − λ
V (X)σX ,
(1.1)
HΛ (σ) = −
x∈BΛ
X⊂Λ
where J1 , J2 are two positive constants, representing the couplings in the horizontal and vertical directions, ǫx = ǫx (σ) := σz
σz′ for x = {z, z ′ }; the spin
Λ
variable σ belongs to ΩΛ := {±1} , and σX := x∈X σx ; V is a finite range,
translationally invariant, even interaction, obtained by periodizing in the horizontal direction a Λ-independent, translationally invariant, potential on Z2 ;
finally, λ is the strength of the interaction, which can be of either sign and, for
most of the discussion below, the reader can think of as being small, compared
to J1 , J2 , but independent of the system size. In the following, we shall refer to
model (1.1) with λ = 0 as to the ‘interacting’ model, in contrast with the standard nearest-neighbor model, which we will refer to as the ‘non-interacting’,
one of several terminological conventions motivated by analogy with quantum
field theory. The Hamiltonian defines a Gibbs measure ·β,Λ depending on the
inverse temperature β > 0, which assigns to any F : ΩΛ → R the expectation
value
e−βHΛ (σ) F (σ)
Λ
.
(1.2)
F β,Λ := σ∈Ω
−βHΛ (σ)
σ∈ΩΛ e
The truncated correlations, or cumulants, of the energy observable ǫx , denoted
ǫx1 ; · · · ; ǫxn β,Λ , are given by
∂n
log eA1 ǫx1 +···An ǫxn β,Λ
.
ǫx1 ; · · · ; ǫxn β,Λ :=
∂A1 · · · ∂An
A1 =···=An =0
(1.3)
For the formulation of the main result, let us fix once and for all an interaction
V with the properties spelled out after (1.1), and assume that J1 /J2 and L/M
belong to a compact K ⊂ (0, +∞). We let tl := tl (β) := tanh βJl , with l = 1, 2,
and recall that in the non-interacting case, λ = 0, the critical temperature
βc (J1 , J2 ) is the unique solution of t2 (β) = (1 − t1 (β))/(1 + t1 (β)). Note that
there exists a suitable compact K ′ ⊂ (0, 1) such that whenever J1 /J2 ∈ K
and β ∈ [ 21 βc (J1 , J2 ), 2βc (J1 , J2 )], then t1 , t2 ∈ K ′ . From now on, we will
1 If
z = ((z)1 , (z)2 ) ∈ Λ has horizontal coordinate (z)1 = L, we use the convention that
z + ê1 ≡ (1, (z)2 ).
1064
G. Antinucci et al.
Ann. Henri Poincaré
think K, K ′ to be fixed once and for all. Moreover, we parameterize the Gibbs
measure in terms of tl as follows:
·β,Λ ≡ ·λ,t1 ,t2 ;Λ .
Given these premises, we are ready to state the main result proven in [4].
Theorem 1.1. Fix V as discussed above. Fix J1 , J2 so that J1 /J2 belongs to the
compact K introduced above. There exist λ0 > 0 and analytic functions βc (λ),
t∗1 (λ), Z1 (λ), Z2 (λ), defined for |λ| ≤ λ0 , such that, for any finite cylinder Λ
with L/M ∈ K and any m-tuple x = (x1 , . . . xm ) of distinct elements of BΛ ,
with m1 horizontal elements, m2 vertical elements, and m = m1 + m2 ≥ 2,
m2
m1
Z2 (λ)
ǫx1 ; . . . ; ǫxm λ,t1 (λ),t2 (λ);Λ = Z1 (λ)
ǫx1 ; . . . ; ǫxm 0,t∗1 (λ),t∗2 (λ);Λ
+RΛ (x),
(1.4)
where t1 (λ) := tanh(βc (λ)J1 ), t2 (λ) := tanh(βc (λ)J2 ) and t∗2 (λ) := (1 −
t∗1 (λ))/(1 + t∗1 (λ)). Moreover, denoting by δ(x) the tree distance of x, i.e.,
the cardinality of the smallest connected subset of BΛ containing the elements
of x, and by d = d(x) the minimal pairwise distance among the midpoints of
the edges in x and the boundary of Λ, for all θ ∈ (0, 1) and ε ∈ (0, 1/2) and a
suitable Cθ,ε > 0, the remainder RΛ can be bounded as
m
|RΛ (x)| ≤ Cθ,ε
|λ|m!
1
dm+θ
d
δ(x)
2−2ε
.
(1.5)
As a corollary of this theorem, one readily obtain the existence and
explicit structure of the scaling limit for the ‘energy sector’ of the interacting model, with quantitative estimates on the speed of convergence; see [4,
Corollary 1.2] and Appendix D.
The proof of Theorem 1.1 is based on a multiscale analysis of the generating function of the energy correlations, formulated in the form of a Grassmann
(Berezin) integral. While the strategy of this proof is based on the same general ideas used in [18] in the translationally invariant setting, that is, on the
methods of the fermionic constructive RG, the presence of boundaries introduces several technical and conceptual difficulties, whose solution requires to
adapt, improve and generalize the ‘standard’ RG procedure (e.g., in the definition of the ‘localization procedure’, in the way in which the kernels of the
‘effective potentials’ are iteratively bounded and in which the resulting bounds
are summed over the label specifications, etc.) As discussed in [4, Section 3.1],
which we refer to for additional details, we expect that understanding how to
implement RG in the presence of boundaries or, more in general, of defects
breaking translational invariance, will have an impact on several related problems, such as the computation of boundary critical exponents in models in the
Luttinger liquid universality class, the Kondo problem, the Casimir effect, and
the phenomenon of many-body localization.
In this paper we give a full presentation of some of the key ingredients
required in the proof of our main result, namely:
Vol. 23 (2022) Non-integrable Ising Models in Cylindrical Geometry
1065
1. exact solution of the nearest neighbor model on the cylinder in its Grassmann formulation, including multiscale bounds on the bulk and edge
parts of the fermionic Green’s function (Sect. 2);
2. reformulation of the generating function of energy correlations of the
interacting model as a Grassmann integral (Sect. 3);
3. tree expansion and iterative bounds on the kernels of the effective potentials of the interacting theory in the full plane limit, including the computation and proof of analyticity of the interacting critical temperature
(Sect. 4).
The other ingredients, including most of the novel aspects of the RG construction in finite volume, such as the definition of the localization procedure
in finite volume, the norm bounds on the edge part of the effective potentials and the asymptotically sharp estimates on the correlation functions in
the cylinder, are deferred to [4]; see the end of [4, Section 3.1] for a detailed
summary and roadmap of the proof of Theorem 1.1.
Before starting the technical presentation, let us anticipate in little more
detail the contents of the following sections, thus clarifying the main results
of this paper.
Section 2: exact solution of the model in the cylinder. The multiscale construction of the interacting theory in the domain Λ requires a very fine control of the
non-interacting model at the critical point, and, in particular, of the structure
of its fermionic Green’s function, which we call the ‘propagator’; the propagator is nothing but the inverse of a signed adjacency matrix A, whose definition
we recall in Sect. 2 below [26, Chapter IX]. The key properties we need, and
we prove in Sect. 2 below (with some—important!—technical aspects of the
proofs deferred to Appendices A, B and C ), see, in particular, Eq. (2.2.14)
and Proposition 2.3 below, are the following:
• multiscale decomposition of the propagator and bulk–edge decomposition
of the single-scale propagator;
• exponentially decaying pointwise bounds on the bulk and edge parts
of the single-scale propagators, with optimal dimensional bounds (with
respect to the scale index) on their L∞ norms and on their decay rates;
• Gram representation2 of the bulk and edge parts of the single-scale propagators, with optimal dimensional bounds (with respect to the scale index)
on the norms of the Gram vectors.
In reference with the second item, let us remark that the exponential decay
needed (and proved below) for the propagator between two points z, z ′ ∈ Λ,
is in terms of the ‘right’ distance between z and z ′ , namely: the standard
Euclidean distance on the cylinder between z and z ′ in the case of the bulk
part of the single-scale propagator; the Euclidean distance on the cylinder
between z, z ′ and the boundary of Λ, in the case of the edge part of the
single-scale propagator. In particular, the exponential decay of the edge part
2 We
say that a matrix g admits a Gram representation, if its elements gi,j can be written
as the scalar product of two vectors ui and vj in a suitable Hilbert space.
1066
G. Antinucci et al.
Ann. Henri Poincaré
of the single-scale propagator in the distance of z, z ′ from the boundary of
Λ is of crucial importance for proving improved dimensional bounds on the
finite-size corrections to the thermodynamic and correlation functions of the
interacting model, which are systematically used in the conclusion of the proof
of Theorem 1.1 in the companion paper [4], see [4, Section 4].
The proof we give of these key properties is based on an exact diagonalization of the signed adjacency matrix A in terms of the roots of a set of
polynomials (this calculation first appeared in [23], and a similar calculation
for a rectangle appears in [25]). It is unlikely that such an explicit diagonalization can be obtained in more general domains than the torus, the straight
cylinder or the rectangle. Therefore, in order to generalize Theorem 1.1 to more
general domains, it would be desirable to prove the properties summarized in
the three items above via a more robust method, not based on an explicit
diagonalization of A. It remains to be seen whether the methods of discrete
holomorphicity, which allowed to prove the convergence of the propagator in
general domains to an explicit, conformally covariant, limiting function [10],
may allow one to prove the desired properties in general domains.
Section 3: Grassmann representation of the generating function. In Sect. 3,
we turn our attention to the generating function for the energy correlations
βJj(x) + Ax ǫx + βλ
exp
V (X)σX . (1.6)
ZΛ (A) :=
σ∈ΩΛ
x∈BΛ
X⊂Λ
that, if computed at a configuration A such that
Ax is equal to Ai for x = xi
and zero otherwise, reduces to the combination eA1 ǫx1 +···An ǫxn β,Λ appearing
in (1.3), up to an overall multiplicative constant, independent of A.
In Proposition 3.1 and Eq.(3.22) (adapting a similar result for the torus
in [18]), we show that the correlations without repeated bonds are the same
as those obtained by replacing ZΛ (A) with a Grassmann integral of the form
(1)
W(A )
∗
ΞΛ (A) := e
Pc∗ (Dφ)Pm
(Dξ)eV (φ,ξ,A ) ,
(1.7)
∗
are Gaussian Grassmann measures associated with the critwhere Pc∗ and Pm
ical, non-interacting Ising model at parameters t∗1 , t∗2 := (1 − t∗1 )/(1 + t∗1 ),
with t∗1 a free parameter. Moreover, W(A) is a multilinear function of A
and V (1) (φ, ξ, A) is a Grassmann polynomial whose coefficients are multilinear functions of A, both of which are defined in terms of explicit, convergent,
expansions. As a corollary of Lemma 3.2, we additionally prove that the ‘kernels’ of W(A) and V (1) (φ, ξ, A) (i.e., the coefficients of their expansions in
A, φ, ξ, thought of as functions of the positions of the components of A, φ, ξ on
the cylinder) can be naturally decomposed into sums of a ‘bulk’ part (equal,
essentially, to their infinite plane limit restricted to the cylinder, with the
appropriate boundary conditions) plus an ‘edge’ part (their boundary corrections), exponentially decaying in the appropriate distances. In particular, the
edge part of the kernels decays exponentially (on the lattice scale) away from
Vol. 23 (2022) Non-integrable Ising Models in Cylindrical Geometry
1067
the boundary, a fact that will play a major role in the control of the boundary
corrections to the correlation functions in [4].
Let us remark that, in addition to t∗1 and to the inverse temperature β,
the representation (1.7) has another free parameter, Z (entering the definition of V (1) ); while, for the validity of (1.7), these parameters can be chosen
arbitrarily in certain intervals, in order for this representation to produce a convergent expansion for the critical energy correlations of the interacting model,
uniformly in the system size, we will need to fix t∗1 , β, Z appropriately (a posteriori, they will be fixed uniquely by our construction, see below). Parameters
of this kind are known as counterterms in the RG terminology.
Section 4: the RG expansion for the effective potentials in the full plane limit.
Equation (1.7) is the starting point for a multiscale expansion, which is fully
presented in the companion paper [4], see in particular [4, Section 3], but which
we summarize here in order to provide the context for Sect. 4, where we carry
out an auxiliary expansion for the full plane limit of the kernels of the ‘effective
potentials’. Such an auxiliary expansion, among other things, fixes the values
of t∗1 , β, Z which are actually used in Theorem 1.1, see Sect. 4.5 below.
The goal is to iteratively compute (1.7) in terms of a sequence of effective
potentials, defined as follows: at the first step we let
(1)
W (0) (A )+V (0) (φ,A )
∗
∝ Pm
(Dξ)eV (φ,ξ,A ) ,
(1.8)
e
where ∝ means ‘up to a multiplicative constant independent of A’; the polynomials W (0) , V (0) are specified uniquely by the normalization W (0) (0) =
V (0) (0, A) = 0.
(0)
We are left with computing the integral of eV (φ,A ) with respect to the
Gaussian integration Pc∗ (Dφ) with propagator g∗c . As anticipated above, in
0
Sect. 2.2 we decompose the critical propagator g∗c as g(≤h) + j=h+1 g(j) , for
any h < 0; correspondingly, in light of the addition formula for Grassmann
integrals (see, e.g., [21, Proposition 1]), we introduce the sequences P (≤h) and
P (h) of Gaussian Grassmann integrations, whose propagators are g(≤h) and
g(h) , respectively, and satisfy, for any Grassmann function f ,
P (≤h) (Dφ)f (φ) = P (≤h−1) (Dφ)P (h) (Dϕ)f (φ + ϕ).
(1.9)
We can then iteratively define V (h) and W (h) with W (h) (0) = V (h) (0, A) ≡ 0
and
(h−1)
(h)
(A )+V (h−1) (φ,A )
eW
∝ P (h) (Dϕ)eV (φ+ϕ,A ) .
(1.10)
The iteration continues until the scale h∗ = −⌊ log2 (min{L, M })⌋ is reached,
at which point we let
(h∗ −1)
∗
(h∗ )
(A )
∝ P (≤h ) (Dφ)eV (φ,A ) ,
(1.11)
eW
1068
G. Antinucci et al.
Ann. Henri Poincaré
giving
ΞΛ (A) ∝ exp W(A) +
0
h=h∗ −1
W
(h)
(A) .
(1.12)
In order to obtain bounds on the kernels of W (h) (A) leading to an expansion
for the energy correlations that is uniform in the system size, at each step it
is necessary to isolate from V (h) the contributions that tend to expand (in an
appropriate norm) under iterations: these, in the RG terminology, are the relevant and marginal terms, which we collect in the so-called local part of V (h) ,
denoted by LV (h) . In other words, at each step of the iteration, we rewrite
V (h) = LV (h) + RV (h) , where, in our case, LV (h) includes: three terms that
are quadratic in the Grassmann variables and independent of A, depending on
a sequence of h-dependent parameters which we denote υ = {(νh , ζh , ηh )}h≤1
and call the running coupling constants; and two terms that are quadratic in
the Grassmann variables and linear in A, depending on another sequence of
effective parameters, {Z1,h , Z2,h }h≤0 , called the effective vertex renormalizations. Moreover, RV (h) is the so-called irrelevant, or renormalized, part of the
effective potential, which is not the source of any divergence.
Such a decomposition corresponds to a systematic reorganization, or
‘resummation’, of the expansions arising from the multiscale computation of
the generating function. The goal will be to show that, by appropriately choosing the parameters t∗1 , β, Z, which the right side of (1.7) depends on (and which
are related via a simple invertible mapping to the initial values of the running
coupling constants, ν0 , ζ0 , η0 ), the whole sequence υ remains bounded, uniformly in h∗ ; see Sect. 4.5. Under these conditions, we will be able to show
that the resulting expansions for multipoint energy correlations are convergent,
uniformly in h∗ . Our estimates are based on writing the quantities involved
as sums over terms indexed by Gallavotti-Nicolò (GN) trees [13–15], which
emerge naturally from the multiscale procedure; the relevant aspects of the
definitions of the GN trees will be reviewed in Sect. 4.3 below.
In order to obtain L, M independent values of these parameters, we study
the iteration in the limit L, M → ∞ in Sect. 4; we can also restrict to A = 0,
since this already includes all of the potentially divergent terms. This would
superficially appear to involve a number of complications such as defining an
infinite-dimensional Grassmann integral, but in fact the multiscale computation of the generating function, when understood as an iteration for the kernels
(h)
of V (h) , denoted by VΛ , has a perfectly straightforward infinite-volume version, which is stated and analyzed in Sect. 4. The convergence as L, M → ∞
(h)
(h)
of the finite volume kernels VΛ to the solution V∞ of the infinite-volume
recursive equations for the kernels is one of the main subjects of [4], especially
[4, Section 3].
Section 4 is a reformulation of [18, Section 3]. We nonetheless present
it at length, partly because the treatment of the propagator on the cylinder
in Sect. 2 imposes a different choice of variables which makes the translation
of some statements awkward, but mainly in order to take the opportunity to
Vol. 23 (2022) Non-integrable Ising Models in Cylindrical Geometry
1069
make a number of technical improvements and simplify some unnecessarily
obscure aspects of what is already a complicated argument.
Previously, e.g., in [3,15,16,18], the localization operator (and consequently the remainder) was defined in terms of the Fourier transform of the
functions involved. This has the advantage of providing a simple procedure for
parametrizing the local part of the effective potential by a finite number of
running coupling constants, but is quite difficult to apply to non-translationinvariant systems (in [3] this led to a peculiar restriction on the dependence of
the interaction on the system size). Moreover, it makes the treatment of finite
size corrections awkward and leads to a convoluted definition of the derivative operators in the remainder (see [6] and [19] for the treatment of finitesize corrections on a finite torus via the ‘standard’ definition of localization
operator). To deal with this, in Sect. 4.2 we introduce a localization operator
defined directly in terms of lattice functions and write the remainder in terms
of discrete derivatives using a lattice interpolation procedure. Such definitions
naturally admit finite volume counterparts, discussed in [4, Section 3.1].
The strategy used to estimate the interpolation factors in the above
cited works also involves decomposing them into components which can be
matched with propagators; this involves a number of complications, since it
cannot be done in a strictly iterative fashion (this is the problem discussed
in [6, Section 3.3]). When we handle this issue in Sect. 4.4, see in particular
Proposition 4.6, we instead show iteratively that the coefficient function of the
effective interaction satisfies a norm bound morally equivalent to exponential
decay in the position variables (associated with exponential decay in the scaledecomposed propagators, see Proposition 2.3); this then makes it possible to
bound the contribution of the interpolation operator immediately, avoiding
technical issues such as the ‘accumulation of derivatives’ (see [6, Section 3.3]
and [15, end of Section 8.4]) or the proof that the Jacobian associated with the
change of variables arising from the interpolation procedure is equal to 1 (see
[6, (3.119)]). In these aspects, the strategy used in this paper for iteratively
estimating the kernels of the effective potentials overlaps with [20], which was
developed in parallel with the present work.
2. The Nearest-Neighbor Model
In this section we review some aspects of the exact solution of the nearestneighbor model (λ = 0), which will play a central role in the multiscale computation of the generating function for the energy correlations of the nonintegrable model, to be discussed in the following sections. In particular, after
having recalled the Grassmann representation of the partition function, we
explain how to diagonalize the Grassmann action; next, we compute the Grassmann propagator and define its multiscale decomposition, to be used in the
following; finally, we compute the scaling limit of the propagator, with quantitative bounds on the remainder.
1070
G. Antinucci et al.
Ann. Henri Poincaré
2.1. Diagonalization of the Free Action
2.1.1. Introduction to the Grassmann Variables and Representation. Let us
recall the form of the Hamiltonian HΛ (σ) (1.1) in the integrable case λ = 0:
HΛ (σ) = −
2
Jl
l=1
σz σz+êl ,
z∈Λ
with the understanding that σz+ê2 = 0 for z = (z1 , z2 ) such that3 z2 = M and
σz+ê1 = σz+(1−L)ê1 for z such that z1 = L.
As is well known [26, Chapter
VI.3], the partition function at inverse
−βHΛ (σ)
temperature β > 0, ZΛ =
e
, can be written as a Pfaffian,
σ∈ΩΛ
which admits the following representation in terms of Grassmann variables,
see, e.g., [27] or [16, Appendix A1]:
(2.1.1)
ZΛ =2LM (cosh βJ)L(2M −1) DΦ eSt1 ,t2 (Φ) ,
where Φ = {(H z , Hz , V z , Vz )}z∈Λ is a collection of 4LM Grassmann variables
(we will also use the notation {Φi }i∈I for I a suitable label set with 4LM
elements), DΦ denotes the Grassmann ‘differential’,
DΦ =
dH z dHz dV z dVz ,
z∈Λ
and
St1 ,t2 (Φ) :=
(t1 H z Hz+ê1 + t2 V z Vz+ê2 + H z Hz + V z Vz + V z H z
z∈Λ
+Vz H z + Hz V z + Vz Hz )
(2.1.2)
where tl = tanh βJl for l = 1, 2, and H(L+1,(z)2 ) , V((z)1 ,M +1) should be interpreted as representing −H(1,(z)2 ) and 0, respectively. The identification of
H(L+1,(z)2 ) with −H(1,(z)2 ) corresponds to anti-periodic boundary conditions in
the horizontal direction for the Grassmann variables: these are the right boundary conditions to consider for a cylinder with an even number of sites in the
periodic direction, see [26, Eq.(2.6d)]. For later reference, we let Ex = H z Hz+ê1
for a horizontal edge x with endpoints z, z + ê1 , and Ex = V z Vz+ê2 for a vertical edge x with endpoints z, z + ê2 . Sometimes, we will call {(H z , Hz )}z∈Λ
the horizontal variables, and {(V z , Vz )}z∈Λ the vertical ones.
The quadratic form St1 ,t2 (Φ) can be written as St1 ,t2 (Φ) = 21 (Φ, AΦ)
for a suitable 4LM × 4LM anti-symmetric matrix A (here (·, ·) indicates the
standard scalar product
for vectors whose components are labeled by indices
in I, i.e., (Φ, AΦ) = i,j∈I Φi Aij Φj ). In terms of this matrix A, (2.1.1) can
be rewritten as
ZΛ = 2LM (cosh βJ1 )
3 In
LM
(cosh βJ2 )
L(M −1)
PfA.
this section we shall denote the components of z ∈ Λ by z1 , z2 . We warn the reader that
in the following sections the symbols z1 , z2 will mostly be used, instead, for the first two
elements of an n-ple z in Λn or in (Z2 )n , for which we will use the notation z = (z1 , . . . , zn ).
Vol. 23 (2022) Non-integrable Ising Models in Cylindrical Geometry
1071
[We recall that the Pfaffian of a 2n × 2n antisymmetric matrix A is defined as
1
PfA := n
(−1)π Aπ(1),π(2) ...Aπ(2n−1),π(2n) ;
(2.1.3)
2 n! π
where the sum is over permutations π of (1, . . . , 2n), with (−1)π denoting the
signature. One of the properties of the Pfaffian is that (PfA)2 = detA.] For
later purpose, we also need to compute the averages of arbitrary monomials
in the Grassmann variables Φi , with i ∈ I. These can all be reduced to the
computation of the inverse of A, thanks to the ‘fermionic Wick rule’:
1
1
(2.1.4)
DΦ Φi1 · · · Φim e 2 (Φ,AΦ) = PfG ,
Φi1 · · · Φim :=
PfA
where, if m is even, G is the m × m matrix with entries
Gjk = Φij Φik = −[A−1 ]ij ,ik
(2.1.5)
(if m is odd, the r.h.s. of (2.1.4) should be interpreted as 0). Often Φi Φj is
referred to as the (ij component of the) propagator of the Grassmann field Φ,
or as the covariance of DΦeS(Φ) ; such a form (with a quadratic function in the
exponent) is known as a Grassmann Gaussian measure.
In the following sections, we shall explain how to compute the Pfaffian
of A and its inverse A−1 , via a block diagonalization procedure.
2.1.2. Diagonalization of St 1 ,t 2 .
Horizontal direction diagonalization and Schur reduction. By exploiting the
periodic boundary conditions in the horizontal direction, we can block diagonalize the Grassmann action by performing a Fourier transform in the same
direction: for each z2 ∈ {1, 2, . . . , M } we define
Hz2 (k1 ) =
L
ik1 z1
e
H(z1 ,z2 ) ,
H z2 (k1 ) =
z1 =1
Vz2 (k1 ) =
L
DL :=
in terms of which
1
St1 ,t2 (Φ) =
L
eik1 z1 H (z1 ,z2 ) ,
z1 =1
ik1 z1
e
z1 =1
with k1 ∈ DL , where
L
V(z1 ,z2 ) ,
V z2 (k1 ) =
L
(2.1.6)
ik1 z1
e
V (z1 ,z2 ) ,
z1 =1
π(2m − 1)
L
L
: m = − + 1, · · · ,
L
2
2
k1 ∈DL
z2 =1,...,M
,
(2.1.7)
(1 + t1 e−ik1 )H z2 (−k1 )Hz2 (k1 ) + V z2 (−k1 )Vz2 (k1 )
+t2 V z2 (−k1 )Vz2 +1 (k1 ) + V z2 (−k1 )H z2 (k1 ) + Vz2 (−k1 )H z2 (k1 )
+Hz2 (−k1 )V z2 (k1 ) + Vz2 (−k1 )Hz2 (k1 ) .
(2.1.8)
[Note that as a consequence of the convention that V(z1 ,M +1) = 0, VM +1 (k1 )
should also be interpreted as 0.] The terms in the second line of (2.1.8), mixing
1072
G. Antinucci et al.
Ann. Henri Poincaré
the horizontal with the vertical variables, can be eliminated by a linear transformation corresponding to a Schur reduction of the coefficient matrix (cf. [26,
p. 120]):
ξ
(1 + t1 eik1 )−1 −(1 + t1 eik1 )−1 φ+,z2 (k1 )
(k )
H z2 (k1 )
= +,z2 1 +
,
ξ−,z2 (k1 )
(1 + t1 e−ik1 )−1 (1 + t1 e−ik1 )−1 φ−,z2 (k1 )
Hz2 (k1 )
φ
(k )
V z2 (k1 )
= +,z2 1 ;
φ−,z2 (k1 )
Vz2 (k1 )
(2.1.9)
Defining
a related set of Grassmann variables on Λ by φω,z =
1
e−ik1 z1 φω,z2 (k1 ) and analogously for ξω,z , we then obtain
k
∈D
L
1
L
L
ξ+,z
s+ (z1 − y) −s+ (z1 − y) φ+,(y,z2 )
Hz
,
=
+
φ−,(y,z2 )
s− (z1 − y) s− (z1 − y)
ξ−,z
Hz
y=1
φ+,z
Vz
=
,
(2.1.10)
φ−,z
Vz
e−ik1 z1
where s± (z1 ) := L1 k1 ∈DL 1+t
±ik1 . By the Poisson summation formula,
1e
π
dk1 e−ik1 z1
.
(−1)n s∞,± (z1 + nL), with s∞,± (z) =
s± (z1 ) =
±ik1
−π 2π 1 + t1 e
n∈Z
(2.1.11)
It is straightforward to check, via a complex shift of the path of integration over k1 , that s∞,± (and, therefore, s± ) decays exponentially in z1 ; more
precisely, |s∞,± (z1 )| ≤ e−α|z1 | (1 − t1 eα )−1 for any α ∈ [0, − log t1 ), and
|s± (z1 ) − s∞,± (z1 )| ≤ e−αL (1 − t1 eα )−1 whenever |z1 | ≤ L/2.
In terms of the new variables, the Grassmann action reads St1 ,t2 (Φ) =
Sm (ξ) + Sc (φ) (the labels ‘m’ and ‘c’ stand for ‘massive’ and ‘critical’, for
reasons that will become clear soon), where
Sm (ξ) =
M
1
(1 + t1 e−ik1 )ξ+,z2 (−k1 )ξ−,z2 (k1 ),
L
z =1
k1 ∈DL
(2.1.12)
2
M
1
− b(k1 )φ+,z2 (−k1 )φ−,z2 (k1 )+t2 φ+,z2 (−k1 )φ−,z2 +1 (k1 )
Sc (φ) =
L
k1 ∈DL z2 =1
i
i
− ∆(k1 )φ+,z2 (−k1 )φ+,z2 (k1 ) + ∆(k1 )φ−,z2 (−k1 )φ−,z2 (k1 ) ,
2
2
(2.1.13)
with
∆(k1 ) :=
2t1 sin k1
,
|1 + t1 eik1 |2
b(k1 ) :=
1 − t21
,
|1 + t1 eik1 |2
(2.1.14)
where, as a consequence of the convention used above for Vz , the term in Sc (φ)
involving φM +1,− (k1 ) should be interpreted as being equal to zero. Since Sm
and Sc involve independent sets of Grassmann variables, the Gaussian integral
Vol. 23 (2022) Non-integrable Ising Models in Cylindrical Geometry
1073
appearing in the partition function factors into a product of two integrals, and
the propagators associated with the two terms can be calculated separately.
The ‘massive’ propagator The calculations for Sm are trivial. Let the antisymmetric matrix Am be defined by Sm (ξ) = 21 (ξ, Am ξ). Recall that Sm was
defined in (2.1.12), from which
PfAm =
(1 + t1 eik1 )M ,
k1 ∈DL
and the propagator is given by the appropriate entry of A−1
m , which in the
form used in (2.1.12) is block-diagonal with 2 × 2 blocks such that
ω
ξω,z2 (k1 )ξω′ ,z2′ (k1′ ) = Lδz2 ,z2′ δω,−ω′ δk1 ,−k1′
.
1 + t1 eiωk1
Therefore, going back to x-space,
ξω,z ξω′ ,z′ = ω δω,−ω′ sω (z1 − z1′ ) δz2 ,z2′ ,
(2.1.15)
where sω (z1 ) was defined right after (2.1.10). For later reference, the matrix
formed by the elements in (2.1.15) will be referred to as the massive propagator
and denoted by
0
s+ (z1 − z1′ )
′
.
(2.1.16)
gm (z, z ) = δz2 ,z2′
−s− (z1 − z1′ )
0
Recalling that s± decays exponentially, see comments after (2.1.11), we see
that gm (z, z ′ ) decays exponentially as well, and so corresponds to a massive
field in the language of quantum field theory.
The ‘critical’ propagator The antisymmetric matrix Ac defined by Sc (φ) =
1
2 (φ, Ac φ) can be placed into an explicit block-diagonal form by an ansatz
resembling a Fourier sine transform with shifted frequencies; this involves a
lengthy but elementary calculation which is detailed in Appendix A. Here we
simply state the result for the critical case
1 − t1
1 − t2
t 1 t2 + t 1 + t 2 = 1 ⇔ t2 =
⇔ t1 =
,
(2.1.17)
1 + t1
1 + t2
which is the only case of relevance for the present work4 . In this case we obtain
g (z, z ′ ) g+− (z, z ′ )
φ+,z φ+,z′ φ+,z φ−,z′
= gc (z, z ′ ) (2.1.18)
= ++
g−+ (z, z ′ ) g−− (z, z ′ )
φ−,z φ+,z′ φ−,z φ−,z′
where
1
L k ∈D
′
1
e−ik1 (z1 −z1 )
2NM (k1 , k2 )
1
L k2 ∈QM (k1 )
−ik2 (z2 +z2′ ) ĝ++ (k1 , k2 )
−ik2 (z2 −z2′ )
× e
ĝ(k1 , k2 ) − e
ĝ−+ (k1 , k2 )
gc (z, z ′ ) :=
4 The
ĝ+− (k1 , −k2 )
e2ik2 (M +1) ĝ−− (k1 , k2 )
(2.1.19)
case t2 < (1 − t1 )/(1 + t1 ), corresponding to the paramagnetic phase can be considered
in the same way, while the case t2 > (1 − t1 )/(1 + t1 ) is somewhat more complicated since
one of the frequencies may be complex [23].
1074
G. Antinucci et al.
with
Ann. Henri Poincaré
ĝ++ (k1 , k2 ) ĝ+− (k1 , k2 )
ĝ−+ (k1 , k2 ) ĝ−− (k1 , k2 )
1
−(1 − t21 )(1 − B(k1 )e−ik2 )
−2it1 sin k1
:=
+2it1 sin k1
D(k1 , k2 ) (1 − t21 )(1 − B(k1 )eik2 )
ĝ(k1 , k2 ) :=
(2.1.20)
where
D(k1 , k2 ) := 2(1 − t2 )2 (1 − cos k1 ) + 2(1 − t1 )2 (1 − cos k2 ),
(2.1.21)
ik1 2
B(k1 ) := t2
|1 + t1 e |
,
1 − t21
(2.1.22)
QM (k1 ) is the set of solutions of the following equation, thought of as an
equation for k2 at k1 fixed, in the interval (−π, π):
sin k2 (M + 1) = B(k1 ) sin k2 M,
(2.1.23)
and
NM (k1 , k2 ) =
d
dk2
(B(k1 ) sin k2 M − sin k2 (M + 1))
B(k1 ) cos k2 M − cos k2 (M + 1)
.
(2.1.24)
Remark 2.1. From the above formula it is immediately clear that
ĝ++ (k1 , k2 ) = ĝ++ (k1 , −k2 ) = −ĝ++ (−k1 , k2 ) = ĝ−− (−k1 , k2 ) (2.1.25)
and
ĝ+− (k1 , k2 ) = ĝ+− (−k1 , k2 ) = −ĝ−+ (k1 , −k2 );
(2.1.26)
furthermore, Eq. (2.1.23) is equivalent to
ĝ+− (k1 , k2 ) = −e−2ik2 (M +1) ĝ−+ (k1 , k2 ),
(2.1.27)
which therefore holds for all q ∈ QM (k1 ). Moreover, NM (k1 , k2 ) = NM (−k1 , k2 )
= NM (k1 , −k2 ). As we will see in Sect. 2.4 below, these relationships are closely
related to the symmetries of the Ising model on a cylindrical lattice.
Remark 2.2. The definition (2.1.19) can be extended to all z, z ′ ∈ R2 ; then in
particular, using the relationships listed in the previous remark, we have
g++ ((z1 , 0) , z ′ ) = g++ (z, (z1′ , 0)) = g+− ((z1 , 0) , z ′ ) = g−+ (z, (z1′ , 0)) = 0,
(2.1.28)
and
g+− (z, (z1′ , M + 1)) = g−+ ((z1 , M + 1) , z ′ ) = g−− ((z1 , M + 1) , z ′ )
= g−− (z, (z1′ , M + 1)) = 0,
(2.1.29)
for all z, z ′ .
Vol. 23 (2022) Non-integrable Ising Models in Cylindrical Geometry
1075
2.2. The Critical Propagator: Multiscale Decomposition and Decay Bounds
In this section, we decompose gc (z, z ′ ) into a sum of terms satisfying bounds
which are the main inputs of the multiscale expansion.
2.2.1. Multiscale and Bulk/Edge Decompositions. Let
fη (k1 , k2 ) := D(k1 , k2 )e−ηD(k1 ,k2 ) ,
(2.2.1)
with D(k1 , k2 ) as in (2.1.21). Note that
∞
fη (k1 , k2 ) dη = 1
(2.2.2)
0
as long as k1 and k2 are not both integer multiples of 2π (and so whenever k2 ∈ QM (k1 )). Comparing with Eq. (2.1.20), we see immediately that
fη (k1 , k2 )ĝ(k1 , k2 ) is an entire function of both k1 and k2 . The reader may find
it helpful in what follows to bear in mind that, for large η, fη (k1 , k2 ) is peaked
in a region where D(k1 , k2 ) is of the order η −1 , and so k1 , k2 are of order η −1/2 .
Thanks to (2.2.2), fη induces the following multiscale decomposition of gc
defined in (2.1.19). Let h∗ := −⌊ log2 (min{L, M })⌋; then, for any h∗ ≤ h < 0,
0
g(j) (z, z ′ ),
(2.2.3)
g[η] (z, z ′ ) dη,
(2.2.4)
g[η] (z, z ′ ) dη for h∗ < h < 0,
∞
g[η] (z, z ′ ) dη,
g(≤h) (z, z ′ ) :=
(2.2.5)
gc (z, z ′ ) = g(≤h) (z, z ′ ) +
j=h+1
where
(0)
g
g(h) (z, z ′ ) :=
′
(z, z ) :=
1
0
2−2h
2−2h−2
(2.2.6)
2−2h−2
and
[η]
g
′
(z, z ) :=
1
L
1
−ik1 (z1 −z1′ )
e
fη (k1 , k2 )
2NM (k1 , k2 )
ĝ+− (k1 , −k2 )
−ik2 (z2 −z2′ )
−ik2 (z2 +z2′ ) ĝ++ (k1 , k2 )
× e
ĝ(k1 , k2 ) − e
.
2ik2 (M +1)
ĝ−+ (k1 , k2 ) e
ĝ−− (k1 , k2 )
k1 ∈DL k2 ∈QM (k1 )
(2.2.7)
Note that the single-scale propagators preserve the cancellations at the boundary spelled out in Remark 2.2 above, namely, denoting the components of g(h)
(h)
by gωω′ , with ω, ω ′ ∈ {±}, in analogy with (2.1.18),
(h)
(h)
(h)
(h)
g++ ((z1 , 0) , z ′ ) = g++ (z, (z1′ , 0)) = g+− ((z1 , 0) , z ′ ) = g−+ (z, (z1′ , 0)) = 0,
(h)
(h)
(h)
g+− (z, (z1′ , M + 1)) = g−+ ((z1 , M + 1) , z ′ ) = g−− ((z1 , M + 1) , z ′ )
(h)
= g−− (z, (z1′ , M + 1)) = 0,
(2.2.8)
1076
G. Antinucci et al.
Ann. Henri Poincaré
(≤h)
and analogously for gωω′ . Note also that, taking L, M → ∞, the cutoff prop[η]
agator gc (z, z ′ ) tends to its infinite-plane counterpart, provided that z, z ′ are
chosen ‘well inside the cylinder’; in particular, if zL,M := (L/2, ⌊M/2⌋), then
lim
g[η] (zL,M + z, zL,M + z ′ )
dk1 dk2 −i[k1 (z1 −z1′ )+k2 (z2 −z2′ )]
e
fη (k1 , k2 )ĝ(k1 , k2 )
(2π)2
L,M →∞
=
(2.2.9)
[−π,π]2
[η]
(z − z ′ ).
=: g∞
For later purposes, we need to decompose the cutoff propagator g[η] into a
‘bulk’ part which is minimally sensitive to the size and shape of the cylinder,
plus a remainder which we call the ‘edge’ part. The bulk part is simply chosen
[η]
to be the restriction of g∞ to the cylinder, with the appropriate (anti-periodic)
boundary conditions in the horizontal direction:
[η]
[η]
gB (z, z ′ ) := sL (z1 − z1′ ) g∞
(perL (z1 − z1′ ) , z2 − z2′ )
∈ {1, . . . , L},
⎧
′
⎪
⎨+1, |z1 − z1 | < L/2
sL (z1 − z1′ ) := 0,
|z1 − z1′ | = L/2
⎪
⎩
−1, |z1 − z1′ | > L/2
where, recalling that
and
(2.2.10)
z1 , z1′
z1
1
+
perL (z1 ) := z1 − L
.
L
2
(2.2.11)
(2.2.12)
The edge part is, by definition, the difference between the full cutoff propagator
and its bulk part:
[η]
[η]
gE (z, z ′ ) := g[η] (z, z ′ ) − gB (z, z ′ ).
(h)
(2.2.13)
(h)
Using these expressions, we define gB and gE via the analogues of (2.2.4)–
(2.2.5), with the subscript c replaced by B and E, respectively. As a consequence, for any h∗ ≤ h < 0,
gc (z, z ′ ) = g(≤h) (z, z ′ ) +
0
(j)
(j)
(gB (z, z ′ ) + gE (z, z ′ )).
(2.2.14)
j=h+1
As already observed in Remark 2.2, all the functions involved in this identity
can be naturally extended to all x, y ∈ R2 (and, therefore, in particular, to
all z, z ′ ∈ Z2 ), by interpreting the right side of (2.2.7), etc., as a function on
R2 × R2 .
2.2.2. Decay Bounds and Gram Decomposition: Statement of the Main
Results. Given the multiscale and bulk–edge decomposition (2.2.14), we now
intend to prove suitable decay bounds for the single-scale bulk and edge propagators, as well as to show the existence of an inner product representation
(‘Gram representation’) thereof. These will be of crucial importance in the
non-perturbative multiscale bounds on the partition and generating functions,
Vol. 23 (2022) Non-integrable Ising Models in Cylindrical Geometry
1077
discussed in the rest of this work, and they are summarized in the following
proposition.
2
Given a function f : Z2 → C (or a function f : Λ2 → C extend2 2
able to (Z ) , in the sense explained after (2.2.14)), we let ∂1,j be the discrete derivative in direction j with respect to the first argument, defined by
∂1,j f (z, z ′ ) := f (z + êj , z ′ ) − f (z, z ′ ), with êj the j-th Euclidean basis vector;
an analogous definition holds for ∂2,j .
Proposition 2.3. There exist constants c, C such that, for any integer h in
[h∗ + 1, 0], any r = (r1,1 , r1,2 , r2,1 , r2,2 ) ∈ Z4+ , and any z, z ′ ∈ Λ2 ,
1.
∂ r g(h) (z, z ′ ) ≤ C 1+|r |1 r!2(1+|r |1 )h e−c2
h
z−z ′ 1
(2.2.15)
r (h)
where the matrix norm in the left side (recall that ∂ g (x, y) is a 2 × 2
matrix) is the max norm, i.e., the maximum over the matrix elements,
2
2
r
∂ r := i,j=1 ∂i,ji,j , r! = i,j=1 ri,j !, and z1 = |perL (z1 )| + |z2 |, see
(2.2.12),
Moreover, if z, z ′ ∈ Λ are such that |perL (z1 − z1′ )| < L/2 − |r1,1 | − |r2,1 |,
2.
(h)
∂ r gE (z, z ′ ) ≤ C 1+|r |1 r!2(1+|r |1 )h e−c2
h
dE (z,z ′ )
(2.2.16)
where dE (z, z ′ ) := min{|perL (z1 − z1′ )| + min{z2 + z2′ , 2(M + 1) − z2 −
z2′ }, L − |perL (z1 − z1′ )| + |z2 − z2′ |}.
Finally, there exists a Hilbert space HLM with inner product (·, ·) including
(h)
(h)
(≤h)
(≤h)
elements γω,s,z , γ̃ω,s,z , γω,s,z , γ̃ω,s,z (for s = (s1 , s2 ) ∈ Z2+ , x ∈ Λ) such that
whenever h∗ ≤ h ≤ 0,
′
(h)
(h)
(h)
3. ∂ (s,s ) gωω′ (z, z ′ ) ≡ γ̃ω,s,z , γω′ ,s ′ ,z′ and
′
(≤h)
(≤h)
(≤h)
∂ (s,s ) gωω′ (z, z ′ ) ≡ γ̃ω,s,z , γω′ ,s ′ ,z′ , and
(h) 2 (h) 2 (≤h) 2 (≤h) 2
4. γω,s,z , γ̃ω,s,z , γω,s,z , γ̃ω,s,z ≤ C 1+|s|1 s!2h(1+2|s|1 ) where | · |
is the norm generated by the inner product (·, ·).
Combining points 3 and 4, we see that
Corollary 2.4. For all z, z ′ ∈ Λ, r ∈ Z4+ , and h∗ ≤ h ≤ 0,
∂ r g(≤h) (z, z ′ ) ≤ C 1+|r |1 r!2(1+|r |1 )h .
(h)
(h)
(2.2.17)
Remark 2.5. Since gB = g(h) − gE , from items 1 and 2 it follows that the
(h)
bulk, single-scale, propagator gB satisfies the same estimate as (2.2.15) for all
x, y ∈ Λ allowed in Item 2, i.e., whenever |perL (z1 − z1′ )| < L/2 − |r1,1 | − |r2,1 |.
The latter restriction on z, z ′ just comes from the requirement that the discrete
derivatives do not act on the discontinuous functions sL and perL entering the
(h)
(h)
definition of gB (and, therefore, of gE ); in fact, one can easily check from
the proof that, if r = 0, then (2.2.16) is valid for all z, z ′ ∈ Λ, without further
restrictions.
1078
G. Antinucci et al.
Ann. Henri Poincaré
Remark 2.6. All the estimates stated in the proposition are uniform in L, M ,
therefore, they remain valid for the L, M → ∞ limit of the propagators. In
(h)
(≤h)
particular, g∞ (z − z ′ ) and g∞ (z − z ′ ) satisfy the same estimates as (2.2.15)
and (2.2.17), respectively. Similarly, the Gram representation stated in items
(h)
3 and 4 is also valid for g∞ . Therefore, if |z1 − z1′ | ≤ L/2 − |s1 | − |s′1 |, also
′
′
′
(h)
(h)
∂ (s,s ) gE (x, y) = ∂ (s,s ) g(h) (z, z ′ ) − ∂ (s,s ) g∞ (z − z ′ ) admits a Gram representation, with qualitatively the same Gram bounds (and, of course, such
a representation can be extended by anti-periodicity to all z, z ′ such that
|perL (z1 − z1′ )| ≤ L/2 − |s1 | − |s′1 |).
Remark 2.7. If we rename the massive propagator in (2.1.16) as gm (z, z ′ ) =:
(1)
(1)
(1)
g(1) (z, z ′ ) and then use it to define g∞ , gB , and gE in the same way that we
did above with g[η] , it is straightforward to check that the estimates in items
1 and 2 of the proposition remain valid for h = 1. Similarly, the reader can
check that the proof of items 3 and 4 given below can be straightforwardly
applied to the case h = 1, as well.
Remark 2.8. Again since D(k1 , k2 ) is exactly the denominator in the definition (2.1.20), it is easily seen that fη (k1 , k2 )ĝ(k1 , k2 ) is an entire function of
t1 . All of the bounds in Proposition 2.3 are obtained by writing the relevant
quantities as absolutely convergent integrals or sums in k1 , k2 , and η; since
these bounds are locally uniform in t1 as long as it is bounded away from 0
and 1, we also see that all of the propagators are analytic functions of t1 with
all other arguments held fixed.
The proof of Proposition 2.3 is in Appendix B.
2.3. Asymptotic Behavior of the Critical Propagator
Although our main result, Theorem 1.1, involves correlation functions for a
finite lattice, since the continuum limit of the energy correlation functions of
the non-interacting model is well understood [11,24], as a result we also obtain
a characterization of the scaling limit [4, Corollary 1.2]. For completeness, we
give here a description of the scaling limit of the critical propagator, from
which the non-interacting energy correlation functions are easily calculated.
We rescale the lattice as follows: fix two positive constants ℓ1 , ℓ2 (no
condition on the ratio ℓ1 /ℓ2 ), and let L = 2⌊a−1 ℓ1 /2⌋, M = ⌊a−1 ℓ2 ⌋ for a > 0
the lattice mesh. Let Λa := aΛ and Λℓ1 ,ℓ2 the continuum cylinder. We also let
· indicate the Euclidean distance on the cylinder Λℓ1 ,ℓ2 . Given z, z ′ ∈ Λℓ1 ,ℓ2 ,
we let
gc,a (z, z ′ ) = a−1 gc (⌊a−1 z⌋, ⌊a−1 z ′ ⌋).
(2.3.1)
The main result of this section concerns the limiting behavior of gc,a as a → 0.
Proposition 2.9. Given ℓ1 , ℓ2 > 0, there exist C, c > 0 such that for all x, y ∈
Λℓ1 ,ℓ2 such that z = z ′ and a > 0 for which a(min{ℓ1 , ℓ2 , z − z ′ })−1 ≤ c,
gc,a (z, z ′ ) = gscal (z, z ′ ) + Ra (z, z ′ ),
(2.3.2)
Vol. 23 (2022) Non-integrable Ising Models in Cylindrical Geometry
and Ra (z, z ′ ) ≤ Ca min{ℓ1 , ℓ2 , z − z ′ }
′
gscal (z, z ) =
(−1)
n
scal
g∞ (z1
−
′
z1
−2
+ n1 ℓ1 , z2 −
1079
, where
′
z2
+ 2n2 ℓ2 )
n ∈Z2
+
g2scal (z1 − z1′ + n1 ℓ1 , z2 + z2′ + 2n2 ℓ2 )
−g1scal (z1 − z1′ + n1 ℓ1 , z2 + z2′ + 2n2 ℓ2 )
−g2scal (z1 − z1′ + n1 ℓ1 , z2 + z2′ + 2n2 ℓ2 ) g1scal (z1 − z1′ + n1 ℓ1 , z2 + z2′ + 2(n2 − 1)ℓ2 )
!
,
(2.3.3)
and where, letting
g
scal
dk1 dk2 −ik1 z1 −ik2 z2 −ik1
1
e
(z1 , z2 ) :=
t2 (1 − t2 )
(2π)2
k12 + k22
2
R
z1
1
= −
,
2πt2 (1 − t2 ) z12 + z22
(2.3.4)
z2
z1
we denoted g1scal (z1 , z2 ) := g scal ( 1−t
, z2 ), g2scal (z1 , z2 ) := g scal ( 1−t
, z1 ),
2 1−t1
1 1−t2
and
scal
g1 (z1 , z2 )
g2scal (z1 , z2 )
scal
g∞ (z1 , z2 ) := scal
.
(2.3.5)
g2 (z1 , z2 ) −g1scal (z1 , z2 )
The proof of Proposition 2.9 is given in Appendix C. It is easy to see
from the definition of gscal
that its entries vanish for z2 = 0 and/or z2 = ℓ2
c
and/or z2′ = 0 and/or z2′ = ℓ2 in a fashion analogous to the one discussed in
Remark 2.2.
2.4. Symmetries of the Propagator
Note that the action St1 ,t2 (Φ) of Eq. (2.1.2) is unchanged by the substitutions
H z → iHθ1 z ,
Hz → iH θ1 z ,
with θ1 (z1 , z2 ) := (L + 1 − z1 , z2 ), or
H z → −iH θ2 z ,
Hz → iHθ2 z ,
V z → iV θ1 z ,
V z → iVθ2 z ,
Vz → −iVθ1 z
(2.4.1)
Vz → iV θ2 z
(2.4.2)
where θ2 (z1 , z2 ) := (z1 , M + 1 − z2 ). These transformations, of course, correspond to the reflection symmetries of the Ising model on a cylinder. In terms
of the φ, ξ variables, it is easy to see from Eq. (2.1.10) that these substitutions
are equivalent to
φ±,z → Θ1 φ±,z := ±iφ±,θ1 z ,
ξ±,z → Θ1 ξ±,z := iξ∓,θ1 z
(2.4.3)
ξ±,z → Θ2 ξ±,z := ∓iξ±,θ2 z .
(2.4.4)
and
φ±,z → Θ2 φ±,z := iφ∓,θ2 z ,
With a little more notation we can write this more compactly: letting φω,z
denote φ+,x , φ−,x , ξ+,x , ξ−,x for ω = 1, −1, +i, −i, respectively, and letting
θj ω := (−1)j+1 ω for ω ∈ C, Eqs. (2.4.3) and (2.4.4) can be combined into
Θj φω,z := iαj,ω φθj ω,θj x
(2.4.5)
where αj,ω is −1 if (j = 1 and ω = −1) or (j = 2 and ω = i), and 1 otherwise. Since these transformations act on the vector Ψ as orthogonal matrices,
this is equivalent to the symmetry of the coefficient matrix A (and therefore
1080
G. Antinucci et al.
Ann. Henri Poincaré
its inverse) under the associated similarity transform, and since gc is just a
diagonal block of A−1 we have
−gc;++ gc;+−
gc (z, z ′ ) =
(2.4.6)
(θ1 z, θ1 z ′ )
gc;−+ −gc;−−
and
gc;−− gc;−+
gc (z, z ) = −
(θ2 z, θ2 z ′ ) .
gc;+− gc;++
′
(2.4.7)
These relationships can also be recovered from Eq. (2.1.19), using the observations on ĝ∞ in Remark 2.1. This latter point is helpful because, since fη is even
[η]
in both k1 and k2 , it also applies to gc . Taking the appropriate L, M → ∞
limit we also obtain
#
#
" [η]
" [η]
[η]
[η]
g∞;−− g∞;−+
−g∞;++ g∞;+−
[η]
(z1 , −z2 ) ,
(−z1 , z2 ) = − [η]
g∞ (z1 , z2 ) =
[η]
[η]
[η]
g∞;+− g∞;++
g∞;−+ −g∞;−−
(2.4.8)
[η]
gB
has the symmetries (2.4.6) and (2.4.7). Applying
which also implies that
the differences and integrals in the relevant definitions we see that
(h)
(≤h)
(h)
(h)
Lemma 2.10. gc , gc , gB , and gE all have the symmetries (2.4.6)
(h)
and (2.4.7) for any h∗ ≤ h ≤ 0, and g∞ has the symmetries (2.4.8) for
any h ≤ 0.
(1)
For gc ≡ gm we have
gm;−− gm;−+
−gm;++ gm;+−
′
′
(θ1 z, θ1 z ) =
(θ2 z, θ2 z ′ )
gm (z, z ) = −
gm;+− gm;++
gm;−+ −gm;−−
(2.4.9)
(1)
(1)
(1)
which similarly extends to g∞ , gB , and gE .
3. Grassmann Representation of the Generating Function
In this section we rewrite the generating function of the energy correlations for
the Ising model (1.1) with finite range interactions as an interacting Grassmann
integral, and we set the stage for the multiscale integration thereof, to be
discussed in the following sections. The estimates in this and in the following
section are uniform for J1 /J2 , L/M ∈ K and t1 , t2 ∈ K ′ , but may depend
upon the choice of K, K ′ , with K, K ′ the compact sets introduced before the
statement of Theorem 1.1. As anticipated there, we will think of K, K ′ as being
fixed once and for all and, for simplicity, we will not track the dependence
upon these sets in the constants C, C ′ , . . . , c, c′ , . . . , κ, κ′ , . . ., appearing below.
Unless otherwise stated, the values of these constants may change from line to
line.
Our goal is to show that the generating function (1.6) of the energy
correlations can be replaced, for the purpose of computing multipoint energy
Vol. 23 (2022) Non-integrable Ising Models in Cylindrical Geometry
1081
correlations at distinct edges, by the Grassmann generating function (1.7).
Our main representation result for the Taylor coefficients at A = 0 of the
logarithm of ZΛ (A), analogous to [18, Proposition 1], is the following.
Proposition 3.1. For any translation invariant interaction V of finite range,
there exists λ0 = λ0 (V ) such that, for any |λ| ≤ λ0 (V ),
∂
∂
∂
∂
$
...
log ZΛ (A)
...
log ΞΛ (A)
=
∂Ax
∂Ax
∂Ax
∂Ax
1
A =0
n
1
A =0
n
as long as n ≥ 2 and the xj are distinct, where
$ Λ (A) := eW(A ) DΦ eSt1 ,t2 (Φ)+V(Φ,A )
Ξ
(3.1)
where St1 ,t2 was defined in (2.1.2) and, recalling that Ex is the Grassmann
binomial defined after (2.1.2):
1.
V(Φ, A) =
(1 − t2j(x) )Ex Ax +
WΛint (X, Y )
Ex
Ax
X,Y ⊂BΛ
X =∅
x∈BΛ
x∈X
x∈Y
(3.2)
≡ Bfree (Φ, A) + V int (Φ, A)
where, for any n ∈ N, m ∈ N0 , and suitable positive constants C, c, κ,
|WΛint (X, Y )|ecδ(X∪Y ) ≤ C m+n |λ|max(1,κ(m+n))
sup
x0 ∈BΛ
X,Y ⊂BΛ : x0 ∋X
|X|=n, |Y |=m
(3.3)
and δ(X), for X ⊂ BΛ , denotes the size of the smallest Z ⊃ X which is
the edge set of a connected subgraph of GΛ .
2.
W(A) =
Y ⊂BΛ
|Y |≥2
wΛ (Y )
Ax
where, for any m ∈ N, and the same C, c, κ as above,
|wΛ (Y )|ecδ(Y ) ≤ C m |λ|max(1,κm) .
sup
x0 ∈BΛ
(3.4)
x∈Y
(3.5)
Y ⊂BΛ :
x0 ∋Y, |Y |=m
3. WΛint , wΛ , considered as functions of λ, t1 , and t2 , can be analytically continued to any complex λ, t1 , t2 such that |λ| ≤ λ0 and |t1 |, |t2 | ∈ K ′ , with
K ′ the same compact set introduced before the statement of Theorem 1.1,
and the analytic continuations satisfies the same bounds above.
Proof. The proof is basically the same as [18, Proposition 1], so we refer to
that for the details. Note that the restriction in [18] to a pair interaction is
unimportant, since any even interaction of the form V (X)σX with X ⊂ Λ can
always be written as a product of factors of ǫx , in analogy with the rewriting
σz σz′ = 21 (Uz,z′ + Dz,z′ ) discussed after [18, Equation (2.8)]; the only difference
1082
G. Antinucci et al.
Ann. Henri Poincaré
in the current setting is that the ‘strings’ graphically associated with Uz,z′ and
Dz,z′ , see [18, Figure 3], are replaced by other figures, whose specific shape
depends on V (X) and that one should use t1 or t2 in place of t as appropriate5 .
Note that the set of strings associated with a pair interaction, or the set of more
general figures associated with a generic even interaction, is, or can be chosen
to be, invariant under the basic symmetries of the model, namely horizontal
translations, and horizontal and vertical reflections; therefore, in the following,
we shall assume that such a graphical representation is invariant under these
symmetries.
By proceeding as in [18] we get the analogue of [18, Eq.(2.20)], namely
WΛ (X, Y )
Ex
ϕT (Γ)
Ax
ζG (γ) =
W(A) + V(Φ, A) =
Γ∈CΛ
γ∈Γ
X,Y ⊂BΛ
x∈X
x∈Y
(3.6)
where CΛ is the set of multipolygons in Λ, ϕT is the Mayer’s coefficient, and
ζG is the activity of the polygon γ, which is a polynomial in the Ex , Ax for
edges x in γ (for more details about the notation and more precise definitions,
we refer to [18]). The terms with X = ∅ contribute to W(A) (that is, we let
wΛ (Y ) := WΛ (∅, Y )), while those with X = ∅ contribute to V(Φ, A) (note
$ Λ (A) of order 2 or
that, for the purpose of computing the derivatives of log Ξ
more, the terms with |X| = 0 and |Y | = 0, 1 can be dropped from the definition
of W(A), and we do so). The explicit computation of the term independent
of λ, which has X = Y ∈ B, leads to the decomposition in Eq. (3.2). The
bounds Eqs. (3.3) and (3.5) follow directly from the bounds in [18], see, e.g.,
[18, Eq.(2.25)] and following discussion.
Finally, the analyticity property (which was used implicitly in [18]) follows by noting that we have defined all of the quantities of interest as uniformly
absolutely convergent sums of terms which are themselves analytic functions
of λ, t1 , t2 as long as the absolute values of these parameters belong to the
appropriate intervals.
With a view toward the analysis of finite size effects in [4] (and, in particular, toward the claims done in [4, Section 2.2] after the statement of [4,
Proposition 2.5]), it is convenient to decompose the kernel WΛint of V int (Φ, A)
into a ‘bulk’ plus an ‘edge’ part. This requires a bit of notation. Note that
any subset X of Λ with horizontal diameter smaller than L/2 can be identified (non-uniquely, of course) with a subset of Z2 with the same diameter and
‘shape’ as X; we call X∞ ⊂ Z2 one of these arbitrarily chosen representatives6
5 We
take the occasion to point out that [18, Figure 4] contains a mistake: the string S2
depicted there is not allowed by the conventions of [18]: the shape S2 can only be obtained
as the union of two appropriate strings.
6 For instance, given X = {z , . . . , z }, recalling that (z ) ∈ {1, 2, . . . , L} and (z ) ∈
n
1
i 1
i 2
{1, 2, . . . , M }, we can let X∞ = {y1 , . . . , yn } be the set of points such that: (1) the vertical
coordinates are the same as those of z , i.e., (yi )2 = (zi )2 , ∀i = 1, . . . , n; (2) the horizontal
coordinate of y1 is the same as z1 , i.e., (y1 )1 = (z1 )1 ; (3) all the other horizontal coordinates
are the same modulo L, i.e., (yi )1 = (zi )1 mod L, ∀i = 2, . . . , n; (4) the specific values of (yi )2
Vol. 23 (2022) Non-integrable Ising Models in Cylindrical Geometry
1083
of X, and we shall use an analogous convention for the subsets of BΛ with
horizontal diameter smaller than L/2.
Lemma 3.2. Under the same assumptions of Proposition 3.1, the kernel WΛint
of V int (Φ, A) can be decomposed as
int
WΛint (X, Y ) = W∞
(X∞ , Y∞ ) 1(diam1 (X ∪ Y ) ≤ L/3) + WEint (X, Y )
=: WBint (X, Y ) + WEint (X, Y ),
(3.7)
where diam1 is the horizontal diameter on the cylinder Λ; X∞ , Y∞ ⊂ B :=
BZ2 are two representatives of X, Y , respectively, such that X∞ ∪ Y∞ is a
representative of X ∪Y , in the sense defined before the statement of the lemma;
int
is a function, independent of L, M , invariant under translations and
W∞
under reflections about either coordinate axis, which satisfies the same weighted
L1 bound (3.3) as WΛint . Moreover, for any n ∈ N and m ∈ N0 , WEint satisfies
1
|WEint (X, Y )|ecδE (X∪Y ) ≤ C m+n |λ|max(1,κ(m+n)) ,
(3.8)
L
X,Y ⊂BΛ
|X|=n, |Y |=m
with the same C, c, κ as in Proposition 3.1, where δE (X) is the cardinality of
the smallest connected subset of BΛ which includes X and either touches the
boundary of the cylinder7 , or its horizontal diameter is larger than L/3.
Proof. In order to obtain the decomposition (3.7), let
int
W∞
(X∞ , Y∞ ) :=
lim
L,M →∞
WΛint (X∞ + zL,M , Y∞ + zL,M ),
(3.9)
where zL,M = (L/2, ⌊M/2⌋) and X∞ + zL,M is the translate of X∞ by zL,M ;
note that this limit is well defined thanks to the fact that WΛint can be expressed
in terms of a sum like [18, Eq.(2.21)], which is exponentially convergent, see
int
satisfies the analogue of (3.6), that is
[18, Eq.(2.25)]. The kernel W∞
Γ∈C∞
T
⎛
ϕ (Γ) ⎝
γ∈Γ
ζG (γ) −
γ∈Γ
⎞
ζG (γ)λ=0 ⎠ =
X,Y ⊂B
int
W∞
(X, Y )
x∈X
Ex
Ax ,
x∈Y
(3.10)
2
with C∞ the set of multipolygons on Z , and the activity ζG (γ) the same as
the one in (3.6), provided that γ is considered now as a polygon in Z2 , rather
than in Λ (note that such identification is possible as long as γ does not wrap
int
around the cylinder). Moreover, W∞
is translation invariant, and, letting
int
WEint (X, Y ) := WΛint (X, Y ) − 1(diam1 (X ∪ Y ) ≤ L/3)W∞
(X∞ , Y∞ ),
(3.11)
the contribution to the first term in the right side of all multipolygons in CΛ
with horizontal diameter ≤ L/3 cancels with their counterparts in C∞ from
for i ≥ 2 are chosen in such a way that the horizontal distances between the corresponding
pairs in X and X∞ are the same, if measured on the cylinder Λ or on Z2 , respectively.
7 We say that X ⊂ B ‘touches the boundary of the cylinder Λ’, if at least one of the edges
Λ
in X has an endpoint whose vertical coordinate is either equal to 1 or to M .
1084
G. Antinucci et al.
Ann. Henri Poincaré
the second term in the right side. Each of the remaining multipolygons either
comes from the first term in the right side and involves a multipolygon in
CΛ whose support has horizontal diameter larger than L/3, or comes from the
second term in the right side and involves a multipolygon in C∞ whose support
contains a set Z∞ which is the representative (in the sense explained before
the statement of the lemma) of a connected subset ZΛ of BΛ that contains
X ∪Y and touches the boundary of Λ; in either case the number of edges in the
support of such a multipolygon is at least δE (X ∪ Y ), from which the bound
(3.8) follows.
In the following, we will wish to work in the φ, ξ variables introduces
in Sect. 2.1.2; applying the change of variables (2.1.10), with some abuse of
notation we rewrite V(Φ, A) in (3.2) as:
V(φ, ξ, A) =
BΛfree ((ω, z), x)φ(ω, z)Ax
x∈BΛ (ω ,z )∈O 2 ×Λ2
+
WΛint ((ω, z), x)φ(ω, z)A(x)
n∈2N (ω ,z )∈O n ×Λn x∈Bm
Λ
m∈N0
=: B free (φ, ξ, A) + V int (φ, ξ, A),
(3.12)
where N and N0 are the sets of positive ad nonnegative integers, respectively,
O := {1, −1, i, −i}, and, for ω ∈ On , z ∈ Λ, we denote
φ(ω, z) =
n
φωj ,zj
(3.13)
j=1
m
with φ±i,z = ξ±,z , and similarly A(x) := j=1 Axj (for x = ∅, we interpret
A(∅) = 1). The decay properties of s± noted after Eq. (2.1.11) together with
Eq. (3.3) imply a similar decay property for the new coefficients:
(3.14)
sup sup
ecδ(z ,x) BΛfree ((ω, z), x) ≤ C,
ω ∈O 2 x∈BΛ
z ∈Λ2
and
sup sup
ω ∈O n z1 ∈Λ
sup
z2 ,...,zn ∈Λ
sup
ω ∈O n x1 ∈BΛ
ecδ(z ) WΛint ((ω , z ), ∅) ≤ C n |λ|max(1,κn)
x2 ,...,xm ∈BΛ z ∈Λn
ecδ(z ,x ) WΛint ((ω , z ), x) ≤ C n+m |λ|max(1,κ(n+m)) .
(3.15)
Note that, with this rewriting in terms of the φ, ξ variables, recalling that
St1 ,t2 (Φ) = Sm (ξ) + Sc (φ), see (2.1.12)–(2.1.13), and denoting Pc (Dφ) :=
Dφ eSc (φ) / Pf(Ac ), Pm (Dξ) := Dξ eSm (ξ) / Pf(Am ) (here Ac and Am are the
Vol. 23 (2022) Non-integrable Ising Models in Cylindrical Geometry
1085
two 2|Λ| × 2|Λ| anti-symmetric matrices associated with the Grassmann quadratic forms Sc (φ) and Sm (ξ), respectively), the Grassmann generating func$ Λ (A) in (3.1) can be rewritten as
tion Ξ
$ Λ (A) ∝ eW(A ) Pc (Dφ) Pm (Dξ)eV(φ,ξ,A ) ,
Ξ
(3.16)
where ∝ means ‘up to a multiplicative constant independent of A’. In view of
these rewritings, Proposition 3.1 implies [4, Proposition 2.5] as an immediate
corollary.
Of course, the bulk–edge decomposition of Lemma 3.2 implies an analogous decomposition for the kernel of V int (φ, ξ, A), which reads as follows:
int
((ω, z ∞ ), x∞ )
WΛint ((ω, z), x) = (−1)α(z ) 1(diam1 (z, x) ≤ L/3) W∞
+ WEint ((ω, z), x)
=:
WBint ((ω, z), x)
+
(3.17)
WEint ((ω, z), x),
where, for any z with diam1 (z) ≤ L/3,
α(z) =
#{zi ∈ z : (zi )1 ≤ L/3},
0,
if maxzi ,zj ∈z {(zi )1 − (zj )1 } ≥ 2L/3,
otherwise.
(3.18)
and
int
W∞
((ω, z), x) :=
lim
L,M →∞
WΛint ((ω, z + zL,M ), x + zL,M ).
(3.19)
The factor (−1)α(z ) in front of the first term in the right side of (3.17), in light
int
of the antiperiodicity of the φ, ξ fields, guarantees that W∞
is translation
invariant (in both coordinate directions), and that both WBint and WEint are
invariant under simultaneous translations of z and x in the horizontal direction, with anti-periodic and periodic boundary conditions in z and x, respectively. In terms of this new notation, Eq.(3.8) implies that, for any n ∈ N and
m ∈ N0 ,
1
sup
|WEint ((ω, z), x)|ecδE (z ,x) ≤ C m+n |λ|max(1,κ(m+n)) , (3.20)
L ω ∈On
n
z ∈Λ
x∈Bm
Λ
with δE (z, x) is the ‘edge’ tree distance of (z, x), i.e., the cardinality of the
smallest connected subset of BΛ that includes x, touches the points of z
and either touches the boundary of the cylinder or it has horizontal diameter
larger than L/3. Of course, BΛfree admits a similar bulk–edge decomposition:
free
+ BEfree with
in analogy with (3.7), BΛfree = BB
1
sup
|BEfree (ω, z, x)|ecδE (z ,x) ≤ C.
(3.21)
L ω ∈O2
2
z ∈Λ
x∈BΛ
Before concluding this section, let us comment on the connection between
(3.16) and (1.7). Fix once and for all a neighborhood U ⊂ R of 1 not containing
0; say, for definiteness, U := {z ∈ R : |z − 1| ≤ 1/2}. For any Z ∈ U and
1086
G. Antinucci et al.
Ann. Henri Poincaré
∗
t∗1 ∈ K ′ , we let t∗2 := (1−t∗1 )/(1+t∗1 ) and let Sc∗ (φ) = 12 (φ, A∗c φ) (resp. Sm
(ξ) =
1
∗ ∗
∗
ξ))
be
obtained
from
S
(resp.
S
)
by
replacing
t
,
t
with
t
,
t2 in
(ξ,
A
c
m
1
2
m
1
2
∗
Eq. (2.1.13) (resp. Eq. (2.1.12)). We also let Pc∗ (Dφ) := DφeSc (φ) / Pf(A∗c ),
∗
∗
(Dξ) := DξeSm (ξ) / Pf(A∗m ). Given these definitions, in (3.16) we first rescale
Pm
the φ and ξ variables by Z −1/2 , then multiply and divide the Grassmann
∗
∗
integrand by eSc (φ)+Sm (ξ) , thus getting
(1)
W(A )
∗
$
Ξ(A) ∝ e
Pc∗ (Dφ)Pm
(Dξ)eV (φ,ξ,A )
(3.22)
with
∗
V (1) (φ, ξ, A ) := Z −1 Sc (φ) − Sc∗ (φ) + Z −1 Sm (ξ) − Sm
(ξ) + V(Z −1/2 φ, Z −1/2 ξ, A )
=: Nc (φ) + Nm (ξ) + V(Z −1/2 φ, Z −1/2 ξ, A ).
(3.23)
This proves (1.7) and puts us in the position of setting the multiscale computation of the sequence of effective potentials, whose infinite plane counterparts
are constructed and bounded in the next section. For later reference, we note
that, in light of (2.1.12), (2.1.13), (3.12), V (1) (φ, ξ, A) can be written as:
(1)
V (1) (φ, ξ, A) =
WΛ ((ω, z), x)φ(ω, z)A(x),
n∈2N, (ω ,z )∈O n ×Λn x∈Bm
Λ
m∈N0
(3.24)
for an appropriate kernel, which inherits its properties from those of Sc ,
(1)
Sm and V. With no loss of generality, we can assume that WΛ is antisymmetric under simultaneous permutations of ω and z, symmetric under
permutations of x, invariant under simultaneous translations of z and x in
the horizontal direction (with anti-periodic and periodic boundary conditions
in z and x, respectively), invariant under the reflection symmetries induced
by the transformations Ax → Aθl x and φω,z → Θl φω,z , see (2.4.3)–(2.4.4).
From now on, with some abuse of notation, given ω = (ω1 , . . . , ωn ) ∈ On
and z = (z1 , . . . , zn ) ∈ Λn , we shall identify the pair (ω, z) with the n-ple
((ω1 , z1 ), . . . , (ωn , zn )) ∈ (O × Λ)n .
4. The Renormalized Expansion in the Full-Plane Limit
In this section we construct the sequence of effective potentials (see the last
part of Sect. 1) in the infinite volume limit and derive weighted L1 bounds for
their kernels, in a form appropriate for the subsequent generalization to the
finite cylinder, discussed in [4, Section 3]. The construction of this section will
allow us, in particular, to fix the free parameters β, Z, t∗1 , which the Grassmann
integral in the right side of (3.22) depends on, in such a way that the sequence
of running coupling constants goes to zero exponentially fast in the infrared
limit; see Sect. 4.5 below.
As anticipated at the end of Sect. 1, here we limit ourselves to construct
the sequence of effective potentials at A = 0, so, for lightness of notation,
we denote by V (h) (φ) := V (h) (φ, 0) the effective potentials with h ≤ 0 at zero
Vol. 23 (2022) Non-integrable Ising Models in Cylindrical Geometry
1087
external fields (similarly, we let V (1) (φ, ξ) = V (1) (φ, ξ, A)). In light of (1.8) and
(1.10), these effective potentials are iteratively defined via
) ∗
(1)
log Pm
(Dξ) eV (φ,ξ)
if h = 1,
(h−1)
) (h)
V
(φ) = const. +
LV (h) (φ+ϕ)+RV (h) (φ+ϕ)
log P (Dϕ) e
if h ≤ 0,
(4.1)
where the const. is fixed so that V (h) (0) = 0, for all h ≤ 0, and LV (h) + RV (h)
is an equivalent rewriting of V (h) , to be defined (in the full plane limit) below.
Expanding the exponential and the logarithm in the right side of (4.1) allows
us to rewrite
⎞
⎛
1
(4.2)
E∗ ⎝V (1) (φ, ·); · · · ; V (1) (φ, ·)⎠
V (0) (φ) = const. +
+,
s! m *
s≥1
and, for h ≤ 0,
V (h−1) (φ) = const. +
⎛
s times
1
s!
s≥1
⎞
E(h) ⎝LV (h) (φ + ·) + RV (h) (φ + ·); · · · ; LV (h) (φ + ·) + RV (h) (φ + ·)⎠ ,
+,
*
s times
(4.3)
E∗m
(h)
where
(resp. E ) denotes the truncated expectation [15, Eq.(4.13)] with
∗
respect to the Grassmann Gaussian integration Pm
(resp. P (h) ). Expanding the effective potentials in terms of their kernels, in analogy with (3.24),
Eqs. (4.2)–(4.3) allow us to iteratively compute the kernels of V (h) , for all
h ≤ 0. For instance, at the first step, using (3.24), (4.2) and the BBFK formula (for Battle, Brydges, Federbush, Kennedy) for the Grassmann truncated
(1)
(1)
expectations [1,7–9], we find that, denoting by VΛ (Ψ) = WΛ (Ψ, ∅) with
n
Ψ = ((ω1 , z1 ), . . . , (ωn , zn )) ∈ ∪n∈2N (O × Λ) =: M1,Λ the kernel of V (1) (φ, ξ),
(0)
the kernel VΛ of V (0) (φ) satisfies, for any Ψ = ((ω1 , z1 ), . . . , (ωn , zn )) ∈
∪n∈2N ({+, −} × Λ)n ,
(0)
VΛ (Ψ)
∞
1
=
s!
s=1
(Ψ)
(1)
GT (Ψ̄1 , . . . , Ψ̄s )
Ψ1 ,...,Ψs ∈M1,Λ T ∈S(Ψ̄1 ,...,Ψ̄s )
⎛
α(Ψ; Ψ1 , . . . , Ψs ) ⎝
s
j=1
⎞
(1)
VΛ (Ψj )⎠ ,
(4.4)
where
• the symbol (Ψ) on the second sum means that the sum runs over all
ways of representing Ψ as an ordered sum of s (possibly empty) tuples,
Ψ′1 ⊕ · · · ⊕ Ψ′s = Ψ, and over all tuples M1,Λ ∋ Ψj ⊇ Ψ′j (here ⊕ indicates
concatenation of ordered tuples); for each such term in the second sum,
1088
G. Antinucci et al.
Ann. Henri Poincaré
we denote by Ψ̄j := Ψj \ Ψ′j and by α(Ψ; Ψ1 , . . . , Ψs ) the sign of the
permutation from Ψ1 ⊕ · · · ⊕ Ψs to Ψ ⊕ Ψ̄1 ⊕ · · · ⊕ Ψ̄s (here ⊕ indicates
concatenation of ordered tuples);
• S(Ψ̄1 , . . . , Ψ̄s ) denotes the set of all the ‘spanning trees’ on Ψ̄1 , . . . , Ψ̄s ,
that is, of all the sets T of ordered pairs (f, f ′ ), with f ∈ Ψ̄i , f ′ ∈ Ψ̄j
and i < j, whose corresponding graph GT = (V, ET ), with vertex set
V = {1, . . . , s} and edge set ET = {(i, j) ∈ V 2 : ∃(f, f ′ ) ∈ T with f ∈
Qi , f ′ ∈ Qj }, is a tree graph (for s = 1, we let S(Ψ̄1 ) ≡ {∅});
(1)
• GT (Ψ̄1 , . . . , Ψ̄s ) is different from zero only if Ψ̄j ∈ ∪2N0 ({+i, −i} × Λ)n
for all j = 1, . . . , s, and, if s > 1, only if Ψ̄j = ∅ for all j = 1, . . . , s; more
(1)
precisely: if s = 1 and Ψ̄1 = ∅, then G∅ (∅) = 1; if s = 1 and Ψ̄1 = ∅,
(1)
(1)
then G∅ (Ψ̄1 ) = Pf GΨ̄1 , where, given a pair ℓ = ((ω, z), (ω ′ , z ′ )) of
(1)
(1)
(1)
distinct elements of Ψ̄1 , GΨ̄1 ℓ = gℓ := g−iω,−iω′ (z, z ′ )8 ; if s > 1 and
Ψ̄j = ∅ for all j = 1, . . . , s, then
(1)
GT (Ψ̄1 , . . . , Ψ̄s )
:= αT (Ψ̄1 , . . . , Ψ̄s )
"
(1)
gℓ
ℓ∈T
#
(1)
PΨ̄1 ,...,Ψ̄s ,T ( dt) Pf GΨ̄1 ,...,Ψ̄s ,T (t) ,
(4.5)
where
– αT (Ψ̄1 , . . . , Ψ̄n ) is the sign of the permutation from Ψ̄1 ⊕ · · · ⊕ Ψ̄s
to T ⊕ (Ψ̄1 \ T ) ⊕ · · · ⊕ (Ψ̄s \ T );
– t = {ti,j }1≤i,j≤s , and PΨ̄1 ,...,Ψ̄s ,T ( dt) is a probability measure with
support on a set of t such that ti,j = ui · uj for some family of
vectors ui = ui (t) ∈ Rs of unit norm;
s
(1)
– letting 2q =
i=1 |Ψ̄i |, GΨ̄1 ,...,Ψ̄s ,T (t) is an antisymmetric (2q −
2s + 2) × (2q − 2s + 2) matrix, whose off-diagonal elements are given
(1)
(1)
by GΨ̄1 ,...,Ψ̄s ,T (t) f,f ′ = ti(f ),i(f ′ ) gℓ(f,f ′ ) , where f, f ′ are elements
of the tuple (Ψ̄1 \ T ) ⊕ · · · ⊕ (Ψ̄s \ T ), and i(f ) is the integer in
{1, . . . , s} such that f is an element of Ψ̄i \ T .
(1)
Recalling that gℓ
(1)
and VΛ
(1)
(1)
admit infinite volume limits gℓ,∞ and V∞ , respec(0)
tively, in the sense of (2.2.9) and (3.19), from (4.4) it follows that VΛ admits
an infinite volume limit as well, equal to the ‘obvious’ analogue of the right side
of (4.4), namely, the expression obtained from that one by replacing: M1,Λ by
(1)
(1)
M1,∞ := ∪n∈2N (O × Z2 )n ; GT by GT,∞ (the latter being defined analogously
(1)
to the former, with gℓ
(1)
VΛ
(1)
by V∞ .
(1)
gσ,σ′ (z, z ′ ) with σ, σ ′ ∈ {+, −}
∗
′
gm (z, z ), where g∗m (z, z ′ ) the same as
8 Here
(1)
replaced by gℓ,∞ in all the involved expressions); and
are the elements of the 2 × 2 matrix g(1) (z, z ′ ) ≡
in (2.1.16) with t1 replaced by t∗1 .
Vol. 23 (2022) Non-integrable Ising Models in Cylindrical Geometry
1089
Proceeding inductively in h ≤ 0, one finds that (4.3) implies a repre(h)
sentation of the kernel VΛ of V (h) analogous to (4.4). Also in that case,
(h)
the resulting formula for VΛ admits a natural infinite volume limit. In this
(h)
way, we obtain a recursive equation for the infinite plane kernels, denoted V∞ ,
whose definition and solution is described below. Convergence of the finite vol(h)
ume kernels VΛ to their infinite volume counterparts, with optimal bounds
on the norm of the finite size corrections, is deferred to [4, Section 3]. A key
point in the derivation of bounds on the kernels that are uniform in the scale
label h is the definition of an appropriate action of the L and R operators,
as well as of their infinite volume counterparts, L∞ and R∞ . As anticipated
above, these operators allow us to isolate the potentially divergent part of the
kernels, LV (h) (the ‘local’ contributions, parametrized at any given scale, by a
finite number of ‘running coupling constants’) from a remainder RV (h) , which
is ‘dimensionally better behaved’ than LV (h) ; in order for the remainder to be
shown to satisfy ‘improved dimensional bounds’, it is necessary to rewrite it in
an appropriate, interpolated, form, involving the action of discrete derivatives
on the Grassmann fields.
The plan of the incoming subsections is the following: in Sect. 4.1 we
describe the representation of the effective potentials in the infinite volume
limit and introduce the notion of equivalent kernels; in Sect. 4.2 we define
the operators L∞ and R∞ ; in Sect. 4.3 we derive the solution to the recursive equations for the infinite volume kernels in terms of a tree expansion;
in Sect. 4.4, we use such a tree expansion to derive weighted L1 bounds on
the kernels; importantly, these bounds depend upon the sequence of running
coupling constants, and they imply analyticity of the kernels provided such a
sequence is uniformly bounded in the scale label; in Sect. 4.5, as a corollary
of the weighted L1 bounds of the previous subsection, we prove a fixed point
theorem, which allows us to fix the free parameters Z, β, t∗1 in such a way that
the flow of the running coupling constants is, in fact, uniformly bounded in
h, as desired: even more, the running coupling constants go to zero exponentially fast as h → −∞, a consequence of the irrelevance of the quartic effective
interaction in the theory at hand.
4.1. Effective Potentials and Kernels: Representation and Equivalence
In this subsection we define the effective potential in the full plane in terms
of equivalence classes of kernels V (Ψ), namely, of real-valued functions play(0)
(1)
ing the same role as the coefficient functions V∞ (Ψ) and V∞ (Ψ) introduced
above. This points of view avoids defining an infinite-dimensional Grassmann
algebra. The equivalence relation among kernels, to be introduced momentarily, generalizes the relationships which hold between different ways of writing
the coefficients of a given Grassmann polynomial.
As mentioned above, in order to obtain bounds on the kernels of the
effective potentials which are uniform in the scale label, we will need to group
some of the Grassmann fields into discrete derivatives; we will mainly use
the directional derivative ∂j φω,z := φω,z+êj − φω,z (note that this is the same
1090
G. Antinucci et al.
Ann. Henri Poincaré
convention used in Sect. 2.2.2). We consequently consider kernels which specify
when and how this is done, and in particular define the equivalence relationship
with this in mind.
Let Λ∞ := Z2 , let B denote the set of nearest neighbor edges of Λ∞ , and
let D := {D ∈ {0, 1, 2}2 : D1 ≤ 2}. Let M∞ = ∪n∈2N ({+, −} × D × Λ∞ )n
be the set of field multilabels. for some n ∈ 2N, such that Di 1 ≤ 2. We can
think of any Ψ = ((ω1 , D1 , z1 ), . . . , (ωn , Dn , zn )) ∈ M∞ as indexing a formal
Grassmann monomial φ(Ψ) given by
φ(Ψ) = ∂ D1 φω1 ,z1 · · · ∂ Dn φωn ,zn ,
(4.1.1)
where, denoting Di = ((Di )1 , (Di )2 ) ∈ D, we let
(Di )1 (Di )2
φωi ,zi ,
∂2
∂ Di φωi ,zi := ∂1
with ∂1 and ∂2 the (right) discrete derivatives introduced above. In the
following, with some abuse of notation, any element Ψ ∈ M∞ of length
|Ψ| = n will be denoted indistinctly by Ψ = ((ω1 , D1 , z1 ), . . . , (ωn , Dn , zn ))
or Ψ = (ω, D, z), with the understanding that ω = (ω1 , . . . , ωn ), etc.
We will call a function V : M∞ → R a kernel function, let Vn denote its
restriction to field multilabels of length n, and let Vn,p be the restriction of
Vn to field multilabels with D1 = p. Thinking of such a V as the coefficient
function of a formal Grassmann polynomial
V (Ψ)φ(Ψ)
(4.1.2)
V(φ) =
Ψ∈M∞
suggests an equivalence relationship corresponding to manipulations allowed
by the anticommutativity of the Grassmann variables and by the definition of
discrete derivative.
More precisely, we say that V is equivalent to V ′ , and write V ∼ V ′ , if
either:
1. V ′ is obtained from V by permuting the arguments and changing the sign
according to the parity of the permutation;
2. V ′ is obtained from V by writing out the action of a derivative: that
is, there exist n ∈ 2N, i ∈ {1, . . . , n} and j ∈ {1, 2} such that, letting
−
D+
i,j = (D1 , . . . , Di−1 , Di +êj , Di+1 , . . . , Dn ) and z i,j = (z1 , . . . , zi−1 , zi −
êj , zi+1 , . . . , zn ),
⎧
⎪
⎨0
′
−
+
Vn,p
(ω , D , z ) = Vn (ω , D +
i,j , z i,j ) − Vn (ω , D i,j , z )
⎪
⎩
+
−
Vn (ω , D , z )+Vn (ω , D i,j , z i,j )−Vn (ω , D +
i,j , z )
if
if
if
(Di )j = 2,
(Di )j = 1,
(Di )j = 0,
(4.1.3)
while Vm′ = Vm for all m ∈ 2N \ {n};
3. V ′ is obtained from V by adding an arbitrary kernel V ∗ that is different from zero only for arguments with common repetition, that is,
V ∗ (ω, D, z) = 0 unless there is some i = j such that (ωi , Di , zi ) =
(ωj , Dj , zj );
Vol. 23 (2022) Non-integrable Ising Models in Cylindrical Geometry
1091
or V ′ is obtained from V by a countable sequence of such elementary operations
and of convex combinations thereof. Moreover, we assume that the equivalence
′
relation ∼ is preserved by linear
i.e., if Vα ∼ Vα for all α in the
combinations,
′
countable index set I, then α∈I Vα ∼
α∈I Vα . We will call the equivalence classes generated by ∼ potentials and often specify them by formal sums
like (4.1.2).
Remark 4.1. The operation in item 2 can be thought of as a form of ‘integration by parts’. The kernels equivalent to zero, V ∼ 0, correspond to what are
known as ‘null fields’ in the literature on conformal field theories.
4.2. Localization and Interpolation
In this section we define the operators L∞ and R∞ acting on kernels indexed
by field multilabels in M∞ and show several estimates related to R∞ . We
recall that, given a kernel V , the symbol Vn,p denotes its restriction to field
multilabels of length n, such that D1 = p.
The operator L∞ . First of all, we let
n
(4.2.1)
L∞ (Vn,p ) := 0, if 2 − − p < 0.
2
In the RG jargon, the combination 2 − n2 − p is the scaling dimension of Vn,p
and will reappear below, for example in Lemma 4.7; in this sense, (4.2.1) says
that the local part of the terms with negative scaling dimension (the irrelevant
terms, in the RG jargon) is set equal to zero.
There are only a few cases for which 2 − n2 − p ≥ 0, namely (n, p) =
(2, 0), (2, 1), (4, 0). In these cases, the action of L∞ on Vn,p is non-trivial and
$ which
will be defined in terms of other basic operators, the first of which is L,
$ n,p ) = (LV
$ )n,p ≡ LV
$ n,p , with
is defined as: L(V
⎛
$ n,p (ω , D , (z1 , . . . , zn )) := ⎝
LV
n
j=2
⎞
δzj ,z1 ⎠
Vn,p (ω , D , (z1 , y2 , . . . , yn )).
y2 ,...,yn ∈Λ∞
(4.2.2)
If (n, p) = (2, 1), (4, 0), we let
$ 2,1 ),
L∞ (V2,1 ) := A(LV
$ 4,0 ),
L∞ V4,0 := A(LV
(4.2.3)
where A is the operator that antisymmetrizes with respect to permutations
and symmetrizes with respect to reflections in the horizontal and vertical directions9 . A first important remark, related to the definitions (4.2.3), is that, if
V2,1 is invariant under translations and under the action of A, then
L∞ (V2,1 ) = c1 Fζ,∞ + c2 Fη,∞ ,
9 The
(4.2.4)
reflection transformations in the infinite plane act on the Grassmann fields in the
same way as (2.4.5), with the difference that the reflections θ1 (x1 , x2 ) and θ2 (x1 , x2 ) are
replaced by their infinite-plane analogues, namely by θ̃1 (x1 , x2 ) = (−x1 , x2 ) and θ̃2 (x1 , x2 ) =
(x1 , −x2 ), respectively.
1092
G. Antinucci et al.
Ann. Henri Poincaré
for some real numbers c1 , c2 and Fζ,∞ , Fη,∞ the A-invariant kernels associated
with the potentials
φω,z d2 φ−ω,z ,
ωφω,z d1 φω,z ,
Fη,∞ (φ) :=
Fζ,∞ (φ) :=
ω=± z∈Λ∞
ω=± z∈Λ∞
with dj the symmetric discrete derivatives, acting on the Grassmann fields as
dj φω,z = 12 (∂j φω,z + ∂j φω,z−êj ). A second and even more important remark
$ 4,0 is
is that, due to the fact that ω only assumes two values and that LV
supported on z such that z1 = z2 = z3 = z4 , one has
L∞ (V4,0 ) = 0,
(4.2.5)
a cancellation that will play an important role in the following.
In order to define the action of L∞ on V2,0 , we want to obtain a kernel
$ 2,0 , denoted by (RV
$ )2,1 , which is supported
function equivalent to V2,0 − LV
on arguments with an additional derivative. As we will see, the definition of
$ will also play a central role in the definition of the operator R∞ below. We
R
rewrite
$ 2,0 (ω, 0, z)]φ(ω, 0, z)
[V2,0 (ω, 0, z) − LV
z ∈Λ2∞
=
z ∈Λ2∞
=
z ∈Λ2∞
V2,0 (ω, 0, z)[φ(ω, 0, z) − φ(ω, 0, (z1 , z1 ))]
V2,0 (ω, 0, z)φω1 ,z1 (φω2 ,z2 − φω2 ,z1 ).
(4.2.6)
We now intend to write the difference φω1 ,z1 (φω2 ,z2 − φω2 ,z1 ) as a sum of terms
′
of the form φω1 ,z1 ∂ D φω2 ,y , with D′ 1 = 1, over the sites y on a path from z1
to z2 . To do this we must specify which path is to be used.
For each z, z ′ ∈ Λ∞ , let γ(z, z ′ ) = (z, z1 , z2 , . . . , zn , z ′ ) be the shortest
path obtained by going first horizontally and then vertically from z to z ′ . Note
that γ is covariant under the symmetries of the model on the infinite plane,
i.e.,
Sγ(z, z ′ ) = γ(Sz, Sz ′ )
(4.2.7)
where S : Λ∞ → Λ∞ is some composition of translations and reflections parallel to the coordinate axes. Given z, z ′ two distinct sites in Λ∞ , let INT(z, z ′ )
be the set of (σ, (D1 , D2 ), (y1 , y2 )) ≡ (σ, D, y) ∈ {±} × {0, ê1 , ê2 }2 × Λ2∞ such
that: (1) y1 = z, (2) D1 = 0, (3) y2 , y2 + D2 ∈ γ(z, z ′ ), (4) σ = + if y2 precedes
y2 + D2 in the sequence defining γ(z, z ′ ), and σ = − otherwise. In terms of
this definition, one can easily check that (4.2.6) can be rewritten as
V2,0 (ω, 0, z)
σφ(ω, D, y)
(4.2.6) =
z ∈Λ2∞
(σ,D ,y )∈INT(z )
(4.2.8)
(1)
≡
y ∈Λ2∞ D
$ )2,1 (ω, D, y)φ(ω, D, y),
(RV
Vol. 23 (2022) Non-integrable Ising Models in Cylindrical Geometry
1093
where, if z = (z1 , z1 ), the sum over (σ, D, y) in the first line should be interpreted as being equal to zero (in this case, we let INT(z1 , z1 ) be the empty
set). In going from the first to the second line, we exchanged the order of
(p)
summations over z and y; moreover,
D denotes the sum over the pairs
D = (D1 , D2 ) such that D1 = p, and
$ )2,1 (ω, D, y) :=
(RV
σV2,0 (ω, 0, z).
(4.2.9)
σ,z :
(σ,D ,y )∈INT(z )
$ 2,0 . We
$ )2,1 ∼ V2,0 − LV
From the previous manipulations, it is clear that (RV
are finally ready to define the action of L∞ on V2,0 :
$ 2,0 + L(
$ RV
$ )2,1 ).
L∞ (V2,0 ) := A(LV
(4.2.10)
L∞ (V2,0 ) = c0 Fν,∞ + c1 Fζ,∞ + c2 Fη,∞ ,
(4.2.11)
Note that, if V2,0 is invariant under translations and under the action of A,
then
for some real numbers c0 , c1 , c2 , and Fν,∞ the A-invariant kernel associated
with the potential
1
Fν,∞ (φ) :=
ωφω,z φ−ω,z ,
(4.2.12)
2 ω=±
z∈Λ∞
while we recall that Fζ,∞ , Fη,∞ were defined right after (4.2.4). Summarizing,
in view of Eq. (4.2.5),
⎧
$
⎪
if (n, p) = (2, 0),
⎨A (LV2,0 )
$
$
$
(L∞ V )n,p = A(LV2,1 + L(RV )2,1 ) if (n, p) = (2, 1),
(4.2.13)
⎪
⎩
0
otherwise.
The operator R∞ . We now want to define an operator R∞ such that
R∞ V ∼ V − L∞ V for kernels V that are invariant under translations and
under the action of A. First of all, we let
R∞ (Vn,p ) = (R∞ V )n,p := Vn,p ,
or n = 2 and p ≥ 3.
if: n ≥ 6,
or n = 4 and p ≥ 2,
(4.2.14)
Moreover, we let
(R∞ V )2,0 = (R∞ V )2,1 = (R∞ V )4,0 := 0.
(4.2.15)
The only remaining cases are (n, p) = (2, 2), (4, 1). For these values of (n, p),
(R∞ V )n,p is defined in terms of an interpolation generalizing the definition
$ )2,1 in (4.2.9). As a preparation for the definition, we first introduce
of (RV
$
(RV )n,p for (n, p) = (2, 2), (4, 1). For this purpose, we start from the analogues
of (4.2.6), (4.2.8) in the case that (2, 0) is replaced by (n, p) = (2, 1), (4, 0): for
such values of (n, p) we write
1094
G. Antinucci et al.
z ∈Λn
∞
=
$ n,p (ω, D, z)]φ(ω, D, z)
[Vn,p (ω, D, z) − LV
z ∈Λn
∞
=
Ann. Henri Poincaré
Vn,p (ω, D, z)[φ(ω, D, z) − φ(ω, D, (z1 , z1 , . . . , z1 ))]
Vn,p (ω, D, z)
z ∈Λn
∞
σφ(ω, D + D ′ , y).
(4.2.16)
(σ,D ′ ,y )∈INT(z )
In the last expression, if n = 2, then INT(z) is the same defined after (4.2.7);
if n = 4, then INT(z) is the set of (σ, (D1 , . . . , D4 ), (y1 , . . . , y4 )) ≡ (σ, D, y) ∈
{±} × {0, ê1 , ê2 }4 × Λ4∞ such that: either y1 = y2 = y3 = z1 , D1 = D2 =
D3 = 0, and (σ, (0, D4 ), (z1 , y4 )) ∈ INT(z1 , z4 ); or y1 = y2 = z1 , y4 = z4 ,
D1 = D2 = D4 = 0, and (σ, (0, D3 ), (z1 , y3 )) ∈ INT(z1 , z3 ); or y1 = z1 ,
y3 = z3 , y4 = z4 , D1 = D3 = D4 = 0, and (σ, (0, D2 ), (z1 , y2 )) ∈ INT(z1 , z2 ).
By summing (4.2.16) over D and exchanging the order of summations over z
and y, we find
(p)
z ∈Λn
∞
=
Vn,p (ω, D, z)
with
σφ(ω, D + D ′ , y)
(σ,D ′ ,y )∈INT(z )
D
(p+1)
y ∈Λn
∞
D
(4.2.17)
$ )n,p+1 (ω, D, y)φ(ω, D, y).
(RV
$ )n,p+1 (ω, D, y) :=
(RV
′
σ,z ,D :
(σ,D ′ ,y )∈INT(z )
σVn,p (ω, D − D ′ , z).
(4.2.18)
We are now ready to define:
$ )2,2 + (R(
$ RV
$ ))2,2 ),
(R∞ V )2,2 := A(V2,2 + (RV
$ )4,1 ).
(R∞ V )4,1 := A(V4,1 + (RV
(4.2.19)
Summarizing,
(R∞ V )n,p
⎧
0
⎪
⎪
⎪
⎨A(V + (RV
$ )2,2 + (R(
$ RV
$ ))2,2 )
2,2
=
$
⎪A(V4,1 + (RV )4,1 )
⎪
⎪
⎩
Vn,p
if (n, p) = (2, 0), (2, 1), (4, 0),
if (n, p) = (2, 2),
if (n, p) = (4, 1),
otherwise
(4.2.20)
From the previous manipulations and definitions, it is clear that, if V is invariant under translations and under the action of A, then V − L∞ V ∼ R∞ V .
For later use, given D = (D1 , . . . , Dn ) with D1 = p, we let R∞ V D be the
restriction of (RV )n,p to that specific choice of derivative label.
Remark 4.2. From the definitions of L∞ and R∞ , it also follows that, if V
is invariant under translations and under the action of A, then L∞ (L∞ V ) =
L∞ V and R∞ (L∞ V ) = 0, two properties that will play a role in the following.
Vol. 23 (2022) Non-integrable Ising Models in Cylindrical Geometry
1095
Norm bounds. Let us conclude this section by a couple of technical estimates,
which relate a suitable weighted norm of R∞ V to that of V , and will be
useful in the following. Suppose that V is translationally invariant. Let, for
any κ ≥ 0,
(p)
eκδ(z ) sup |Vn,p (ω, D, z)|,
(4.2.21)
Vn,p (κ) := sup
ω
D
z ∈Λn
∞:
z1 fixed
where the label (p) on the sup over D indicates the constraint that D1 = p,
and δ(z) is the tree distance of z. With these definitions,
Lemma 4.3. For any positive ǫ,
$ )n,p (κ) ≤ (n − 1)ǫ−1 Vn,p−1 (κ+ǫ) if
(RV
(n, p) = (2, 2), (4, 1),
(4.2.22)
$ RV
$ ))2,2 (κ) ≤ ǫ−2 V2,0 (κ+2ǫ) .
(R(
(4.2.23)
As a consequence, noting that AVn,p (κ) ≤ Vn,p (κ) and recalling the
definitions (4.2.20), we find that
(R∞ V )2,2 (κ) ≤ V2,2 (κ) + ǫ−1 V2,1 (κ+ǫ) + ǫ−2 V2,0 (κ+2ǫ) ,
−1
(R∞ V )4,1 (κ) ≤ V4,1 (κ) + 3ǫ
V4,0 (κ+ǫ) .
(4.2.24)
(4.2.25)
In the following, we will use bounds of this kind in order to evaluate the size
of the renormalized part of the effective potential on scale h. In such a case,
both κ and ǫ will be chosen of the order 2h .
Proof. In order to prove (4.2.22), note that it follows directly from the defini$ that
tion of R
(p)
$ )n,p (κ) ≤ sup
(RV
eκδ(z ) |Vn,p−1 (ω, D − D ′ , y)|.
sup
ω
z ∈Λn
∞:
z1 fixed
D
σ,y ,D ′ :
(σ,D ′ ,z )∈INT(y )
(4.2.26)
If we now exchange the order of summations over z and y, we find
(p−1)
$ )n,p (κ) ≤ sup
(RV
eκδ(z ) sup |Vn,p−1 (ω, D, y)|.
ω
y ∈Λn
σ,z ,D ′ :
∞:
y1 fixed (σ,D ′ ,z )∈INT(y )
D
(4.2.27)
Now note that δ(z) ≤ δ(y) and that | INT(y)| ≤ (n − 1)δ(y), so that
(p−1)
$ )n,p (κ) ≤ (n − 1) sup
(RV
eκδ(y ) δ(y) sup |Vn,p−1 (ω, D, y)|
ω
y ∈Λn
∞:
y1 fixed
n−1
≤
Vn,p−1 κ+ǫ ,
ǫ
D
(4.2.28)
1096
G. Antinucci et al.
Ann. Henri Poincaré
where in the last step we used the fact that δ ≤ eǫδ /ǫ, for any ǫ > 0. A two-step
iteration of the bound (4.2.22) proves (4.2.23).
Similar estimates are valid for more general values of (n, p), but
Lemma 4.3 includes all the cases which are relevant to the present work.
The running coupling constants. At each scale h ≤ 0 we represent the infinite
volume effective potential, as arising from the iterative application of (4.3) in
the infinite volume limit, in the form (4.1.2), namely:
(h)
(h)
V∞
V∞
(Ψ)φ(Ψ).
(4.2.29)
(φ) =
Ψ∈M∞
(h−1)
(h)
For each h ≤ 0, in order to compute V∞
from V∞ via the infinite volume
(h)
(h)
(h)
limit of (4.3), we decompose V∞ ∼ L∞ V∞ + R∞ V∞ . Note that the kernel
(h)
L∞ V∞ , in light of (4.2.4), (4.2.5) and (4.2.4), takes the form
(h)
L∞ V∞
= 2h νh Fν,∞ + ζh Fζ,∞ + ηh Fη,∞ =: υh · F∞ ,
(4.2.30)
for three real constants νh , ζh , ηh , called the running coupling constants at
by the fact that the Fν,∞ =
scale h. The factor 2h in front of νh is motivated
(Fν,∞ )2,0 has scaling dimension 2 − n2 − p(n,p)=(2,0) = 1, see (4.2.1), see also
(4.3.10) below.
4.3. Trees and Tree Expansions
In this section, we describe the expansion for the kernels of the effective potentials, as it arises from the iterative application of Eq. (4.3). As anticipated
above, it is convenient to graphically represent the result of the expansion in
terms of GN trees. At the first step, recalling (3.23) and denoting by Nc (Ψ)
and Nm (Ψ) the kernels of Nc (φ) and Nm (ξ), respectively, we reorganize the
(0)
expression for V∞ obtained by taking the infinite volume of Eq. (4.4) to
obtain, for any Ψ = (ω, D, z) ∈ M∞ such that D = 0 (which we identify
with the corresponding element (ω, z) of ∪n∈2N ({+, −} × Λ∞ )n ),
(0)
int
V∞
(Ψ) = Nc (Ψ) + Z −|Ψ|/2 V∞
(Ψ) +
⎛
(1)
int
Z −|Ψ1 |/2 V∞
(Ψ1 )G∅,∞ (Ψ1 \ Ψ)
Ψ1 ∈M1,∞ :
Ψ1 ⊃Ψ
⎞
(Ψ)
s
∞
1
−|Ψ
|/2
int
j
⎝
Nm (Ψj ) + Z
+
V∞ (Ψj ) ⎠ α(Ψ; Ψ1 , . . . , Ψs )
s! Ψ ,...,Ψ ∈M
s=2
j=1
1
s
1,∞
(1)
×
GT ,∞ (Ψ̄1 , . . . , Ψ̄s ),
(4.3.1)
T ∈S(Ψ̄1 ,...,Ψ̄s )
int
is the infinite volume limit of the kernel of V int (φ, ξ) := V int (φ, ξ, 0),
where V∞
(1)
and we recall that in the first (resp. last) line, the factor GT,∞ (Ψ1 \ Ψ) (resp.
(1)
GT,∞ (Ψ̄1 , . . . , Ψ̄s )) is different from zero only if Ψ1 \ Ψ is an element (resp.
Ψ̄1 , . . . , Ψ̄s are elements) of ∪n∈2N ({+i, −i} × Λ∞ )n . For the definition of Ψ̄j ,
see the first item in the list after Eq. (4.4).
Vol. 23 (2022) Non-integrable Ising Models in Cylindrical Geometry
1097
Figure 1. Graphical interpretation of (4.3.1)
Similarly, at the following steps, for any h ≤ 0 and any Ψ ∈ M∞ , we
(h−1)
can write V∞
(Ψ), as computed via the infinite volume analogue of (4.3),
as follows:
(h−1)
V∞
(h)
(Ψ) ∼ υh · F∞ (Ψ) + R∞ V∞ (Ψ) +
(h)
(h)
R∞ V∞ (Ψ1 )G∅,∞ (Ψ1 \ Ψ)
Ψ1 ∈M1,∞ :
Ψ1 ⊃Ψ
⎞
⎛
(Ψ)
∞
s
1
(h)
⎝
υh · F∞ (Ψj )+R∞ V∞ (Ψj ) ⎠ α(Ψ; Ψ1 , . . . , Ψs )
+
s! Ψ ,...,Ψ ∈M
s=2
j=1
1
s
1,∞
(h)
GT ,∞ (Ψ̄1 , . . . , Ψ̄s ),
(4.3.2)
×
T ∈S(Ψ̄1 ,...,Ψ̄s )
(h)
where GT,∞ (Ψ̄1 , . . . , Ψ̄s ) is the infinite volume analogue of the function defined
in (4.5), differing from it for an important feature (besides the ‘obvious’ one
(h)
(h)
that GT,∞ is defined in terms of the infinite plane propagators gℓ,∞ rather
than those on the cylinder): since now the field multilabels Ψi have the form
(ω i , D i , z i ), with D i different from 0, in general, the infinite plane propagators
(h)
(h)
gℓ,∞ , with ℓ = ((ωi , Di , zi ), (ωj , Dj , zj )), entering the definition of GT,∞ should
D
(h)
now be interpreted as ∂zDi i ∂zj j gωi ωj (zi , zj ).
We graphically interpret (4.3.1) as in Fig. 1. On the right-hand side of
the first line, we have drawn a series of diagrams consisting of a root at scale
1, which we will usually denote v0 , connected to s other vertices (which we
will call endpoints) at scale 2 which are of two different types: , which we call
counterterm endpoints and which represent Nc or Nm , and , called interaction
int
endpoints and representing V∞
. In the first two terms (in which there is no
(1)
(1)
factor GT,∞ , because it is ‘trivial’, i.e., it equals G∅,∞ (∅) = 1) the root is
drawn simply as the end of a line segment (we will say it is undotted); in the
other terms (including all those with s > 1) we draw a dot to indicate the
(1)
presence of a non-trivial GT,∞ and α factors and additional sums.
(0)
(0)
(0)
In order to iterate the scheme, we decompose V∞ as V∞ ∼ L∞ V∞ +
(0)
(0)
R∞ V∞ and graphically represent L∞ V∞ = υ0 · F∞ by a counterterm vertex
(0)
at scale 1, and R∞ V∞ by a big vertex
at scale 1, as indicated in the
second line of Fig. 1. Next, using the conventions of Fig. 1, we graphically
(−1)
represent V∞ , computed by (4.3.2) with h = 0, as described in Fig. 2.
1098
G. Antinucci et al.
Ann. Henri Poincaré
Figure 2. Graphical interpretation of (4.3.2) with h = 0
In passing from the first to the second line of Fig. 2 we expanded the big
(0)
vertex on scale 1, which represents R∞ V∞ , by using the first line of Fig. 1,
with an additional label R∞ on the vertices on scale 1, to represent the action
of R∞ .
The graphical equations in Figs. 1 and 2 are the analogues of the graphical
equations in [15, Figures 6-7], which contains a more detailed discussion of
some aspects of this construction. By iterating the same kind of graphical
equations on lower scales, expanding the big vertices
until we are left with
(h)
endpoints all of type or , we find that V∞ can be graphically expanded in
terms of trees of the kind depicted in Fig. 3, with the understanding that in
principle there should be a label R∞ at all the intersections of the branches
with the vertical lines, with the sole exception of v0 ; however, by convention,
in order not to overwhelm the figures, we prefer not to indicate these labels
(h)
explicitly. We call such trees ‘GN trees’ and denote by T∞ , with h ≤ 0, the
set of GN trees with root v0 on scale h + 1. We call ‘vertices’ of a GN tree the
root v0 , all its dotted nodes, and its endpoints.
We introduce some conventions and observations about the set of GN
trees:
• The root v0 is the unique leftmost vertex of the tree. Its degree (number
of incident edges) must be at least 1, i.e., v0 cannot be an endpoint. It
may or may not be dotted; in order for v0 not to be dotted, its degree
must be 1.
• Vertices, other than the root, with exactly one successor, are called ‘trivial’.
• Interaction endpoints can only be on scale 2. Counterterm endpoints
can be on all scales ≤ 2; if such an endpoint is on a scale h < 2, then
it must be connected to a non-trivial vertex on scale h − 1. [The reason is
the following: if this were not case, then there would be an R∞ operator
Vol. 23 (2022) Non-integrable Ising Models in Cylindrical Geometry
1099
v0
h+1
2
(h)
Figure 3. Example of a tree in T∞ . As explained in the
text, one should imagine that a label R, indicating an action
of the R operator, is present at all the intersections of the
branches with the vertical lines, with the only exception of
v0 . In order not to overwhelm the figures, these labels are left
implicit
acting on the value of the endpoint, but this would annihilate it, because
(h)
a endpoint on scale h < 2 corresponds to L∞ V∞ , and the definitions
(h)
of L∞ , R∞ are such that R∞ (L∞ V∞ ) ∼ 0, see Remark 4.2.]
(h)
(h)
In terms of these trees, we shall write the expansion for V∞ = V∞ [υ],
thought of as a function of υ := {(νh , ζh , ηh )}h≤0 , as
(h)
V∞
W∞ [υ; τ ].
(4.3.3)
[υ] ∼
(h)
τ ∈T∞
In order to write W∞ [υ; τ ] more explicitly, we need to specify some additional
(h)
notations and conventions about GN trees. Let τ ∈ T∞ := ∪h≤0 T∞ and
v0 = v0 (τ ) be its root. Then:
• We let V (τ ) be the set of vertices, Ve (τ ) ⊂ V (τ ) the set of endpoints,
and V0 (τ ) := V (τ ) \ Ve (τ ). We also let V ′ (τ ) := V (τ ) \ {v0 } and V0′ (τ ) :=
V0 (τ ) \ {v0 }.
• Given v ∈ V (τ ), we let hv be its scale.
1100
G. Antinucci et al.
Ann. Henri Poincaré
• v ≥ w or ‘v is a successor of w’ means that the (unique) path from v to
v0 passes through w. Obviously, v > w means that v is a successor of w
and v = w.
• ‘v is an immediate successor of w’, denoted v ⊲ w, means that v ≥ w,
v = w, and v and w are directly connected. For any v ∈ τ , Sv is the set
of w ∈ τ such that w ⊲ v.
• For any v > v0 , we denote by v ′ the unique vertex such that v ⊲ v ′ .
(h −1)
• Subtrees: for each v ∈ V0 (τ ), let τv ∈ T∞ v
denote the subtree consisting of the vertices with w ≥ v.
Next, we need to attach labels to their vertices, in order to distinguish the
various contributions to the kernels arising from the different choices of the
sets Ψi , etc., in (4.3.1), (4.3.2), also keeping track of the order in which they
appear. In particular, with each v ∈ V (τ ) we associate a set Pv of field labels,
sometimes called the set of external fields, whose elements carry two informations: their position within an ordered list which they belong to, and their
ω index; more precisely, the family P = {Pv }v∈V (τ ) is characterized by the
following properties, which correspond to properties of the iteration of the
kernel:
• |Pv | is always even and positive. If v is a endpoint, then |Pv | = 2.
• If v is an endpoint of τ , then Pv has the form {(j, 1, ω1 ), . . . , (j, 2n, ω2n )},
where j is the position of v in the ordered list of endpoints, and ωl ∈
{+, −, i, −i}, if hv = 2, while ωl ∈ {+, −}, if hv < 2. Given f = (j, l, ωl ),
we let o(f ) = (j, l) and ω(f ) :=
. ωl .
• If v is not an endpoint, Pv ⊂ w∈Sv Pw .
.
• If v ∈ V0 (τ ), we let Qv :=
w∈Sv Pw \ Pv be the set of contracted fields.
If v is dotted, then we require |Qv | ≥ 2 and |Qv | ≥ 2(|Sv | − 1); and
conversely Qv is empty if and only if v = v0 and
0
. v0 is /not dotted.
• If hv = 1 and v is not an endpoint, then Qv = w∈Sv f ∈ Pw ω(f ) ∈
{+i, −i} (all and only massive fields are integrated on scale 1).
For τ ∈ T∞ , we denote by P(τ ) be the set of allowed P = {Pv }v∈V (τ ) .
We also denote by ω v the tuple of components ω(f ), with f ∈ Pv , and by
ω v Q the restriction of ω v to any subset Q ⊆ Pv . Note that the definitions
imply that for v, w ∈ τ such that neither v ≥ w or v ≤ w (for example when
v ′ = w′ but v = w), Pv and Pw are disjoint, as are Qv and Qw .
Next, given P ∈ P(τ ), for all v ∈ V0 (τ ) we define sets Tv ,
0
/
Tv = (f1 , f2 ) , . . . , f2|Sv |−3 , f2|Sv |−2 ⊂ Q2v ,
called spanning trees associated with v, characterized by the following
properties: if w(f ) denote the (unique) w ∈ Sv for which f ∈ Pw ,
′
then (f, f ′ ) ∈ Tv /⇒ w(f ) = w(f ′ ) and
/
00 o(f ) < o(f ); moreover,
{w(f1 ), w(f2 )} , . . . , w(f2|Sv |−3 ), w(f2|Sv |−2 ) is the edge set of a tree with
vertex set Sv . We denote by S(τ, P ) the set of allowed T = {Tv }v∈V0 (τ ) .
Finally, for each v ∈ V (τ ), we denote by Dv a map Dv : Pv → D = {D ∈
{0, 1, 2}2 : D1 ≤ 2}; the reader should think that a derivative operator
∂ Dv (f ) acts on the field labeled f . We denote by D(τ, P ) the set of families
Vol. 23 (2022) Non-integrable Ising Models in Cylindrical Geometry
1101
of maps D = {Dv }v∈V (τ ) . We also denote by D v the tuple of components
Dv (f ), with f ∈ Pv , and by D v Q the restriction of D v to any subset Q ⊆ Pv .
Additionally, if a map z : Pv → Λ∞ is assigned, we denote by z v the tuple of
components z(f ), with f ∈ Pv , and by Ψv = Ψ(Pv ) := (ω v , D v , z v ) the field
multilabel associated with ω v , D v , z v ; moreover, if v ∈ V0 (τ ) and also the
maps z : Pw → Λ∞ , for all w ∈ Sv , are
assigned, for each w ∈ Sv we denote
by Ψ̄w = Ψ(Pw \ Pv ) = (ω w P \P , D w P \P , z w P \P ) the restriction of Ψw
w
v
w
v
w
v
to Pw \ Pv (here z w is the restriction of z w to the subset Q ⊂ Pw ).
Q
In terms of these definitions, we write W∞ [υ; τ ] in the right side of (4.3.3)
as
W∞ [υ; τ ] =
W∞ [υ; τ, P , T , D],
(4.3.4)
P ∈P(τ ) T ∈S(τ,P ) D∈D(τ,P )
where W∞ [υ; τ, P , T , D] is the translationally
invariant kernel inductively
1
defined as follows: letting D ′v0 := v∈Sv D v P for hv0 < 1 and D ′v0 := 0
v0
0
for hv0 = 1,
α
v0
W∞ [υ; τ, P , T , D](ω 0 , D 0 , z 0 ) = 1(ω 0 = ω v0 )1 D 0 = D v0 = D ′v0
|Sv0 |!
(h )
(4.3.5)
)
G v0 (Ψ̄ , . . . , Ψ̄
K
(Ψ ),
×
z:Pv0 ∪Qv0 →Λ∞
z 0 =z v0
Tv0 ,∞
v1
v sv
v,∞
0
v
v∈Sv0
where αv0 = α(Ψv0 ; Ψv1 , . . . , Ψvsv ), cf. (4.3.1), and we recall that, if |Sv0 | = 1,
0
(h
)
v0
then Tv0 = ∅. In this case, if Ψv1 = Ψv0 , then G∅,∞
(∅) should be interpreted
as being equal to 1; this latter case is the one in which, graphically, v0 is not
dotted. In the second line of (4.3.5), if hv0 = 1,
⎧
⎪
if v is of type and v0 is undotted,
⎨Nc (Ψ)
Kv,∞ (Ψv ) = Nm (Ψv )
if v is of type and v0 is dotted,
⎪
⎩ −|Ψv |/2 int
Z
V∞ (Ψv ) if v is of type ,
(4.3.6)
while, if hv0 < 1,
Kv,∞ (Ψv ) :=
⎧
⎪
⎪
⎪υhv0 · F∞ (Ψv )
⎪
⎨R N (Ψ )
∞ c
v
−|Ψv |/2
⎪
Z
⎪
⎪
⎪
⎩W
int
R∞ V∞
(Ψv )
∞ [υ; τv , P v , T v , D v ](Ψv )
if v ∈ Ve (τ ) is of type
and hv = hv0 +1
if v ∈ Ve (τ ) is of type
and hv = 2
if v ∈ Ve (τ ) is of type
if v ∈ V0 (τ ),
(4.3.7)
where, in the last line of (4.3.7), letting P v (resp. T v , resp. Dv ) be the
′
′
restriction of P (resp. T , resp. D) to the subtree τv , and
1 Dv := {Dv } ∪
′
′
{Dw }w∈V (τ ):w>v0 (here Dv is the map such that D v := w∈Sv D w P ), we
v
denoted
(4.3.8)
W ∞ [υ; τv , P v , T v , Dv ] := R∞ W∞ [υ; τv , P v , T v , D′v ] ,
Dv
1102
G. Antinucci et al.
Ann. Henri Poincaré
and we recall that the definition of R∞ V D was given a few lines after (4.2.20).
(h)
(h)
The inductive proof that
V∞ = V∞ [υ], as iteratively computed by (4.3.1)–
(4.3.2), is equivalent to τ ∈T∞(h) W∞ [υ; τ ], with W∞ [υ; τ ] as in (4.3.4), (4.3.5),
is straightforward and left to the reader.
Remark 4.4. Given τ ∈ T∞ and P ∈ P(τ ), we say that D is ‘allowed’ if
W∞ [υ; τ, P , T , D] ∼ 0. With some abuse of notation, from now on we will redefine D(τ, P ) to be the set of allowed D for a given τ and P . Of course, such
a redefinition has no impact on the validity of (4.3.4). If D is allowed, then
it must satisfy a number of constraints. In particular, if v ∈ V0 (τ ), w ∈ Ve (τ )
, then Dv (f) ≥ Dw (f ). Moreover, letting, for any v ∈ V0 (τ ),
and f ∈ Pv ∩ Pw
Rv := D v 1 − w∈Sv D w P 1 ≡ D v 1 − D ′v 1 , one has
v
⎧
2 ′2
2D v 2 = 0
⎪
2,
|P
|
=
2
and
v
⎨
2 21
2 2
Rv = 1, |Pv | = 2 and 2D ′v 21 = 1 or |Pv | = 4 and 2D ′v 21 = 0
⎪
⎩
0, otherwise,
(4.3.9)
with the exception of v0 , for which Rv0 ≡ 0 (in other words D v0 = D ′v0 , see
(4.3.5)). Finally, the combination
|Pv |
− D v 1 ,
(4.3.10)
2
known as the scaling dimension of v, is ≤ −1 for all v ∈ V0′ (τ ), and for all
v ∈ Ve (τ ) such that hv = 2 and hv′ < 1. As we shall see below, see in particular
the statement of Lemma 4.7, the fact that d(Pv , D v ) ≤ −1 for all such vertices
guarantees that the expansion in GN trees is convergent uniformly in hv0 .
d(Pv , D v ) := 2 −
Remark 4.5. With the other arguments fixed, the number of choices of D for
which W∞ [υ; τ, P , T , D] does not vanish is no more than 10|V (τ )| : there is a
choice of at most 10 possible values for each endpoint10 , and then the other
values are fixed except for a choice of up to 10 possibilities each time that R∞
acts non-trivially on a vertex v ∈ V0′ (τ ), i.e., each time that, for such a vertex,
Rv > 0, see (4.3.9).
4.4. Bounds on the Kernels of the Full Plane Effective Potentials
In this subsection we show that the norm of the kernels W∞ [υ; τ, P , T , D] is
summable over τ, P , T , D, provided that the elements of the sequence υ are
bounded and sufficiently small. We measure the size of W∞ [υ; τ, P , T , D] in
terms of the weighted norm (4.2.21), with κ = c20 2hv0 , where hv0 is the scale of
10 The
value 10 bounds the number of different terms that the operator R∞ produces when
it acts non-trivially on an interaction endpoints. In fact, the cases in which R∞ acts nontrivially are those listed in the right side of (4.2.20) with (n, p) = (2, 2), (4, 1). If (n, p) =
(2, 2), the number of possible values taken by Dv is 10 (one derivative in direction i ∈
{1, 2} on the first Grassmann field and one derivative in direction j ∈ {1, 2} on the second
Grassmann field, etc.); if (n, p) = (4, 1), the number of possible values taken by Dv is 8,
which is smaller than 10 (one derivative in direction i ∈ {1, 2} on the k-the Grassmann field,
with k ∈ {1, 2, 3, 4}).
Vol. 23 (2022) Non-integrable Ising Models in Cylindrical Geometry
1103
the root of τ and c0 is the minimum between the constant c in Proposition 2.3
and half the constant c in (3.14), (3.15), (3.20), (3.21). With some abuse of
notation, we let
W∞ [υ; τ, P , T , D]hv0 := W∞ [υ; τ, P , T , D]( c0 2hv0 ) .
(4.4.1)
2
The first, basic, bound on the kernels W∞ [υ; τ, P , T , D] is provided by the
following proposition. We recall that we assumed once and for all that Z ∈ U ,
with U = {z ∈ R : |z −1| ≤ 1/2}, and that t∗1 , t1 , t2 ∈ K ′ , with K ′ the compact
set defined before the statement of Theorem 1.1.
Proposition 4.6. Let W∞ [υ; τ, P , T , D] be inductively defined as in (4.3.5).
There exist C, κ, λ0 > 0 such that, for any τ ∈ T∞ , P ∈ P(τ ), T ∈ S(τ, P ),
D ∈ D(τ, P ), and |λ| ≤ λ0 ,
W∞ [υ; τ, P , T , D]hv0
⎛
≤ C v∈Ve (τ ) |Pv | ⎝
1
2( 2 |Qv |+
w∈Sv
D w |Qv 1 −Rv +2−2|Sv |)hv
|Sv |!
v∈V0 (τ )
×
v∈Ve (τ )
1
2(hv −1)(2− 2 |Pv |−D v 1 ) ǫhv −1
|λ|max{1,κ|Pv |}
if v is of type
if v is of type
⎞
⎠
(4.4.2)
where ǫh := max{|νh |, |ζh |, |ηh |} if h ≤ 0 and ǫ1 := max {Nc 2 , Nm 2 }.
Proof. Let us first consider the case hv0 = 1, in which case, using (4.3.5) and
(4.3.6), we find
1
W∞ [υ; τ, P , T , 0]hv0 ≤
|Sv0 |!
z:∪v∈Sv Pv →Λ∞ :
0
z(f1 ) fixed
(1)
×ec0 δ(z v0 ) GTv ,∞ (Ψ̄v1 , . . . , Ψ̄vsv ) ·
0
0
⎧
⎪
if v is of type and v0 is undotted,
⎨|Nc (Ψv )|
·
and v0 is dotted,
|Nm (Ψv )|
if v is of type
⎪
v∈Sv0 ⎩|Z|−|Ψv |/2 |V int (Ψ )| if v is of type ,
v
∞
(4.4.3)
(1)
where f1 is the first element of Pv0 . By using the definition of GT,∞ and the
property (PfM )2 = det M , valid for any antisymmetrix matrix M , we find
3
(1)
(1)
(1)
G
|g | sup | det G
(Q1 , . . . , Qs ) ≤
(t)|, (4.4.4)
T,∞
ℓ,∞
ℓ∈T
t
Q1 ,...,Qs ,T,∞
(1)
so that, thanks to items 1,3,4 of Proposition 2.3 (which apply to g∞ by
Remark 2.6), and to the Gram-Hadamard inequality [15, Appendix A.3],
which allows one to bound the determinant of any matrix M with elements
1104
G. Antinucci et al.
Mi,j = (γi , γ̃j ) as | det M | ≤
|γi | |γ̃i |,
⎛
(1)
s−1 ⎝
G
T,∞ (Q1 , . . . , Qs ) ≤ C
Ann. Henri Poincaré
i
e−2c0 z(f )−z(f
(f,f ′ )∈T
⎛
×⎝
f ∈∪i Qi \T
≤ (C ′ )
If we now note that
δ(z 0 ) ≤
(f,f ′ )∈Tv0
i
|Qi |
⎛
⎝
(1)
|γω(f ),0,z(f ) |
·
′
⎞
)1 ⎠
e−2c0 z(f )−z(f
(f,f ′ )∈T
z(f ) − z(f ′ )1 +
⎞1/2
(1)
|γ̃ω(f ),0,z(f ) |⎠
′
⎞
)1 ⎠
.
δ(z v ),
(4.4.5)
(4.4.6)
v∈Sv0
and plug Eq. (4.4.5) into (4.4.3), we find
W∞ [υ; τ, P , T , 0]hv0 ≤
C
⎧
⎨(Nc )2 (c0 )
⎪
×
(Nm )2 (c0 )
⎪
v∈Sv0 ⎩(V int )
∞ |Pv | (c0 )
v∈Sv
0
|Pv |
|Sv0 |!
−c0 z1
e
z∈Λ∞
|Sv0 |−1
if v is of type and v0 is undotted,
if v is of type and v0 is undotted,
if v is of type ,
(4.4.7)
which immediately implies the desired bound for hv0 = 1, because of the
(1)
definition of ǫ1 and (V∞ )|Pv | (c0 ) ≤ C |Pv | |λ|max{1,κ|Pv |} , see (3.15) and recall
the definition of c0 at the beginning of this subsection.
Next, we consider the case hv0 < 1, in which case Kv,∞ is defined by
(4.3.7). We proceed similarly: we start from (4.3.5) and use the analogue of
(4.4.4)–(4.4.5), namely
(hv )
(hv0 )
G 0 (Q1 , . . . , Qs ) ≤
|
|g
T,∞
ℓ,∞
(f,f ′ )∈T
⎛
⎝
f ∈∪i Qi \T
(h
)
(h
⎞1/2
)
|γω(fv0),D(f ),z(f ) | · |γ̃ω(fv0),D(f ),z(f ) |⎠
1
≤ (C2hv0 ) 2
i
|Qi | hv0
2
f ∈∪i Qi
D(f )1
⎛
⎝
(f,f ′ )∈T
e−c0 2
hv
0 z(f )−z(f ′ )
1
⎞
⎠,
(4.4.8)
where in the second inequality we used again the bounds in items 1,4 of
Prop.2.3. Using (4.4.8) and, again, (4.4.6), we obtain the analogue of (4.4.7),
W∞ [υ; τ, P , T , D]hv
0
Vol. 23 (2022) Non-integrable Ising Models in Cylindrical Geometry
≤
·
1
|Sv0 |!
v∈Sv
0
(C2
hv
0
)
|Qv |
0
2
2
hv
0
v∈Sv
0
D v |Q
v
⎧
(υhv · F∞ )2, D v 1 hv
⎪
⎪
0
0
⎪
⎪
⎨(R N )
∞ c 2, D v 1 hv
0
−|Ψv |/2
int
⎪
(R∞ V∞
)|Pv |, D v
⎪|Z|
⎪
⎪
⎩
(W ∞ [υ; τv , P v , T v , D v ])|Pv |,
0
1
·
e
c
h
− 0 2 v0
2
z∈Λ∞
1
hv
0
D v 1 hv0
z 1
|Sv
1105
0
|−1
·
if v ∈ Ve (τ ) is of type
and hv = hv0 + 1
if v ∈ Ve (τ ) is of type
and hv = hv0 + 1
if v ∈ Ve (τ ) is of type
if v ∈ V0 (τ ),
(4.4.9)
The terms in the second line can be bounded as follows:
• If v ∈ Ve (τ ) is of type
and hv = hv0 + 1,
(υhv0 · F∞ )2,D v 1 hv0 ≤ C
2hv0 |νhv0 |
if D v 1 = 0
max{|ζhv0 |, |ηhv0 |} if D v 1 = 1
≤ C ′ 2(hv −1)(2−
|Pv |
2 −D v 1 )
ǫhv −1
(4.4.10)
where we recall that ǫh = max{|νh |, |ζh |, |ηh |} and, in passing from the
first to the second line, we used the fact that |Pv | = 2 and hv0 = hv − 1,
|Pv |
so that 2(hv −1)(2− 2 −D v 1 ) is equal to 2hv0 , if D v 1 = 0, and is equal
to 1, if D v 1 = 1.
and hv = hv0 + 1, using Lemma 4.3 and the
• If v ∈ Ve (τ ) is of type
definition of ǫ1 we have
(R∞ Nc )2,D v 1 hv0 ≤ (R∞ Nc )2,D v 1 0 ≤ 2CNc 1 ≤ 2Cǫ1
= C2(hv −1)(2−
|Pv |
2 −D v 1 )
ǫ1 ,
where the last equality holds trivially since we necessarily have hv = 2
and |Pv | = 2, and (R∞ Nc )2,D v 1 vanishes unless D v 1 = 2.
• If v ∈ Ve (τ ) is of type (and, therefore, hv = 2), then
int
int
(R∞ V∞
)|Pv |,D v 1 0 ≤ C |Pv | |λ|max{1,κ|Pv |} ,
)|Pv |,D v 1 hv0 ≤ (R∞ V∞
thanks to Lemma 4.3 and Eq.(3.15).
• If v ∈ V0 (τ ), recalling the definition of W ∞ [υ; τv , P v , T v , Dv ], see
Eq. (4.3.8), and the bounds on the norm of R∞ , see (4.2.22)–(4.2.25),
we find
(W ∞ [υ; τv , P v , T v , Dv ])|Pv |,D v 1 hv0
′
≤ C2−hv (D v 1 −D v 1 ) (W∞ [υ; τv , P v , T v , D′v ])|Pv |,D ′v 1 hv .
(4.4.11)
Recalling the definition of D ′v , that is D ′v = w∈Sv D w P , as well as the
v
one of Rv , see the line before (4.3.9), we recognize that D v 1 −D ′v 1 =
Rv . Note also that W∞ [υ; τv , P v , T v , D′v ] coincides with its restriction
(W∞ [υ; τv , P v , T v , D′v ])|Pv |,D ′v 1 , so that the second line of (4.4.11) can
be rewritten more compactly as C2−hv Rv W∞ [υ; τv , P v , T v , D′v ]hv .
1
1106
G. Antinucci et al.
Plugging these bounds in (4.4.9), noting that
we find
W∞ [υ; τ, P , T , D]
hv0
≤
1
|Sv0 |!
C
|Qv0 |
+|Sv0 |
2
⎧
|Pv |
⎪2(hv −1)(2− 2 −D v 1 ) ǫhv −1
⎨
·
C |Pv | |λ|max{1,κ|Pv |}
⎪
v∈Sv0 ⎩ −hv Rv
W∞ [υ; τv , P v , T v , D′v ]
2
2
hv
Ann. Henri Poincaré
z∈Λ∞
hv0 (
e−
c0
2
|Qv0 |
+ v∈Sv
2
0
2hv0 z1
≤ C2−2hv0 ,
D v |Qv 1 +2−2|Sv0 |)
0
·
if v ∈ Ve (τ ) is of type
if v ∈ Ve (τ )is of type
if v ∈ V0 (τ ),
(4.4.12)
Now, in the last line, for v ∈ Sv0 ∩V0 (τ ), we iterate the bound, and we continue
to do so until we reach all the endpoints. By doing so, recalling that Rv0 = 0,
we obtain the desired bound, (4.4.2), provided that
C
v∈V0 (τ ) (
|Qv |
2 +|Sv |)
≤ (C ′ )
v∈Ve (τ )
|Pv |
.
(4.4.13)
In order to prove this, note that, for any dotted v ∈ V0 (τ ), |Sv | ≤ 1 + |Q2v | ≤
|Qv |, because |Qv | ≥ max{2, 2(|Sv |−1)}; moreover, ifv0 is not dotted, |Sv0 | = 1
and |Qv0 | = 0. Therefore, recalling also that |Qv | = w∈Sv |Pw |−|Pv |, we find
C
v∈V0 (τ ) (
|Qv |
2 +|Sv |)
which implies (4.4.13).
3
≤ C 1+ 2
v∈V0
|Qv |
3
≤ C 1+ 2
v∈Ve (τ )
|Pv |
,
Next, we rearrange (4.4.2) in a different form, more suitable for summing
over GN trees and their labels.
Lemma 4.7. Under the same assumptions as Proposition 4.6,
1
2hv0 d(Pv0 ,D v0 )
W∞ [υ; τ, P , T , D]hv0 ≤ C v∈Ve (τ ) |Pv |
|Sv0 |!
×
v∈V ′ (τ )
1 (hv −hv′ )d(Pv ,D v )
2
|Sv |!
v∈Ve (τ )
ǫhv −1
|λ|max{1,κ|Pv |}
(4.4.14)
if v is of type
if v is of type
where d(Pv , D v ) = 2− |P2v | −D v 1 is the scaling dimension of v, see (4.3.10).
Note that, as observed in Remark 4.4, the scaling dimensions appearing
at exponent in the product over v ∈ V ′ (τ ) are all negative, with the exception
of the case that v is an endpoint such that hv′ = hv − 1. Note, however, that in
such a case 2(hv −hv′ )d(Pv ,D v ) ≤ 2, which is a constant that can be reabsorbed
in C v∈Ve (τ ) |Pv | , up to a redefinition of the constant C.
Proof. First note that, for all v ∈ V0 (τ ), |Qv | = w∈Sv |Pw | − |Pv |, so that,
recalling that Rv = D v 1 − w∈Sv D w Pv 1 , we can rewrite the factor
1
( 2 |Qv |+ w∈Sv D w |Qv 1 −Rv +2−2|Sv |)hv
in (4.4.2) as
v∈V0 (τ ) 2
1
1
v0
w∈Sv |Pw |− 2 |Pv |+ w∈Sv D w 1 −D v 1 −2(|Sv |−1))
v∈V0 (τ ) ( 2
2h
·
1
1
v∈V0 (τ ) (hv −hv0 )( 2
w∈Sv |Pw |− 2 |Pv |+ w∈Sv D w 1 −D v 1 −2(|Sv |−1))
·2
. (4.4.15)
Vol. 23 (2022) Non-integrable Ising Models in Cylindrical Geometry
1107
Now, the factor in the first line can be further rewritten by noting that:
|Pw | − |Pv | =
(4.4.16)
|Pv | − |Pv0 |,
v∈V0 (τ )
w∈Sv
w∈Sv
v∈V0 (τ )
v∈Ve (τ )
D v 1 − D v 1
=
v∈Ve (τ )
D v 1 − D v0 1 , (4.4.17)
(|Sv | − 1) = |Ve (τ )| − 1.
(4.4.18)
v∈V0 (τ )
(The first two identities are ‘obvious’, due to the telescopic structure of the
summand; the latter identity can be easily proved by induction.) Therefore,
1
v∈V0 (τ ) ( 2
2hv0
w∈Sv
hv0 (2−
=2
|Pw |− |P2v | +
w∈Sv
|Pv |
0 −D
v0 1 )
2
D w 1 −D v 1 −2(|Sv0 |−1))
−hv0 (2− |P2v | −D v 1 )
2
v∈Ve (τ )
.
(4.4.19)
Similarly, the exponent of the factor in the second line of (4.4.15) can be
rewritten as
1
1
|Pw | − |Pv |
(hw − hw′ )
2
2
v∈V0 (τ )
+
(hw − hw′ )
w∈V0 (τ )
w>v0
+
=
v∈V0 (τw )
1
1
|Pw | − |Pv |
2
2
w∈Sv
D w 1 − D v 1 − 2(|Sv0 | − 1)
w∈Sv
D w 1 − D v 1 − 2(|Sv0 | − 1)
w∈Sv
=
w∈Sv
w∈V0 (τ )
v0 <w≤v
|Pw |
− D w 1 −
(hw − hw′ ) 2 −
2
w∈V0 (τ )
w>v0
v∈Ve (τw )
|Pv |
2−
− D v 1
2
.
(4.4.20)
Using (4.4.19) and (4.4.20) and recalling that 2 −
we rewrite (4.4.15) as
hv0 d(Pv0 ,D v0 )
(4.4.15) = 2
v∈V0′ (τ )
(hv −hv′ )d(Pv ,D v )
2
|Pv |
2
− D v 1 ≡ d(Pv , D v ),
1108
G. Antinucci et al.
−(hv0 +
2
v∈Ve (τ )
= 2hv0 d(Pv0 ,D v0 )
Ann. Henri Poincaré
v0 <w<v
w∈V0
(τ ) (hw −hw′ ))d(Pv ,D v )
2(hv −hv′ )d(Pv ,D v )
v∈V0′ (τ )
2−hv′ d(Pv ,D v ) .
v∈Ve (τ )
By using this rewriting in (4.4.2) and noting that
⎛
1
2(hv −1)(2− 2 |Pv |−D v 1 ) ǫhv −1
⎝
2−hv′ d(Pv ,D v ) ·
|λ|max{1,κ|Pv |}
v∈Ve (τ )
⎛
ǫhv −1
if
≤⎝
2(hv −hv′ )d(Pv ,D v ) ·
|Pv |
max{1,κ|Pv |}
2
|λ|
if
v∈V (τ )
e
hv d(Pv ,D v )
(since if v is an endpoint of type , then 2
obtain the desired estimate, (4.4.14).
(4.4.21)
⎞
if v is of type ⎠
if v is of type
⎞
v is of type ⎠
vis of type
(4.4.22)
−|Pv |
≥ 2
), we readily
As announced above, the bound (4.4.14) is written in a form suitable for
summing over the GN trees and their labels, as summarized in the following
lemma.
Lemma 4.8. Under the same assumptions as Proposition 4.6, for any ϑ ∈
(h)
(0, 1), there exists Cϑ > 0 such that, letting T∞;(N,M ) denote the subset of
(h)
T∞ whose trees have N endpoints of type and M endpoints of type ,
W∞ [υ; τ, P , T , D]h+1
2−ϑh
(h)
τ ∈T∞;(N,M ) P ∈P(τ ) T ∈S(τ,P ) D∈D(τ,P )
|Pv0 |=n
D v0 1 =p
≤ CϑN +M |λ|N
′
max 2−ϑh ǫh′
h<h′ ≤1
M
(4.4.23)
2h·d(n,p) .
Remark 4.9. (Short memory property). The fact that this estimate holds with
the factors of 2−ϑh included indicates that the contribution of trees covering
a large range of scales are exponentially suppressed, a behavior referred to in
previous works (e.g., [18]) as the ‘short memory property’. As we shall see
below, taking advantage of this there is a way to choose the free parameters
Z, t∗1 , β such that |ǫh | ≤ Kϑ |λ|2ϑh , see Propositions 4.10 and 4.11 below. Under
this condition, Lemma 4.8 implies that
Vol. 23 (2022) Non-integrable Ising Models in Cylindrical Geometry
(h)
τ ∈T∞;(N,M ) P ∈P(τ ) T ∈S(τ,P ) D∈D(τ,P )
|Pv0 |=n
D v0 1 =p
1109
W∞ [υ; τ, P , T , D]h+1
≤ CϑN +M |λ|N +M 2hd(n,p) 2ϑh ,
(4.4.24)
for all ϑ < 1 and N + M > 0. The factor 2ϑh in the right side is called the
‘short-memory factor’.
Summing over N + M ≥ 1, this immediately implies
2
2
2
2
≤ Cϑ |λ|2hd(n,p) 2ϑh .
(4.4.25)
2(W∞ )n,p 2
h+1
Proof. Thanks to (4.4.14), the left side of (4.4.23) can be bounded from above
by
2h·d(n,p) |λ|N
′
max
2−ϑh ǫh′
′
h<h ≤1
D∈D(τ,P ) D
v0 1 =p
2ϑhv
v∈Ve (τ )
M
2−ϑh
1
⎛
|Sv0 |!
C
(h)
τ ∈T∞;(N,M ) P ∈P(τ )
|Pv0 |=n
⎝
v∈V ′ (τ )
|Pv |
v∈Vep (τ )
T ∈S(τ,P )
⎞
1 (hv −hv′ )d(Pv ,D v ) ⎠
2
|Sv |!
(4.4.26)
|λ|[κ|Pv |−1]+ ,
v∈Ve (τ )
v of type
v of type
where, in the last factor, [·]+ indicates the positive part. Note that, if either
v ∈ V0′ (τ ) or v is an endpoint such that hv′ < hv −1, then the scaling dimension
of v can be bounded uniformly in D v , i.e., d(Pv , D v ) ≤ min{−1, 2 − |Pv |/2};
if v is an endpoint such that hv′ = hv − 1, then the factor2(hv −hv′ )d(Pv ,D v ) is
|P |
smaller than 2 (and, therefore, it can be reabsorbed in C v∈Vep (τ ) v up to a
redefinition of the constant C). Moreover, recall that the number of elements
of D(τ, P ) is bounded by 10|V (τ )| , see Remark 4.5. Finally, the number of
elements of S(τ, P ) is bounded by
|P |
|Sv |!,
(4.4.27)
|S(τ, P )| ≤ C v∈Vep (τ ) v
v∈V0 (τ )
see, e.g., [15, Lemma A.5]. Therefore, putting these observations together, we
see that (4.4.26) can be bounded from above by
2
h·d(n,p)
×
|λ|
v∈V ′ (τ )
N
2
max 2
h<h′ ≤1
−ϑh′
ǫh′
M
2
ϑ(hmax
−h)
τ
v∈Ve (τ )
v of type
′
(C )
P ∈P(τ )
(h)
τ ∈T∞;(N,M )
(hv −hv′ ) min{−1,2− |Pv | }
2
2
ϑhv
|λ|
v∈Vep (τ )
|Pv |
[κ|Pv |−1]+
×
.
v∈Ve (τ )
v of type
(4.4.28)
1110
G. Antinucci et al.
Ann. Henri Poincaré
We can cancel the factor of 2−ϑh with a product of factors leading to an
endpoint, and simplify the remaining bounds, to get
−ϑh
(hv −hv′ ) min{−1,2− |P2v | }
2
2ϑhv
2
v∈V ′ (τ )
≤
(hv −hv′ )(ϑ−1) |P6v |
2
v∈V ′ (τ )
v∈Ve (τ )
v of type
.
This leaves a variety of exponential factors which make it possible to control
the sum over P , giving
′
(hv −hv′ )(ϑ−1) |P6v |
v∈Vep (τ ) |Pv |
(C )
2
v∈V ′ (τ )
(h)
τ ∈T∞;(N,M ) P ∈P(τ )
|λ|[κ|Pv |−1]+ ≤ (C ′′ )N +M ,
v∈Ve (τ )
v of type
see, e.g., [15, App.A.6.1], from which (4.4.23) follows.
We conclude this subsection by noting that, in order for the right side of
the bound (4.4.23) to be summable over N, M uniformly in h, we need that ǫh
is bounded and small, uniformly in h. In view of Lemma 4.8, this condition is
(h)
sufficient for the whole sequence of kernels V∞ , h ≤ 1, to be well defined. In
the next subsection, we study the iterative definition of the running coupling
constants and prove that they in fact remain bounded and small, uniformly in
h, provided that the counterterms υ1 are properly fixed.
4.5. Beta Function Equation and Choice of the Counterterms
The definition of the running coupling constants, (4.2.30), combined with the
GN tree expansion for the effective potentials implies that the running coupling
constants υ = {(νh , ζh , ηh )}h≤0 satisfy the following equation, for all h ≤ 0:
2h νh Fν,∞ + ζh Fζ,∞ + ηh Fη,∞ =
L∞ W∞ [υ; τ ].
(4.5.1)
(h)
τ ∈T∞
More explicitly, using the definitions of Fν,∞ , Fζ,∞ , Fη,∞ (see the lines after
(4.2.4) and after (4.2.11)), for any z ∈ Λ∞ ,
νh = 2−h (2ω)
L∞ W∞ [υ; τ ] (ω, 0, z), (−ω, 0, z) ,
(h)
ζh = 4ω
τ ∈T∞
(h)
τ ∈T∞
ηh = 4
(h)
τ ∈T∞
L∞ W∞ [υ; τ ] (ω, 0, z), (ω, ê1 , z) ,
L∞ W∞ [υ; τ ] (ω, 0, z), (−ω, ê2 , z) .
(4.5.2)
Vol. 23 (2022) Non-integrable Ising Models in Cylindrical Geometry
1111
In view of the iterative definition of the kernels W∞ [υ; τ ], see Eqs. (4.3.4) and
(4.3.5), the right sides of these three equations can be naturally thought of as
functions of υ or, better yet, of the restriction of υ to the scales larger than h.
(1)
Therefore, these are recursive equations for the components of υ: given V∞
∗
(and, in particular, given t1 , β, Z), one can in principle construct the whole
sequence υ. More precisely, since the definition of the right sides of Eq. (4.5.2)
requires the summations over τ to be well defined, in view of Lemma 4.8, these
recursive equations allow one to construct the running coupling constants only
for the scales h such that maxh′ >h ǫh′ is small enough. As we shall soon see,
given a sufficiently small λ, the quantity maxh′ >h ǫh′ stays small, uniformly
in h, only for a special choice of the free parameters t∗1 , β, Z; to understand
the appropriate choice of these parameters, it is helpful first to isolate those
trees with only counterterm vertices, whose contribution in (4.5.2) is especially
simple:
(h)
1. For h < 0, T∞;(0,1) consists of two trees, one with a single counterterm endpoint v on scale hv = 2 > hv0 + 1, for which L∞ W∞ [υ, τ ] =
L∞ R∞ Nc = 0, and one with a single counterterm endpoint on scale
hv = hv0 + 1, for which we have
L∞ W∞ [υ, τ ] = 2h νh Fν,∞ + ζh Fζ,∞ + ηh Fη,∞ ,
(h)
2. For h = 0, T∞;(0,1) consists of only a single tree τ . This tree gives a
contribution involving Nc which can be calculated quite explicitly from
Eq. (3.23) and Eq. (2.1.13) taking advantage of the fact that the latter
is written in terms of the Fourier transformed fields:
L∞ W∞ [υ, τ ] = L∞ Nc
1
t1
t∗1
1 − t1
1
Fζ,∞
Fν,∞ +
−
t2 −
=
Z
1 + t1
Z (1 + t1 )2
(1 + t∗1 )2
t
t∗
2
− 2 Fη,∞
+
(4.5.3)
2Z
2
(recall t∗2 = (1 − t∗1 )/(1 + t∗1 )) which we can put in the same form as the
other scales by letting
t1
1
1 − t1
1
t∗1
t2 −
, ζ1 :=
2ν1 :=
−
,
Z
1 + t1
Z (1 + t1 )2
(1 + t∗1 )2
t∗
t2
− 2.
(4.5.4)
η1 :=
2Z
2
(h)
3. The contributions from the trees in T∞,(0,M ) (i.e., those with no interaction endpoints and M counterterm endpoints) vanish for all h ≤ 0
and M ≥ 2, as can be seen as follows. In this case, each vertex must be
assigned exactly 2 field labels, and for any endpoint v the components
of ω v must be both imaginary (corresponding to Nm , i.e., to hv = 2,
hv′ = 1) or both real (in all the other cases). If any endpoints of the
first type appear there is no allowed assignment of field labels to the
corresponding tree (since the ξ fields this case represents can neither be
1112
G. Antinucci et al.
Ann. Henri Poincaré
external fields nor be contracted with φ fields to form a connected diagram). In any case, all of the internal field labels are included in spanning
trees, and the local part of the contribution can be made quite explicit.
(h)
For example, denoting by g∞ (z − z ′ ) ω,ω′ the elements of the transla(h)
tion invariant infinite volume propagator g∞ (z − z ′ ), see (2.2.4), (2.2.5),
(2.2.9), the local part of the term with D v = 0 for all vertices (which
contributes to νh ) is a sum of terms proportional to
(h1 )
2)
g∞
(z1 − z2 ) −ω ,ω g(h
∞ (z2 − z3 ) −ω ,ω
1
2
2
3
z2 ,...,zM ∈Z2
M −1 )
· · · g(h
(zM −1 − zM ) −ω
∞
M −1 ,ω1
(h2 )
(h1 )
M −1 )
= ĝ∞ −ω ,ω (0) ĝ∞ −ω ,ω (0) · · · ĝ(h
∞
−ω
1
2
2
3
M −1 ,ω1
(0),
(4.5.5)
for some tuple (ω1 , . . . , ωM −1 ) ∈ {+, −}M −1 ; from Eqs. (2.2.9), (2.2.4)
and (2.2.5), the Fourier transform ĝ(h) in the right side satisfies
2−2h
(h)
(4.5.6)
ĝ (k) =
ĝ[η] (k) dη, ĝ[η] (k) = e−ηD(k) D(k)ĝ(k)
2−2h−2
(except for h = 0, where the lower limit of integration is 0), with D
defined in Eq. (2.1.21), from which it is evident that D(0) = 0, and
in consequence that the right-hand side of Eq. (4.5.5) is likewise zero.
The same argument holds, mutatis mutandis, when D v does not vanish
for all vertices of the tree, with some discrete derivatives appearing in
Eq. (4.5.5).
In view of these properties, we can write
ν
νh = 2νh+1 + Bh+1
[υ],
ζ
ζh = ζh+1 + Bh+1
[υ],
η
ηh = ηh+1 + Bh+1
[υ],
(4.5.7)
♯
for all h ≤ 0, where the functions Bh+1
[υ] are given by the restrictions of the
(h)
sums on the right-hand sides of Eq. (4.5.2) to trees in some T∞,(M,N ) with
♯
N ≥ 1. The functions Bh+1
[υ], with ♯ ∈ {ν, ζ, η}, are called the components
of the beta function, and (4.5.7) are called the beta function flow equations for
the running coupling constants; note that, even if not explicitly indicated, the
♯
functions Bh+1
[υ], in addition to υ, depend analytically upon λ, t∗1 , β, Z. For
♯
later reference, we let Bh♯ [υ; τ ] be the contribution to Bh+1
[υ] associated with
the GN tree τ . In view of Lemma 4.8, we find that, for any |Z − 1| ≤ 1/2,
any |t∗1 |, |t1 |, |t2 | ∈ K ′ , any ϑ ∈ (0, 1) there exists Cϑ > 0 such that, if |λ| and
maxh′ >h {ǫh′ } are small enough, then
max |Bh♯ [υ]| ≤ Cϑ |λ|2ϑh ,
♯∈{ν,ζ,η}
∀ h ≤ 0.
(4.5.8)
Proposition 4.10. For any ϑ ∈ (0, 1), there exist Kϑ , λ0 (ϑ) > 0 and functions
ν1 (λ; t∗1 , β, Z), ζ1 (λ; t∗1 , β, Z), η1 (λ; t∗1 , β, Z), υ(λ; t∗1 , β, Z) := {(νh (λ; t∗1 , β, Z),
ζh (λ; t∗1 , β, Z), ηh (λ; t∗1 , β, Z))}h≤0 , analytic in |λ| ≤ λ0 (ϑ), |Z − 1| ≤ 1/2,
Vol. 23 (2022) Non-integrable Ising Models in Cylindrical Geometry
1113
|t∗1 | ∈ K ′ and β such that |t1 |, |t2 | ∈ K ′ (recall that ti = tanh(βJi )), for all
λ, t∗1 , β, Z in such an analyticity domain:
1. ν1 (λ; t∗1 , β, Z), ζ1 (λ; t∗1 , β, Z), η1 (λ; t∗1 , β, Z) and the components of
υ(λ; t∗1 , β, Z) satisfy (4.5.7), for all h ≤ 0;
2. For all h ≤ 1,
ǫh (λ; t∗1 , β, Z) := max{|νh (λ; t∗1 , β, Z)|, |ζh (λ; t∗1 , β, Z)|,
|ηh (λ; t∗1 , β, Z)|} ≤ Kϑ |λ| 2ϑh .
(4.5.9)
Note that, at this point, we do not prove that ν1 (λ; t∗1 , β, Z), ζ1 (λ; t∗1 , β, Z),
η1 (λ; t∗1 , β, Z) satisfy Eq. (4.5.4); this comes later.
Proof. For simplicity, we do not track the dependence of the constants and
of the norms upon ϑ which, within the course of this proof, we assume
to be a fixed constant in (0, 1). In order to construct υ̃(λ; t∗1 , β, Z) :=
{(νh (λ; t∗1 , β, Z), ζh (λ; t∗1 , β, Z), ηh (λ; t∗1 , β, Z))}h≤1 , we first note that the equations for νh , ζh , ηh in (4.5.7) imply that, for k < h ≤ 1, νh = 2k−h νk −
j−h−1 ν
Bj [υ], ζh = ζk − k<j≤h Bjζ [υ], and ηh = ηk − k<j≤h Bjη [υ]. If
k<j≤h 2
we send k → −∞ and impose that ǫk := max{|νk |, |ζk |, |ηk |} → 0 as k → −∞,
we get
⎧
j−h−1 ν
⎪
Bj [υ],
⎨νh = − j≤h 2
ζ
(4.5.10)
ζh = − j≤h Bj [υ],
⎪
⎩
ηh = − j≤h Bjη [υ],
which we regard as a fixed point equation υ̃ = T [υ̃] for a map T on the
space of sequences Xε := {υ̃ = {(νh , ζh , ηh )}h≤1 : υ̃ ≤ ε}, with υ̃ =
suph≤1 {2−ϑh ǫh } and ε a sufficiently small constant.
We now intend to prove that T is a contraction on Xε and, more precisely,
that: (1) the image of Xε under the action of T is contained in Xε ; (2) T [υ̃] −
T [υ̃ ′ ] ≤ (1/2) υ̃ − υ̃ ′ for all υ̃, υ̃ ′ ∈ Xε . Once T is proved to be a contraction,
it follows that it admits a unique fixed point in Xε , which corresponds to the
desired sequence υ̃(λ; t∗1 , β, Z). The analyticity of υ̃(λ; t∗1 , β, Z) follows from the
analyticity of the components of the beta function with respect to λ, t∗1 , β, Z
and υ that, in turn, follows from the absolute summability of its tree expansion,
which is a power series in λ, with coefficients analytically depending upon
t∗1 , t1 , t2 , Z, with ti = tanh(βJi ).
The fact that the image of Xε under the action of T is contained in Xε
follows immediately from Eq. (4.5.8). In order to prove that T [υ̃] − T [υ̃ ′ ] ≤
(1/2) υ̃ − υ̃ ′ , we rewrite the ν-component of T [υ̃]−T [υ̃ ′ ] at scale h as a linear
interpolation
1
d ν
νh − νh′ = −
Bj [υ̃(t); τ ] dt,
(4.5.11)
2j−h−1
0 dt
(j−1)
j≤h τ ∈T∞
′
where υ̃(t) = υ̃ + t(υ̃ − υ̃ ′ ), and similarly for the ζ- and η-components. When
the derivative with respect to t acts on the tree value Bjν [υ(t); τ ], it has the
effect of replacing one of the factors νh (t), or ζh (t), or ηh (t), associated with
1114
G. Antinucci et al.
Ann. Henri Poincaré
d
d
one of the counterterm endpoints, by dt
νh (t) = νh −νh′ , or dt
ζh (t) = ζh −ζh′ , or
d
′
dt ηh (t) = ηh − ηh , respectively. Therefore, we get the analogue of Eq. (4.5.8):
if |λ| and υ̃(t) are sufficiently small, then
d ♯
max
Bh [υ(t); τ ]
dt
♯∈{ν,ζ,η}
(h−1)
τ ∈T∞
≤ C|λ|2ϑh max
{|νh′ − νh′ ′ |, |ζh′ − ζh′ ′ |, |ηh′ − ηh′ ′ |}
′
h ≥h
(4.5.12)
for all h ≤ 1. Plugging this estimate into Eq. (4.5.11) and its analogues for
ζh − ζh′ and ηh − ηh′ , we readily obtain the desired estimate, T [υ̃] − T [υ̃ ′ ] ≤
(1/2) υ̃ − υ̃ ′ , for λ0 and ε sufficiently small.
We now need to show that it is possible to choose the free parameters β, Z, t∗1 in such a way that the functions ν1 (λ; t∗1 , β, Z), ζ1 (λ; t∗1 , β, Z),
η1 (λ; t∗1 , β, Z) constructed in Proposition 4.10 satisfy Eq. (4.5.4); that is, for
given J1 , J2 and λ, there exists a critical value of the inverse temperature β
for which the above expansion for the kernels of the infinite plane effective
potentials is convergent, with dressed parameters t∗1 and Z. The desired result
is summarized in the following proposition.
Proposition 4.11. For any J1 , J2 satisfying the conditions of Theorem 1.1,
and any ϑ ∈ (0, 1), there exist λ0 (ϑ) > 0 and functions t∗1 (λ), βc (λ),
Z(λ), analytic in |λ| ≤ λ0 (ϑ), such that Eq. (4.5.4) holds, with t1 =
tanh(βc (λ)J1 ), t2 = tanh(βc (λ)J2 ), and (ν1 , ζ1 , η1 ) = ((ν̃1 (λ), ζ̃1 (λ), η̃1 (λ)),
with ν̃1 (λ) = ν1 (λ; t∗1 (λ), βc (λ), Z(λ)) (here ν1 (λ; t∗1 , β, Z) is the same as in
Proposition 4.10), and similarly for ζ̃1 (λ) and η̃1 (λ). Correspondingly, the flow
of running coupling constants with initial datum (ν̃1 (λ), ζ̃1 (λ), η̃1 (λ)), generated by the flow equations (4.5.7), is well defined for all h ≤ 0 and satisfies
(4.5.9).
Proof. The result is a direct consequence of the analytic implicit function
theorem: we intend to fix t∗1 = t∗1 (λ), β = βc (λ), Z = Z(λ) in such a way
that Eq. (4.5.4) holds, with ν1 = ν1 (λ; t∗1 , β, Z), ζ1 = ζ1 (λ; t∗1 , β, Z), η1 =
η1 (λ; t∗1 , β, Z), the same functions as in Proposition 4.10. With this in mind,
we recast the system of equations (4.5.4) in the following form (recall once
more that ti = tanh(βJi ) and t∗2 = (1 − t∗1 )/(1 + t∗1 )):
2Zν1 (λ; t∗1 , β, Z) − tanh βJ2 + e−2βJ1 = 0,
4Zt∗1
+ e−4βJ1 − 1 = 0,
4Zζ1 (λ; t∗1 , β, Z) +
(1 + t∗1 )2
1 − t∗1
2Zη1 (λ; t∗1 , β, Z) + Z
− tanh βJ2 = 0.
1 + t∗1
(4.5.13)
(4.5.14)
(4.5.15)
Note that, by Proposition 4.10, everything appearing in Eqs. (4.5.13) to
(4.5.15) is analytic, so if we succeed in showing that the implicit function
theorem applies, the resulting solution will be analytic in λ, as desired. Note
that, at λ = 0, we have ν1 (0; t∗1 , β, Z) = ζ(0; t∗1 , β, Z) = η1 (0; t∗1 , β, Z) = 0, so
in this case the system (4.5.13)–(4.5.15) is solved by β = βc (J1 , J2 ) (with
Vol. 23 (2022) Non-integrable Ising Models in Cylindrical Geometry
1115
βc (J1 , J2 ) the critical temperature
of the nearest neighbor model, see the
lines after (1.3)), t∗1 = tanh βc (J1 , J2 )J1 and Z = 1. Note also that the
partial derivatives of ν1 (λ; t∗1 , β, Z), ζ1 (λ; t∗1 , β, Z), η1 (λ; t∗1 , β, Z) with respect
to t∗1 , β and Z vanish at λ = 0, so the determinant of the Jacobian of the
with respect to t∗1 , β, Z, computed at λ = 0 and
system (4.5.13)–(4.5.15)
∗
(t1 , β, Z) = (tanh βc (J1 , J2 )J1 , βc (J1 , J2 ), 1), equals
0
−J2 sech2 (βc J2 ) − 2J1 e−2βc J1
0
1−t∗1
4t∗
1
4 (1+t∗ )3
−4J1 e−4βc J2
∗ )2
(1+t
1
1
1−t∗
1
− (1+t2 ∗ )2
−J2 sech2 (βc J2 )
∗
1+t
1
1 + (t∗1 )2
= −4(J2 sech2 (βc J2 ) + 2J1 e−2βc J1 )
(1 + t∗1 )4
1
(4.5.16)
with βc = βc (J1 , J2 ) and t∗1 = tanh(βc (J1 , J2 )J1 ); the right-hand side is evidently nonzero, and, therefore, the analytic implicit function theorem applies,
implying the desired claim.
This concludes the construction of the sequence of effective potentials in
the infinite volume limit, uniformly in the scale label, with optimal bounds
on the speed at which, after proper rescaling, such effective potentials go to
zero as h → −∞. This result, and the methods introduced to prove it, is a
key ingredient in the proof of Theorem 1.1, for whose completion we refer the
reader to [4].
Acknowledgements
We thank Hugo Duminil-Copin for several inspiring discussions. This work has
been supported by the European Research Council (ERC) under the European Union’s Horizon 2020 research and innovation programme (ERC CoG
UniCoSM, grant agreement No. 724939 for all three authors and also ERC
StG MaMBoQ, grant agreement No. 802901 for R.L.G.). G.A. acknowledges
financial support from the Swiss Fonds National. A.G. acknowledges financial
support from MIUR, PRIN 2017 project MaQuMA PRIN201719VMAST01.
Funding Open access funding provided by Università degli Studi Roma Tre
within the CRUI-CARE Agreement.
Open Access. This article is licensed under a Creative Commons Attribution 4.0
International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and
indicate if changes were made. The images or other third party material in this article
are included in the article’s Creative Commons licence, unless indicated otherwise
in a credit line to the material. If material is not included in the article’s Creative
Commons licence and your intended use is not permitted by statutory regulation
or exceeds the permitted use, you will need to obtain permission directly from the
1116
G. Antinucci et al.
Ann. Henri Poincaré
copyright holder. To view a copy of this licence, visit http://creativecommons.org/
licenses/by/4.0/.
Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
A. Diagonalization of the Matrix Ac
In this section we compute the propagator of φ and the Gaussian integral
associated with Sc . For this purpose, we first need to block diagonalize the
coefficient matrix, which we do by using a transformation which can be thought
of as a Fourier sine transformation with modified frequencies. We write
1
−i∆(k1 )
−b(k1 ) + t2 τ φ+ (k1 )
φ+ (−k1 )
,
Sc (φ) =
φ− (k1 )
φ− (−k1 )
i∆(k1 )
b(k1 ) − t2 τ T
2L
k1 ∈DL
1
φ+ (k1 )
φ+ (−k1 )
=:
,
(A.1)
, Ãc (k1 )
φ− (k1 )
φ− (−k1 )
2L
k1 ∈DL
where φω (k1 ), with ω = ±, is the column vector whose components are
φω,z2 (k1 ) with z2 = 1, . . . , M , and τ is the M × M shift matrix τz2 ,z2′ :=
δz2 +1,z2′ , that is,
⎤
⎡
0
1
0
⎥
⎢
..
⎥
⎢0
.
0
1
⎥
⎢
⎥
⎢
.
.
.
.
..
..
..
..
⎥
⎢0
⎥.
⎢
τ =⎢
⎥
..
..
..
⎢
.
.
.
1
0⎥
⎥
⎢
⎥
⎢
..
⎣
.
0
0
1⎦
0
0
0
For brevity, we will write Ãc = Ãc (k1 ), ∆ = ∆(k1 ), b = b(k1 ) since dependence on k1 plays no role in the next several pages. It is helpful to begin by
diagonalizing the real symmetric matrix
+
B̃c
0
2
Ãc =
,
0
B̃c−
where B̃c+ is the M × M tri-diagonal matrix
+
B̃c
⎡
bt2
−∆2 − b2 − t22
⎢
bt2
−∆2 − b2 − t22
⎢
⎢
⎢
⎢
bt2
=⎢
⎢
⎢
⎢
⎢
⎣
⎤
bt2
−∆2 − b2 − t22
..
.
..
.
..
.
bt2
..
.
−∆ − b2 − t22
bt2
2
bt2
−∆2 − b2
⎥
⎥
⎥
⎥
⎥
⎥.
⎥
⎥
⎥
⎥
⎦
Note that all the diagonal entries are equal to −∆2 −b2 −t22 apart from the last
one, which equals −∆2 − b2 . The block B̃c− is obtained from B̃c+ by reversing
Vol. 23 (2022) Non-integrable Ising Models in Cylindrical Geometry
1117
the order of the rows and columns. B̃cω , with ω = ±, can each be thought of
as a discrete Laplacian with mixed boundary conditions, which suggests the
ansatz
⎤
⎡
αk2 eik2 + βk2 e−ik2
i2k2
⎢
+ βk2 e−i2k2 ⎥
vk2 = ⎣αk2 e
(A.2)
⎦,
..
.
for their eigenvectors. In fact, we see that vk2 is an eigenvector of B̃c+ iff the
system of equations
(−∆2 − b2 − t22 )(αk2 eik2 + βk2 e−ik2 ) + bt2 (αk2 ei2k2 + βk2 e−i2k2 )
= λk2 (αk2 eik2 + βk2 e−ik2 )
(A.3)
bt2 (αk2 ei(z2 −1)k2 + βk2 e−ik2 (z2 −1) ) + (−∆2 − b2 − t22 )(αk2 eik2 z2 + βk2 e−ik2 z2 )
+ bt2 (αk2 eik2 (z2 +1) + βk2 e−ik2 (z2 +1) )
= λk2 (αk2 eik2 z2 + βk2 e−ik2 z2 ),
1 < z2 < M
(A.4)
bt2 (αk2 eik2 (M −1) + βk2 e−ik2 (M −1) ) + (−∆2 − b2 )(αk2 eik2 M + βk2 e−ik2 M )
= λk2 (αk2 eik2 M + βk2 e−ik2 M )
(A.5)
are all satisfied. Equation (A.4) is solved by choosing
λk2 = bt2 (eik2 + e−ik2 ) + (−∆2 − b2 − t22 ),
which reduces the other two conditions to
bt2 (αk2 + βk2 ) = 0,
ik2 (M +1)
bt2 (αk2 e
(A.6)
−ik2 (M +1)
+ βk2 e
)−
t22 (αk2 eik2 M
−ik2 M
+ βk2 e
) = 0. (A.7)
The first condition implies βk2 = −αk2 , which can be used to rewrite Eq. (A.7)
as
sin k2 (M + 1) = B(k1 ) sin k2 M
(A.8)
where for brevity we have introduced
B(k1 ) :=
|1 + t1 eik1 |2
t2
= t2
,
b(k1 )
1 − t21
(A.9)
(cf. Eq. (2.1.22)). We call Q+
M (k1 ) the set of the solutions of (2.1.23) with
ℜk2 ∈ [0, π].
Restricting to the critical case (2.1.17),
B(k1 ) = 1 −
2t1 t2
(1 − cos k1 ) =: 1 − κ(1 − cos k1 ),
1 − t21
(A.10)
so that 0<B(k1 )<1 for k1 ∈ DL . Equation (2.1.23) is equivalent to
tan k2 (M + 1) =
B(k1 ) sin k2
,
B(k1 ) cos k2 − 1
(A.11)
1118
G. Antinucci et al.
Ann. Henri Poincaré
which for 0 <B(k1 ) < 1 has a unique real solution in each interval In :=
π
1
M +1 (n + 2 , n + 1), n = 0, . . . , M − 1, since the left-hand side increases monotonically from −∞ to 0, while the right-hand side is negative and decreasing.
Thus, all M eigenvectors of B̃c+ (and, as a consequence, of B̃c− ) correspond to
real solutions of this form by
⎤
⎡
sin k2
:
⎢ sin 2k2 ⎥
2
⎥
⎢
=
u+
⎥
⎢
..
k2
NM (k1 , k2 ) ⎣
⎦
.
sin k2 M
⎤⎞
⎛
⎡
sin k2 M
:
⎜
⎢sin k2 (M − 1)⎥⎟
2
⎥⎟
⎢
⎜
⎥⎟ ,
⎜and, respectively, u−
⎢
..
k2 =
NM (k1 , k2 ) ⎣
⎦⎠
⎝
.
sin k2
(A.12)
where
NM (k1 , k2 ) := 2
M
sin2 k2 x = M +
x=1
1 1 sin(2M + 1)k2
−
2 2
sin k2
(A.13)
so that the eigenvectors are normalized.
To obtain Eq. (2.1.24), we note that since q satisfies (A.8), we have
sin [k2 (M + 1) ± k2 M ] = sin k2 (M + 1) cos k2 M ± sin k2 M cos k2 (M + 1)
= [B(k1 ) cos k2 M ± cos k2 (M + 1)] sin k2 M,
and so we can rewrite NM (k1 , k2 ) as
NM (k1 , k2 ) =
=
B(k1 )M cos k2 M − (M + 1) cos k2 (M + 1)
B(k1 ) cos k2 M − cos k2 (M + 1)
d
dk2
(B(k1 ) sin k2 M − sin k2 (M + 1))
B(k1 ) cos k2 M − cos k2 (M + 1)
.
(A.14)
We now return to Ãc . Equation (2.1.23) is equivalent to
and therefore
beik2 (M +1) − t2 eik2 M = be−ik2 (M +1) − t2 e−ik2 M ,
(A.15)
ik2 (M +1)
− t2 eik2 M )u−
(b − t2 τ T )u+
k2 ,
k2 = −(be
and
ik2 (M +1)
− t2 eik2 M )u+
(−b + t2 τ )u−
k2
k2 = (be
whenever k2 ∈ QM (k1 ). Combining this with the definition of Ãc (k1 ) in
Eq. (A.1), we see that
+
−i∆u+
uk2
k2
Ãc
,
=
0
−(beik2 (M +1) − t2 eik2 M )u−
k2
ik (M +1)
0
(be 2
− t2 eik2 M )u+
k2 ,
and
Ãc − =
(A.16)
u k2
i∆u−
k2
Vol. 23 (2022) Non-integrable Ising Models in Cylindrical Geometry
1119
+
0
uk2
or in other words, the change of variables induced by
and − puts
u k2
0
Ãc in block-diagonal form, with 2 × 2 blocks
g̃
−i∆
e−ik2 (M +1) (b − t2 eik2 )
−eik2 (M +1) (b − t2 e−ik2 )
i∆
1
−2it1 sin k1
e−ik2 (M +1) (1 − t21 )(1 − B(k1 )eik2 )
,
=
ik
(M
+1)
2
−ik
2
2
2
ik
(1 − t1 )(1 − B(k1 )e
)
2it1 sin k1
|1 + t1 e 1 | −e
−1
(k1 , k2 ) :=
(A.17)
recalling the definitions (2.1.14) and (2.1.22). This block-diagonalization
implies that
=
PfAc =
det g̃−1 (k1 , k2 ),
(A.18)
k1 ∈DL k2 ∈Q+ (k1 )
M
where explicitly, using the criticality condition (2.1.17),
det g̃−1 (k1 , k2 ) =
2(1 − t2 )2 (1 − cos k1 ) + 2(1 − t1 )2 (1 − cos k2 )
|1 + t1 eik1 |
2
.
(A.19)
Note that this determinant vanishes iff k1 = k2 = 0 mod 2π (in particular, it
is positive if k1 ∈ DL ).
Concerning the propagator, denoting the inverse of (A.17) by
g̃(k1 , k2 ) :=
:=
g̃++ (k1 , k2 ) g̃+− (k1 , k2 )
g̃−+ (k1 , k2 ) g̃−− (k1 , k2 )
1
D(k1 , k2 )
−e−ik2 (M +1) (1 − t21 )(1 − B(k1 )eik2 )
2it1 sin k1
,
eik2 (M +1) (1 − t21 )(1 − B(k1 )e−ik2 )
−2it1 sin k1
(A.20)
where D(k1 , k2 ) is defined as in (2.1.21) and, letting
φ̃k2 ,ω (k1 ) :=
M
φω,x2 (k1 )uω
k2 (z2 ),
z2 =1
we have, for
k1 , k1′
∈ DL ,
k2 , k2′
′
∈ Q+
M , and ω, ω ∈ {±},
(A.21)
φ̃k2 ,ω (k1 )φ̃k2′ ,ω′ (k1′ ) = −Lδk1 ,−k1′ δk2 ,k2′ g̃ωω′ (k1 , k2 ),
so that, in terms of φω,z = L1 k1 ∈DL e−ik1 z1 φω,z2 (k1 ),
′
1
ω′ ′
φω,z φω′ ,z′ = −
g̃ωω′ (k1 , k2 )e−ik1 (z1 −z1 ) uω
k2 (z2 )uk2 (z2 )
L
+
k1 ∈DL k2 ∈Q (k1 )
M
≡ gωω′ (z, z ′ ).
u±
k2 (z2 )
(A.22)
Then recalling the definition of
and the identity (A.15), Eqs. (2.1.18)
and (2.1.19) follow by writing out the sines in terms of complex exponentials
+
and relabeling the sum in terms of QM (k1 ) := Q+
M (k1 ) ⊔ (−QM (k1 )).
1120
G. Antinucci et al.
Ann. Henri Poincaré
B. Proof of Proposition 2.3
For the proof of items 1 and 2 it is convenient to start by proving their analogues for the infinite volume limit propagators.
[η]
(h)
[η]
Decay bounds on g∞ , g∞ . Recall that g∞ was defined in (2.2.9). We intend
to prove that, for all x ∈ Z2 ,
[η]
(z) ≤ C 1+r+s ×
∂1r ∂2s g∞
r!s!η −
e−|z|1
3+r+s
2
e−η
−1/2
|z|1
η ≥ 1,
0 ≤ η ≤ 1,
if
if
(B.1)
where ∂j is the discrete derivative with respect to the j-th coordinate. By using
(h)
(B.1) in the definition of g∞ , namely
⎧ 1
⎪
[η]
⎪
g∞
(z) dη,
if h = 0,
⎨
(h)
0
g∞ (z) = 2−2h
(B.2)
⎪
⎪
[η]
⎩
g∞ (z) dη, if h < 0,
2−2h−2
we obtain the analogue of (2.2.15),
(h)
∂1r ∂2s g∞
(z) ≤ C 1+r+s r!s!2(1+r+s)h e−c2
h
|z|1
,
(B.3)
for all z ∈ Z2 and h ≤ 0.
In order to prove (B.1), we start from the explicit expression of the function in the left side,
∂1r ∂2s g[η]
∞ (z)
dk1 dk2
e−i(k1 z1 +k2 z2 ) (e−ik1 − 1)r (e−ik2 − 1)s e−ηD(k1 ,k2 ) M (k1 , k2 )
,
=
(2π)2
[−π,π]2
(B.4)
where D(k1 , k2 ) is as in Eq. (2.1.21) (cf. Eq. (2.2.1)) and
M (k1 , k2 ) := D(k1 , k2 )ĝ(k1 , k2 )
−2it1 sin k1
−(1 − t21 )[1 − e−ik2 B(k1 )]
=
.
(1 − t21 )[1 − eik2 B(k1 )]
2it1 sin k1
(B.5)
Note that the integrand in Eq. (B.4) is periodic with period 2π both in k1 and
in k2 , and is entire in both arguments. Therefore, we can shift both variables
in the complex plane, k1 → k1 − ia and k2 → k2 − ib, to obtain
[η]
∂1r ∂2s g∞
(z) = e−az1 −bz2
e−i(k1 x1 +k2 x2 ) (e−ik1 −a − 1)r (e−ik2 −b − 1)s
[−π,π]2
−ηD(k1 −ia,k2 −ib)
×e
(B.6)
dk1 dk2
.
M (k1 − ia, k2 − ib)
(2π)2
Vol. 23 (2022) Non-integrable Ising Models in Cylindrical Geometry
1121
We now pick a = α sign(z1 ) and b = α sign(z2 ), with α = min{η −1/2 , 1}, and
take the absolute value, thus getting
2
2
[η]
(z) ≤ e−α|z|1 e−2η[(1−t1 ) +(1−t2 ) ](cosh α−1)
∂1r ∂2s g∞
×
e−ηD(k1 ,k2 ) (αeα + |e−ik1 − 1|)r
(B.7)
[−π,π]2
(αeα + |e−ik2 − 1|)s M0 (k1 , k2 , α)
dk1 dk2
,
(2π)2
where M0 (k1 , k2 , α) = max|a|=|b|=α M (k1 − ia, k2 − ib) and we used the fact
that, if |a| = |b| = α, then
|D(k1 − ia, k2 − ib)| ≥ D(k1 , k2 ) + 2[(1 − t1 )2 + (1 − t2 )2 ](cosh α − 1) (B.8)
and
|e−ik1 −a − 1| ≤ αeα + |e−ik1 − 1|,
|e−ik2 −b − 1| ≤ αeα + |e−ik2 − 1|.
(B.9)
Now, if η ≤ 1 and, therefore, α = 1, then (B.7) immediately implies
[η]
that ∂1r ∂2s g∞ (z) ≤ C 1+r+s e−|z|1 , as desired. If η ≥ 1 and, therefore, α = η −1/2 , we make the following observations: (i) the factor
2
2
−1/2
−1)
e−2η[(1−t1 ) +(1−t2 ) ](cosh η
is bounded from above uniformly in η; (ii)
if −π ≤ k1 , k2 ≤ π, then D(k1 , k2 ) ≥ c(k12 + k22 ), |e−ik1 − 1| ≤ C|k1 |,
|e−ik2 − 1| ≤ C|k2 | and M0 (k1 , k2 , η −1/2 ) ≤ C(η −1/2 + |k1 | + |k2 |). By using
these inequalities in (B.7), we find
−1/2
[η]
|z|1
(z) ≤ C 1+r+s e−η
∂1r ∂2s g∞
2
2
×
e−cη(k1 +k2 ) (η −1/2 + |k1 |)r (η −1/2 + |k2 |)s
(η
R2
−1/2
+ |k1 | + |k2 |) dk1 dk2 ,
(B.10)
and expanding the powers in the integrand we obtain a sum of Gaussian integrals which reduce to (B.1) for this case as well.
[η]
(h)
[η]
Decay bounds on gE , gE , and proof of items 1 and 2. Recall that gE (z, z ′ ) =
[η]
[η]
g[η] (z, z ′ ) − gB (z, z ′ ), with g[η] as in (2.2.7) and gB as in (2.2.10). We focus
′
on the case that z1 − z1 = ±L/2 (recall that in our conventions z1 , z1′ ∈
{1, . . . , L}), in which the function sL in (2.2.10) is equal to ±1; the comple[η]
mentary case, z1 − z1′ = ±L/2, in which gB (z, z ′ ) = 0, can be treated in a way
analogous to the discussion below, and is left to the reader. If z1 − z1′ = ±L/2,
by using the anti-periodicity of the propagator in the horizontal direction, we
can reduce without loss of generality to the case z1 − z1′ = perL (z1 − z1′ ) (i.e.,
[η]
[η]
−L/2 < z1 − z1′ < L/2), in which gB (z, z ′ ) = g∞ (z − z ′ ), and we shall do so
in the following. Therefore, in this case,
[η]
gE (z, z ′ ) =
k1 ∈DL k2 ∈QM (k1 )
1
[η]
[η]
G+ (k1 , k2 ; z, z ′ ) − G− (k1 , k2 ; z, z ′ )
2LNM (k1 , k2 )
1122
G. Antinucci et al.
Ann. Henri Poincaré
′
−g[η]
∞ (z − z ),
(B.11)
where, recalling the definitions of fη in (2.2.1) and of ĝ(k1 , k2 ) and ĝωω′ (k1 , k2 )
in (2.1.20),
[η]
′
′
[η]
′
′
G+ (k1 , k2 ; z, z ′ ) := e−ik1 (z1 −z1 ) e−ik2 (z2 −z2 ) fη (k1 , k2 )ĝ(k1 , k2 ),
G− (k1 , k2 ; z, z ′ ) := e−ik1 (z1 −z1 ) e−ik2 (z2 +z2 ) fη (k1 , k2 )
ĝ++ (k1 , k2 )
ĝ+− (k1 , −k2 )
,
ĝ−+ (k1 , k2 ) e2ik2 (M +1) ĝ−− (k1 , k2 )
(B.12)
which are entire functions of k1 , k2 , and 2π-periodic in both variables. We
intend to prove that, for all z, z ′ ∈ Λ,
3+|r |1
[η]
∂ r gE (z, z ′ ) ≤ C 1+|r |1 ×
r!η − 2 e−cη
′
e−c dE (z,z )
−1/2
dE (z,z ′ )
∗
1 ≤ η ≤ 2−2h ,
0 ≤ η ≤ 1,
if
if
(B.13)
where we recall that h∗ = −⌊log2 min{L, M }⌋. Recalling also the relationship
(h)
[η]
between gE and gE , this implies
(h)
∂ r gE (z, z ′ ) ≤ C 1+|r|1 r!e−c2
h
dE (z,z ′ )
2−2h
η −(3+|r |1 )/2 dη
2−2h−2
(1+|r |1 )h −c2h dE (z,z ′ )
≤ (C ′ )1+|r |1 r!2
e
,
(B.14)
′
(0)
for h∗ < h < 0, and ∂ r gE (z, z ′ ) ≤ C 1+|r|1 e−c dE (z,z ) for h = 0.
(h)
(h)
Recalling that g(h) = g∞ + gE and noting that dE (z, z ′ ) ≥ z − z ′ 1 ,
inequalities (B.3) and (B.14) also imply that g(h) satisfies a bound of the form
(2.2.15).
In order to prove (B.13), we start from (B.11). Recalling that QM (k1 )
is the set of roots of B(k1 ) sin k2 M − sin k2 (M + 1), with k2 ∈ (−π, π], that
[η]
NM (k1 , k2 ) is given by (A.14), and that G♯ (k1 , k2 ; z, z ′ ) are entire and 2πperiodic, for ♯ ∈ {±}, we can rewrite
k2 ∈QM (k1 )
1
=
2
>
C
1
[η]
G (k1 , k2 ; z, z ′ )
2NM (k1 , k2 ) ♯
B(k1 ) cos k2 M − cos k2 (M + 1) [η]
dk2
G (k1 , k2 ; z, z ′ )
,
B(k1 ) sin k2 M − sin k2 (M + 1) ♯
2πi
(B.15)
where C is the boundary of the rectangle in the complex plane of vertices
−π − ib, π − ib, π + ib, −π + ib, b > 0, traversed counterclockwise. We rewrite
R± (k1 , k2 )e±2ik2 (M +1)
B(k1 ) cos k2 M − cos k2 (M + 1)
= ∓i 1 + 2
,
B(k1 ) sin k2 M − sin k2 (M + 1)
1 − R± (k1 , k2 )e±2ik2 (M +1)
(B.16)
Vol. 23 (2022) Non-integrable Ising Models in Cylindrical Geometry
1123
where
1 − B(k1 )e∓ik2
,
1 − B(k1 )e±ik2
(B.17)
Rσ (k1 , k2 )e2iσk2 (M +1)
.
1 − Rσ (k1 , k2 )e2iσk2 (M +1)
(B.19)
R± (k1 , k2 ) :=
and using this and noting that the contributions of the left and right sides of
C in the integral in Eq. (B.15) cancel by the periodicity of the integrand, we
then have
1
[η]
G (k1 , k2 ; z, z ′ )
2NM (k1 , k2 ) ♯
k2 ∈QM (k1 )
π+iσb
dk2
[η]
=
(B.18)
Aσ (k1 , k2 )G♯ (k1 , k2 ; z, z ′ )
2π
σ=0,±1 −π+iσb
where A0 (k1 , k2 ) ≡ 1 and, if σ = ±1,
Aσ (k1 , k2 ) :=
We now need to sum (B.18) over k1 ∈ DL . Notice that the right side of (B.18)
is analytic in k1 in a sufficiently small strip around the real axis (quantitative
bounds on the width of the analyticity strip will follow) and is 2π-periodic
in k1 . Given any function F (k1 ) that is 2π-periodic and analytic in a strip of
width 2b > 0 around the real axis, we have
>
π+iσ′ b
dk1 F (k1 )
dk1
1
A′σ′ (k1 ) F (k1 )
=
,
F (k1 ) =
−ik1 L
L
2π
1
+
e
2π
′
−π+iσ b
C
′
k1 ∈DL
σ =0,±1
(B.20)
A′0 (k1 )
′
iσ ′ k1 L
A′σ′ (k1 )
iσ ′ k1 L
where
≡ 1 and, if σ = ±1,
= −e
/(1 + e
); moreover
C is the same contour defined after (B.15). Using (B.11), (B.18) and (B.20),
we obtain
dk1 dk2 r [η]
[η]
∂ r gE (z, z ′ ) = −
∂ G− (k1 , k2 ; z, z ′ )
(2π)2
[−π,π]2
+
♯=±
♯
∗
[η]
A′σ′ (k1 + iσ ′ b) Aσ (k1 + iσ ′ b, k2 + iσb) ∂ r G♯
σ,σ ′ =0,±
(k1 + iσ ′ b, k2 + iσb; z, z ′ ) ,
(B.21)
′
where the ∗ on the sum indicates the constraint that (σ, σ ) = (0, 0), and
[η]
∂ r G♯ (k1 , k2 ; z, z ′ ) = (e−ik1 − 1)r1,1 (eik1 − 1)r2,1 (e−ik2 − 1)r1,2
[η]
(e♯ik2 − 1)r2,2 G♯ (k1 , k2 ; z, z ′ ).
[η]
(B.22)
Now, by using the definition of G− and by proceeding as in the proof of (B.1),
we see that the first term in the right side of (B.21) admits the same bound as
[η]
g∞ (z − z ′ ), see (B.1), with the only difference that |z − z ′ |1 should be replaced
1124
G. Antinucci et al.
Ann. Henri Poincaré
by |z1 − z1′ | + min{z2 + z2′ , 2(M + 1) − z2 − z2′ } ≤ dE (z, z ′ ). Therefore, the first
term in the right side of (B.21) satisfies (B.13) as desired.
[η]
Let us now prove that the contribution to [∂ r gE (z, z ′ )]ωω′ from the second line of (B.21) satisfy (B.13). For this purpose, if σ ′ = 0, we shift k1 in the
complex plane as k1 → k1 −i b sign(z1 −z1′ ). If σ = 0, we shift k2 in the complex
plane as k2 → k2 − iτ b, with τ = ±, its specific valued depending on ♯ and
on the matrix element (ω, ω ′ ) we are looking at; more precisely, τ = τ♯,(ωω′ ) ,
with
⎧
′
⎪
⎨−sign(z2 − z2 ) if ♯ = +
τ♯,(ωω′ ) := −1
(B.23)
if ♯ = − and (ω, ω ′ ) = (−, −)
⎪
⎩
′
+1
if ♯ = − and (ω, ω ) = (−, −).
Once these complex shifts are performed, we bound the contribution to
[η]
[∂ r gE (z, z ′ )]ωω′ from the second line of (B.21) by the sum over ♯ and over
σ, σ ′ (with (σ, σ ′ ) = (0, 0)) of:
dk1 dk2 ′
Aσ′ (k1 + iσ̃ ′ b) Aσ (k1 + iσ̃ ′ b, k2 + iσ̃b)
2
(2π)
[−π,π]2
r [η]
[∂ G♯ (k1 + iσ̃ ′ b, k2 + iσ̃b; z, z ′ )]ωω′ ,
σ
τ♯,(ωω′ )
that, if σ ′ =
0, then
where σ̃ =
if σ = 0
, and σ̃ ′ =
if σ = 0
(B.24)
σ′
−sign(z1 − z1′ )
if σ ′ = 0
. Note
if σ ′ = 0
e−bL
.
(B.25)
1 − e−bL
If σ = 0, we recall that Aσ (k1 , k2 ) is given by (B.19), with Rσ (k1 , k2 ) as in
(B.17). We claim that, if b ≤ c0 , with c0 sufficiently small, and k1 , k2 real,
then
Rσ (k1 + iσ̃ ′ b, k2 + iσ̃b) ≤ eCb .
(B.26)
|A′σ′ (k1 + iσ̃ ′ b)| ≤
for some C > 0; this will be proved momentarily, after (B.29). We now pick
b = c0 min{1, η −1/2 }, so that, using (B.26), for σ =
0, M sufficiently large,
and k1 , k2 real,
eCb
e−2b(M +1) .
(B.27)
1 − e−bM
If we now use (B.25) and (B.27) in the second line of (B.21) and we estimate
[η]
the integral of ∂ r G♯ (k1 + iσ̃ ′ b, k2 + iσ̃b; z, z ′ ) via the same strategy used in
the proof of (B.1), see Eqs. (B.7) to (B.10), we find that, for η ≤ 1,
|Aσ (k1 + iσ̃ ′ b, k2 + iσ̃b)| ≤
′
(B.24) ≤ C 1+|r |1 e−c0 dE (z,z ) ,
(B.28)
∗
where dE was defined after (2.2.16), while, if 1 ≤ η ≤ 2−2h ,
(B.24) ≤ C 1+|r |1 e−c0 η
−1/2
≤ (C ′ )1+|r |1 e−c0 η
dE (z,z ′ )
−1/2
2
2
intR2 dk1 dk2 e−cη(k1 +k2 ) (|k1 | + |k2 | + η −1/2 )1+|r |1
dE (z,z ′ )
r !η −
3+|r |1
2
.
(B.29)
Vol. 23 (2022) Non-integrable Ising Models in Cylindrical Geometry
1125
This completes the proof that the terms in the second line of (B.21) satisfy
(B.13), with c = c0 , provided that the bound (B.26) holds.
Proof of (B.26). We will prove the following version of (B.26): if σ = 0, b ≤ c0
with c0 sufficiently small and a := σ̃ ′ b with σ̃ ′ = ±1, then
|Rσ (k1 +ia, k2 + iσb)|2 ≤ e2b (1 + C0 b),
(B.30)
where C0 can be chosen
0
/
C0 = max 8κ(1 + π 2 ), 128π 4 κ(1 − 2κ)−2 ,
(B.31)
and κ is the same as in (A.10). By using the definition of Rσ , (B.17), one sees
that (B.30) is equivalent to
e−2b + ρ2 − 2ρe−b cos(σk2 − β) ≤ (1 + C0 b) 1 + ρ2 e−2b − 2ρe−b cos(σk2 + β) ,
(B.32)
where ρ = ρ(k1 , a) := |B(k1 +ia)| and β = β(k1 , a) := Arg(B(k1 +ia)). As
shown below, if −π ≤ k1 ≤ π and |a|= b ≤ c0 with c0 sufficiently small, then
ρ ≤ 1 + κ[(1 + π 2 )b2 − k12 /π 2 ] ,
|β| ≤
4κ
|k1 | · b .
1 − 2κ
(B.33)
By rearranging the terms in the two sides, one sees that (B.32) is equivalent
to
(ρ2 − 1)(1 − e−2b ) ≤ 4ρe−b sin(σk2 ) sin β + C0 b[(1 − ρe−b )2
+2ρe−b 1 − cos(σk2 + β) ].
(B.34)
By using the first bound in (B.33), the fact that |a| = b ≤ c0 with c0 sufficiently
small, and the bound 1 − cos(σk2 + β) ≥ (σk2 + β)2 /π 2 valid for β sufficiently
small, it is straightforward to check that the left side of (B.34) is smaller or
equal than 4κb[(1+π 2 )b2 −k12 /π 2 ], while the right side is greater or equal than
b2
2
+ 2 ρe−b (|k2 | − |β|)2 .
−4ρe−b |k2 | · |β| + C0 b
2
π
Therefore, (B.34) is a consequence of
4κ(1 + π 2 )b3 + 4ρe−b |k2 | · |β| ≤ 4κbk12 /π 2 + C0
b3
2
+ C0 b 2 ρe−b (|k2 | − |β|)2 .
2
π
(B.35)
Now, the first term in the left side of (B.35) is smaller than the third term
in the right side, C0 b3 /2, because C0 ≥ 8κ(1 + π 2 ), see (B.31). By using the
second bound in (B.33) and the fact that ρe−b ≤ 1 for b small enough (thanks
to the first bound in (B.33)), we see that, if |k2 | ≤ 1−2κ
4π 2 |k1 |, then the second
term in the left side of (B.35) is smaller than the first term in the right side.
In the complementary case, |k2 | ≥ 1−2κ
4π 2 |k1 | (which implies, in particular, that
|k2 | ≥ 2|β|, thanks to the second bound in (B.33)), then the second term in
1126
G. Antinucci et al.
Ann. Henri Poincaré
16κ
the left side of (B.35) is bounded from above by 1−2κ
ρe−b b|k2 | · |k1 |, while the
−b
2
0b
last term in the right side is bounded from below by C
2π 2 ρe |k2 | ; now,
16κ
C0 b −b
ρe−b b|k2 | · |k1 | ≤
ρe |k2 |2
1 − 2κ
2π 2
⇔
|k2 | ≥
32π 2 κ
|k1 |,
C0 (1 − 2κ)
4
−2
which is verified for |k2 | ≥ 1−2κ
, see
4π 2 |k1 |, because C0 ≥ 128π κ(1 − 2κ)
(B.31). In conclusion, (B.35) is always verified and, as a consequence, (B.34)
(and, therefore, (B.30)) is, as desired.
We are left with proving the validity of (B.33) for |a|= b small enough.
By definition
ρeiβ = 1 − κ + κ cos k1 cosh a−iκ sin k1 sinh a,
(B.36)
so that, using 1 − cos k1 ≥ 2k12 /π 2 and the fact that, for |a|= b small, cosh a ≤
1 + b2 and | sinh a| ≤ 2b,
ρ ≤ 1 + κ(b2 − 2k12 /π 2 + 2b|k1 |).
Using 2b|k1 | ≤ π 2 b2 + k12 /π 2 , we get the first of (B.33). Finally, from (B.36),
we find
κ| sin k1 sinh a|
2κb|k1 |
|β| ≤
≤
.
1 − κ + κ cos k1 cosh a
1 − κ(2 + b2 )
Now, picking b2 smaller than (1 − 2κ)/2κ, we find that β satisfies the second
of (B.33).
Gram representation: proof of items 3 and 4. Recall that
′
(h)
∂ (s,s ) gωω′ (z, z ′ )
=
dη
′
1
[η]
∂ (s,s ) G+,ωω′ (k1 , k2 ; z, z ′ )
2LNM (k1 , k2 )
Ih
k1 ∈DL k2 ∈QM (k)
(s,s ′ ) [η]
G−,ωω′ (k1 , k2 ; z, z ′ ) ,
(B.37)
−∂
where I0 = [0, 1), and Ih = [2−2h−2 , 2−2h ) for all h∗ ≤ h < 0. We recall
′
[η]
that ∂ (s,s ) G♯ is given by (B.22), with s playing the role of (r1,1 , r1,2 ) and
s′ playing the role of (r2,1 , r2,2 ). In the following, we will exhibit a Gram
[η]
decomposition separately for the two terms in (B.37) corresponding to G+ and
[η]
G− , which will immediately imply a Gram decomposition for the combination
of the two.
We rewrite
′
[η]
1Ih (η) ∂ (s ,s ) G♯,ωω′ (k1 , k2 ; z, z ′ ) =
4
σ=1
(h)
γ̃♯,ω,s ,z (k1 , k2 , η)
∗
σ
(h)
γ♯,ω′ ,s ′ ,z′ (k1 , k2 , η)
σ
,
(B.38)
where 1Ih is the characteristic function of the interval Ih , and
(h)
(h)
γ♯,ω,s,z (k1 , k2 , η) σ , γ̃♯,ω,s,z (k1 , k2 , η) σ are the components of the following
Vol. 23 (2022) Non-integrable Ising Models in Cylindrical Geometry
4-vectors:
(h)
γ̃♯,+,s,z (k1 , k2 , η) = 1Ih (η) eik1 z1 +ik2 z2 (eik1 − 1)s1 (eik2 − 1)s2
⎡=
∗ ⎤
ĝ♯,++ (k1 , k2 )
⎢=
∗ ⎥
⎢
⎥
⎢ ĝ♯,+− (k1 , k2 ) ⎥ ,
⎢
⎥
⎣
⎦
0
0
(h)
γ̃♯,−,s,z (k1 , k2 , η) = 1Ih (η) eik1 z1 +ik2 z2 (eik1 − 1)s1 (eik2 − 1)s2
⎤
⎡
0
⎥
⎢
0
⎢=
∗ ⎥
⎥
⎢
⎢ ĝ♯,−+ (k1 , k2 ) ⎥ ,
⎣=
∗ ⎦
ĝ♯,−− (k1 , k2 )
3
fη (k1 , k2 )
3
fη (k1 , k2 )
(h)
γ♯,+,s,z (k1 , k2 , η) = 1Ih (η) eik1 z1 +♯ik2 z2 (eik1 − 1)s1 (e♯ik2 − 1)s2
⎡=
⎤
ĝ♯,++ (k1 , k2 )
⎢
⎥
0
⎢ =
⎥
⎣ ĝ♯,−+ (k, q) ⎦ ,
0
(h)
1127
γ♯,−,s,z (k1 , k2 , η) = 1Ih (η) eik1 z1 +♯ik2 z2 (eik1 − 1)s1 (e♯ik2 − 1)s2
⎡
⎤
0
=
⎢ ĝ♯,+− (k1 , k2 )⎥
⎥,
⎢
⎣
⎦
0
=
ĝ♯,−− (k1 , k2 )
3
fη (k1 , k2 )
3
fη (k1 , k2 )
where if ♯ = +, then ĝ+,ωω′ (k1 , k2 ), with ω, ω ′ ∈ {±}, are the components of
the 2×2 matrix ĝ+ (k1 , k2 ) ≡ ĝ(k1 , k2 ), see (2.1.20); if ♯ = −, then ĝ♯,ωω′ (k1 , k2 ),
with ω, ω ′ ∈ {±}, are the components of
ĝ++ (k1 , k2 )
ĝ+− (k1 , −k2 )
,
ĝ− (k1 , k2 ) ≡
ĝ−+ (k1 , k2 ) e2ik2 (M +1) ĝ−− (k1 , k2 )
=
cf. (C.6). The square roots g♯,ωω′ (k1 , k2 ) of the complex numbers g♯,ωω′ (k1 , k2 )
are all defined by the same (arbitrarily chosen) branch.
In conclusion, in light of (B.38), (B.37) can be rewritten as
′
(h)
∂ (s,s ) gωω′ (z, z ′ )
∞
dη
=
0
k1 ∈DL k2 ∈QM (k1 )
(h)
γ♯,ω′ ,s ′ ,z′ (k1 , k2 , η) σ
4
∗
1
(h)
♯ γ̃♯,ω,s,z (k1 , k2 , η) σ
2LNM (k1 , k2 ) σ=1
♯=±
(h)
(h)
(h)
(h)
=: γ̃+,ω,s,z ⊗ ê1 + γ̃−,ω,s,z ⊗ ê2 , γ+,ω′ ,s ′ ,z′ ⊗ ê1 − γ−,ω′ ,s ′ ,z′ ⊗ ê2
1128
G. Antinucci et al.
Ann. Henri Poincaré
(h)
(h)
≡ γ̃ω,s,z
, γω′ ,s ′ ,z′ ,
(B.39)
where in the last line ê1 , ê2 are the elements of the standard Euclidean basis of
R2 . We can adapt all of the preceding discussion to g(≤h) simply by replacing
Ih with [2−2h−2 , ∞); this concludes the proof of item 3.
In order to prove the bounds in item 4, we first note that the definitions
(h)
(h)
given above for γ̃ω,s,z , γω,s,z immediately imply
1
(h) 2 (h) 2
dη
γ̃ω,s,z , γω,s,z ≤
LN
(k
M 1 , k2 )
Ih
k∈DL k2 ∈QM (k)
(B.40)
ik
e 1 − 12s1 eik2 − 12s2 · ·fη (k1 , k2 )
|gωω′ (k1 , k2 )|.
ω,ω ′ =±
Now, recall that the set DL consists of points in [−π, π] that are equi-spaced
at a mutual distance 2π/L, and that the set QM (k) consists of points in
[−π, π] that are almost equi-spaced at a mutual distance π/(M + 1) (more
precisely, recall that there is exactly one point of QM (k) in every interval
1
π
M +1 (n + 2 , n + 1), n = 0, . . . , M − 1, and exactly one point in every interval
1
π
11
M +1 (−n − 1, −n − 2 ), n = 0, . . . , M − 1). Note also that NM (k1 , k2 ) ≥ M ,
and that the summand in (B.40) is continuous, so we can bound (B.40) by a
Riemann sum and obtain
(h) 2 (h) 2
dk1 dk2 eik1
dη
γ̃ω,s,z , γω,s,z ≤ C
Ih
[−π,π]2
2s2
2s1
|gωω′ (k1 , k2 )|
−1 eik2 − 1 fη (k1 , k2 )
≤ (C ′ )1+s1 +s2
Ih
dη
ω,ω ′ =±
dk1 dk2
[−π,π]2
2
2
min(1, |k1 |2s1 |k2 |2s2 e−cη(k1 +k2 ) (|k1 | + |k2 |))
≤ (C ′′ )1+s1 +s2 s1 !s2 !
min(1, η −s1 −s2 −3/2 ) dη
Ih
≤ (C ′′′ )1+2s1 +2s2 s1 !s2 ! 2h(1+2s1 +2s2 ) .
(B.41)
Similarly
(≤h) 2 (≤h) 2
′′ 1+s1 +s2
s1 !s2 !
γ̃ω,s,z , γω,s,z ≤(C )
∞
η −s1 −s2 −3/2 dη
2−2h−2
′′′ 1+2s1 +2s2
s1 !s2 ! 2h(1+2s1 +2s2 ) ,
≤ (C )
(B.42)
and these bounds constitute item 4.
11 In
fact, by Eq. (A.13) and the definition of QM (k1 ), one has NM (k1 , k2 ) − M = (1 −
B(k1 ) cos k2 )/(B 2 (k1 ) − 2B(k1 ) cos k2 + 1) ≥ 0.
Vol. 23 (2022) Non-integrable Ising Models in Cylindrical Geometry
1129
C. Proof of Proposition 2.9
Recall that Λ is the discrete cylinder of sides L = 2⌊a−1 ℓ1 /2⌋ and M + 1 =
⌊a−1 ℓ2 ⌋+1 and gscal (z, z ′ ) the scaling limit propagator (2.3.3) in the continuum
cylinder Λℓ1 ,ℓ2 of sides ℓ1 , ℓ2 . In order to emphasize its dependence upon the
sides of the cylinder, let us denote the scaling limit propagator in Λℓ1 ,ℓ2 by
gscal (ℓ1 , ℓ2 ; z, z ′ ). Note that, upon rescaling by ξ > 0, this propagator satisfies:
ξgscal (ℓ1 , ℓ2 ; ξ z, ξ z ′ ) = gscal (ξ −1 ℓ1 , ξ −1 ℓ2 ; z, z ′ ).
′
(C.1)
′
′
We will prove that, for any z, z ∈ Λ such
√ that z ′= z ′, and
√ any w, w ∈
′
Λa−1 ℓ1 ,a−1 ℓ2 such that w = w , w − z ≤ 2 and w − z ≤ 2,
2
2
2gc (z, z ′ ) − agscal (ℓ1 , ℓ2 ; aw, aw′ )2 ≤ C(min{L, M, z − z ′ })−2 , (C.2)
provided that min{L, M, z − z ′ } is sufficiently large. Proposition 2.9 readily
follows from (C.2), simply by rescaling by a−1 . Note that, thanks to (C.1),
ag
; aw, aw′ ) =√gscal (a−1 ℓ1 , a−1 ℓ2 ; w, w′ ); note also that |a−1 ℓ1 − L| ≤
√scal (ℓ1 , ℓ2−1ℓ
2 2 and |a 2 −M | ≤ 2. By using the explicit√expression of the scaling
limit
√
propagator (2.3.3) and the fact that w − z ≤ 2 and w′ − z ′ ≤ 2, we find
that gscal (a−1 ℓ1 , a−1 ℓ2 ; w, w′ ) − gscal (L, M + 1; z, z ′ ) ≤ C(min{L, M, z −
z ′ })−2 . Therefore, in order to prove (C.2), it is enough to show that, for
min{L, M, z − z ′ } large,
2
2
2gc (z, z ′ ) − gscal (L, M + 1; z, z ′ )2 ≤ C(min{L, M, z − z ′ })−2 , (C.3)
which is what we will prove )in the rest of this appendix.
∞
Recall that gc (z, z ′ ) = 0 g[η] (z, z ′ ) dη, with
1
[η]
G (k1 , k2 ; z, z ′ ), (C.4)
g[η] (z, z ′ ) =
♯
2LNM (k1 , k2 ) ♯
♯=±
k1 ∈DL k2 ∈QM (k1 )
[η]
G♯
where
were defined in (B.12). Similarly, gscal (L, M + 1; z, z ′ ) =
(L, M + 1; z, z ′ ) dη, with
[η]
gscal (L, M + 1; z, z ′ )
♯
:=
♯=±
π
k1 ∈ L
(2Z+1) k2 ∈ 2(Mπ+1) (2Z+1)
)∞
0
[η]
gscal
1
[η]
G
(k1 , k2 ; z, z ′ ),(C.5)
2L(M + 1) scal;♯
where d(k1 , k2 ) := (1 − t2 )2 k12 + (1 − t1 )2 k22 and
[η]
′
′
[η]
′
′
Gscal;+ (k1 , k2 ; z, z ′ ) :=e−ik1 (z1 −z1 )−ik2 (z2 −z2 )−ηd(k1 ,k2 )
−2it1 k1 −(1 − t21 )ik2
,
−(1 − t21 )ik2
2it1 k1
Gscal;− (k1 , k2 ; z, z ′ ) :=e−ik1 (z1 −z1 )−ik2 (z2 +z2 )−ηd(k1 ,k2 )
−2it1 k1
(1 − t21 )ik2
.
−(1 − t21 )ik2 e2ik2 (M +1) 2it1 k1
(C.6)
We rewrite
[η]
[η]
[η]
gscal (L, M + 1; z, z ′ ) − g[η] (z, z ′ ) = R1 (z, z ′ ) + R2 (z, z ′ ),
(C.7)
1130
G. Antinucci et al.
where
[η]
R1 (z, z ′ )
:=
♯=±
−
with BL,M :=
Moreover,
π
L (2Z
[η]
R2 (z, z ′ )
1
♯
2L(M + 1)
"
Gscal;♯ (k1 , k2 ; z, z ′ )
(k1 ,k2 )∈BL,M
π
2(M +1) (2Z
1
:=
♯
L
♯=±
−
#
[η]
G♯ (k1 , k2 ; z, z ′ )
(k1 ,k2 )∈DL,M
+ 1) ×
Ann. Henri Poincaré
k1 ∈DL
"
k2 ∈QM (k1 ,k2 )
,
(C.8)
+ 1), and DL,M := DL × D2(M +1) .
k2 ∈D2(M +1)
1
2(M + 1)
#
1
[η]
G (k1 , k2 ; z, z ′ ).
2NM (k1 , k2 ) ♯
(C.9)
The first remainder term. We consider the contribution to gscal (L, M +
[η]
1; z, z ′ ) − gc (z, z ′ ) from R1 first. Examining the definitions (B.12) and (C.6),
[η]
we see that, if (k1 , k2 ) ∈ DL,M , each matrix element of Gscal;♯ (k1 , k2 ; z, z ′ ) −
[η]
2
G♯ (k1 , k2 ; z, z ′ ) is bounded in absolute value by C|k|2 e−cη|k| , with |k|2 =
k12 + k22 , for some C, c > 0; if (k1 , k2 ) ∈ BL,M \ DL,M , each matrix ele2
[η]
ment of Gscal;♯ (k1 , k2 ; z, z ′ ) is bounded in absolute value by C|k|e−cη|k| .
[η]
These bounds are sufficient for performing the integral of R1 (z, z ′ ) over
η ≥ (min{L, M })2 . In fact, for such values of η,
1
2L(M + 1)
|k|2
|k|
2
(k1 ,k2 )∈BL,M
e−cη|k| ·
if max{|k1 |, |k2 |} < π
≤ Cη −2 ,
otherwise
(C.10)
for some C > 0, so that
∞
[η]
′
R1 (z, z )dη ≤ C
∞
(min{L,M })2
(min{L,M })2
η −2 dη ≤ C ′ (min{L, M })−2 .
(C.11)
[η]
In order to bound the contribution from the integral of R1 (z, z ′ ) over
[η]
η ≤ (min{L, M })2 , we need to rewrite R1 (z, z ′ ) as a suitable integral in the
complex plane, in analogy with what we did in Appendix B. More precisely,
by using (B.20) and its analogue for the sums over k2 , we find that the matrix
[η]
elements of R1 (z, z ′ ) can be rewritten as:
[η]
R1 (z, z ′ ) ωω′ =
♯
·
♯=±
·
∞+iσ1 b
−∞+iσ1 b
dk1
2π
σ1 ,σ2 =0,±
∞+iσ2 b
−∞+iσ2 b
dk2 ′
A (k1 ) A′′σ2 (k2 ) Gscal;♯ (k1 , k2 ; z, z ′ ) ωω′
2π σ1
Vol. 23 (2022) Non-integrable Ising Models in Cylindrical Geometry
−
π+iσ1 b
dk1
2π
−π+iσ1 b
π+iσ2 b
−π+iσ2 b
1131
!
dk2 ′
[η]
′
′′
A (k1 ) Aσ2 (k2 ) G♯ (k1 , k2 ; z, z ) ωω′ ,
2π σ1
(C.12)
where b will be conveniently fixed below, A′σ (k) was defined after (B.20) and
A′′σ is defined analogously: A′′0 (k) ≡ 1 and, if σ = ±, A′′σ (k) = −e2iσk(M +1) /(1+
e2iσk(M +1) ). We now proceed as described before and after (B.23): if σ1 = 0,
we shift k1 in the complex plane as k1 → k1 − ib sign(z1 − z1′ ); if σ2 = 0,
depending on the values of ♯, ω, ω ′ , we shift k2 → k2 − iτ♯,(ωω′ ) b, with τ♯,(ωω′ )
as in (B.23). Next, we combine the third line of (C.12) with the contribution
to the integral in the second line from the region max{|ℜk1 |, |ℜk2 |} ≤ π. After
[η]
these manipulations, we find that R1 (z, z ′ ) ωω′ can be further rewritten as
[η]
R1 (z, z ′ )
ωω ′
=
♯
σ1 ,σ2 =0,±
♯=±
∞+iσ̃1 b
−∞+iσ̃1 b
dk1
2π
∞+iσ̃2 b
−∞+iσ̃2 b
[η]
Gscal;♯ (k1 , k2 ; z, z ) ωω′ − G♯ (k1 , k2 ; z, z ′ ) ωω′
Gscal;♯ (k1 , k2 ; z, z ′ ) ωω′
·
′
dk2 ′
A (k1 ) A′′
σ2 (k2 ) ·
2π σ1
if max{|ℜk1 |, |ℜk2 |} ≤ π
if max{|ℜk1 |, |ℜk2 |} > π
(C.13)
where σ̃1 = −sign(z1 −z1′ ), if σ1 = 0, and σ̃1 = σ1 , otherwise; and σ̃2 = τ♯,(ωω′ ) ,
if σ2 = 0, and σ̃2 = σ2 , otherwise. We now pick b = η −1/2 and notice that, if
ℑk1 = σ̃1 η −1/2 and ℑk2 = σ̃2 η −1/2 , with η ≤ (min{L, M })2 , the integrand in
the right side of (C.12) is bounded in absolute value by
Ceη
−1/2
e−η
−1/2
z−z ′ −cη|k|2
|k|2 + η −1
|k| + η −1/2
·
if max{|ℜk1 |, |ℜk2 |} ≤ π
if max{|ℜk1 |, |ℜk2 |} > π
for some C, c > 0. Therefore, recalling that z − z ′ ≫ 1,
[η]
R1 (z, z ′ )
+
− 21 η −1/2 z−z ′
≤ Ce
"
2
dk e−cη|k| (|k|2 + η −1 )
[−π,π]2
2
#
1
dk e−cη|k| (|k| + η −1/2 ) ≤ C ′ e− 2 η
R2 \[−π,π]2
−1/2
z−z ′
′
(η −2 +e−c η η −3/2 ).
(C.14)
1
−1/2
′
′
z−z −c η
Note that e− 4 η
≤ Ce−c
2
η ≤ (min{L, M }) , we find:
0
(min{L,M })2
[η]
′′
z−z ′ 2/3
, so that, by integrating over
R1 (z, z ′ )dη ≤ C(z − z ′ −2 + z − z ′ −1 e−c
≤ C ′ z − z ′ −2 .
′′
z−z ′ 2/3
)
(C.15)
1132
G. Antinucci et al.
Combining this with (C.11), we find that
z ′ })−2 .
)∞
0
Ann. Henri Poincaré
[η]
R1 (z, z ′ )dη ≤ C(min{L, M, z−
The second remainder term. Let us now consider the contribution to
[η]
gscal (L, M + 1; z, z ′ ) − gc (z, z ′ ) from R2 . By using (B.18) and the analogue of
[η]
(B.20) for the sums over k2 in D2(M +1) , we rewrite R2 (z, z ′ ) as
1 π+iσb dk2
[η]
R2 (z, z ′ ) :=
♯
L
2π
♯=±
k1 ∈DL σ=± −π+iσb
[η]
× A′′σ (k2 ) − Aσ (k1 , k2 ) G♯ (k1 , k2 ; z, z ′ ),
(C.16)
where A′′σ and Aσ were defined after (C.12) and in (B.19), respectively. Note
that, for σ = ±,
A′′σ (k2 ) − Aσ (k1 , k2 ) = −
=−
e2iσk(M +1)
Rσ (k1 , k2 )e2iσk2 (M +1)
−
2iσk(M
+1)
1+e
1 − Rσ (k1 , k2 )e2iσk2 (M +1)
2(1 + Rσ (k1 , k2 ))e2iσk2 (M +1)
.
(1 + e2iσk2 (M +1) )(1 − Rσ (k1 , k2 )e2iσk2 (M +1) )
(C.17)
Recalling the definition (B.17) of Rσ , we have
2
2
2
1 − B(k1 ) cos(k2 )
≤ C |k1 | + |k2 | + b
|1 + Rσ (k1 , k2 )| = 2
iσk
1 − B(k1 )e 2
b
(C.18)
for |ℑk1 | ≤ σℑk2 = b positive and sufficiently small; and recalling the bound
(B.26) on Rσ , which remains valid with iσ̃b replaced by iσ̃b′ , |b′ | ≤ b, the
denominator of (C.17) is bounded from below as
|1 − Rσ (k1 , k2 )e2iσk2 (M +1) ||1 + e2iσk2 (M +1) | ≥ (1 − e−b(M +1))
2
(C.19)
for M larger than some constant and |ℑk1 | ≤ σℑk2 = b positive and sufficiently
small. We now proceed slightly differently, depending on whether η is larger
or smaller than (ℓ(z, z ′ ))2 , with ℓ(z, z ′ ) := max{min{L, M }, z − z ′ }.
The case of η smaller than (ℓ(z, z ′ ))2 . In this case, we rewrite (C.16) by
using (B.20); if σ ′ = 0, we perform the complex shift k1 → k1 − ib sign(z1 − z1′ ),
thus getting (letting σ̃ ′ = σ ′ , if σ ′ = ±, and σ̃ ′ = −sign(z1 − z1′ ), if σ ′ = 0)
π+iσ̃′ b dk1 π+iσb dk2
[η]
A′σ′ (k1 )
♯
R2 (z, z ′ ) :=
2π
′ b 2π
−π+iσ̃
−π+iσb
♯=± σ ′ =0,± σ=±
[η]
A′′σ (k2 ) − Aσ (k1 , k2 ) G♯ (k1 , k2 ; z, z ′ ).
(C.20)
We now bound the integrand by its absolute value, by using, in particular,
[η]
(C.17)–(C.18), and by estimating the matrix elements of G♯ in the same
way as we did several times above and in Appendix B. We thus get, for b =
Vol. 23 (2022) Non-integrable Ising Models in Cylindrical Geometry
1133
c0 min{1, η −1/2 } with c0 sufficiently small,
′
2
2
Ce−bz−z
(|k| + b)3 −η|k|2
2 [η]
′ 2
e
dk
2R2 (z, z )2 ≤
(1 − e−bL )(1 − e−bM )2 [−π,π]2
b
′
Cb4 e−bz−z
≤
.
(1 − e−bL )(1 − e−bM )2
(C.21)
If we now integrate this inequality with respect to η, for 0 ≤ η ≤ ℓ(z, z ′ ),
recalling that ℓ(z, z ′ ) = max{min{L, M }, z − z ′ }, we find, for z − z ′ 1,
1
(ℓ(z,z′ ))2 2
2
′
2 [η]
′ 2
e−c0 z−z dη
2R2 (z, z )2 dη ≤ C
0
0
+C
(min{L,M })2
η −2 e−c0 η
−1/2
z−z ′
dη
1
+1(z − z ′ > min{L, M })
z−z′ 2
−1/2
C
z−z ′
η −1/2 e−c0 η
dη
LM 2 (min{L,M })2
′
1
z − z ′
≤ C ′ e−c0 z−z +
+
z − z ′ 2
LM 2
′′
C
≤
.
(C.22)
(min{L, M, z − z ′ })2
The case of η larger than ℓ(z, z ′ ). In this case we go back to the representation (C.16) (no rewriting of the sum over k1 in terms of an integral in the
complex plane).
We proceed slightly differently for the diagonal and off-diagonal elements
[η]
of R2 . Let us begin with the diagonal terms. Note that the diagonal elements
[η]
of G♯ have the form
′
±2it1 e−ik1 (z1 −z1 ) e−ik2 Z2 e−ηD(k1 ,k2 ) sin k1
where Z2 is either z2 − z2′ , z2 + z2′ , or z2 + z2′ − 2M − 2. We thus see that each
[η]
diagonal element of R2 is given by a sum of four terms (due to the sums over
♯ and σ) of the form
±
2it1
L
π+iσb
−π+iσb
k1 ∈DL
−ik (z −z′ ) −ik Z −ηD(k ,k )
1
1
2 2
1
2
1 e
A′′
e
σ (k2 ) − Aσ (k1 , k2 ) e
dk2
2π
(C.23)
′
π+iσb
2(1 + Rσ (k1 , k2 ))eiσk2 Z2,σ sin k1 (z1 − z1′ )e−ηD(k1 ,k2 ) sin k1
2t1
=±
L k ∈D −π+iσb
(1 − Rσ (k1 , k2 )e2iσk2 (M +1) )(1 + e2iσk2 (M +1) )
sin k1
1
L
dk2
,
2π
where in passing from the first to the second line we used (C.17) and the
′
fact that Rσ (k1 , k2 ) is even in k1 . Moreover, in the second line, Z2,σ
is either
′
′
2(M + 1) − σ(z2 − z2 ), 2(M + 1) − σ(z2 + z2 ), or 2(M + 1)(1 + σ) − σ(z2 + z2′ );
1134
G. Antinucci et al.
Ann. Henri Poincaré
′
in any case, Z2,σ
≥ 2. We can then use this, together with Inequalities (C.18)
and (C.19) and the observation that | sin k1 (z1 − z1′ )| ≤ |k1 | · z − z ′ to obtain,
for ω = ± and b = c0 η −1/2 , with η ≥ ℓ(z, z ′ ),
C z − z′
[η]
R2 (z, z ′ ) ωω ≤
L
′
≤
k1 ∈DL
C z−z η
M2
π
dk2
−π
η 1/2 k12 + k22 + η −1
2
2
|k1 |2 e−cη(k1 +k2 )
(1 − e−c0 η−1/2 M )2
−3/2
(C.24)
,
and thus, recalling that ℓ(z, z ′ ) = max{min{L, M }, z − z ′ },
∞
Cz − z ′
C
[η]
′
R2 (z, z ) ωω dη ≤ 2
≤ 2,
(ℓ(z,z′ ))2
M ℓ(x, y)
M
(C.25)
which is of the desired order.
[η]
The off-diagonal elements of G♯ are equal, up to a sign, to
′
2(1 − t21 )e−ik1 (z1 −z1 ) e−ik2 Z2 e−ηD(k1 ,k2 ) (1 − B(k1 )e±ik2 ),
where as before Z2 is either z2 − z2′ , z2 + z2′ , or z2 + z2′ − 2M − 2. Noting that
R± (k1 , k2 ) = R∓ (k1 , −k2 ), we rewrite
π−ib
[η]
dk2
A′′− (k2 ) − A− (k1 , k2 ) G♯ (k1 , k2 ; z, z ′ )
2π
−π−ib
π+ib
[η]
dk2
(C.26)
A′′+ (k2 ) − A+ (k1 , k2 ) G♯ (k1 , −k2 ; z, z ′ )
=
2π
−π+ib
[η]
and so each off-diagonal element of R2 can be written as a sum of two terms
(due to the sum over ♯) of the form
π+ib
′
1
2(1 + R+ (k1 , k2 ))e2ik2 (M +1)
e−ik1 (z1 −z1 ) e−ηD(k1 ,k2 )
L k ∈D −π+ib (1 − R+ (k1 , k2 )e2ik2 (M +1) )(1 + e2ik2 (M +1) )
1
L
× e−ik2 Z2 (1 − B(k1 )e±ik2 ) + eik2 Z2 (1 − B(k1 )e∓ik2 )
up to uninteresting coefficients; then noting that
1
2
dk2
2π
e−ik2 Z2 (1 − B(k1 )e±ik2 ) + eik2 Z2 (1 − B(k1 )e∓ik2 )
= cos k2 Z2 − B(k1 ) cos k2 (Z2 ∓ 1)
= (1 − cos k2 ) cos k2 Z2 ∓ sin k2 sin k2 Z2 + [1 − B(k1 )] cos k2 (Z2 ∓ 1),
we obtain, for ω = ± and b = c0 η −1/2 with η ≥ ℓ(z, z ′ ), noting also that
|Z2 | ≤ 2M ,
2
C|Z | π
η 1/2 k12 + k22 + η −1
2
2
[η]
2
′
e−cη(k1 +k2 )
dk2
R2 (z, z ) ω,−ω ≤
−1/2 M 2
−c
η
0
L
(1 − e
)
−π
k ∈D
1
C −3/2
η
,
≤
M
L
(C.27)
Vol. 23 (2022) Non-integrable Ising Models in Cylindrical Geometry
from which
1135
∞
C
[η]
′
,
R2 (z, z )
dη ≤
(ℓ(z,z′ ))2
M ℓ(z, z ′ )
ω,−ω
(C.28)
which is again of the desired order. Combining this with (C.25) and (C.22),
2
) ∞ 2 [η]
we find that 0 2R2 (z, z ′ )2 dη ≤ C(min{L, M, z − z ′ })−2 . Together with
[η]
the bound on R1 , see the line after (C.15), this concludes the proof of (C.3)
and, therefore, of Proposition 2.9.
D. Non-interacting Correlation Functions in the Scaling Limit
In this appendix, we explain how to express the scaling limit of the noninteracting correlation function appearing in Theorem 1.1 in terms of the propagators studied in Appendix C, thus proving, in particular, [4, Eq.(1.12)]. For
notational simplicity, in this appendix we let t∗1 (λ) ≡ t1 and t∗2 (λ) ≡ t2 =
(1 − t1 )/(1 + t1 ). By using the Grassmann representation of Proposition 3.1 in
the case λ = 0, we find that, for the lattice of unit mesh and any m-tuple of
distinct edges x1 , . . . , xm , with m ≥ 2,
ǫx1 ; · · · ; ǫxm 0,t1 ,t2 ;Λ
=
∂m
log
∂Ax1 · · · ∂Axm
St1 ,t2 (Φ)+
DΦe
2
x∈BΛ (1−tj(x) )Ex Ax
A=0
(D.1)
(note that the expectation in the left side is the truncated one). Introducing
the rescaled energy observable εaℓ (z) := a−1 σz σz+aêℓ , rescaling the lattice by
a factor of a and passing over to the non-truncated expectation, we obtain
a
εl1 (z1 ) · · · εalm (zm ) 0,t ,t ;Λa
1 2
= a−m (1 − t21 )m1 (1 − t22 )m2 Ex(z1 ,l1 ) − Ex(z1 ,l1 ) · · · Ex(zm ,lm )
− Ex(zm ,lm ) ,
(D.2)
where, in the right side: Ex(z,1) = H z Hz+aê1 and Ex(z,2) = V z Vz+aê2 ; the
symbol (·) indicates normalized Grassmann measure
Sta1 ,t2
a
)
Sa
(Φ)
DΦe t1 ,t2
(·)
)
Sa
(Φ) ,
DΦe t1 ,t2
with
the same as (2.1.2) on the rescaled lattice Λ .
Recall the transformation (2.1.10) relating the variables {H z , Hz , V z ,
ω∈{±}
Vz }z∈Λ to {φω,z , ξω,z }z∈Λ , from which we see that, if x is a vertical
edge of endpoints z, z + aê2 , then Ex = φ+,z φ−,z+aê2 , while, if x is a horizontal edge of endpoints
z, z + aê1 , then
(with obvious notation) Ex =
s+ ∗(φ+ −φ− )(z) s− ∗(φ+ +φ− )(z +aê1 ) plus terms involving the ‘massive’
variables {ξω,z }z∈Λ,ω∈{±} .
The reader can convince herself that, for the purpose of computing the
limit a → 0+ of (D.2), in the right side of (D.2) we can freely replace Ex by the
following local expressions in the Grassmann ‘massless’ variables: φ+,z φ−,z , if
x is a vertical edge of endpoints z, z + aê2 (note that φ+,z φ−,z is obtained from
φ+,z φ−,z+aê2 by ‘localizing’ the second field at the same position of the first
1136
G. Antinucci et al.
Ann. Henri Poincaré
one); and (1 + t1 )−2 (φ+,z − φ−,z )(φ+,z + φ−,z ) = 2(1 + t1 )−2 φ+,z φ−,z , if x is a
vertical edge of endpoints z, z + aê1 (note that (1 + t1 )−2 (φ+,z − φ−,z
)(φ+,z +
φ−,z ) is obtained from s+ ∗ (φ+− φ− )(z) s− ∗ (φ+ + φ− )(z + aê1 ) by localizing s− ∗ (φ+ + φ− )(z + aê1 ) at z, and by replacing the non-local, exponentially decaying, kernels s± (z1 ) by their local counterparts, namely c0 δz1 ,0 ,
L
with c0 = limL→∞ y=1 s± (y) = (1 + t1 )−1 ). It is, in fact, easy to check that
the difference between the exact expression of Ex and such a ‘local approximations’ is of higher order in a and its contribution to the correlation function
vanishes in the limit a → 0. Therefore,
lim+ εal1 (z1 ) · · · εalm (zm ) 0,t1 ,t2 ;Λa
a→0
= lim+ a−m
a→0
2(1 − t21 )
(1 + t1 )2
m1
(1 − t22 )m2 : φ+,z1 φ−,z1 : · · · : φ+,zm φ−,zm :,
(D.3)
where : φ+,z φ−,z : denotes the difference φ+,z φ−,z − φ+,z φ−,z . Note that
2(1−t21 )
(1+t1 )2
= 2t2 . The Grassmann average in the right side of (D.3) can be
expressed in terms of the fermionic Wick rule or, equivalently, in terms of
the Pfaffian of the 2m × 2m anti-symmetric matrix Ma (z), whose elements,
labeled by the indices (1, +), (1, −), . . . , (m, +), (m, −), are equal to
φω,zi φω′ ,zj
if i = j,
a
M (z) (i,ω)(j,ω′ ) =
0
otherwise.
In view of Proposition 2.9, lima→0 a−1 φω,z φω′ ,z′ = gscal (z, z ′ ) ωω′ and,
therefore,
lim εal1 (z1 ) · · · εalm (zm ) 0,t ,t ;Λa = (2t2 )m1 (1 − t22 )m2 Pf(M(z)), (D.4)
a→0+
1
2
with
M(z) (i,ω),(j,ω′ ) =
gscal (zi , zj ) ωω′
0
if i = j,
otherwise,
(D.5)
Since gscal is covariant under rescaling, see (C.1), the scaling limit (D.4) is,
as well. Note that rescalings are, together with translations and parity, the
only conformal transformations from finite cylinders to finite cylinders or,
equivalently, from a finite circular annulus to a finite circular annulus: in
fact, it is well known [5,28] that an annulus {z ∈ C : r < |z| < R}
can be conformally mapped to another one only if the two annuli have
1
log(R/r); moreover, any automorphism of the annulus
the same modulus 2π
{z ∈ C : r < |z| < R} is either a rotation z → zeiθ or a rotation followed by
an inversion z → Rr/z. Equivalently, in terms of finite cylinders, this classical result of complex analysis implies that the only conformal transformations
from finite cylinders to finite cylinders are uniform rescaling, translations and
parity.
Vol. 23 (2022) Non-integrable Ising Models in Cylindrical Geometry
1137
References
[1] Abdesselam, A., Rivasseau, V.: Explicit fermionic tree expansions. Lett. Math.
Phys. 44, 77–88 (1998)
[2] Aizenman, M., et al.: Emergent planarity in two-dimensional Ising models with
finite range interactions. Invent. Math. 216, 661–743 (2019)
[3] Antinucci, G.: Interacting fermions on the half-line: boundary counterterms
and boundary corrections. PhD thesis, Università degli Studi Roma Tre (2016).
arXiv: 1701.05356
[4] Antinucci, G., Guiliani, A., Greenblatt, R.L.: Energy correlations of nonintegrable Ising models: the scaling limit in the cylinder (2020). arXiv:
2006.04458
[5] Astala, K., et al.: Schottky’s theorem on conformal mappings between annuli: a
play of derivatives and integrals. Contemp. Math. 455, 35–39 (2008)
[6] Benfatto, G., Mastropietro, V.: Renormalization group, hidden symmetries and
approximate ward identities in the XYZ model. Rev. Math. Phys. 13, 1323–1435
(2001)
[7] Brydges, D.: A short course on cluster expansions. In: Osterwalder, K., Stora, R.
(eds.) Phénomènes critiques, systèmes aléatoires, théories de jauge. Les Houches
Summer School Session, vol. 43, pp. 129–183. North-Holland, Amsterdam (1986)
[8] Brydges, D., Federbush, P.: A new form of the Mayer expansion in classical
statistical mechanics. J. Math. Phys. 19, 2064 (1978)
[9] Brydges, D.C., Kennedy, T.: Mayer expansions and the Hamilton–Jacobi equation. J. Stat. Phys. 48, 19–49 (1987)
[10] Chelkak, D., Smirnov, S.: Universality in the 2D Ising model and conformal
invariance of fermionic observables. Invent. Math. 189, 1–66 (2009)
[11] Chelkak, D., Cimasoni, D., Kassel, A.: Revisiting the combinatorics of the 2D
Ising model. Annales de l’Institut Henri Poincaré D 4, 309–385 (2017)
[12] Chelkak, D., Hongler, C., Izyurov, K.: Conformal invariance of spin correlations
in the planar Ising model. Ann. Math. 181, 1087–1138 (2015)
[13] Gallavotti, G.: Renormalization theory and ultraviolet stability for scalar fields
via renormalization group methods. Rev. Mod. Phys. 57, 471–562 (1985)
[14] Gallavotti, G., Nicolò, F.: Renormalization theory in four-dimensional scalar
fields, part I. Commun. Math. Phys. 100, 545–590 (1985)
[15] Gentile, G., Mastropietro, V.: Renormalization group for one-dimensional
fermions. A review on mathematical results. Phys. Rep. 352, 273–438 (2001)
[16] Giuliani, A., Mastropietro, V.: Anomalous universality in the anisotropic
Ashkin–Teller model. Commun. Math. Phys. 256, 681–735 (2005)
[17] Giuliani, A.: Conformal invariance and renormalization group. In: Markov Processes and Related Fields (to appear)
[18] Giuliani, A., Greenblatt, R.L., Mastropietro, V.: The scaling limit of the energy
correlations in non-integrable Ising models. J. Math. Phys. 53, 095214 (2012)
[19] Giuliani, A., Mastropietro, V.: Universal finite size corrections and the central
charge in non-solvable Ising models. Commun. Math. Phys. 324, 179–214 (2013)
[20] Giuliani, A., Mastropietro, V., Rychkov, S.: Gentle introduction to rigorous
renormalization group: a worked fermionic example. J. High Energy Phys. 2021,
26 (2021)
1138
G. Antinucci et al.
Ann. Henri Poincaré
[21] Giuliani, A., Mastropietro, V., Toninelli, F.L.: Height fluctuations in interacting
dimers. Annales de l’Institut Henri Poincaré - Probabilités et Statistiques 53,
98–168 (2017)
[22] Giuliani, A., Mastropietro, V., Toninelli, F.L.: Non-integrable dimers: universal fluctuations of tilted height profiles. Commun. Math. Phys. 377, 1883–1959
(2020)
[23] Greenblatt, R.L.: The Ising model on a cylinder: universal finite size corrections
and diagonalized action (2014). arXiv: 1409.1852
[24] Hongler, C.: Conformal invariance of Ising model correlations. PhD thesis, Université de Genève (2010)
[25] Hucht, A.: The square lattice Ising model on the rectangle I: finite systems. J.
Phys. A Math. Theor. 50, 065201 (2017)
[26] McCoy, B., Wu, T.: The Two-Dimensional Ising Model. Harvard University
Press, Harvard (1973)
[27] Samuel, S.: The use of anticommuting variable integrals in statistical mechanics.
I. The computation of partition functions. J. Math. Phys. 21, 2806 (1980)
[28] Schottky, F.: Über die conforme Abbildung mehrfach zusammenhängender
ebener Flächen. Journal für die reine und angewandte Mathematik 83, 300–351
(1877)
[29] Smirnov, S.: Conformal invariance in random cluster models. I. Holmorphic
fermions in the Ising model. Ann. Math. 172, 1435–1467 (2010)
Giovanni Antinucci
Section de mathématiques
Université de Genève
2-4 rue du Lièvre, 1211 Genève 4
Geneva
Switzerland
Alessandro Giuliani and Rafael L. Greenblatt
Dipartimento di Matematica e Fisica
Università degli Studi Roma Tre
L.go S. L. Murialdo 1
00146 Roma
Italy
Alessandro Giuliani
Centro Linceo Interdisciplinare Beniamino Segre
Accademia Nazionale dei Lincei, Palazzo Corsini
Via della Lungara 10
00165 Roma
Italy
e-mail: giuliani@mat.uniroma3.it
Vol. 23 (2022) Non-integrable Ising Models in Cylindrical Geometry
Present Address
Rafael L. Greenblatt
Scuola Internazionale Superiore di Studi Avanzati (SISSA)
Mathematics Area, Via Bonomea 265
34136 Trieste
Italy
Communicated by Vieri Mastropietro.
Received: April 21, 2021.
Accepted: August 23, 2021.
1139