TRIVIALITY OF THE SCALING LIMITS OF CRITICAL ISING AND

φ4 MODELS WITH EFFECTIVE DIMENSION AT LEAST FOUR

ROMAIN PANIS

Abstract. We prove that any scaling limit of a critical reflection positive Ising or φ4
model of effective dimension deff at least four is Gaussian. This extends the recent
breakthrough work of Aizenman and Duminil-Copin [ADC21]— which demonstrates the
corresponding result in the setup of nearest-neighbour interactions in dimension four—
to the case of long-range reflection positive interactions satisfying deff = 4. The proof
relies on the random current representation which provides a geometric interpretation
arXiv:2309.05797v1 [math.PR] 11 Sep 2023

of the deviation of the models’ correlation functions from Wick’s law. When d = 4,
long-range interactions are handled with the derivation of a criterion that relates the
speed of decay of the interaction to two different mechanisms that entail Gaussianity:
interactions with a sufficiently slow decay induce a faster decay at the level of the model’s
two-point function, while sufficiently fast decaying interactions force a simpler geometry
on the currents which allows to extend nearest-neighbour arguments. When 1 ≤ d ≤ 3
and deff = 4, the phenomenology is different as long-range effects play a prominent role.

Contents
1. Introduction 1
Ackowledgements 12
Notations 12
2. The Griffiths-Simon class of measures 12
3. Reflection positivity 13
4. Random current representation 24
5. Reflection positive Ising models satisfying deff > 4 28
6. Reflection positive Ising models in dimension d = 4 31
7. Reflection positive Ising models in dimension 1 ≤ d ≤ 3 satisfying deff = 4 51
8. Extension of the results to models in the Griffiths-Simon class 65
Appendix A. Spectral representation of reflection positive Ising models 73
Appendix B. The backbone representation of the Ising model 75
Appendix C. Properties of currents for models of the Ising-type in the GS class 76
Appendix D. Triviality and finiteness of the Bubble diagram 77
References 78

1. Introduction
1.1. Motivation. We are interested in ferromagnetic real-valued spin models on Zd that
arise in statistical mechanics. Mathematically, these models can be seen as probability
measures on spin configurations τ : Zd → R formally given by
1
Z  X  Y
⟨F (τ )⟩β = F (τ ) exp β Jx,y τx τy dρ(τx ), (1.1)
Z
x,y∈Zd x∈Zd

2020 Mathematics Subject Classification. 60G60, 60K35, 82B20, 82B27.

Research supported by the Swiss National Science Foundation and the NCCR SwissMAP.
1
2 R. PANIS

where β > 0 is the inverse temperature, Jx,y ≥ 0 are (possibly long-range) interactions,
Z is a normalisation constant, and dρ is a single-site probability measure. Of particular
interest to us are the Ising model which corresponds to choosing dρ to be the uniform
measure on {−1, +1}, and the φ4 model which corresponds to confining the spins in a
quartic potential given by
1 −gφ4 −aφ2
dρ(φ) = e dφ, (1.2)
zg,a
with g > 0, a ∈ R and zg,a a normalisation constant, and where dφ is the Lebesgue
measure on R.
The study of these models plays a key role in two distinct, yet interacting, research
areas: constructive Euclidean field theory and statistical mechanics.
Constructive Euclidean field theory aims at constructing random distributions on Rd ,
with a particular focus on interacting, or non-trivial, field theories. This contrasts with
the study of Gaussian fields, which have a trivial correlation function structure in the
sense that it is entirely determined by their two-point function via Wick’s law. A natural
attempt to build non-Gaussian field theories is to try to define a measure on the set of
functions Rd → R whose averages are given by
1
Z Y
⟨F (Φ)⟩ = F (Φ) exp (−H(Φ)) dΦx ,
Z d x∈R
with
Z h i
H(Φ) := A|∇Φ(x)|2 + B|Φ(x)|2 + P (Φ(x)) dx,
Rd
where A, B > 0 and P is an even polynomial of degree 4 with a strictly positive leading
coefficient. This choice corresponds to what would be the definition of the φ4 field theory
on Rd . Due to the lack of a natural Lebesgue measure on infinite dimensional spaces,
the above quantity is ill-defined. However, it is still possible to make sense of it using
a pair of ultraviolet (short distance) and infrared (long distance) cutoffs. Highlights of
this approach include the rigorous construction of the φ4 measure, with infrared cutoff,
in dimension two by Nelson [Nel66], and in dimension three by Glimm and Jaffe [GJ73].
These works were later extended to the infinite volume limit [MS76, GJ12]. A few years
after these first results, Aizenman [Aiz82] and Fröhlich [Frö82] showed that φ4 is not a
good candidate to construct interacting field theories when d ≥ 5. In their works, they
proved that any field obtained as a scaling limit of critical Ising or φ4 models in dimension
d ≥ 5 is Gaussian. These papers, and other subsequent works [Sok82, AG83, DCCF83,
GK85, FMRS87, HT87, BBS14], provided strong heuristics that the same result should
hold in dimension d = 4. It was not until very recently that these heuristics were confirmed
by the breakthrough work of Aizenman and Duminil-Copin [ADC21].
Constructive Euclidean field theory is also closely related to constructive quantum field
theory (CQFT). Indeed, the Osterwalder-Schrader Theorem [OS73, OS75] provides a way
to build quantum field theories in the sense proposed by Wightman [Wig56] from Euclidean
field theories. We refer to [GJ12, ADC21, Aiz21] for a more complete description of the
CQFT point of view.
From the perspective of statistical mechanics, the Ising model and the φ4 model are
among the simplest examples which exhibit a phase transition1 at a critical parameter
βc ∈ (0, ∞). As proved in [ADCS15, GPPS22], this phase transition is continuous for
reflection positive interactions2, and one of the main challenges of the field is to understand
the nature of their scaling limits at criticality.
1The conditions J has to satisfy for a phase transition to occur are recalled below.
2This result holds for all reflection positive interactions if d ≥ 3, and is restricted to some interactions
when d ∈ {1, 2}.
 TRIVIALITY OF CRITICAL ISING AND φ4 MODELS WITH deff ≥ 4 3

The connection between the Ising model and the φ4 model is predicted to be very
rich: they are believed to belong to the same universality class. Renormalisation group
heuristics (see [Gri70, Kad93] or the recent book [BBS19]) predict that at their respective
critical points many of their properties (e.g. critical exponents) coincide exactly. Hints
of these deep links where established by Griffiths and Simon in [SG73], where they show
that the φ4 model emerges as a certain near-critical scaling limit of a collection of mean-
field Ising models. This permits to transfer rigorously many useful properties of the Ising
model, such as correlation inequalities, to the φ4 model. In the other direction, the Ising
model can be obtained as a limit of φ4 using the following limit
δ−1 + δ1 1 2 2
= lim e−g(φ −1) dφ.
2 g→∞ zg,−2g
The high dimension triviality results mentioned above are related to the simplicity of
the critical exponents of these models, suggesting that for d ≥ 4, they must take their
mean-field values. Rigorous results in that direction have been obtained in [Aiz82, AG83,
AF86, AF88, BBS19, MPS23]. What appeared to be a negative result from the perspective
of constructive Euclidean field theory is positive in the framework of statistical mechanics
as it provides information at criticality for a wide class of non-integrable models.
The main step in the proof of Aizenman and Fröhlich in dimension d ≥ 5 is the derivation
of the so-called tree diagram bound through geometric representations, and the use of
reflection positivity [FSS76, FILS78] to argue that the four-point Ursell function’s scaling
limit always vanishes at criticality. Although initially presented in the case of nearest-
neighbour interactions Jx,y = 1|x−y|1 =1 , these methods are robust and extend to more
general (in particular long-range) reflection positive interactions. However, this is no
longer true in dimension four, where only the case of nearest-neighbour interactions is
treated [ADC21].
The interest in the study of long-range interactions comes from the fact that the rate of
decay of the interactions may change the effective dimension of the model by increasing it,
meaning that one can recover high-dimensional features in some well-chosen one, two or
three dimensional systems. An observation of this phenomenon was made for algebraically
decaying long-range interactions of the form 1/rd+α by Fischer, Ma and Nickel [FMN72]
using renormalisation group heuristics. They noted that the parameter α had the effect
of changing the value of the upper critical dimension3 into dc (α) = min(2α, 4), suggesting
that the effective dimension of the model should be given by deff (α) = d/(1 ∧ (α/2)) (see
Figure 1). This was later studied by Aizenman and Fernández [AF88], and lead to the
observation that some Ising models in dimension 1 ≤ d ≤ 3 present trivial scaling limits
at criticality, which is not expected in the case of nearest-neighbour interactions4. Other
rigorous results were obtained through lace expansion methods [HvdHS08, CS15, CS19].
Conversely, if the interaction decays fast enough, the upper critical dimension of the model
is unchanged (this corresponds to α ≥ 2 in the example above). The prediction is that
only two situations may occur: either the interaction decays very fast and we expect to
fall into the universality class of the nearest-neighbour models, or the decay is exactly fast
enough for additional logarithmic corrections to appear. The latter scenario was shown to
occur [CS19] in dimension d ≥ 4, for (sufficiently spread out) interactions decaying such
as 1/rd+2 .

3That is, the dimension above which the system simplifies drastically, adopting Gaussian features.
This is also the dimension above which the bubble diagram converges, which is an indicator of mean-field
behaviour as recalled below.
4Indeed, non-triviality of the nearest-neighbour Ising model has been proven for d = 2 in [Aiz82], while
the case d = 3 remains open. Let us mention that the recent conformal bootstrap approach to the study
of the critical 3d Ising model strongly supports this conjecture [RSDZ17].
4 R. PANIS

deff α

2d
deff = α Non-trivial
2

deff = d Trivial
1
d
1
2

0
2
α 1 2 3 4 d

Figure 1. Left: The graph of α 7→ deff (α) for the interaction J given
by Jx,y = C|x − y|−d−α1 (for d ≥ 2). The transition between the regime
deff (α) > d and deff = d occurs at α = 2. At this point, we expect logarith-
mic corrections at the level of the decay of the critical two-point function.
Right: A summary of the expected behaviour of the critical scaling limits
for the same interaction J. The red region (including segments and points),
correspond to interactions which are expected to yield trivial scaling limits.
The results of this paper concern the study of the “marginal” cases that
separate the two phases.

Finally, let us briefly mention that long-range interactions of the above type have been
used to conduct rigorously the so-called “(dc − ε)-expansions”— motivated by Wilson and
Fisher through renormalisation group heuristics [WF72]— which give a precise under-
standing of the critical exponents of these models below the upper critical dimension, see
[FMN72, SYI72, BDH98, Sla18].
The goal of this paper is threefold. First, we prove that reflection positive Ising or
4
φ models with effective dimension strictly above four are trivial. This revisits some of
the results of [Aiz82, AF88] together with the notion of effective dimension, and provides
explicit examples of trivial models in dimensions one, two, and three. Second, we extend
the results of [ADC21] to near-critical and critical reflection positive Ising and φ4 models
in dimension d = 4 beyond the nearest-neighbour case. In particular, this case contains
algebraically decaying interactions (as above) with α > 0. The result was already known
[AF88] for α ∈ (0, 2) but is new in the case α ≥ 2. Third, we prove triviality of the scaling
limits of one, two, and three dimensional reflection positive Ising models with effective
dimension four. This is the main novelty of the paper. Such examples of models can be
obtained choosing α = d/2 above for 1 ≤ d ≤ 3. Our results apply to a wide class of
single-site measures called the Griffiths-Simon class of measures (see Section 2), which in
particular contains the examples mentioned below (1.1), and which can be recovered from
weak limits of Ising-type single site measures.
As in [Aiz82, ADC21], we use the random current representation of the Ising model
which enables, by means of the switching lemma, to express the correlation functions’
deviation from Wick’s law in terms of intersection probabilities of two independent random
currents with distanced sources.
When d = 4, two situations may occur. First, the interaction’s decay may be “slow”, in
which case we observe a decay of the model’s two-point function which is slightly better
than the one obtained for nearest-neighbour interactions (see Corollary 3.10). We can
then conclude using the tree diagram bound obtained in [Aiz82]. In particular, this first
case contains reflection positive interactions of algebraic decay with α ∈ (0, 2]. Second, the
 TRIVIALITY OF CRITICAL ISING AND φ4 MODELS WITH deff ≥ 4 5

decay of the interaction may be “too fast”, in which case we observe no improvement at
the level of the decay of the two-point function. As explained above, this case corresponds
to the situation where we expect the model to behave like a nearest-neighbour one, this
corresponds to choosing α > 2 above. We follow the strategy of [ADC21] and improve the
tree diagram bound. The proof goes by arguing that in dimension four, just like random
walks, if two independent random currents intersect at least once, they must re-intersect
a large number of times. By the mean of a multi-scale analysis, the authors of [ADC21]
showed that intersections occur with large probability in a density of (well-chosen) scales.
This essentially required three tools: regularity properties for the model’s two-point func-
tion, a proof of the fact that intersections happen with (uniform) positive probability on
all scales, and a mixing statement which allows to argue that intersections at different
scales are roughly independent events. However, in the case of long-range interactions,
these steps fail. Indeed, the extension of the proof to the general setup requires an adap-
tation of the reflection positivity arguments to the case of arbitrary interactions which
builds on a different view point on the spectral analysis of these models (see Section 3 and
Appendix A). This viewpoint was already introduced in [BC96, Ott19]. Then, long-range
interactions may have the effect of making intersections less likely as it becomes possible
to “jump” scales. Finally, long-range interactions may create more dependency between
pieces of the current at different scales. We solve these problems by arguing that the
currents do not jump above a (4 − ε)-dimensional annulus with very high probability (see
Section 6.2). As it turns out, this is enough to (essentially) recover the same geometric
properties of currents as in the nearest-neighbour case.
When 1 ≤ d ≤ 3 and deff = 4, the above improvements are not sufficient. The main
reason is that in this precise regime, the decay of the interaction is too slow to exclude
jumps above (d − ε)-dimensional annuli (see Remark 6.10). This additional difficulty is
treated by going one step further in the analysis of the currents (see Section 7.2). As a
byproduct of our methods, we obtain a (quantitative) mixing statement that is valid for
all models of effective dimension at least four (see Section 7.3).
To study the near-critical regime, it is important to introduce a typical length below
which the model essentially behaves like a critical one. It is tempting to use the correlation
length ξ(β) defined for β < βc , by
−1
log⟨τ0 τne1 ⟩β

ξ(β) := − lim .
n→∞ n

However, this quantity is not relevant in the case of long-range interactions since one may
have ξ(β) = ∞ for all β < βc (see [NS98, Aou21, AIOV21]). Another contribution of
this paper is the introduction of the sharp length L(β) (defined in Section 3.6), whose
definition is inspired by [DCT16].
We first prove an improved tree diagram bound for the Ising model, and then extend it
to the φ4 model (and more generally every model in the Griffiths-Simon class) using its
viewpoint as a generalised Ising model.
Let us mention that we also expect a direct analysis, meaning at the level of the φ4
model and without any mention of the Ising model, to be possible with the use of the
random tangled current representation of φ4 recently introduced in [GPPS22].

1.2. Definitions and statement of the results. We start by stating the results for the
case of the Ising model.

1.2.1. Results for the Ising model. In what follows, Λ is a finite subset of Zd . Let J =
(Jx,y ){x,y}⊂Zd be an interaction (or a collection of coupling constants) and h ∈ R. For
6 R. PANIS

σ = (σx )x∈Λ ∈ {±1}Λ , introduce the Hamiltonian

X X
HΛ,J,h (σ) := − Jx,y σx σy − h σx ,
{x,y}⊂Λ x∈Λ

and define the associated finite volume Gibbs equilibrium measure ⟨·⟩Λ,J,h,β at inverse
temperature β ≥ 0 to be the probability measure under which, for each F : {±1}Λ → R,
1 X

⟨F ⟩Λ,J,h,β := F (σ) exp −βHΛ,J,h (σ) ,
Z(Λ, J, h, β)
σ∈{±1}Λ

where
X

Z(Λ, J, h, β) := exp −βHΛ,J,h (σ) ,
σ∈{±1}Λ

is the partition function of the model. We make the following assumptions on the inter-
action J:
(A1) Ferromagnetic: For all x, y ∈ Zd , Jx,y ≥ 0,
(A2) Locally finite: For any x ∈ Zd ,
X
|J| := sup Jx,y < ∞,
x∈Zd y∈Zd

(A3) Translation invariant: For all x, y ∈ Zd , Jx,y = J0,y−x ,

(A4) Irreducible: For all x, y ∈ Zd , there exist x1 , . . . , xk ∈ Zd such that

Jx,x1 Jx1 ,x2 . . . Jxk−1 ,xk Jxk ,y > 0,

(A5) Reflection positive: see Section 3.

We postpone the definition of reflection positivity to Section 3.1, but to fix the ideas the
reader might keep in mind the following examples5 of interactions which satisfy (A1)-(A5):
(1) (nearest-neighbour interactions) Jx,y = C 1|x−y|1 ,1 for C > 0,
(2) (exponential decay / Yukawa potentials) Jx,y = C exp(−µ|x − y|1 ) for µ, C > 0,
(3) (algebraic decay) Jx,y = C|x − y|−d−α
1 for α, C > 0.
Using Griffiths’ inequalities [Gri67], one can obtain the associated infinite volume Gibbs
measure by taking weak limits of ⟨·⟩Λ,J,h,β as Λ ↗ Zd . We denote the limit by ⟨·⟩J,h,β .
For convenience, in what follows, we omit the mention of the interaction in the notation
of the Gibbs measures.
In dimensions d > 1, the model exhibits a phase transition for the vanishing of the
spontaneous magnetisation. That is, if

m∗ (β) := lim ⟨σ0 ⟩β,h ,

h→0+

then, βc := inf{β > 0, m∗ (β) > 0} ∈ (0, ∞). The above assumptions guarantee [Fis67]
that βc > 0 (in fact βc ≥ |J|−1 ), while Peierls’ celebrated argument [Pei36] yields the
bound βc < ∞. In dimension d = 1, the phase transition occurs [Dys69] under the
additional assumption that Jx,y ≍ |x − y|−1−α with α ∈ (0, 1]. We now assume that h = 0.
Our results concern the nature of the scaling limits at6, or near the critical parameter βc .

5Here, |.| refers to the ℓ1 norm on Rd .

1
6At a parameter β < β , finiteness of the susceptibility [ABF87] implies that the scaling limit is the
c
(Gaussian) white noise distribution [New80].
 TRIVIALITY OF CRITICAL ISING AND φ4 MODELS WITH deff ≥ 4 7

To determine the nature of the scaling limit, we look at the joint distribution of the
smeared observables 7 given for β > 0 and L ≥ 1 by
1 x
X  
Tf,L,β (σ) := p f σx ,
ΣL (β) x∈Zd L

where f ranges over the set C0 (Rd ) of continuous, real valued, and compactly supported
functions, and where
X
2 X
ΣL (β) := σx β
= ⟨σx σy ⟩β , with ΛL := [−L, L]d ∩ Zd .
x∈ΛL x,y∈ΛL

Definition 1.1. A discrete system as above is said to converge in distribution to a scaling

limit if the collection of random variables (Tf,L,β (σ))f ∈C0 (Rd ) converges in distribution (in
the sense of finite dimensional distributions) as L goes to infinity. Using Kolmogorov’s
extension theorem and the separability of C0 (Rd ), we can represent any scaling limit as a
random field.
Our first result concerns the study of models of effective dimension deff > 4. We
postpone the precise definition of effective dimension to Section 5 and illustrate this con-
cept using the example of algebraically decaying reflection positive interactions mentioned
above, i.e. Jx,y = C|x − y|1−d−α for α, C > 0. In that case, we will see that deff ≥ 1∧(α/2)
d

so that the hypothesis deff > 4 corresponds to d − 2(α ∧ 2) > 0.

Theorem 1.2. Let d ≥ 1. Let J be the interaction defined for x ̸= y ∈ Zd by Jx,y =
C0 |x − y|1−d−α where C0 , α > 0. We also assume that d − 2(α ∧ 2) > 0. There exists
C = C(C0 , d), γ = γ(d) > 0 such that for all β ≤ βc , L ≥ 1, f ∈ C0 (Rd ) and z ∈ R,
!
z2
⟨exp (zTf,L,β (σ))⟩β − exp ⟨Tf,L,β (σ)2 ⟩β
2
C(β −4 ∨ β −2 )∥f ∥4∞ rfγ z 4
!
z2
≤ exp ⟨T (σ)2 ⟩β ,
2 |f |,L,β Ld−2(α∧2)
 
where ∥f ∥∞ = supx∈Rd |f (x)| and rf = max{r ≥ 0, ∃x ∈ Rd , |x| = r, f (x) ̸= 0} ∨ 1 .
As a consequence, for β ≤ βc , every sub-sequential scaling limit (in the sense of Defini-
tion 1.1) of the model is Gaussian.
We now move the focus to the case d = 4. As explained in the introduction, and
discussed in Section 4, the Gaussian behaviour of the model can be seen at the level of
the four-point Ursell function [New75, Aiz82] defined for all x, y, z, t ∈ Zd by
U4β (x, y, z, t) := ⟨σx σy σz σt ⟩β − ⟨σx σy ⟩β ⟨σz σt ⟩β − ⟨σx σz ⟩β ⟨σy σt ⟩β − ⟨σx σt ⟩β ⟨σy σz ⟩β .
In dimension d > 4, Aizenman [Aiz82] was able to conclude the triviality of the scaling
limits using the tree diagram bound mentioned above,
|U4β (x, y, z, t)| ≤ 2
X
⟨σx σu ⟩β ⟨σy σu ⟩β ⟨σz σu ⟩β ⟨σt σu ⟩β , (1.3)
u∈Zd

7Note that for f = 1[−1,1]d ,

⟨Tf,L,β (σ)2 ⟩β = 1,
and more generally for f ̸= 0, one has 0 < cf ≤ ⟨Tf,L,β (σ)2 ⟩β ≤ Cf < ∞, which means that the following
quantity is bounded away from 0 and ∞ by constants that only depend on f . This indicates that this is
the scaling that is the most likely to yield interesting limits.
8 R. PANIS

together with the crucial input of reflection positivity, which implies (with the Messager-
Miracle-Solé inequalities [MMS77]) the infrared bound [FSS76, FILS78],
C
⟨σx σy ⟩βc ≤ , (1.4)
|x − y|d−2
where |.| denotes the infinite norm on Rd . As noticed in [Aiz82], the relevant question
is to see whether |U4β (x, y, z, t)|/⟨σx σy σz σt ⟩β vanishes or not, as the mutual distance
L(x, y, z, t) := minu̸=v∈{x,y,z,t} |u − v| between x, y, z and t goes to infinity but the dis-
tances between the pairs are all of the same order. The proof can be summed up by the
following (incomplete) argument: assume that β = βc and that the two-point function is of
comparable order for pairs of points at comparable distance; for a set of points x, y, z, t at
mutual distance of order L, the sum of the right-hand side of (1.3) is of order O(L8−3d ),
and we expect the four-point function ⟨σx σy σz σt ⟩βc to be of order the product of two
two-point functions, hence of order at least L4−2d . As a result, we have

|U4βc (x, y, z, t)|

= O(L4−d ). (1.5)
⟨σx σy σz σt ⟩βc
The above bound is clearly inconclusive in the case d = 4. However, in the case of nearest-
neighbour interactions, (1.3) was improved by a logarithmic factor to obtain Gaussianity.
The case of long-range interactions is more subtle since we do not necessarily expect
any improvement in the tree diagram bound in dimension 4. As it turns out, we do not
need any such improvement when the decay of the interaction is sufficiently slow so that
the decay of the model’s two-point function is faster than (1.4).
To determine whether this decay is fast enough or not, it is (almost) enough to look at
whether the following quantity is finite or not:
X
m2 (J) := |x|2 J0,x .
x∈Zd

When m2 (J) = ∞, the decay of the interaction is slow enough to conclude using 1.3.
Theorem 1.3. Let d = 4. Assume that J satisfies (A1)-(A5), and that m2 (J) = ∞.
Then, for all β ≤ βc , f ∈ C0 (Rd ) and z ∈ R,
!
z2
lim ⟨exp (zTf,L,β (σ))⟩β − exp ⟨Tf,L,β (σ)2 ⟩β = 0.
L→∞ 2
As a consequence, for β ≤ βc , every sub-sequential scaling limit of the model is Gaussian.
Remark 1.4. As we will see in Section 5, the rate of convergence to 0 can be expressed
in terms of
X
|x|2 J0,x .
|x|≤k

For instance, in the case of J defined by Jx,y = C|x − y|−d−2

1 , one can check that m2 (J) =
∞. The rate of convergence to 0 is then given by C/ log L.
We now discuss the case m2 (J) < ∞. In fact, we will have to restrict to interactions J
satisfying the following additional condition, which is slightly stronger:
(A6) There exist C, ε > 0 such that for all k ≥ 1,
X C
|x|2 J0,x ≤ . (1.6)
|x|=k
k 1+ε
 TRIVIALITY OF CRITICAL ISING AND φ4 MODELS WITH deff ≥ 4 9

As explained above, in this case we expect that the mechanism which leads to Gaussianity
is the same as for the nearest-neighbour case. Hence, we first prove an improved tree
diagram bound. The quantity L(β) was briefly mentioned above and will be introduced
in Section 3.6.
Theorem 1.5 (Improved tree diagram bound for d = 4). Let d = 4. Assume that J
satisfies (A1)-(A6). There exist c, C > 0 such that, for all β ≤ βc , for all x, y, z, t ∈ Z4
at mutual distance at least L of each other with 1 ≤ L ≤ L(β),
C
|U4β (x, y, z, t)| ≤
X
⟨σx σu ⟩β ⟨σy σu ⟩β ⟨σz σu ⟩β ⟨σt σu ⟩β ,
BL (β)c 4 u∈Z

where BL (β) is the bubble diagram truncated at distance L defined by

X
BL (β) := ⟨σ0 σx ⟩2β .
x∈ΛL

It is predicted, for an interaction J satisfying (A1)-(A6), that the bubble diagram

diverges at criticality. This improves the O(1) of (1.5) to a O(BL (β)−c ).
Remark 1.6. As noticed in [Aiz82, AF86], the bubble condition
B(βc ) < ∞, (1.7)
implies that some of the model’s critical exponents take their mean-field value. It is also
possible to show that the bubble condition (together with some monotonicity properties of
the two-point function), implies triviality of the scaling limits. We provide a proof of this
fact in Appendix D.
Remark 1.7. The reason why we restrict to interactions satisfying (A6) is technical and
will become more transparent in Section 6.2. All the examples of reflection positive inter-
actions we will encounter either satisfy (A6) or m2 (J) = ∞ (when d = 4). We strongly
believe that the methods below are sufficient to conclude provided we know a sufficiently
good estimate on quantities like (1.6).
One can check that the result still holds (see Remark 6.7) if one replaces (A6) by
(A6′ ) There exist C, ε > 0 such that for all k ≥ 1,
X C
|x|2 J0,x ≤ . (1.8)
|x|=k
k(log k)1+ε

A natural question would be to ask what is the optimal condition J should satisfy when
m2 (J) < ∞. For instance, is there is any reflection positive interaction that does not
satisfy (A6′ ) but for which m2 (J) < ∞? The methods below should be optimal whenever
the interaction J is such that the two-point function satisfies ⟨σ0 σx ⟩βc ≍ |x|−2 as |x| → ∞.
As pointed out in [Uch98, Section 7], the Green function associated to J has this decay
provided
1
X  
J0,x =O 2
.
|x|≥k
k log k

If we expect the two-point function to behave, at criticality, like the Green function of the
random walk associated to J (as suggested by the infrared bound in Proposition 3.4), then
the above appears to be the optimal condition. This leaves open the possibility of finding a
reflection positive positive interaction which satisfies (A6′ ) with ε = 0, but for which we
still expect the triviality result to hold.
With the improved tree diagram bound, we can obtain a formulation of triviality similar
to the one obtained in Theorem 1.3.
10 R. PANIS

Corollary 1.8. Let d = 4. Assume that J satisfies (A1)-(A6). There exist C, c, γ > 0
such that, for all β ≤ βc , 1 ≤ L ≤ L(β), f ∈ C0 (Rd ), and z ∈ R,
C∥f ∥4∞ rfγ z 4
! !
z2 z2
⟨exp (zTf,L,β (σ))⟩β − exp ⟨Tf,L,β (σ)2 ⟩β ≤ exp ⟨T (σ)2 ⟩β .
2 2 |f |,L,β (log L)c
As a consequence, for β = βc , every sub-sequential scaling limit of the model is Gaussian.
We now turn to the case of low dimensional models of effective dimension equal to
four. As above, we illustrate the result by focusing on the case of algebraically decaying
reflection positive interactions. The situation of interest corresponds to choosing α = d/2
for 1 ≤ d ≤ 3. More general versions of the following statements can be found in Section
7 (see Theorem 7.1 and Corollary 7.3).
For technical reasons (see Remark 7.2), the following statement is not as quantitative
as the above ones, and for simplicity we restrict it to β = βc .
Theorem 1.9 (Improved tree diagram bound for 1 ≤ d ≤ 3). Let 1 ≤ d ≤ 3. Let J be the
−3d/2
interaction defined for x ̸= y ∈ Zd by Jx,y = C0 |x − y|1 (i.e. α = d/2) where C0 > 0.
There exist C > 0 and a function ψ : R → R>0 which satisfies ψ(t) → ∞ as t → ∞, such
that, for all x, y, z, t ∈ Zd at mutual distance at least L of each other,
C
|U4βc (x, y, z, t)| ≤
X
⟨σx σu ⟩βc ⟨σy σu ⟩βc ⟨σz σu ⟩βc ⟨σt σu ⟩βc .
ψ(BL (βc )) d
u∈Z

Remark 1.10. In fact, for d = 1, the result is much stronger and we recover the improve-
ment of order O(BL (β)−c ) obtained when d = 4. The precise statements will be given in
Section 7.
We can still deduce a triviality statement from this improved tree diagram bound. It
involves the so-called renormalised coupling constant. We begin with a definition. For
σ > 0, we define the correlation length of order σ by: for β < βc ,
1/σ
|x|σ ⟨σ0 σx ⟩β
P
x∈Zd
ξσ (β) := ,
χ(β)
where χ(β) := x∈Zd ⟨σ0 σx ⟩β . As it turns out, the above quantity is well-defined when
P

Jx,y = C|x − y|−d−α as soon as σ < α (see for instance [NS98, Aou21, AIOV21]). Also,
by the results of [ABF87], one has ξσ (β) → ∞ as β → βc . For such a σ > 0, we introduce
another convenient measure of the interaction called the renormalised coupling constant
of order σ and defined for β < βc by:
1
U4β (0, x, y, z).
X
gσ (β) := −
χ(β)2 ξ σ (β)d
x,y,z∈Zd

The vanishing of the above quantity is known to imply triviality of the scaling limits of
the model (see [New75, Theorem 11] or [Aiz82, Sok82, AG83]).
Corollary 1.11. We keep the assumptions of Theorem 1.9. Then, for σ ∈ (0, d/2),
lim gσ (β) = 0.
β↗βc

As a consequence, for β = βc , every sub-sequential scaling limit of the model is Gaussian.

1.2.2. Results for the φ4 model. We now extend the above results to the φ4 model. These
results also extend to models in the Griffiths-Simon class of measures, whose definition is
postponed to the next section. We refer to Section 8 for the general statement of triviality
for these models.
 TRIVIALITY OF CRITICAL ISING AND φ4 MODELS WITH deff ≥ 4 11

We start with a proper definition of the φ4 model. Let ρ be given by (1.2). As for
the Ising model, the ferromagnetic φ4 model on Λ is defined by the finite volume Gibbs
equilibrium state: for F : RΛ → R,
1
Z

Y
⟨F (φ)⟩Λ,ρ,β = F (φ) exp −βHΛ,J (φ) dρ(φx ),
Z(Λ, ρ, β) x∈Λ
where Z(Λ, ρ, β) is the partition function and
X
HΛ,J (φ) := − Jx,y φx φy .
{x,y}⊂Λ

We call ⟨·⟩ρ,β the model’s infinite volume Gibbs measure. It is also possible to introduce a
critical parameter βc (ρ), together with a sharp length L(ρ, β), a four-point Ursell function
U4ρ,β , and a renormalised coupling constant gσ (ρ, β). The extension of the results concern-
ing models of effective dimension deff > 4 is quite straightforward and will be discussed in
Section 5 (see Remark 5.7). We focus on the results for 1 ≤ d ≤ 4 with deff = 4.
Theorem 1.12. Let d = 4. Assume that J satisfies (A1)-(A5), and that m2 (J) = ∞.
Then, for all β ≤ βc (ρ), f ∈ C0 (Rd ) and z ∈ R,
!
z2
lim ⟨exp (zTf,L,β (φ))⟩ρ,β − exp ⟨Tf,L,β (φ)2 ⟩ρ,β = 0.
L→∞ 2
As a consequence, for β = βc , every sub-sequential scaling limit of the model is Gaussian.
For the φ4 model, the tree diagram bound takes a slightly different form.
Theorem 1.13. Let d = 4. Assume that J satisfies (A1)-(A6). There exist c, C > 0
such that, for all β ≤ βc (ρ), for all x, y, z, t ∈ Z4 at mutual distance at least L of each
other with 1 ≤ L ≤ L(ρ, β),

|U4ρ,β (x, y, z, t)|

c X
B0 (ρ, β)
 X
≤C ⟨τx τu ⟩ρ,β βJu,u′ ⟨τu′ τy ⟩ρ,β ⟨τz τu ⟩ρ,β βJu,u′′ ⟨τu′′ τt ⟩ρ,β .
BL (ρ, β)
u∈Z4 u′ ,u′′ ∈Z4

Corollary 1.14. Let d = 4. Assume that J satisfies (A1)-(A6). Consider a φ4 model

on Z4 with coupling constants J. There exist C, c, γ > 0 such that, for all β ≤ βc (ρ),
1 ≤ L ≤ L(ρ, β), f ∈ C0 (Rd ), and z ∈ R,
!
z2
⟨exp (zTf,L,β (φ))⟩ρ,β − exp ⟨Tf,L,β (φ)2 ⟩ρ,β
2
C∥f ∥4∞ rfγ z 4
!
z2
≤ exp ⟨T (φ)2 ⟩ρ,β .
2 |f |,L,β (log L)c
As a consequence, for β = βc , every sub-sequential scaling limit of the model is Gaussian.
We may also extend the results obtained for the Ising model for 1 ≤ d ≤ 3 and deff = 4.
The corresponding modifications of Theorem 1.9 and Corollary 1.11 will be stated in
Section 8. Their main consequence for the φ4 model is stated below.
Theorem 1.15. Let 1 ≤ d ≤ 3. Let J be the interaction defined for x ̸= y ∈ Zd by
−3d/2
Jx,y = C|x − y|1 (i.e. α = d/2) where C > 0. Then, for σ ∈ (0, d/2),
lim gσ (ρ, β) = 0.
β↗βc (ρ)

As a consequence, for β = βc (ρ), every sub-sequential scaling limit of the model is Gauss-
ian.
12 R. PANIS

Organisation of the paper. In Section 2, we define the Griffiths-Simon class of single-

site measures to which our results apply. In Section 3, we recall the definition of reflection
positivity and present the main properties it implies for the models under consideration
(monotonicity of the two-point function, infrared bound, etc). The main result of this
section is the derivation of the existence of regular scales in Propositions 3.28 and 3.29.
In Section 4, we provide the basic knowledge on the random current representation of the
Ising model and explain the heuristics it provides on Gaussianity of the scaling limits.
We introduce the notion of effective dimension and prove a generalisation of Theorem 1.2
(see Theorem 5.5) to models with effective dimension deff > 4 in Section 5. Then, in
Section 6, we prove Theorems 1.3 and 1.5, together with Corollary 1.8. The handling of
long-range interactions is performed in Section 6.2. In Section 7, we prove more general
versions of Theorem 1.9 and Corollary 1.11 (see Theorem 7.1 and Corollary 7.3). The
main modifications in comparison to the d = 4 case are treated in Section 7.2. In Section
8, we extend the results to all the models introduced in Section 2.
In Appendix A, we provide the proofs of the main spectral tools we will use for reflection
positive models in the Griffiths-Simon class (see Theorem 3.11 and Proposition 3.16). In
Appendix B, we recall the definition of the backbone representation of the Ising model
which will be useful in our proofs. In Appendix C, we recall some useful bounds for
the probability of connectivity events for the random current representation of Ising-type
models in the Griffiths-Simon class. Finally, in Appendix D, we provide an alternative
proof of some of our results in the case where the bubble condition (1.7) is satisfied.

Ackowledgements
We warmly thank Hugo Duminil-Copin for suggesting the problem, for stimulating
discussions, and for constant support. We thank Sébastien Ott for pointing us references
concerning the spectral representation of the Ising model. We thank Lucas D’Alimonte,
Piet Lammers, Trishen Gunaratnam, and Christoforos Panagiotis for numerous valuable
comments and suggestions on a first version of the paper.

Notations
We write a point x ∈ Rd as x = (x1 , . . . , xd ) and denote by ej the unit vector with
xj = 1. We will use the following notations for the standard norms on Rd : for x ∈ Rd ,
|x|1 := |x1 | + . . . + |xd |, ∥x∥22 := x21 + . . . + x2d , and |x| := max1≤i≤d |xi |. Finally, for k ≥ 1,
denote Λk := [−k, k]d ∩ Zd .
If (an )n≥0 , (bn )n≥0 ∈ (R∗+ )N , we will write an ≳ bn (resp. an ≍ bn ) if there exists
C1 = C1 (d) (resp. C1 = C1 (d), C2 = C2 (d) > 0) such that for all n ≥ 1, bn ≤ C1 an
(resp. C1 an ≤ bn ≤ C2 bn ). We will also use Landau’s formalism and write an = O(bn )
(resp. an = o(bn )) if there exists C = C(d) > 0 such that for all n ≥ 1, an ≤ Cbn (resp.
limn→∞ an /bn = 0).

2. The Griffiths-Simon class of measures

In this section, we define the proper class of single-site measures to which our results
apply.

Definition 2.1 (The GS class of measures). A Borel measure ρ on R is said to belong to

the Griffiths-Simon (GS) class of measures if it satisfies one of the following conditions:
(i) there exist an integer N ≥ 1, a renormalisation constant Z > 0, and sequences
2
(Ki,j )1≤i,j≤N ∈ (R+ )N and (Qn )1≤n≤N ∈ RN such that for every F : R → R+
 TRIVIALITY OF CRITICAL ISING AND φ4 MODELS WITH deff ≥ 4 13

bounded and measurable,

 
N N
!
1
Z X X X
F (τ )dρ(τ ) = F Qn σn exp  Ki,j σi σj  ,
R Z n=1 i,j=1
σ∈{±1}N

(ii) the measure ρ can be presented as a weak limit of probability measures of the above
type, and it is of sub-Gaussian growth: for some α > 2,
Z
α
e|τ | dρ(τ ) < ∞.
R
Measures that satisfy (i) are said to be of the “Ising type”.
The following result was proved in [Gri69, SG73] to extend the Lee-Yang theorem,
together with Griffiths’ correlation inequalities, to the φ4 model on Zd . We sketch its
proof for sake of completeness.
Proposition 2.2 (The φ4 measure belongs to the GS class, [SG73]). Let g > 0 and a ∈ R.
The probability measure ρg,a on R given by
1 −gφ4 −aφ2
dρg,a (φ) = e dφ,
zg,a
where zg,a is a renormalisation constant, belongs to the GS class.
Proof. Let N ≥ 1. Let ge = (12g)−1/4 and a
e = 2age2 . Define the coupling constants
1 a
 
cN := geN −3/4 , 1− √
e
dN := .
N N
Define the Ising Gibbs measure µN on the complete graph KN , with Hamiltonian given
by
X
HN (σ) := −dN σi σj .
{i,j}⊂KN
PN
Let ρN be the law of the random variable ΦN := cN i=1 σi . Then, ρN converges weakly
to ρg,a . □

3. Reflection positivity
In this section, we define reflection positivity and gather all the properties it implies in
our setup (monotonicity, infrared bound, gradient estimates, etc). We refer to the review
[Bis09] or the original papers [FSS76, FILS78], for more information on this notion.
The end-goal is to derive the existence of regular scales for general reflection positive
interactions (see Propositions 3.28 and 3.29). Most of the results are classical and their
proofs in the case of nearest-neighbour ferromagnetic (n.n.f) models in the GS class were
already derived in [ADC21] using the spectral representation of the Ising model through
the lens of transfer matrices. This approach is not optimal in the most general setup. Our
viewpoint will be that of self-adjoint operators in infinite dimensional spaces, which allows
to import general results from [Hal13].
The following statements apply to both Ising and φ4 systems. To unify the notations,
we refer to the spin or field variables by the symbol τ , with an a-priori spin distribution
dρ(τ ) in the GS class which is supported on a set S ⊂ R. The expectation value functional
with respect to the Gibbs measure, or functional integral, for a system in a domain Λ, is
denoted ⟨·⟩Λ,ρ,β . We denote by ⟨·⟩ρ,β the state’s natural infinite volume limit. We also
denote by βc (ρ) the critical inverse temperature, and ξ(ρ, β) the correlation length. We
sometimes omit ρ in the notations when it is clear from context.
We will use the following notation for the model’s two-point function,
Sρ,β (x) := ⟨τ0 τx ⟩ρ,β .
14 R. PANIS

Also, introduce the finite-volume susceptibility, for L ≥ 1,

X
χL (ρ, β) := Sρ,β (x),
x∈ΛL

and define the susceptibility to be χ(ρ, β) := limL→∞ χL (ρ, β). Finally, recall that |J| =
P
x∈Zd J0,x .

3.1. Definition of reflection positivity. Let d ≥ 1. Consider the torus TL := (Z/LZ)d

with L ≥ 2 an even integer. The torus is endowed with a natural reflection symmetry along
hyperplanes H which are orthogonal to one of the lattice’s directions. The hyperplane
H either passes trough sites of TL or through mid-edges, and H divides the torus into
−
two pieces T+L and TL . The two pieces are disjoint for mid-edges reflections and satisfy
−
T+ ±
L ∩ TL = H for site reflections. Denote by A the algebra of all real valued functions f
±
that depend only on the spins in TL . Denote by Θ the reflection map associated with H;
it naturally acts on A± : for all f ∈ A± ,
Θ(f )(τ ) := f (Θ(τ )), ∀τ ∈ S TL .
d ×Zd
If J = (Jx,y )x,y∈Zd ∈ (R+ )Z , we can view it as an interaction J (L) on TL by setting
X
(L)
Jx,y := Jx,y+Lz .
z∈Zd
d d
Definition 3.1 (Reflection positivity). Let J = (Jx,y )x,y∈Zd ∈ (R+ )Z ×Z be an interac-
tion. The measure ⟨·⟩TL ,ρ,β = ⟨·⟩TL ,ρ,J (L) ,β is called reflection positive (RP) with respect to
Θ, if for all f, g ∈ A+ ,
⟨f · Θ(g)⟩TL ,ρ,β = ⟨Θ(f ) · g⟩TL ,ρ,β ,
and,
⟨f · Θ(f )⟩TL ,ρ,β ≥ 0.
We say that J is reflection positive if for all L ≥ 2 even, the associated measure ⟨·⟩TL ,ρ,β
is reflection positive with respect to Θ for all such reflections Θ.
Before discussing the interest of studying such interactions let us briefly mention some
examples (for more details see [FSS76, FILS78, AF88, Bis09]):
(i) (nearest-neighbour interactions) Jx,y = 1|x−y|1 =1 ,
(ii) (exponential decay / Yukawa potentials) Jx,y = C exp(−µ|x − y|1 ) for µ, C > 0,
(iii) (power law decay) Jx,y = C|x − y|−d−α1 for α, C > 0,
Qd 2 τ
(iv) Jx,y = 1/ i=1 (|xi − yi | + ai ) , for τi ≥ 0 and ai ∈ R.
i

The last example above can be found in [FILS78, AF88] and is an example of a d-
dimensional RP interaction constructed as the product of 1-dimensional RP interactions.
Furthermore, models of the GS class whose couplings are linear combinations with posi-
tive coefficients of the couplings mentioned above are also reflection-positive. Note that
all finite range reflection positive interaction satisfy J0,x = 0 whenever |x| > 1 (see p.12 of
[FILS78] for more information).
We now turn to the main consequences of reflection positivity. In what follows, we
consider reflection positive models with ρ in the GS class. We fix J satisfying (A1)-(A5).

3.2. Messager-Miracle-Solé inequalities. The Messager-Miracle-Solé inequality pro-

vides monotonicity properties for Sρ,β in the case of reflection positive interactions (see
Figure 2 for an illustration of this result).
 TRIVIALITY OF CRITICAL ISING AND φ4 MODELS WITH deff ≥ 4 15

Proposition 3.2 (MMS inequalities, [Heg77, MMS77, Sch77]). Let Λ be a region endowed
with reflection symmetry with respect to a plane P. Let A, B be two sets of points on the
same side of the reflection plane. If Θ is the reflection with respect to P,
* + * +
Y Y Y Y
τx τx ≥ τx τx .
x∈A x∈B Λ,ρ,β x∈A x∈Θ(B) Λ,ρ,β

Moreover, in the infinite volume limit Λ ↗ Zd this result can be extended to reflections with
respect to hyperplanes passing through sites or mid-edges; and to reflections with respect to
diagonal hyperplanes or more precisely, reflections changing only two coordinates xi and
xj which are sent to xi ± L and xj ∓ L respectively, for some L ∈ Z.
As a result, we get the following monotonicity property for reflection positive models in
the GS class.
Corollary 3.3 (Monotonicity of the two-point function). Let d ≥ 1 and β > 0. Then,
(i) for all 1 ≤ j ≤ d, the sequence (Sρ,β (kej ))k≥0 is decreasing,
(ii) for x ∈ Zd ,
Sρ,β ((|x|, 0⊥ )) ≥ Sρ,β (x) ≥ Sρ,β ((|x|1 , 0⊥ )), (MMS1)
where 0⊥ ∈ Zd−1 is the null vector. In particular, for all x, y ∈ Zd with d|x| ≤ |y|,
Sρ,β (x) ≥ Sρ,β (y). (MMS2)

P
P∗

P ∗ (y)

x P(x)

Figure 2. An illustration of Proposition 3.2. The two reflection planes

are represented by the green and red dashed lines. The MMS inequalities
state that ⟨τ0 τx ⟩ρ,β ≥ ⟨τ0 τP(x) ⟩ρ,β and ⟨τ0 τy ⟩ρ,β ≥ ⟨τ0 τP ∗ (y) ⟩ρ,β .

3.3. The infrared bound. The second interesting property we can get on Sρ,β using
reflection positivity is a quantitative estimate of its decay up to the critical point. This
relies on the infrared bound which is recalled below.
16 R. PANIS

We still work on the d-dimensional torus TL (with L even). In view of the model’s
translation invariance, it is natural to introduce the Fourier transform of Sρ,β ,
(L) X
Sbρ,β (p) := eip·x ⟨τx τy ⟩TL ,ρ,β
x∈TL
 d
(L)
where p ranges overs T∗L := 2πLZ ∩ (−π, π]d . As it turns out, Sbρ,β (p) can be expressed
in terms of the Fourier spin-wave modes, defined as
1 X
τbβ (p) := q eip·x τx .
d
(2L) x∈TL
Indeed, one has for p ∈ T∗L ,
(L)
Sbρ,β (p) = ⟨|τbβ (p)|2 ⟩TL ,ρ,β , (3.1)
(L)
so that in particular Sbρ,β (p) ≥ 0.
The following result was first proved and used in [FSS76, FILS78].
Proposition 3.4 (Infrared bound). For any p ∈ T∗L \ {0},
(L) 1
Sbρ,β (p) ≤ ,
2β|J|(1 − J(p))
b
J0,x
:= x∈Zd eip·x |J|
P
where J(p)
b .

Remark 3.5. Note that

p·x
X  
2
|J|(1 − J(p))
b =2 sin J0,x .
2
x∈Zd

Introduce for β < βc (ρ), and p ∈ (−π, π]d ,

X
Sbρ,β (p) := eip·x Sρ,β (x).
x∈Zd

Note that this quantity is well defined since x∈Zd Sρ,β (x) < ∞ for β < βc (ρ) as proved
P

in [ABF87]. The next result will use the celebrated Simon-Lieb inequality which we now
recall.
Lemma 3.6 (Simon-Lieb inequality, [Sim80, Lie04]). Let d ≥ 1. For every ferromagnetic
model in the GS class on Zd with translation invariant coupling constants, every β > 0,
every finite subset Λ of Zd containing 0, and every x ∈
/ Λ,
X
Sρ,β (x) ≤ Sρ,β (u)βJu,v Sρ,β (x − v).
u∈Λ
v ∈Λ
/

The following result extends the infrared bound to Sbρ,β .

Proposition 3.7. Let β < βc (ρ). Let p ∈ (−π, π]d . Then,
1
Sbρ,β (p) ≤ .
2β|J|(1 − J(p))
b

Proof. If β < βc (ρ), it is classical that there is only one infinite volume equilibrium state
that we denote ⟨·⟩ρ,β . Moreover, for all x ∈ Zd ,
⟨τ0 τx ⟩TL ,ρ,β −→ ⟨τ0 τx ⟩ρ,β . (3.2)
L→∞
 TRIVIALITY OF CRITICAL ISING AND φ4 MODELS WITH deff ≥ 4 17

Fix L even and take limits along sequences of the form (Lk )k≥1 so that T∗L ⊂ T∗Lk for all
k ≥ 1. For p ∈ T∗L , notice that by Fatou’s lemma and (3.2),
X
Sbρ,β (p) + χ(ρ, β) = (1 + cos(p · x))Sρ,β (x)
x∈Zd
 
(Lk ) k
≤ lim inf Sbρ,β (p) + χ(L ) (ρ, β)
1 k
≤ + lim inf χ(L ) (ρ, β),
2β|J|(1 − J(p))
b
k (Lk )
where χ(L ) (ρ, β) := Sbρ,β (0) and χ(ρ, β) := Sbρ,β (0). It then suffices to show that
k
χ(L ) (ρ, β)
goes to χ(ρ, β) as k goes to infinity. Using the Simon-Lieb inequality (as
in [DCT16]), we get that for any K ≥ 0,
k k
χ(L ) (ρ, β) ≤ φeρ,β (ΛK )χ(L ) (ρ, β) + χK (ρ, β), (3.3)
Jx,y ⟨τ0 τx ⟩ρ,β .
P
where φeρ,β (ΛK ) := β x∈ΛK
y ∈Λ
/ K
Then, since χ(ρ, β) < ∞, one has lim supK φeρ,β (ΛK ) = 0. In particular, for every ε > 0,
k 1
lim sup χ(L ) (ρ, β) ≤ χ(ρ, β).
1−ε
Which gives the result. □
As a first consequence of the above result, using that (1 − J(p)) b ≳ |p|2 near 0, we see
that if d ≥ 3, there exists C = C(d) > 0 such that for β < βc (ρ),
C
Z
⟨τ02 ⟩ρ,β = Sbρ,β (p)dp ≤ .
(−π,π]d β|J|
Note that this bound extends to βc (ρ) by continuity. Since β 7→ ⟨τ02 ⟩ρ,β is non-decreasing8,
we also get that for all β ≤ βc (ρ),
C
⟨τ02 ⟩ρ,β ≤ . (3.4)
βc (ρ)|J|
Proposition 3.4 together with the MMS inequalities also yield the following result.
Proposition 3.8 (Infrared bound). Let d ≥ 3. There exists C = C(d) > 0 such that for
every β ≤ βc (ρ), and every x ∈ Zd \ {0},
2
C e−∥p∥2
Z
Sρ,β (x) ≤ d dp.
β|J||x|d −π|x|,π|x| 1 − J(|p|/|x|)
b

In particular, for all x ∈ Zd \ {0},

C
Sρ,β (x) ≤ Sρ,βc (ρ) (x) ≤ . (IRB)
βc (ρ)|J||x|d−2
Proof. We prove the result for β < βc (ρ) and extend it to βc (ρ) with a continuity argument.
Using (MMS2), we get that for some C1 = C1 (d) > 0,
C1 X C1
Sρ,β (x) ≤ Sρ,β (y) ≤ χ (ρ, β). (3.5)
|x|d
−1 −1
|x|d |x|
y∈Ann(|x|(2d) ,|x|d )

We now observe that Proposition 3.7 provides a control on the finite volume susceptibility
χL (ρ, β). Let
2
e−(∥x∥2 /L) Sρ,β (x).
X
χ
eL (ρ, β) :=
x∈Zd

8This is a classical consequence of Griffiths’ inequalities.

18 R. PANIS

There exists C2 = C2 (d) > 0 such that, χL (β) ≤ C2 χ

eL (ρ, β). Using classical Fourier
identities, we get C3 = C3 (d) > 0 such that,
Z
2 ∥p∥2
eL (ρ, β) ≤ C3 Ld
χL (ρ, β) ≤ C2 χ e−L 2 Sbρ,β (p)dp.
(−π,π]d

With the change of variable u = pL, and Proposition 3.4,

1
Z Z
2 ∥p∥2 2
e−L 2 Sbρ,β (p)dp ≤ e−∥u∥2 Sbρ,β (u/L)du
(−π,π]d Ld (−Lπ,Lπ]d
2
C3 e−∥u∥2
Z
≤ du.
β|J|Ld (−Lπ,Lπ]d 1 − J(u/L)
b

The second part of the statement then follows from monotonicity in β of Sβ (x) together
with the observation that 1 − J(k)
b ≳ ∥k∥22 as k → 0. □

The preceding result essentially gives that the decay of the two-point function is gov-
erned by the behaviour of 1 − J(p)
b as p goes to 0. We have the following estimates for the
examples of RP interactions given above,
• Nearest-neighbour interactions or Yukawa potentials: as p → 0,
1 − J(p)
b ≍ |p|2 .
• Power law decay interactions: as p → 0,

2
 |p|
 if α > 2,
1
1 − J(p)
b ≍ |p|2 log |p| if α = 2,
 |p|α

if α ∈ (0, 2).
Together with Corollary 3.8, we get that for algebraic decay interactions,

 |x|−(d−2) if α > 2,
C 
⟨τ0 τx ⟩ρ,β ≤ |x|−(d−2) (log |x|)−1 if α = 2,
β|J|  −(d−α)
 |x| if α ∈ (0, 2).
Remark 3.9. The above bound is also valid for d = 2 in the case α ∈ (0, 2) and for d = 1
with α ∈ (0, 1), since in both cases |p|−α is locally integrable.
Proposition 3.8 also yields the following improvement on the bound of the model’s
two-point function when the interaction J has a slow decay.
Corollary 3.10. Let d = 4. Assume that m2 (J) = ∞. Then, as |x| → ∞,
1
 
⟨τ0 τx ⟩ρ,βc (ρ) = o .
|x|2
3.4. Spectral representation of reflection positive models and applications. In
this subsection, we generalise the spectral representation used in [ADC21]. The following
lines are inspired by [BC96, Ott19]. We fix a reflection positive model on Zd (d ≥ 1) with
an interaction J satisfying (A1)-(A5). The following result is classical for n.n.f models
in the GS class (see [GJ73, Sok82] or [ADC21, Proposition A.6]). We present its proof in
Appendix A.
Theorem 3.11 (Spectral representation). Let d ≥ 1. For every β ≤ βc (ρ) and every
function v : Zd−1 → C in ℓ2 (Zd−1 ), there exists a positive measure µv,β of finite mass
Z ∞ X
dµv,β (a) = vx⊥ vy⊥ Sρ,β ((0, x⊥ − y⊥ )) ≤ ∥v∥22 ⟨τ02 ⟩ρ,β ,
0
x⊥ ,y⊥ ∈Zd−1
 TRIVIALITY OF CRITICAL ISING AND φ4 MODELS WITH deff ≥ 4 19

such that for every n ∈ Z,

Z ∞
e−a|n| dµv,β (a).
X
vx⊥ vy⊥ Sρ,β ((n, x⊥ − y⊥ )) = (3.6)
0
x⊥ ,y⊥ ∈Zd−1

The following result provides a very useful representation of the Fourier transform of
Sρ,β .
Corollary 3.12. Let β < βc (ρ). Let p = (p1 , p⊥ ) ∈ (−π, π]d . There exists a measure
µp⊥ ,β such that
ea − e−a
X Z ∞
ip·x
Sbρ,β (p) := e Sρ,β (x) =
2 dµp⊥ ,β (a),
x∈Zd
0 E1 (p1 ) + ea/2 − e−a/2
where E1 (k) := 2(1 − cos k) = 4 sin2 (k/2).
Remark 3.13. The preceding result is still true under any permutation of the indices.
Proof. Consider for L ≥ 1, x⊥ ∈ Zd−1 and p⊥ ∈ (−π, π]d−1 ,
eip⊥ ·x⊥
vL (x⊥ ) := q 1 (d−1) ,
(d−1) x⊥ ∈ΛL
|ΛL |
(d−1)
where ΛL := [−L, L]d−1 ∩Zd−1 . We apply Theorem 3.11 to the sequence (vL )L≥1 which
yields that for L ≥ 1, and n ∈ Z
Z ∞
1 ip⊥ ·(x⊥ −y⊥ )
e−a|n| dµvL ,β (a).
X
(d−1)
e S ρ,β ((n, x⊥ − y ⊥ )) =
|ΛL | x ,y ∈Λ(d−1) 0
⊥ ⊥ L

Observe that one can rewrite the left-hand side above as

(d−1) (d−1)
|ΛL ∩ (ΛL − z⊥ )|
eip⊥ ·z⊥ Sρ,β ((n, z⊥ ))
X
(d−1)
.
z⊥ ∈Zd−1 |ΛL |
Thus, (use [ABF87] and the fact that β < βc (ρ) to justify that every sum converges
absolutely)
1 ip⊥ ·(x⊥ −y⊥ )
eip⊥ ·z⊥ Sρ,β ((n, z⊥ )).
X X
lim e S ρ,β ((n, x⊥ − y ⊥ )) =
L→∞ |Λ(d−1) | d−1
L (d−1) z⊥ ∈Z
x ,y ∈Λ
⊥ ⊥ L

The moments of a 7→ e−a under µvL ,β converge as L → ∞. Hence, the moment criterion
for the convergence of positive measures over bounded intervals (here [0, 1]) allows to
conclude existence of the (weak) limit limL→∞ µvL ,β =: µp⊥ ,β . We obtained that for any
n ∈ Z,
Z ∞
ip⊥ ·z⊥
e−a|n| dµp⊥ ,β (a).
X
e Sρ,β ((n, z⊥ )) =
0
z⊥ ∈Zd−1

Taking the Fourier transform on each side of the above formula, and evaluating it at p1
yields the desired formula. □
Remark 3.14. We used the fact that for p1 , a ∈ R,
X ea − e−a
eip1 n−a|n| =
2 .
n∈Z E1 (p1 ) + ea/2 − e−a/2
Corollary 3.15. Let β < βc (ρ), d ≥ 1. Then,
(i) Sbρ,β (p1 , . . . , pd ) is monotone decreasing in each |pj | over [−π, π].
(ii) E1 (p1 )Sbρ,β (p) and |p1 |2 Sbρ,β (p) are monotone increasing in |p1 |.
20 R. PANIS

Proof. The result is a direct consequence of Corollary 3.12 and of the monotonicity of the
following functions:

u2
u ∈ [0, π] 7→ E1 (u), u > 0 7→ ,
E1 (u)
and for all a ≥ 0,
E1 (u)
u > 0 7→ .
E1 (u) + (ea/2 − e−a/2 )2
□

The next result, obtained in [ADC21] in the case of nearest-neighbour interactions, will
be useful to derive Theorem 3.18. We derive it using the same methods used to obtain
Theorem 3.11, and postpone the proof to Appendix A.

Proposition 3.16. Let d ≥ 1 and β < βc (ρ). Introduce for p ∈ Rd ,

(mod)
Sbρ,β (p) := Sbρ,β (p) + Sbρ,β (p + π(1, 1, 0, . . . , 0)).

(mod)
Then Sbρ,β is monotone decreasing in |p1 − p2 | with p1 + p2 and (p3 , . . . , pd ) constant.

Corollary 3.17. Let d ≥ 2, β < βc (ρ). There exists C = C(d) > 0 such that for all
p ∈ [−π/2, π/2]d ,
C
Sbρ,β ((|p|, 0⊥ )) ≥ Sbρ,β (p) ≥ Sbρ,β ((|p|1 , 0⊥ )) − .
β
Proof. The first inequality is a direct consequence of the first item of Corollary 3.15. For
the second inequality, notice that an iteration of Proposition 3.16 yields
(mod) (mod)
Sbρ,β (p) ≥ Sbρ,β ((|p|1 , 0⊥ )).

Now recall that by Corollary 3.12, Sbρ,β ≥ 0, and that by the infrared bound, there exists
C > 0 such that for p ∈ [−π/2, π/2]d ,
C
|Sbρ,β (p + π(1, 1, 0, . . . , 0))| ≤ .
β
□

The following result was first derived in [ADC21]. The above work is sufficient to
extends it to general reflection positive interactions.

Theorem 3.18 (Sliding-scale infrared bound, [ADC21, Theorem 5.6]). Let d ≥ 2. There
exists C = C(d) > 0 such that for all β ≤ βc (ρ) and 1 ≤ ℓ ≤ L,

χL (ρ, β) C χℓ (ρ, β)
2
≤ .
L β ℓ2
Remark 3.19. In fact, using the results of Section 3.6, provided that the sharp length
(defined below) satisfies L(ρ, β) ≥ 1, we can replace the C/β by a C in the theorem
above. This minor improvement will be used later to derive bounds with constants that are
independent of β.
 TRIVIALITY OF CRITICAL ISING AND φ4 MODELS WITH deff ≥ 4 21

3.5. Gradient estimates. The following is a consequence of Theorem 3.11. It will play
a crucial role in the proof of existence of regular scales that follows.
Proposition 3.20 (Gradient estimate, [ADC21, Proposition 5.9]). Let d ≥ 1. There
exists C = C(d) > 0 such that for every β ≤ βc (ρ), every x ∈ Zd and every 1 ≤ i ≤ d,
F (|x|)
|Sρ,β (x ± ei ) − Sρ,β (x)| ≤ Sρ,β (x)
|x|
n
ρ,β ( 2 e1 )
 2S 
Sρ,β (dne1 )
where F (n) := C Sρ,β (ne1 ) log Sρ,β (ne1 ) .
The above estimate becomes particularly interesting whenever there exists c0 > 0 such
that,
n
 
Sρ,β (2dne1 ) ≥ c0 Sρ,β e1 .
2
Indeed, in that case, one can find C0 = C0 (c0 , d) > 0 such that for all x ∈ ∂Λn and
1 ≤ i ≤ d,
C0
|Sρ,β (x ± ei ) − Sρ,β (x)| ≤ Sρ,β (x).
|x|
3.6. The sharp length and a lower bound on the two-point function. Since we
work with infinite range interactions, it is possible that ξ(ρ, β) = ∞ throughout the
subcritical phase β < βc (ρ) (this is for instance the case for algebraically decaying RP
interactions [NS98, Aou21]). This forces us to revisit the notion of “typical length” in
these setups. As suggested by the work [DCT16], the quantity defined below is a good
candidate for the typical size of a box in which the model has a critical behaviour.
Definition 3.21 (Sharp length). Let β > 0. Let S be a finite subset of Zd containing 0.
Let
X
φρ,β (S) := β Jx,y ⟨τ0 τx ⟩S,ρ,β .
x∈S
y ∈S
/

Define9 the sharp length of parameter α ∈ (0, 1) by

1 n o
L(α) (ρ, β) := · inf k ≥ 1, ∃S ⊂ Zd with 0 ∈ S, rad(S) ≤ 2k, φρ,β (S) < α ,
2d
where rad(S) := max{|x − y|, x, y ∈ S}, and with the convention that inf ∅ = ∞. We will
set L(ρ, β) := L(1/2) (ρ, β).
Remark 3.22. Using the work of [ABF87], together with the strategy implemented in
[DCT16], we see that for any α ∈ (0, 1),
L(α) (ρ, βc (ρ)) = ∞.
Indeed, using Simon-Lieb inequality as in (3.3), one can show that if S is a finite subset
of Zd containing 0 and satisfying φρ,βc (ρ) (S) < 1, then
|S|⟨τ02 ⟩ρ,βc (ρ)
χ(ρ, βc (ρ)) ≤ .
1 − φρ,βc (ρ) (S)
This is in contradiction with the infiniteness of the susceptibility at criticality. A similar
argument gives that L(α) (ρ, β) increases to infinity as β tends to βc (ρ).
Below L(ρ, β), the two-point function can be lower bounded by an algebraically decaying
function. We start by stating this result in the special case where m2 (J) < ∞.
9The (2d)−1 in the definition of L(α) (ρ, β) is purely technical and will be useful in the proof of Proposition
3.23.
22 R. PANIS

Proposition 3.23 (Lower bound on the two-point function). Let d ≥ 2. Assume that J
satisfies (A1)-(A5) and that m2 (J) < ∞. There exists c = c(d, J) > 0 such that for all
β ≤ βc (ρ), and for all x ∈ Zd satisfying 1 ≤ |x| ≤ L(ρ, β),
c
⟨τ0 τx ⟩ρ,β ≥ .
β|x|d−1
Proof. Let n ≤ (2d)L(ρ, β). By definition of L(ρ, β), one has φρ,β (Λn ) ≥ 1/2. Using
(IRB) and the assumption m2 (J) < ∞, one has for some C1 > 0,
X X X 1
β Jx,y ⟨τ0 τx ⟩ρ,β ≤ C1 n2 J0,x ≤ 4C1 |x|2 J0,x ≤ ,
x∈Λ |x|≥n/2 |x|≥n/2
4
n/2
y ∈Λ
/ n

provided that n ≥ N0 (where N0 only depends on J). Hence, we now additionally assume
that (2d)L(ρ, β) ≥ N0 . We obtained that,
X 1
β Jx,y ⟨τ0 τx ⟩Λn ,ρ,β ≥ .
x∈Λ \Λ
4
n n/2
y ∈Λ
/ n

Then, using (MMS1), for some C2 , C3 > 0,

n X
1 X
≤ C2 βnd−1 ⟨τ0 τ(n/2)e1 ⟩ρ,β J0,y ≤ C3 βm2 (J)nd−1 ⟨τ0 τ(n/2)e1 ⟩ρ,β ,
4 k=0 |y|≥k

so that for some c1 > 0,

c1
⟨τ0 τ(n/2)e1 ⟩ρ,β ≥ .
βnd−1
Hence, using this time (MMS2), there exists c2 > 0 such that, if (2d)−1 N0 ≤ k ≤ L(ρ, β)
and x ∈ ∂Λk ,
c2
⟨τ0 τx ⟩ρ,β ≥ ⟨τ0 τ 2d|x| e ⟩ρ,β ≥ .
2 1 β|x|d−1
We now handle the smaller values of k = |x| by noticing that for 1 ≤ k ≤ (2d)−1 N0 ∧
L(ρ, β), the hypothesis that φρ,β (Λk ) ≥ 21 , together with (MMS1), yield
c3
⟨τ0 τke1 ⟩ρ,β ≥ ,
β
for some c3 = c3 (d, J). This concludes the proof. □
As it turns out, it also possible to extend this result with the following assumption:
there exist c0 , C0 , α > 0 such that,
c0 X C0
1+α
≤ J0,x ≤ 1+α , ∀k ≥ 1. (3.7)
k |x|=k
k

Using Proposition 3.8 we get that reflection positive interactions satisfying the above
assumption also satisfy: there exists C = C(d) > 0 such that for all β ≤ βc (ρ), for all
x ∈ Zd \ {0},
C
⟨τ0 τx ⟩ρ,β ≤ . (3.8)
βc (ρ)|x| d−α∧2 (log |x|)δα,2
The prototypical example of interactions satisfying (A1)-(A5) and (3.7) is given by alge-
braically decaying RP interactions.
Remark 3.24. Note that we may have models for which m2 (J) < ∞ and yet Assumption
3.7 is not satisfied. The next proposition will be useful in the study of models with d ∈
{1, 2, 3} and deff = 4, which do not satisfy m2 (J) < ∞.
 TRIVIALITY OF CRITICAL ISING AND φ4 MODELS WITH deff ≥ 4 23

Proposition 3.25. Let d ≥ 2. Assume that J satisfies (A1)-(A5) and (3.7). There exists
c > 0 which only depend on J and d such that, for all β ≤ βc (ρ), for all 1 ≤ |x| ≤ cL(ρ, β),

 1 if α > 1
c −1
⟨τ0 τx ⟩ρ,β ≥ × (log |x|) if α = 1
β|x|d−1  |x|α−1 if α ∈ (0, 1).
Moreover, the above result still holds for d = 1 and α ∈ (0, 1).
Remark 3.26. The lower bound matches (3.8) for α ∈ (0, 1).
Proof. We proceed like in the proof of Proposition 3.23. Notice that if ε > 0 is sufficiently
small and n ≥ 1,
 
X X X 1
β Jx,y ⟨τ0 τx ⟩ρ,β ≤ C1  ⟨τ0 τx ⟩ρ,β  J0,u ≤ C2 εα∧2 < ,
x∈Λεn x∈Λεn |u|≥n/2
4
y ∈Λ
/ n

where C1 , C2 > 0, and where we used (3.7) together with (3.8) on the second inequality.
For such a choice of ε, we have that for 1 ≤ n ≤ L(ρ, β),
X 1
β Jx,y ⟨τ0 τx ⟩ρ,β ≥ .
x∈Λn \Λεn
4
y ∈Λ
/ n

Using (MMS1), we find that

n X n
1 X X 1
≤ C3 ⟨σ0 σεne1 ⟩β nd−1 J0,u ≤ C4 ⟨σ0 σεne1 ⟩β nd−1 ,
4β k=1 |u|≥k k=1
kα

from which we obtain the desired result. □

3.7. Existence of regular scales. In this subsection, we introduce the notion of regular
scales. These scales will be defined in such a way that, on them, the two-point function
behaves “nicely” i.e. as if we knew that ⟨τ0 τx ⟩ρ,β decayed algebraically fast (for 1 ≤ |x| ≤
L(ρ, β)).
Definition 3.27 (Regular scales). Fix c, C > 0. An annular region Ann(n/2, 8n) is said
to be (c, C)-regular if the following properties hold :
(P1) For every x, y ∈ Ann(n/2, 8n), Sρ,β (y) ≤ CSρ,β (x),
C|x−y|
(P2) for every x, y ∈ Ann(n/2, 8n), |Sρ,β (x) − Sρ,β (y)| ≤ |x| Sρ,β (x),
(P3) χ2n (ρ, β) ≥ (1 + c)χn (ρ, β),
(P4) for every x ∈ Λn and y ∈ / ΛCn , Sρ,β (y) ≤ 12 Sρ,β (x).
A scale k is said to be regular if n = 2k is such that Ann(n/2, 8n) is (c, C)-regular, a
vertex x ∈ Zd will be said to be in a regular scale if it belongs to an annulus Ann(n, 2n)
with n = 2k and k a regular scale.
We can now state the main result of this subsection.
Proposition 3.28 (Existence of regular scales). Let d ≥ 3. Let J satisfy (A1)-(A5)
and m2 (J) < ∞. Let γ > 2. There exist c0 , c1 , C0 > 0 such that for every ρ in the GS
class, every β ≤ βc (ρ), and every 1 ≤ nγ ≤ N ≤ L(ρ, β), there are at least c1 log2 N n
(c0 , C0 )-regular scales between n and N .
The proof of this result can be found in [ADC21]. However, since it is a crucial tool for
what follows, we decide to include it.
24 R. PANIS

Proof. Using the lower bound of Proposition 3.23, together with (IRB), we get the exis-
tence of c1 , c2 > 0 such that
 γ−2  γ−2
c1 c1 N N
 
γ−1 γ−1
2
χN (ρ, β) ≥ N≥ n ≥ c2 χn (ρ, β).
βc (ρ) βc (ρ) n n
Using Theorem 3.18, we find r, c3 > 0 and independent of n, N , such that there are at
least c3 log2 (N/n) scales m = 2k between n and N such that
χrm (ρ, β) ≥ χ16dm (ρ, β) + χm (ρ, β). (3.9)
We prove that such an m is a (c0 , C0 )-regular scale for a good choice of c0 , C0 . Indeed, to
show it satisfies (P1) it is enough10 to show that Sρ,β ( 21 me1 ) ≤ C4 Sρ,β (16dme1 ) for some
constant C4 = C4 (d) > 0. However, one has

(3.10) |Ann(16dm, rm)|Sρ,β (16dme1 ) ≥ χrm (ρ, β) − χ16dm (ρ, β)

≥ χm (ρ, β) ≥ |Λm/(2d) |Sρ,β (me1 /2)
where in the first inequality we used (MMS1) to get that for all x ∈ Ann(16dm, rm) one
has Sρ,β (x) ≤ Sρ,β (16dme1 ), in the second inequality we used (3.9), and in the third one
we used (MMS1) again to argue that for all x ∈ Λm/(2d) one has Sρ,β (x) ≥ Sρ,β ( 21 me1 ).
This gives (P1). Note that (P2) follows from the remark below Proposition 3.20. Now,
using again (3.10) and (MMS2), we get that for every x ∈ Ann(m, 2m) one has
c5
Sρ,β (x) ≥ Sρ,β (16dme1 ) ≥ d χm (ρ, β), (3.11)
m
which implies (P3). Finally, we obtain11 (P4) by observing that for every R, if y ∈
/ ΛdRm
and x ∈ Λm ,
|ΛRm |Sρ,β (y) ≤ χRm (ρ, β) ≤ C6 R2 χm (ρ, β) ≤ C7 R2 md Sρ,β (x), (3.12)
where we used (MMS1) in the first inequality, Theorem 3.18 in the second (more precisely
the improvement discussed in Remark 3.19), and (3.11) in the last one. We obtain the
result by choosing C0 sufficiently large, and c0 sufficiently small. □

Using Proposition 3.25, we can extend the above result to interactions J satisfying (3.7).
Proposition 3.29. Let d ≥ 1. Let J satisfy (A1)-(A5) and (3.7) with α > 0 if d ≥ 3
and α ∈ (0, 1) if d ∈ {1, 2}. Let γ > 2. There exist c, c0 , c1 , C0 > 0 such that for every
γ
ρ in the
 GS class, every β ≤ βc (ρ), and every 1 ≤ n ≤ N ≤ cL(ρ, β), there are at least
c1 log2 Nn (c0 , C0 )-regular scales between n and N .

Proof. We only need to take care of the case d ∈ {1, 2} and α ∈ (0, 1). As noticed above,
in this case Sρ,β (x) ≍ |x|d−α below L(ρ, β). The existence of regular scales in that case is
then a direct consequence of the remark below Proposition 3.20. □

4. Random current representation

Let d ≥ 1. Let J be an interaction on Zd satisfying (A1)-(A5) and let Λ be a finite
subset of Zd .
10This comes from the fact that any x ∈ Ann(m/2, 8m) satisfies

me1
 
Sρ,β (16dme1 ) ≤ Sρ,β (x) ≤ Sρ,β .
2

11This is the only place where the hypothesis d ≥ 3 plays a role.

 TRIVIALITY OF CRITICAL ISING AND φ4 MODELS WITH deff ≥ 4 25

4.1. Definitions and the switching lemma.

Definition 4.1. A current n on Λ is a function defined on the set P2 (Λ) := {{x, y}, x, y ∈
Λ} and taking its values in N = {0, 1, . . .}. We denote by ΩΛ the set of currents on Λ.
The set of sources of n, denoted by ∂n, is defined as
 
 X 
∂n := x∈Λ, nx,y is odd .
 
y∈Λ

Q (βJx,y )nx,y
We also set wβ (n) := {x,y}⊂Λ .
nx,y !
There is a way to expend the correlation functions of the Ising model to relate them to
currents. Indeed, if we use, for σ ∈ {±1}Λ , the expansion
X (βJx,y σx σy )nx,y
exp(βJx,y σx σy ) = ,
nx,y ≥0
nx,y !

we obtain that

Z(Λ, β) = 2|Λ|
X
wβ (n).
∂n=∅

More generally, the correlation functions are given by: for A ⊂ Λ,

P
wβ (n)
⟨σA ⟩Λ,β = P∂n=A , (4.1)
w (n)
∂n=∅ β
Q
where σA := x∈A σx .
A current configuration n with ∂n = ∅ can be seen as the edge count of a multigraph
obtained as a union of loops. Adding sources to a current configuration comes down
to adding a collection of paths connecting pairwise the sources. For instance, a current
configuration with sources ∂n = {x, y} can be seen as the edge count of a multigraph
consisting of a family of loops together with a path from x to y.

Remark 4.2. It is possible to make the above remark rigorous by looking at the so-called
backbone representation of the Ising model, which is a way to expand the correlation func-
tions in terms of a weighted sum over a collection of source-pairing paths. This expansion
will be useful later and we refer to Appendix B for definitions and properties.

As we are about to see, connectivity properties of the multigraph induced by a current

will play a crucial role in the analysis of the underlying Ising model, this motivates the
following definition.

Definition 4.3. Let n ∈ ΩΛ and x, y ∈ Λ.

n
(i) We say that x is connected to y in n and write x ←→ y, if there exists a sequence
of points x0 = x, x1 , . . . , xm = y such that nxi ,xi+1 > 0 for 0 ≤ i ≤ m − 1.
(ii) The cluster of x, denoted by Cn (x), is the set of points connected to x in n.

The main interest of the above expansion lies in the following result that allows to
switch the sources of two currents. This combinatorial result first appeared in [GHS70]
to prove the concavity of the magnetisation of an Ising model with positive external field,
but the probabilistic picture attached to it was popularised in [Aiz82].
26 R. PANIS

Lemma 4.4 (Switching lemma). For any A, B ⊂ Λ and any function F from the set of
currents into R,
X
(SL) F (n1 + n2 )wβ (n1 )wβ (n2 )
n1 ∈ΩΛ : ∂n1 =A
n2 ∈ΩΛ : ∂n2 =B

F (n1 + n2 )wβ (n1 )wβ (n2 )1(n1 +n2 )∈FB ,

X
=
n1 ∈ΩΛ : ∂n1 =A∆B
n2 ∈ΩΛ : ∂n2 =∅

where A∆B = (A ∪ B) \ (A ∩ B) is the symmetric difference of sets and FB is given by

FB = {n ∈ ΩΛ , ∃m ≤ n , ∂m = B}.
In fact, we will also need a slightly different version of the switching lemma, called the
switching principle, whose proof can be found in [ADCTW19, Lemma 2.1]. We use the
representation of a current n ∈ ΩΛ into a multigraph N in which the vertex set is Λ
and where there are exactly nx,y edges between x and y. We will also use the notation
∂N = ∂n.
Lemma 4.5 (Switching principle). For any multigraph M with vertex set Λ, any A ⊂ Λ,
and any function f of a current,
f (N ) = 1∃K⊂M, ∂K=A
X X
f (N ∆K). (SP)
N ⊂M N ⊂M
∂N =A ∂N =∅

The switching lemma provides probabilistic interpretations of several quantities of in-

terest like differences or ratios of correlation functions. The natural probability measures
are defined as follows. If A ⊂ Λ, define a probability measure PA Λ,β on ΩΛ by: for n ∈ ΩΛ ,
wβ (n)
Λ,β [n] := 1∂n=A P
PA ,
∂m=A wβ (m)
and for A1 , . . . , Ak ⊂ Λ, define
PA 1 ,...,Ak
Λ,β := PA Ak
Λ,β ⊗ . . . ⊗ PΛ,β .
1

When A = {x, y}, we will write xy instead of {x, y} in the above notation.
The first consequence of the switching lemma is the following expression of a ratio of
correlation functions in terms of the probability of the occurrence of a certain connectivity
event in a system of random currents. One has,
⟨σA ⟩Λ,β ⟨σB ⟩Λ,β
= PA∆B,∅
Λ,β [n1 + n2 ∈ FB ] , (4.2)
⟨σA σB ⟩Λ,β
where FB was defined in Lemma 4.4. In particular, for 0, x, u ∈ Λ,
⟨σ0 σu ⟩Λ,β ⟨σu σx ⟩Λ,β n1 +n2
= P0x,∅
Λ,β [0 ←→ u]. (4.3)
⟨σ0 σx ⟩Λ,β
Later, it will also be interesting to control two-point connectivity probabilities. One can
prove (see [ADC21, Proposition A.3]) the following result: for every x, u, v ∈ Λ,
n +n ⟨σ0 σu ⟩Λ,β ⟨σu σv ⟩Λ,β ⟨σv σx ⟩Λ,β
(4.4) P0x,∅ 1 2
Λ,β [u, v ←→ 0] ≤ ⟨σ0 σx ⟩Λ,β
⟨σ0 σv ⟩Λ,β ⟨σv σu ⟩Λ,β ⟨σu σx ⟩Λ,β
+ .
⟨σ0 σx ⟩Λ,β
As proved in [ADCS15], the probability measure PA d
Λ,β , for A a finite (even) subset of Z ,
d A
admits a weak limit as Λ ↗ Z that we denote by Pβ . This yields infinite volume versions
of the above results.
 TRIVIALITY OF CRITICAL ISING AND φ4 MODELS WITH deff ≥ 4 27

4.2. The four-point Ursell function. Newman [New75] proved that the triviality of
the scaling limits of the Ising model is equivalent to the vanishing of the scaling limit
of the four-point Ursell function. This result was later quantified by Aizenman [Aiz82,
Proposition 12.1] who obtained the following bound.
Proposition 4.6 (Deviation from Wick’s law). Let d ≥ 2 and n ≥ 2. For all x1 , . . . , x2n ∈
Zd ,

X n
Y
⟨σx1 . . . σx2n ⟩β − ⟨σxπ(2j−1) , σxπ(2j) ⟩β
π pairing of j=1
{1,...,2n}
n−2
3
|U β (xi , xj , xk , xℓ )|
X X Y
≤ ⟨σxπ(2j−1) , σxπ(2j) ⟩β .
2 1≤i<j<k<ℓ≤2n 4 π pairing of j=1
{1,...,2n}\{i,j,k,ℓ}

The switching lemma provides a probabilistic interpretation of the four-point Ursell

function. This was a key step of the proof of triviality in [Aiz82, ADC21]. Although it was
first stated in the case of nearest-neighbour interactions, the proof is valid on any graph
and thus remains valid in the case of general interactions.
Proposition 4.7 (Representation of the four-point Ursell function). For x, y, z, t ∈ Zd ,
U4β (x, y, z, t) = −2⟨σx σy ⟩β ⟨σz σt ⟩β Pxy,zt
β [Cn1 +n2 (x) ∩ Cn1 +n2 (z) ̸= ∅]. (4.5)
This identity might seem tricky to analyse due to the lack of independence between
Cn1 +n2 (x) and Cn1 +n2 (z) but it is possible to show [Aiz82] that
Pxy,zt
β [Cn1 +n2 (x) ∩ Cn1 +n2 (z) ̸= ∅] ≤ Pxy,zt,∅,∅
β [Cn1 +n3 (x) ∩ Cn2 +n4 (z) ̸= ∅].
In particular, if I := Cn1 +n3 (x) ∩ Cn2 +n4 (z), this leads to the following bound,
|U4β (x, y, z, t)| ≤ 2⟨σx σy ⟩β ⟨σz σt ⟩β Pxy,zt,∅,∅
β [|I| > 0]. (4.6)
The random current representation allows us to obtain an expression of U4β in terms of the
probability of intersection of two independent random currents of prescribed sources. As
explained in [Aiz82], the relevant question is then to see whether, in the limit L(x, y, z, t) →
∞, the ratio |U4β (x, y, z, t)|/⟨σx σy σz σt ⟩β vanishes or not.

Heuristic for triviality. We work at β ≤ βc and assume some regularity on the two-
point function in the sense that it takes comparable values for pairs of points at comparable
distances smaller than L(β).
Let us first consider the case of the nearest-neighbour Ising model. If we expect the
intersection properties of two independent random current clusters to behave essentially
like the ones of two independent random walks in Zd conditioned to start and end at x, y
and z, t respectively, we expect the probability on the right-hand side of (4.6) to be very
small in dimension d > 4. Following this analogy, Aizenman [Aiz82] argued the case d > 4
by using a first moment method on |I| which yields the so-called tree diagram bound,
|U4β (x, y, z, t)| ≤ 2
X
⟨σx σu ⟩β ⟨σy σu ⟩β ⟨σz σu ⟩β ⟨σt σu ⟩β . (4.7)
u∈Zd
As discussed in the introduction, (4.7) together with (IRB), imply
|U4β (x, y, z, t)|
= O(L4−d ),
⟨σx σy σz σt ⟩β
where L ≤ L(β) is the mutual distance between x, y, z and t.
28 R. PANIS

In the case of the “marginal” dimension d = 4 the above bound yields no interesting
result and we need to go one step further in the analysis of (4.6). Going back to the analogy
with random walks, it is a well-known result that four is the critical dimension in terms of
intersection for the simple random walk, meaning that two independent (simple) random
walks with starting and ending points at mutual distance L, will intersect with probability
O(1/ log L) while the expected number of points in the intersection will typically be of
order Ω(1) (see [LL10, Chapter 10]). This shows that when two independent (simple)
random walks in dimension four intersect, they do so a logarithmic number of times.
Transposing this idea in the realm of random currents suggests that the probability that
two independent random currents with sources at mutual distance at least L intersect,
but not so many times, should decay as O(1/(log L)c ) for some c > 0. This is indeed the
result that Aizenman and Duminil-Copin obtained to improve by a logarithmic factor the
tree diagram bound.
For general long-range interactions it is possible to extend these ideas. It is well known
that long-range step distributions can virtually “increase” the effective dimension of a
random walk to the point that some low-dimensional random walks start manifesting the
above properties, only observed in dimension d ≥ 4 for the simple random walk. This
observation is made more explicit by the following computation (see Section 5).
|U4β (x, y, z, t)|
= O(L4(d/deff )−d ). (4.8)
⟨σx σy σz σt ⟩β
As a result, models with effective dimension strictly above four are easily shown to be
trivial, this is the content of Section 5. The main contribution of this paper is to treat the
case deff = 4, and to show that we can still improve the tree diagram bound there.

5. Reflection positive Ising models satisfying deff > 4

In this section, we study models of effective dimension deff > 4 and prove a more general
version of Theorem 1.2. As discussed in the introduction, choosing sufficiently slowly
decaying interactions might have the effect of increasing the dimension of the model. As
a result, we expect to find models in low dimensions which admit trivial scaling limits.
These results had already been obtained in [Aiz82, AF88], although not under this slightly
stronger form. We begin with a definition.
Definition 5.1 (Effective dimension). Let d ≥ 1. Assume that J satisfies (A1)-(A4).
The effective dimension deff of the model is related to the critical exponent of the two-point
function. Assume that there exists η ≥ 0 such that, as |x| → ∞,
1
⟨σ0 σx ⟩βc ≍ .
|x|d−2+η+o(1)
The effective dimension deff is then given by
d
deff := .
1 − (η ∧ 2)/2
Remark 5.2. The above formula can be justified using Fourier transform considerations,
see [AF88]. We saw in (3.1) that the spin-wave mode squared averages ⟨|τbβ (p)|2 ⟩TL ,ρ,β —
or thermal averages— coincide with the two-point function’s Fourier transform. As it
turns out, the relevant quantity to look at is often the density of these spin-wave modes—
i.e. dp— expressed as a function of the excitation level which is measured by Sbβ . For the
Gaussian case, one has S(p)
b ≍ p−2 which leads to a density of levels dSb−d/2 . For the case
of the Ising model, if we assume that Sbβc (p) ≍ p−(2−η) for some critical exponent η, we
−d /2
end up with a density dSbβc eff where deff = d/(1 − η/2).
 TRIVIALITY OF CRITICAL ISING AND φ4 MODELS WITH deff ≥ 4 29

Note that for the above definition to make sense, we need the existence of the critical
exponent η, which is expected to hold (but only known to exist in particular cases).
However, we can still bound the effective dimension. Glimm and Jaffe [GJ77] proved that
for reflection positive interactions η < 2, which justifies that deff < ∞ in our setup. For
reflection positive models, (IRB) yields that η ≥ 0, so that
deff ≥ d.

Remark 5.3 (Algebraically decaying RP interactions). In the case of reflection positive

models with coupling constants of algebraic decay, i.e. Jx,y = C|x − y|1−d−α for α, C > 0,
Proposition 3.8 implies that η ≥ |2 − α|+ , which yields,
d
deff ≥ .
1 ∧ (α/2)
As a consequence, if d − 2(α ∧ 2) > 0, one has deff > 4.

More generally, with the definition given above, we see that deff > 4 if
d > 4(1 − η/2).
In what follows, we assume that deff > 4, that is, there exists C > 0 and η ∈ [0, 2) such
that for all x ∈ Zd \ {0},
C
⟨σ0 σx ⟩βc ≤ , (5.1)
|x|d−2+η
where η ≥ 0 is such that d + 2η > 4.

Remark 5.4. Note that the above assumption is automatically satisfied when the inter-
action J satisfies (A1)-(A5) and (3.7) with d − 2(α ∧ 2) > 0.

Theorem 5.5. Let d ≥ 1. Assume that J satisfies (A1)-(A5) and (5.1). There exist
C = C(C, d), γ = γ(d) > 0 such that for all β ≤ βc , L ≥ 1, f ∈ C0 (Rd ) and z ∈ R,
!
z2
⟨exp (zTf,L,β (σ))⟩β − exp ⟨Tf,L,β (σ)2 ⟩β
2
C(β −4 ∨ β −2 )∥f ∥4∞ rfγ z 4
!
z2
≤ exp ⟨T (σ)2 ⟩β .
2 |f |,L,β Ld+2η−4

Remark 5.6. If we assume that the critical exponent η exists, we see that d + 2η − 4 =
d[(4/deff ) − 1], which is consistent with the decay of (4.8).

Proof. Let f ∈ C0 (Rd ), β ≤ βc and z ∈ C. Note that rf ≥ 1. Using Proposition 4.6 one
gets for n ≥ 2 (the inequality being trivial for n = 0, 1),
D E (2n)! D En
(5.2) Tf,L,β (σ)2n − Tf,L,β (σ)2
β 2n n! β
3 D E
≤ (2n)4 T|f |,L,β (σ)2n−4 ∥f ∥4∞ S(β, L, f ),
2 β

where

X U4β (x1 , x2 , x3 , x4 )
S(β, L, f ) := .
x1 ,x2 ,x3 ,x4 ∈Λrf L
ΣL (β)2
30 R. PANIS

z 2n
Multiplying by (2n)! and summing (5.2) over n, one gets,
!
z2
⟨exp (zTf,L,β (σ))⟩β − exp ⟨Tf,L,β (σ)2 ⟩β
2
!
z2
≤ C1 z 4 exp ⟨T (σ)2 ⟩β ∥f ∥4∞ S(β, L, f ).
2 |f |,L,β
Applying the tree diagram bound (4.7), we obtain
X ⟨σx σx1 ⟩β ⟨σx σx2 ⟩β ⟨σx σx3 ⟩β ⟨σx σx4 ⟩β
S(β, L, f ) ≤ 2 .
ΣL (β)2
x∈Zd
x1 ,x2 ,x3 ,x4 ∈Λrf L

Splitting the sum above,

X X
S(β, L, f )/2 ≤ (. . .) + (. . .) .
x∈Λdrf L x∈Λ/ drf L
x1 ,x2 ,x3 ,x4 ∈Λrf L x1 ,x2 ,x3 ,x4 ∈Λrf L
| {z } | {z }
(1) (2)

• Bound on (1). The first term can be written

 4
y∈Λrf L ⟨σx σy ⟩β
P
X ⟨σx σx1 ⟩β ⟨σx σx2 ⟩β ⟨σx σx3 ⟩β ⟨σx σx4 ⟩β X
2
= .
x∈Λdrf L
ΣL (β) x∈Λ
ΣL (β)2
drf L
x1 ,x2 ,x3 ,x4 ∈Λrf L

Noticing that for x ∈ Λdrf L , y∈Λrf L ⟨σx σy ⟩β ≤ χ2drf L (β), using Theorem 3.18 to bound
P

χ2drf L (β) in terms of χL (β), and using that χL (β) ≤ C2 L−d ΣL (β), we get
χL (β)2
(1) ≤ C3 β −4 rf8+d .
Ld
We can then use (5.1) to get the bound χL (β) ≤ C4 L2−η , so that
(1) ≤ C4 β −4 rf8+d L−(d+2η−4) .
• Bound on (2). Combining (MMS2) and the sliding-scale infrared bound of Theorem
3.18, we get that for i = 1, . . . , 4,
C5 C6
(5.3) ⟨σx σxi ⟩β ≤ d
χ|x|/d (β) ≤ χL (β).
|x| βL |x|d−2
2

Bounding the terms indexed by x1 and x2 in the sum using (5.3) and the other two using
(5.1), we get
χL (β)2 1 1
(2) ≤ C7 β −2
X X
2
ΣL (β) |x| 2d−4 4
L |x − x3 |d−2+η |x − x4 |d−2+η
/ drf L x1 ,...,x4 ∈Λrf L
x∈Λ

χL (β)2 1
≤ C8 β −2 rf2d+4 L2d
X
ΣL (β)2 x∈Λ
/ drf L
|x|2d−4+2η

≤ C9 β −2 rf8+d−2η L−(d+2η−4) ,
where we used the inequality |x − xi | ≥ (d − 1)rf L in the second line, and again χL (β) ≤
C2 L−d ΣL (β) in the third line. □
 TRIVIALITY OF CRITICAL ISING AND φ4 MODELS WITH deff ≥ 4 31

Remark 5.7. It is possible— see [Aiz82] or Section 8— to obtain a version of the tree
diagram bound for models in the GS class. In the general setup, it rewrites: for all β > 0,
for all x, y, z, t ∈ Zd ,

|U4ρ,β (x, y, z, t)| ≤ 2

X
⟨τx τu ⟩ρ,β βJu,u′ ⟨τu′ τy ⟩ρ,β ⟨τz τu ⟩ρ,β βJu,u′′ ⟨τu′′ τt ⟩ρ,β . (5.4)
u,u′ ,u′′ ∈Zd

Using the results of Section 3, together with (5.4), we can easily extend the result of this
section to the case of models in the GS class for which the two-point function ⟨τ0 τx ⟩ρ,βc (ρ)
satisfies condition (5.1).

6. Reflection positive Ising models in dimension d = 4

The goal of this section is to prove Theorems 1.3 and 1.5, and Corollary 1.8. The proof
of the first result is a direct consequence of the methods developed in the preceding section.

Proof of Theorem 1.3. Let d = 4. Assume that the interaction J satisfies (A1)-(A5) and
m2 (J) = ∞. Let β ≤ βc . Using Corollary 3.10 and adapting the proof of Section 5 to the
case ⟨σ0 σx ⟩βc = o(|x|2−d ), we get (using the above notations) that S(β, L, f ) = o(1) as
L → ∞. The explicit rate of convergence to 0 is obtained by estimating (see Proposition
3.8)
2
e−∥p∥2
Z
dp,
(−πL,πL]d 1 − J(|p|/L)
b

as L → ∞. For instance, in the case of algebraically decaying RP interactions of the form

Jx,y = C|x − y|−d−2
1 (α = 2), we get a decay of speed O(1/ log L). □

We now turn to the proof of Theorem 1.5. In the rest of the section, we fix d = 4 and
an interaction J satisfying (A1)-(A6). Hence, there exist C, ε > 0 such that for all k ≥ 1,
X C
|x|2 J0,x ≤ . (6.1)
|x|=k
k 1+ε

We expect this case to be more subtle than the m2 (J) = ∞ case since we do not get any
improvement on the effective dimension or on the decay of the two-point function. We
will use the argument of [ADC21] and adapt it to the long range setup. The proof follows
essentially the same steps: we make up for the lack of precise knowledge on the behaviour
of the two-point function at criticality by proving the existence of regular scales in which
the two-point function behave nicely (see Proposition 3.28); then, we introduce a nice local
intersection event which occurs with positive probability; and finally, we obtain a mixing
statement (which is of independent interest) which allows us to argue that intersections
at different scales are roughly independent events. As a result, we are able to prove that
as soon as two independent currents intersect, they intersect a large number of time (see
Proposition 6.2). One difficulty occurs in the process : since the model is long-range, the
clusters of the sources might make big jumps and avoid scales which may drastically reduce
the probability of the intersection event. Furthermore, the infinite range interactions may
be problematic to obtain the mixing statement as they create more correlation between
pieces of the current at different scales. The solution to get rid of that difficulty is to
prove that these scale jumps occur with sufficiently small probability: this is the main
technical point of this section, and the proof will be enabled by the assumption (6.1). As
a consequence, we will be able to argue that the sources’ clusters have a similar geometry
as the one obtained in the nearest-neighbour case. This will be enough to adapt the
multi-scale analysis of [ADC21].
32 R. PANIS

6.1. Proof of Theorem 1.5 conditionally on the clustering bound. We will need
the following deterministic lemma which relates the number of points in a set S ⊂ Zd to
the number of concentric annuli of the form u + Ann(uk , uk+1 ) with u ∈ S it takes to cover
S. For any (possibly finite) increasing sequence U = (uk )k≥0 , any u ∈ Zd , and any K ≥ 0,
define,
Mu (S; U, K) := |{0 ≤ k ≤ K, S ∩ [u + Ann(uk , uk+1 )] ̸= ∅}| .
Lemma 6.1 (Covering Lemma, [ADC21, Lemma 4.2]). With the above notations, for any
sequence U = (uk )k≥1 with u1 ≥ 1 and uk+1 ≥ 2uk ,
|S| ≥ 2minu∈S Mu (S;U ,K)/5 .
Fix D large enough. Define recursively a (possibly finite) sequence L of integers ℓk =
ℓk (β, D) by the formula: ℓ0 = 0 and
ℓk+1 = inf{ℓ ≥ ℓk , Bℓ (β) ≥ DBℓk (β)}.
Note that by (IRB), BL − Bℓ ≤ C0 log(L/ℓ). From this remark and the definition above
one can deduce12 that
Dk ≤ Bℓk (β) ≤ CDk , (6.2)
for every k and some large constant C independent of β, k, and D.
Theorem 1.5 will be a consequence of the following result. Recall that
I = Cn1 +n3 (x) ∩ Cn2 +n4 (z).
Proposition 6.2 (Clustering bound). For D large enough, there exists δ = δ(D) > 0
such that for all β ≤ βc , for all K > 3 with ℓK+1 ≤ L(β), and for all u, x, y, z, t ∈ Z4 with
mutual distance between x, y, z, t larger than 2ℓK ,
Pux,uz,uy,ut
β [Mu (I; L, K) < δK] ≤ 2−δK .
Let us see why this bound implies the improved tree diagram bound.
Proof of Theorem 1.5. Choose D large enough so that Proposition 6.2 holds. Fix x, y, z, t
at mutual distance larger than 2ℓK . Using Lemma 6.1 together with the switching lemma
(SL) we get,

Pxy,zt,∅,∅ Pβxy,zt,∅,∅ [u ∈ I, Mu (I; L, K) < δK]

X
β [0 < |I| < 2δK/5 ] ≤
u∈Z4
X ⟨σx σu ⟩β ⟨σy σu ⟩β ⟨σz σu ⟩β ⟨σt σu ⟩β
= Pux,uz,uy,ut
β [Mu (I; L, K) < δK],
⟨σx σy ⟩β ⟨σz σt ⟩β
u∈Z4
so that, with Proposition 6.2,
X ⟨σx σu ⟩β ⟨σy σu ⟩β ⟨σz σu ⟩β ⟨σt σu ⟩β
Pxy,zt,∅,∅
β [0 < |I| < 2δK/5 ] ≤ 2−δK .
⟨σx σy ⟩β ⟨σz σt ⟩β
u∈Z4

12Indeed, the lower bound is immediate and for the upper bound, write for k ≥ 1,
 
1
Bℓk −1 (β) ≤ DBℓk−1 (β) ≤ DBℓk−1 −1 (β) − CD log 1−
ℓk−1
k−1
X Di
≤ Dk−1 Bℓ1 −1 (β) + C ≤ C1 D k
ℓk−i
i=1

for C1 large enough (independent of D and k). Use (IRB) once again to write,
Bℓk (β) ≤ Bℓk −1 + C log 2 ≤ C1 Dk + C log 2 ≤ C2 Dk .
 TRIVIALITY OF CRITICAL ISING AND φ4 MODELS WITH deff ≥ 4 33

Moreover,
X ⟨σx σu ⟩β ⟨σy σu ⟩β ⟨σz σu ⟩β ⟨σt σu ⟩β
Pxy,zt,∅,∅
β [|I| ≥ 2δK/5 ] ≤ 2−δK/5 ,
⟨σx σy ⟩β ⟨σz σt ⟩β
u∈Z4
which implies that
2
|U4β (x, y, z, t)| ≤
X
⟨σx σu ⟩β ⟨σy σu ⟩β ⟨σz σu ⟩β ⟨σt σu ⟩β .
2δK/5 u∈Z4

Now, if L := 2ℓK , observe that ℓK+1 ≥ L so that by (6.2), BL (β) ≤ BℓK+1 (β) ≤ CDK+1 .
Hence, we may find c > 0 sufficiently small (independent of L and β), such that K ≥
c log BL (β). This gives the result. □
We now turn to the proof of Proposition 6.2. We start by showing that thanks to the
hypothesis (A6), the interaction decays sufficiently fast so that the “jumps” made by the
current are not so problematic. As a byproduct of these estimates, we are able to show
that the clusters do not perform “back and forth” between different scales. This property
was already a key step in the proof of [ADC21].
6.2. Properties of the current. We will use a first moment method and argue that the
expected number of long edges that have a non zero-weight under the current measure
decays quickly in a certain sense. The following results rely on the existence of regular
scales and thus require reflection positivity.
In what follows, we use the notion of backbone. The main properties of this object are
gathered in Appendix B.
First, we prove a bound on the probability that an edge is open in the percolation
configuration deduced from the current measure. Note that this bound is in fact valid on
any graph with any interactions.
Lemma 6.3 (Bound on open edge probability). Let d ≥ 1. Let β > 0. For x, y, u, v ∈ Zd ,
one has
Pxy xy,∅
β [nu,v ≥ 1] ≤ Pβ [nu,v ≥ 1]
!
⟨σx σu ⟩β ⟨σv σy ⟩β ⟨σx σv ⟩β ⟨σu σy ⟩β
≤ βJu,v 2⟨σu σv ⟩β + + .
⟨σx σy ⟩β ⟨σx σy ⟩β
Proof. The first inequality follows from a monotonicity argument. For the second inequal-
ity, we write,

Pxy,∅ xy ∅
β [nu,v ≥ 1] ≤ Pβ [nu,v ≥ 1] + Pβ [nu,v ≥ 1].

Then, we observe that P∅β [nu,v ≥ 1] ≤ βJu,v ⟨σu σv ⟩β , which leads to

⟨σx σy σu σv ⟩β
Pxy
β [nu,v ≥ 1] ≤ βJu,v
⟨σx σy ⟩β
!
⟨σx σu ⟩β ⟨σv σy ⟩β ⟨σx σv ⟩β ⟨σu σy ⟩β
≤ βJu,v ⟨σu σv ⟩β + + ,
⟨σx σy ⟩β ⟨σx σy ⟩β

where we used Lebowitz’ inequality [Leb74] to get U4β (x, y, u, v) ≤ 0 (see also Proposition
4.7). □
We now introduce the event we will be interested in. It is illustrated in Figure 3 below.
Definition 6.4 (Jump event). Let 1 ≤ k ≤ m. We define Jump(k, m) to be the event that
there exist u ∈ Λk and v ∈
/ Λm such that nu,v ≥ 1.
34 R. PANIS

Γ(n)
Λm

Λk

Figure 3. A realisation of the event Jump(k, m) for a current n with

source set ∂n = {0, y}. The bold black path represents the backbone Γ(n).
The dashed curves represent long open edges that jump over the annulus
Ann(k, m).

We now prove that if we consider a current with two sources, and an annulus located
between them, but “far away” from each of them, then with high probability the current
does not “jump over it”. For convenience, we fix one of these sources to be the origin.
Recall that ε > 0 is given by (6.1).

Lemma 6.5 (Jumping a scale is unlikely). Let d = 4. Assume that J satisfies (A1)-(A6).
d
Let ν ∈ (0, 1) be such that ν > d+ε . There exist C, η > 0 such that for all β ≤ βc , for all
d
y ∈ Z in a regular scale with 1 ≤ |y| ≤ L(β), and for all k ≥ 1 such that k 1+4/ε ≤ |y|,

C
P0y,∅
β [Jump(k, k + k ν )] ≤ .
kη

Remark 6.6. If one takes |y| ≥ ℓK+1 , this lemma ensures that the current visits the
annuli Ann(ℓk , ℓk+1 ) for 1 ≤ k ≤ K with high probability. This property will be crucial in
the proof of Proposition 6.2.

Proof. In what follows C = C(d) > 0 may change from line to line. It is sufficient to
bound,

P0y,∅
X
β [nu,v ≥ 1].
u∈Λk , v ∈Λ
/ k+kν
 TRIVIALITY OF CRITICAL ISING AND φ4 MODELS WITH deff ≥ 4 35

Using Lemma 6.3, we have (see Figure 4),

P0y,∅
X X
(6.3) β [nu,v ≥ 1] ≤ 2β Ju,v ⟨σu σv ⟩β
u∈Λk , v ∈Λ
/ k+kν u∈Λk , v ∈Λ
/ k+kν
!
⟨σ0 σu ⟩β ⟨σv σy ⟩β ⟨σ0 σv ⟩β ⟨σu σy ⟩β
+ + =: A1 + A2 + A3 .
⟨σ0 σy ⟩β ⟨σ0 σy ⟩β

• Bound on A1 . Write
X X |v − u|2 J0,v−u
β Ju,v ⟨σu σv ⟩β ≤ C
u∈Λk , v ∈Λ
/ k+kν u∈Λk , v ∈Λ
/ k+kν
|v − u|d
X |v|2 J0,v
≤ Ck d
v ∈Λ
/ kν
|v|d
X
≤ Ck d(1−ν) |v|2 J0,v ≤ Ck d(1−ν)−νε ,
v ∈Λ
/ kν

where we used (IRB) on the first line, translation invariance on the second line, and (6.1)
on the last line.
• Bound on A2 . For the second term, by the lower bound on the two-point function of
Proposition 3.23 (together with the assumption that 1 ≤ |y| ≤ L(β)) one has ⟨σ0 σy ⟩−1β ≤
d−1
Cβ|y| . Using (IRB) and the translation invariance of J,
X ⟨σ0 σu ⟩β ⟨σv σy ⟩β X
β Ju,v ≤ Cβ 2 |y|d−1 J0,v−u ⟨σ0 σu ⟩β ⟨σv σy ⟩β
u∈Λk , v∈Λ|y|/2 (y)
⟨σ0 σy ⟩β u∈Λk
v∈Λ|y|/2 (y)
X J0,v
≤ Ck 2 |y|d−1
v∈Λ2|y|/3 (y)
(|v − y| + 1)d−2
2|y|/3
1
J0,v 1|y−v|=p .
X X
2 d−1
≤ Ck |y|
p=1
pd−2 v∈Λ2|y|/3 (y)

Now, using13 (6.1),

J0,v 1|y−v|=p ≤ Cp|y|−(3+ε)

X

v∈Λ2|y|/3 (y)

Using the assumption that |y| ≥ k 1+4/ε ,

X ⟨σ0 σu ⟩β ⟨σv σy ⟩β Ck 2 log |y| Ck 2 C
Ju,v ≤ ≤ ≤ ε/2 .
u∈Λk , v∈Λ|y|/2 (y)
⟨σ0 σy ⟩β |y|ε |y| ε/2 k

Finally, in the case v ∈

/ Λ|y|/2 (y) ∪ Λk+kν , we may use (P1) (for v ∈ Λ4|y| (y) \ (Λ|y|/2 (y) ∪
Λk+kν )), as well as (MMS2) (for v ∈ / Λ4|y| (y)), to show that ⟨σv σy ⟩β ≤ C0 ⟨σ0 σy ⟩β , so

13Note that this bound is very sub-optimal in some cases but is enough for our purpose. For instance, as-
suming more regularity on the interaction we could replace (6.1) by a bound of the form J0,x ≤ C|x|−d−2−ε
for all x ∈ Zd \ {0}. In that case, we would have J0,v ≲ |y|−d−2−ε for v ∈ Λ2|y|/3 (y) which leads to a bound
of the first contribution of A2 of order k2 /|y|1+ε for any d ≥ 4.
36 R. PANIS

that,
X ⟨σ0 σu ⟩β ⟨σv σy ⟩β X
β Ju,v ≤ C0 β Ju,v ⟨σ0 σu ⟩β
u∈Λk , v ∈Λ
/ |y|/2 (y)∪Λk+kν
⟨σ0 σy ⟩β u∈Λk , v ∈Λ
/ |y|/2 (y)∪Λk+kν
X
≤ Ck 2(1−ν) |v|2 J0,v ≤ Ck 2(1−ν)−νε .
v ∈Λ
/ kν

y y
0 0 0

A1 A2 A3

Figure 4. A graphical representation of the bound (6.3). We represented

each potential contribution to A1 , A2 , A3 (from left to right). The backbone
is the bold path joining 0 and y. Long open edges are the dashed curves.
The largest contribution should come from A2 in which long edges are
induced by the backbone.

• Bound on A3 . Since |y| ≥ k 1+4/ε , one has for u ∈ Λk , ⟨σu σy ⟩β ≤ C0 ⟨σ0 σy ⟩β by the
property (P1) of regular scales. Using this remark, together with (IRB) and (6.1), yields,
X
A3 ≤ 2βC0 Ju,v ⟨σ0 σv ⟩β
u∈Λk , v ∈Λ
/ k+kν
X |v − u|2 J0,v−u 1
≤ C
u∈Λk , v ∈Λ
/ k+kν
|v − u|2 |v|d−2
2
C |v|
 X
≤ sup |u − v|2 J0,v−u .
kd u∈Λk |u − v| u∈Λk , v ∈Λ
/ k+kν
v ∈Λ
/ k+kν
X
≤ Ck 2(1−ν) |v|2 J0,v ≤ Ck 2(1−ν)−νε .
v ∈Λ
/ kν

We then obtain the result taking η sufficiently small, and C large enough. □
Remark 6.7. We could prove a similar result with the assumption (A6) replaced by (A6′ )
as mentioned in the introduction. The only difference is that we now need to consider the
event Jump(k, k + k/(log k)1−ν ) for ν ∈ (0, 1) sufficiently close to 1. Using the exact same
method as above we obtain η, C > 0 such that, when |y| ≥ exp(k 1+4/ε ),
C
P0y,∅
β [Jump(k, k + k/(log k)
1−ν
)] ≤ .
(log k)η
Similar modifications can be applied below.
 TRIVIALITY OF CRITICAL ISING AND φ4 MODELS WITH deff ≥ 4 37

As a corollary of the above result, we can show that the probability that a backbone
does a zigzag between two “distanced” scales is very small. This will be very useful later
to argue that intersection events are (essentially) local.
Definition 6.8 (Zigzag event). For 1 ≤ k ≤ ℓ ≤ M and u, v ∈ Zd , let ZZ(u, v; k, ℓ, M ) be
the event that the backbone of n (with ∂n = {u, v}) goes from u to a point in Ann(ℓ, M ),
then to a point in Λk , before finally hitting v. We let ZZ(u, v; k, ℓ, ∞) be the union of all
ZZ(u, v; k, ℓ, M ) for M ≥ ℓ.
Corollary 6.9 (No zigzag for the backbone). Let d = 4. Assume that J satisfies (A1)-
d
(A6). Let ν ∈ (0, 1) be such that ν > d+ε . There exist C, η > 0 such that for all β ≤ βc ,
d
for all k, ℓ ≥ 1 and y ∈ Z in a regular scale with k 8/(1−ν) ≤ ℓ and ℓ1+4/ε ≤ |y|,
C
P0y
β [ZZ(0, y; k, ℓ, ∞)] ≤ .
ℓη
Proof. Notice that,
ZZ(0, y; k, ℓ, ∞) ⊂ ZZ(0, y; k, ℓ, ℓ + ℓν ) ∪ Jump(ℓ, ℓ + ℓν ).
The chain rule for backbones (see Appendix B), the assumption of regularity made on y,
as well as (IRB), yield
⟨σ0 σv ⟩β ⟨σv σw ⟩β ⟨σw σy ⟩β
P0y
X
ν
β [ZZ(0, y; k, ℓ, ℓ + ℓ )] ≤
v∈Ann(ℓ,ℓ+ℓν )
⟨σ0 σy ⟩β
w∈Λk
k ℓ3+ν
4 C
≤ C ≤ .
ℓ4 ℓ(1−ν)/2
We conclude using Lemma 6.5. □
Remark 6.10. Note that in the above result we heavily relied on the that fact the current
cannot jump over an annulus of dimension strictly smaller than four. We will see that
Lemma 6.5 does not hold anymore in the case deff = 4 and 1 ≤ d ≤ 3 which makes the
study of the zigzag event more complicated.
This second technical result is a small modification of Lemma 6.5 but will be crucial
in the proof of the mixing. It is a little easier since the sources are now located close to
the origin: there is no “long backbone” which might help to jump scales. Note that this
lemma does not rely on the existence of regular scales.
Lemma 6.11. Let d = 4. Assume that J satisfies (A1)-(A6). Let ν ∈ (0, 1) be such that
d
ν > d+ε . There exist C, η > 0 such that for all β ≤ βc , for all n < m ≤ M ≤ k with
1≤M 3/2 ≤ k ≤ L(β), for all x ∈ Λn and all u ∈ Ann(m, M ),
C
Pxu,∅
β [Jump(k, k + k ν )] ≤
.
kη
Proof. We follow the same steps as in the proof of Lemma 6.5. Using Lemma 6.3,

Pxu,∅
X
β [nw,v ≥ 1]
w∈Λk , v ∈Λ
/ k+kν
!
X ⟨σx σw ⟩β ⟨σv σu ⟩β ⟨σx σv ⟩β ⟨σw σu ⟩β
≤ 2β Jw,v ⟨σw σv ⟩β + + .
w∈Λk , v ∈Λ
/ k+kν
⟨σx σu ⟩β ⟨σx σu ⟩β

As for the bound of A1 above,

X
β Jw,v ⟨σw σv ⟩β ≤ C1 k d(1−ν)−νε .
w∈Λk , v ∈Λ
/ k+kν
38 R. PANIS

Using the lower bound of Proposition 3.23 (which is licit because 1 ≤ |x − u| ≤ L(β))
together with (IRB) and (6.1), we get
X ⟨σx σv ⟩β ⟨σw σu ⟩β X
β Jw,v ≤ β 2 C2 M d−1 Jw,v ⟨σx σv ⟩β ⟨σw σu ⟩β
w∈Λk , v ∈Λ
/ k+kν
⟨σx σu ⟩β w∈Λk , v ∈Λ
/ k+kν

≤ C3 M d−1 k 2 k −(d−2)
X
J0,v
v ∈Λ
/ kν

≤ C4 M d−1 k −ν(2+ε) .
Finally, with the same reasoning, we also get
⟨σx σw ⟩β ⟨σv σu ⟩β
≤ C5 M d−1 k −ν(2+ε) .
X
Jw,v
w∈Λk , v ∈Λ
/ k+kν
⟨σx σu ⟩β

The assumption made on ν and the inequality M 3 ≤ k 2 yield the result. □

Similarly, we can rule out the zigzag of the backbone in this setup.
Corollary 6.12. Let d = 4. Assume that J satisfies (A1)-(A6). Let ν ∈ (0, 1) be such
d
that ν > d+ε . There exist C, η > 0 such that for all β ≤ βc , for all n < m ≤ M ≤ k with
1 ≤ M 6/(1−ν) ≤ k ≤ L(β), for all x ∈ Λn and all u ∈ Ann(m, M ),
C
Pxu
β [ZZ(x, u; M, k, ∞)] ≤
.
kη
Proof. We repeat the argument used to prove Corollary 6.9, except that the proof is now
d
easier since u already lies in ΛM . Let ν ∈ (0, 1) be such that ν > d+ε . One has,
ZZ(x, u; M, k, ∞) ⊂ ZZ(x, u; M, k, k + k ν ) ∪ Jump(k, k + k ν ).
The chain rule for backbones, the lower bound of Proposition 3.23, and (IRB), yield
X ⟨σx σv ⟩β ⟨σv σu ⟩β
Pxu ν
β [ZZ(x, u; M, k, k + k )] ≤
v∈Ann(k,k+kν )
⟨σx σu ⟩β

M d−1 k 3+ν C
≤ C ≤ (1−ν)/2 .
k4 k
We conclude using Lemma 6.11.
□
This final technical lemma will be useful to argue that for a current n with ∂n =
{x, y}, the restriction of n to (Γ(n))c (that we denote by n \ Γ(n) below), where Γ(n) is
the backbone of n and Γ(n) is the set of edges revealed during the exploration of Γ(n),
essentially behaves like a sourceless current on a smaller graph. In that case, jump events
should become even more unlikely since there is no backbone to create long connections
anymore.
If n is a current, and E is a set of edges, we let nE be the restriction of n to the edges
in E.
Lemma 6.13. Let d = 4. Assume that J satisfies (A1)-(A6). Let ν ∈ (0, 1) be such that
d
ν > d+ε . There exist C, η > 0 such that for all β ≤ βc , for all k ≥ 1, for all x, y ∈ Zd ,
C
Pxy ν xy,∅ ν
β [n \ Γ(n) ∈ Jump(k, k + k )] ≤ Pβ [(n1 + n2 ) \ Γ(n1 ) ∈ Jump(k, k + k )] ≤ ,
kη
where by (n1 + n2 ) \ Γ(n1 ) we mean the restriction of (n1 + n2 ) to the graph depleted of
the edges belonging to Γ(n1 ).
 TRIVIALITY OF CRITICAL ISING AND φ4 MODELS WITH deff ≥ 4 39

Proof. The first inequality follows by monotonicity. Write A := {(n1 + n2 ) \ Γ(n1 ) ∈

Jump(k, k + k ν )} and for a consistent (see Appendix B for a definition) path γ : x → y,
Aγ := {(n1 + n2 )γ c ∈ Jump(k, k + k ν )}. The idea is to condition on the backbone of n1 .
Going to partition functions14, one has Pxy,∅
β [A] = Zβ
xy,∅
[A]/Zβxy,∅ where,

Zβxy,∅ [A] := wβ (n1 )wβ (n2 )1Γ(n1 )=γ 1A

X X

γ:x→y consistent ∂n1 ={x,y}

∂n2 =∅

wβ ((n1 )γ )wβ ((n1 )γ c )wβ (n2 )1Γ((n1 )γ )=γ 1Aγ

X X
=
γ ∂(n1 )γ ={x,y}
∂(n1 )γ c =∅
∂n2 =∅
xy
[Γ(nγ ) = γ]Zγ∅,∅
X
= Zγ,β c ,Zd ,β [Aγ ]
γ

Zβxy,∅ [Γ(n1 ) = γ]Pγ∅,∅

X
= c ,Zd ,β [Aγ ].
γ

Using Lemma 6.3 as well as Griffiths’ inequality, for any γ as above,

P∅,∅
X
γ c ,Zd ,β
[Aγ ] ≤ 2βJu,v ⟨σu σv ⟩β ≤ Ck d(1−ν)−νε ,
u∈Λk
v ∈Λ
/ k+kν

where the last inequality was obtained in the proof of Lemma 6.5. Hence,
Zβxy,∅ [A] ≤ Ck d(1−ν)−νε Zβxy,∅ ,
d
which yields the result since ν > d+ε . □
As before, having a control of the jump probability over an annulus of dimension < 4
(namely 3 + ν for ν ∈ (0, 1)) has consequences on the geometry of the current. In the
context of Lemma 6.13, it prevents (n1 +n2 )\Γ(n1 ) to connect two (sufficiently) distanced
scales. We begin with a definition.
Definition 6.14 (Crossing event). For 1 ≤ k ≤ ℓ and n a current, we say that n realises
the event Cross(k, L) if n “crosses” Ann(k, ℓ), in the sense that there exists a cluster of n
containing both a point in Λk and in Λcℓ .
Corollary 6.15. Let d = 4. Assume that J satisfies (A1)-(A6). Let ν ∈ (0, 1) be such
d
that ν > d+ε . There exist C, η > 0 such that for all β ≤ βc , for all k, ℓ ≥ 1 with
k 8/(1−ν) ≤ ℓ, for all x, u ∈ Zd ,
xu,∅ C
Pxu
β [n \ Γ(n) ∈ Cross(k, ℓ)] ≤ Pβ [(n1 + n2 ) \ Γ(n1 ) ∈ Cross(k, ℓ)] ≤ η .
ℓ
Proof. The first inequality follows by monotonicity. Notice that,
n o [ n o
(n1 + n2 ) \ Γ(n1 ) ∈ Cross(k, ℓ) ⊂ v ←→ w in (n1 + n2 ) \ Γ(n1 )
v∈Λk , w∈Ann(ℓ,ℓ+ℓν )
n o
∪ (n1 + n2 ) \ Γ(n1 ) ∈ Jump(ℓ, ℓ + ℓν ) .
The second event on the right-hand side above is handled using Lemma 6.13.
To handle the first event, we use the fact that the probability v and w are connected in
(n1 + n2 ) \ Γ(n1 ) can be bounded by ⟨σv σw ⟩2β . Indeed, this result follows from an gener-
alisation of the switching lemma that can be found in [ADCS15, Lemma 2.2]. Proceeding
14One would need to restrict to a finite subset Λ of Zd first, and then take the limit Λ → Zd . We omit
this detail here.
40 R. PANIS

as in the proof of Lemma 6.13, we get

Pxu,∅
β [u ←→ v in (n1 + n2 ) \ Γ(n1 )]
∅,∅
X
≤ Pxu
β [Γ(n) = γ]Pγ c ,Zd ,β [v ←→ w in (m1 + m2 ) \ γ].
γ:x→u consistent

The above-mentioned generalisation of the switching lemma, together with Griffiths’ in-
equality, yield
P∅,∅
γ c ,Zd ,β
[v ←→ w in (m1 + m2 ) \ γ] = ⟨σv σw ⟩γ c ,β ⟨σv σw ⟩β ≤ ⟨σv σw ⟩2β .
As a result,
 
n o
Pxu,∅
[ X
β
 v ←→ w in (n1 + n2 ) \ Γ(n1 )  ≤ ⟨σv σw ⟩2β
v∈Λk , w∈Ann(ℓ,ℓ+ℓν ) v∈Λk
w∈Ann(ℓ,ℓ+ℓν )

k4 C1
≤ C1 ≤ .
ℓ1−ν ℓ(1−ν)/2

This concludes the proof.

□
6.3. Proof of the intersection property. Let d = 4. Assume that J satisfies (A1)-
(A6). Let β ≤ βc . Recall the definition of L = L(β, D) given at the beginning of the
section: ℓ0 = 0 and
ℓk+1 = inf{ℓ ≥ ℓk , Bℓ (β) ≥ DBℓk (β)}. (6.4)
The existence of regular scales and the sliding-scale infrared bound have the following
interesting consequence on how fast the bubble diagram grows from one scale to the
other, it can be see as an improvement over the bound BL (β) − Bℓ (β) ≤ C0 log(L/ℓ).
Lemma 6.16 (Scale to scale comparison of the bubble diagram, [ADC21, Lemma 6.3]).
Let d = 4. There exists C = C(d) > 0 such that for every β ≤ βc , and for every
1 ≤ ℓ ≤ L ≤ L(β),
log2 (L/ℓ)
 
BL (β) ≤ 1 + C Bℓ (β).
log2 (ℓ)
Proof. If N ≥ n and n is a weak regular scale, we may write,

B2N (β) − BN (β) ≤ C1 N −4 χN/d (β)2 ≤ C2 · n−4 χn (β)2 ≤ C3 · n−4 (χ2n (β) − χn (β))2
≤ C4 (B2n (β) − Bn (β)) ,
where we successively used (MMS2), the sliding-scale infrared bound, the property (P3)
of regular scales, and Cauchy–Schwarz inequality. There are log2 (L/ℓ) scales between ℓ
and L, and at least c log2 (ℓ) regular scales between 1 and ℓ. Using the above computation,
X
BL (β) − Bℓ (β) = B2N (β) − BN (β)
N scale between ℓ andL/2
log2 (L/ℓ) X log2 (L/ℓ)
≤ B2n (β) − Bn (β) ≤ Bℓ (β).
c log2 (ℓ) n regular scale
c log2 (ℓ)
between 1 and ℓ
□
The above property has the following important consequence which ensures that the
scales explode sufficiently fast. This will be used later to make sure there is “enough room”
in the annuli Ann(ℓk , ℓk+1 ).
 TRIVIALITY OF CRITICAL ISING AND φ4 MODELS WITH deff ≥ 4 41

Remark 6.17 (Growth of L). Using Lemma 6.16 we get that,

Bℓk (β) C
log2 (ℓk ) ≤ C log2 (ℓk+1 /ℓk ) ≤ log2 (ℓk+1 /ℓk ) ,
Bℓk+1 (β) − Bℓk (β) D−1
so that,
(D−1)/C
ℓk+1 ≥ ℓk .
Recall that the event Jump was defined in Definition 6.18. In the remaining of this
d
section, we fix ν ∈ (0, 1) with ν > d+ε such that Lemmas 6.5, 6.11, and 6.13 hold.
Definition 6.18 (Intersection event). Let k ≥ 1 and y ∈ / Λℓk+2 . A pair of currents
(n, m) with (∂n, ∂m) = ({0, y}, {0, y}) realises the event Ik if the following properties are
satisfied:
(i) The restrictions of n and m to edges with both endpoints in Ann(ℓk , ℓk+1 + ℓνk+1 )
contain a unique cluster “strongly crossing” Ann(ℓk , ℓk+1 ), in the sense that it
contains a vertex in Ann(ℓk , ℓk + ℓνk ) and a vertex in Ann(ℓk+1 , ℓk+1 + ℓνk+1 ).
(ii) The two clusters described in (i) intersect.
Note that the event Ik is measurable in term of edges with both endpoints in the annulus
Ann(ℓk , ℓk+1 + ℓνk+1 ).
The following lemma shows that intersections occur at every scale with a uniformly
positive probability.
Lemma 6.19 (Intersection property). Let d = 4. For D large enough, there exists κ > 0
such that for every β ≤ βc , every k ≥ 2, and every y ∈
/ Λℓk+2 in a regular scale with
1 ≤ |y| ≤ L(β),
P0y,0y,∅,∅
β [(n1 + n3 , n2 + n4 ) ∈ Ik ] ≥ κ.
Proof. We restrict ourselves to the case of y in a regular scale to be able to use the
properties (P1) and (P2). Introduce intermediate scales ℓk ≤ n ≤ m ≤ M ≤ N ≤ ℓk+1
satisfying
8 8
+1 8 8
ℓk1−ν ≥ n ≥ ℓk1−ν , n 1−ν +1 ≥ m ≥ n 1−ν ,
8 8 8 8
M 1−ν +1 ≥ N ≥ M 1−ν , N 1−ν +1 ≥ ℓk+1 ≥ N 1−ν ,
which is possible provided D is large enough by Remark 6.17. Define
M := Cn1 +n3 (0) ∩ Cn2 +n4 (0) ∩ Ann(m, M ).
Using Cauchy–Schwarz inequality,
E0y,0y,∅,∅
β [|M|]2
P0y,0y,∅,∅
β [|M| > 0] ≥ .
E0y,0y,∅,∅
β [|M|2 ]
One has for some c1 > 0,
!2
⟨σ0 σu ⟩β ⟨σu σy ⟩β
E0y,0y,∅,∅
X
β [|M|] =
u∈Ann(m,M )
⟨σ0 σy ⟩β
≥ c1 (BM (β) − Bm−1 (β)),
where we used the regularity to compare ⟨σu σy ⟩β with ⟨σ0 σy ⟩β . Using (4.4), for some
c2 > 0,
!2
⟨σ0 σu ⟩β ⟨σu σv ⟩β ⟨σv σy ⟩β ⟨σ0 σv ⟩β ⟨σv σu ⟩β ⟨σu σy ⟩β
Eβ0y,0y,∅,∅ [|M|2 ]
X
≤ +
u,v∈Ann(m,M )
⟨σ0 σy ⟩β ⟨σ0 σy ⟩β
≤ c2 (BM (β) − Bm−1 (β))B2M (β),
42 R. PANIS

Λℓk+1

ΛN

Ann(m, M )

Γ(n1 )
Λℓk
y

Γ(n2 )

Figure 5. An example of a configuration for which M = ̸ ∅ but Ik does

not occur. n1 + n3 (resp. n2 + n4 ) is drawn in blue (resp. red), and the
backbone of n1 (resp. n2 ) is the blue (resp. red) bold curve joining 0
and y. The outermost grey region represents the annulus Ann(ℓk+1 , ℓk+1 +
ℓνk+1 ). The clusters of the origin Cn1 +n3 (0) and Cn2 +n4 (0) intersect in
Ann(m, M ) thanks to a collection of red loops in (n2 + n4 ) \ Γ(n2 ) which
crosses Ann(M, N ), which means that n2 + n4 realises F3 . As a result,
this intersection is not measurable in terms of edges with both endpoints
in Ann(ℓk , ℓk+1 + ℓνk+1 ).

where we used once again the regularity assumption to compare ⟨σu σy ⟩β and ⟨σv σy ⟩β with
⟨σ0 σy ⟩β . Using Lemma 6.16, we also get,
−1
log2 (ℓk+1 /M ) 1

BM (β) ≥ 1 + C Bℓk+1 (β) ≥ Bℓ (β),
log2 (M ) 1 + C1 k+1
 2 
8
where C is the constant of Lemma 6.16 and C1 = C1 (ν) = C 1−ν +1 − 1 > 0.
Similarly,
1 + C1
Bm−1 (β) ≤ (1 + C1 )Bℓk (β) ≤ Bℓk+1 (β).
D
As a result, we obtain that for D large enough, there exists c3 > 0 such that
P0y,0y,∅,∅
β ̸ 0] ≥ c3 .
[|M| =
To conclude, we must prove uniqueness of the “crossing clusters” in n1 + n3 and n2 + n4 .
We reduce the argument to the one used in the nearest-neighbour case by not allowing
the currents to make big jumps. More precisely, define the jump event J by
[
J := {n1 + n3 ∈ Jump(p, p + pν )} ∪ {n2 + n4 ∈ Jump(p, p + pν )} .
p∈{ℓk ,ℓk+1 }
 TRIVIALITY OF CRITICAL ISING AND φ4 MODELS WITH deff ≥ 4 43

Using Lemma 6.5, we get that for some η > 0 and some constant C2 > 0 (provided
1+4/ε
ℓk+1 ≤ ℓk+2 , which might require to increase D),
C2
P0y,0y,∅,∅ [J] ≤ .
β
ℓηk
Note that on the complement of the above event, the cluster connecting 0 and y in n1 + n3
and n2 + n4 must go through Ann(p, p + pν ) for p ∈ {ℓk , ℓk+1 }. In particular, it satisfies
the “strong crossing” constraint in the definition of Ik . Take D large enough so that there
exists a constant c4 > 0 such that,
P0y,0y,∅,∅
β ̸ 0, Jc ] ≥ c4 .
[|M| =
If the event {M = ̸ ∅} ∩ Jc occurs but not Ik , then (see Figure 5) one of the following
events must occur for the pair (n1 , n3 ) (or (n2 , n4 )):
- F1 := the backbone Γ(n1 ) of n1 does a zigzag between scales ℓk and n, i.e. it
belongs to ZZ(0, y; ℓk , n, ∞),
- F2 := (n1 + n3 ) \ Γ(n1 ) belongs to Cross(n + nν , m),
- F3 := the backbone Γ(n1 ) of n1 does a zigzag between scales N + N ν and ℓk+1 ,
i.e. it belongs to ZZ(0, y; N + N ν , ℓk+1 , ∞),
- F4 := (n1 + n3 ) \ Γ(n1 ) belongs to Cross(M, N ),
- F5 := Jump(n, n + nν ) ∪ Jump(N, N + N ν ).
Note that the event F5 takes into account situations in which (for instance) (n1 + n3 ) \
Γ(n1 ) contains a cluster “almost realising” Cross(M, N ) and Γ(n1 ) “almost” performs
ZZ(0, y; N + N ν , ℓk+1 , ∞), and these two pieces are connected by a long open edge of
Γ(n1 ) \ Γ(n1 ) which jumps over Ann(N, N + N ν ).
Using Corollaries 6.9 and 6.15 we get the existence of C, η > 0 such that,
C C
P0y,∅
β [F1 ] ≤ η , Pβ0y,∅ [F2 ] ≤ η ,
n m
C C
P0y,∅
β [F3 ] ≤ , P0y,∅
β [F4 ] ≤ .
ℓηk+1 Nη
Moreover, as a consequence of Lemma 6.5, we also get that P0y,∅ β [F5 ] ≤ C/n .
η

As a result, taking D large enough, we get that the sum of the probabilities of the five
events for the pairs (n1 , n3 ) and (n2 , n4 ) does not exceed c4 /2 so that, setting κ := c4 /2,
P0y,0y,∅,∅
β [Ik ] ≥ κ.
□
6.4. Proof of the mixing. We now turn to the proof of the mixing property. Recall that
J satisfies (A1)-(A6).
Theorem 6.20 (Mixing property). Let d = 4 and s ≥ 1. There exist γ, c > 0 and
C = C(s) > 0, such that for every 1 ≤ t ≤ s, every β ≤ βc , every 1 ≤ nγ ≤ N ≤ L(β),
every xi ∈ Λn and yi ∈ / ΛN (i ≤ t), and every events E and F depending on the restriction
of (n1 , . . . , ns ) to edges with endpoints within Λn and outside ΛN respectively,

(6.5) Pxβ1 y1 ,...,xt yt ,∅,...,∅ [E ∩ F ] − Pxβ1 y1 ,...,xt yt ,∅,...,∅ [E]Pxβ1 y1 ,...,xt yt ,∅,...,∅ [F ]
−1/2
N

≤ C log .
n
Furthermore, for every x′1 , . . . , x′t ∈ Λn and y1′ , . . . , yt′ ∈
/ ΛN , we have that
−1/2
N

x y1′ ,...,xt yt′ ,∅,...,∅
Pxβ1 y1 ,...,xt yt ,∅,...,∅ [E] − Pβ1 [E] ≤ C log , (6.6)
n
44 R. PANIS

−1/2
N

x′ y1 ,...,x′t yt ,∅,...,∅
Pxβ1 y1 ,...,xt yt ,∅,...,∅ [F ] − Pβ1 [F ] ≤ C log . (6.7)
n
Fix β ≤ βc . Fix two integers t, s satisfying 1 ≤ t ≤ s. Introduce integers m, M such
that n ≤ m ≤ M ≤ N , m/n = (N/n)µ/2 , and N/M = (N/n)1−µ for µ small to be fixed.
For x = (x1 , . . . , xt ) and y = (y1 , . . . , yt ), write:
∅,...,∅
Pxy x1 y1 ,...,xt yt ,∅,...,∅
β := Pβ , Pβxy,∅ := Pxy
β ⊗ Pβ ,

where P∅,...,∅β is the law of a sum of s independent sourceless currents that we denote by
′ ′
(n1 , . . . , ns ).
If p ≥ 1, define for y ∈ / Λ2dp ,
|x − u|
   
Ay (p) := u ∈ Ann(p, 2p) : ∀x ∈ Λp/d , ⟨σx σy ⟩β ≤ 1 + C0 ⟨σu σy ⟩β ,
|y|
where C0 is the constant in the definition of regular scales. Note that if y is in a regular
scale, then Ay (p) = Ann(p, 2p).
Now, introduce the set K of regular scales k between m and M/2 with every element
of K such that the 2k (k ∈ K) differ by a multiplicative constant at least C0 (this will
be useful later to apply (P4)). By the existence of regular scales of Proposition 3.28, we
may assume that |K| ≥ c1 log(N/n) for a sufficiently small c1 = c1 (µ) > 0. Introduce
U := ti=1 Ui , where
Q

1 X 1 ni +n′i
1
X
Ui := [u ←→ xi ],
|K| k∈K Axi ,yi (2k ) k u∈Ayi (2 )

and,
⟨σx σu ⟩β ⟨σu σy ⟩β X
ax,y (u) := , Ax,y (p) := ax,y (u).
⟨σx σy ⟩β u∈Ay (p)

Using the switching lemma (SL), we see that Exy,∅ β [U] = 1. We begin by importing a
concentration inequality whose proof essentially relies on the definition of Ayi (2k ) and the
properties of regular scales15 (see [ADC21, Proposition 6.6]).
Lemma 6.21 (Concentration of U). For all γ > 2, there exists C = C(d, t, γ) > 0 such
that for all n sufficiently large satisfying nγ ≤ N ≤ L(β),
C
Exy,∅
β [(U − 1)2 ] ≤ .
log(N/n)
We now fix γ > 2 (it will be taken large enough later). Using Cauchy–Schwarz inequality
together with Lemma 6.21, we find C1 = C1 (d, t, γ) > 0 such that

Pxy xy,∅
U1[(n1 , . . . , ns ) ∈ E ∩ F ]
 
(6.8) β [E ∩ F ] − Eβ
r
C1
≤ Exy,∅
β [(U − 1)2 ] ≤ p .
log(N/n)
At this stage of the proof, we need to analyse Eβxy,∅ U1[(n1 , . . . , ns ) ∈ E ∩ F ] . By
 

definition of U, this term can be rewritten as a weighted sum of terms of the form
ni +n′
Exy,∅ i=1 1[ui ←→ yi ]1[(n1 , . . . , ns ) ∈ E ∩ F ] where ui ∈ Ayi (2 ) for some ki ∈ K. It
t
Q i
 ki
β
would be very tempting to try to apply the switching lemma directly to turn the measure
Pxy,∅
β into a measure Pxu,uy
β (up to a renormalisation weight). This would suggest that
15This is the only place where we need the property (P4) of regular scales. It also heavily relies on
(P3).
 TRIVIALITY OF CRITICAL ISING AND φ4 MODELS WITH deff ≥ 4 45

the occurrences of E and F are essentially due to (n1 , . . . , ns ) and (n1′ , . . . , ns′ ) respec-
tively under the measure Pxu,uy β . However, the occurrence of E ∩ F is only a function of
(n1 , . . . , ns ) which makes impossible the use of the switching lemma. Nevertheless, we can
still use the switching principle. This motivates the introduction of the following event.

Definition 6.22. Let u = (u1 , . . . , ut ) with ui ∈ Ann(m, M ) for every i. Introduce the
event G(u1 , . . . , ut ) = G(u) defined as follows: for every i ≤ s, there exists ki ≤ ni + ni′
such that ki = 0 on Λn , ki = ni + ni′ outside ΛN , ∂ki = {ui , yi } for i ≤ t, and ∂ki = ∅
for t < i ≤ s.

By the switching principle (SP), one has

ni +n′
 
Pβxy,∅ (n1 , . . . , ns ) ∈ E ∩ F, ui ←→i yi , ∀1 ≤ i ≤ t, G(u)
t
!
axi ,yi (ui ) Pxu,uy (n1 , . . . , ns ) ∈ E, (n1′ , . . . , ns′ ) ∈ F, G(u) .
Y  
= β
i=1

The trivial identity,

Pxu,uy
β [(n1 , . . . , ns ) ∈ E, (n1′ , . . . , ns′ ) ∈ F ]
uy ′ ′
= Pxu
β [(n1 , . . . , ns ) ∈ E]Pβ [(n1 , . . . , ns ) ∈ F ]

suggests to prove that under Pxu,uy

β , the event G(u) occurs with high probability. This
motivates the following result.

Lemma 6.23. Let d = 4. There exist C, ϵ > 0, γ = γ(ϵ) > 0 large enough and µ = µ(ϵ) >
0 small enough such that for every nγ ≤ N ≤ L(β), and every u with ui ∈ Ayi (2ki ) with
ki ∈ K for 1 ≤ i ≤ t,
t
!−1 ϵ
ni +n′i n
  
Pxy,∅ Pxu,uy
Y
β ui ←→ yi , ∀1 ≤ i ≤ t, G(u) axi ,yi (ui ) = β [G(u)c ] ≤C .
i=1
N

Proof. Below, C = C(d) > 0 may change from line to line16. The equality follows from
an application of the switching lemma (SL). If we write G(u) = ∩1≤i≤s Gi (where the
definition of Gi is implicit), then Hi ∩ Fi ⊂ Gi where,

Hi := {Ann(M, N ) is not crossed by a cluster in ni } = {ni ∈

/ Cross(M, N )},

and

Fi := {Ann(n, m) is not crossed by a cluster in ni′ } = {ni′ ∈

/ Cross(n, m)}.

Indeed, if Hi ∩Fi occurs, we may define ki as the sum of the restriction of ni to the clusters
intersecting ΛcN and the restriction of ni′ to the clusters intersecting Λcm . In the following,
we assume that 1 ≤ i ≤ t. The argument for other values of i is easier and follows from
the first case.
Introduce intermediate scales n ≤ r ≤ m ≤ M ≤ R ≤ N with r, R chosen below.
• Bound on Hi . Following the ideas developed in the proof of Lemma 6.19, we define,

Ji := {ni ∈ Jump(R − Rν , R + Rν )}.

16In fact, C will also depend on β (or more precisely on a lower bound on β ).
c c
46 R. PANIS

Notice that17,

Pxu,uy
β [Hic ] ≤ Pxu ν
β [Γ(ni ) ∈ ZZ(xi , ui ; M, R − R , ∞)]

+ Pxu ν xu
β [ni \ Γ(ni ) ∈ Cross(R + R , N )] + Pβ [Ji ].

Assume R = N ι where ι > µ is chosen in such a way that M 6/(1−ν) ≤ R and R8/(1−ν) ≤ N
(note that this is possible if we choose µ sufficiently small and γ sufficiently large since
M ≤ N µ+1/γ ). This choice allows us to use Corollaries 6.12 and 6.15 to get that for some
η > 0,
C C
Pxu ν
β [ni ∈ ZZ(xi , ui ; M, R − R , ∞)] ≤ η
, Pxu ν
β [ni \ Γ(ni ) ∈ Cross(R + R , N )] ≤ .
R Nη
Moreover, to the cost of diminishing the value of ι (and hence also modifying µ and γ)
to ensure that R1+4/ε ≤ N , we may use Lemma 6.5 to argue the existence of η ′ > 0 such
that
C
Pxu
β [Ji ] ≤ .
Rη ′
Putting all the pieces together, we get for some ϵ > 0,
ϵ
n

Pxu,uy
β [Hic ] ≤C .
N
• Bound on Fi . We proceed similarly for Fi although we now encounter an additional
difficulty: we cannot rule out the possibility of ni′ to jump above the annulus Ann(r −
rν , r + rν ), we can only rule it out in the complement of Γ(ni′ ). We set r = (n2 m)1/3 .
Recall that m ≥ N µ/2 . We claim that

(6.9) Pxu,uy
β [Fic ] ≤ Puy ′ ν
β [Γ(ni ) ∈ ZZ(ui , yi ; r + r , m, ∞)]

+ Puy ′ ′ ν uy
β [ni \ Γ(ni ) ∈ Cross(n, r − r )] + Pβ [Ki ],

/ Λr+rν such that (ni′ )a,b ≥ 2 and

where Ki is the event that there exists a ∈ Λr−rν and b ∈
{a, b} ∈ Γ(ni′ ) \ Γ(ni′ ).
Indeed, if none of the events corresponding to the two first probabilities on the right-
hand side of (6.9) occur, the only way we can find a cluster crossing Ann(n, m) is if
Γ(ni′ ) \ Γ(ni′ ) has a long (even) open edge which jumps above Ann(r − rν , r + rν ) (see
Figure 6).
Using (IRB) to get ⟨σui σv ⟩β ≤ Cm−2 and the assumption that ui ∈ Ayi (2ki ) to get
that ⟨σv σyi ⟩β ≤ C⟨σui σyi ⟩β , we obtain
⟨σui σv ⟩β ⟨σv σyi ⟩β
Puy ′
X
ν
β [Γ(ni ) ∈ ZZ(ui , yi ; r + r , m, ∞)] ≤
v∈Λr+rν
⟨σui σyi ⟩β
8
r4 n8/3 N 3γ
≤ C 2 = C 2/3 ≤ C µ .
m m N3
Moreover, using Corollary 6.15 (which requires that n8/(1−ν) ≤ r and hence decreases the
values of µ and 1/γ), there exist ζ > 0 such that
C
Puy ′ ′ ν
β [ni \ Γ(ni ) ∈ Cross(n, r − r )] ≤ .
rζ
17We need to create a “forbidden area” Ann(R − Rν , R + Rν ) between the two parts of the annulus
Ann(M, N ) to rule out the situation in which Γ(ni ) \ Γ(ni ) connects an “almost successful” excursion
of Γ(ni ) (in the sense that it almost crossed Ann(M, R − Rν )), with an “almost successful” excursion of
ni \ Γ(ni ). This is very similar to the argument used in the proof of Lemma 6.19.
 TRIVIALITY OF CRITICAL ISING AND φ4 MODELS WITH deff ≥ 4 47

We conclude the proof with the bound on Ki . If ni′ satisfies that (ni′ )a,b ≥ 2 where
a ∈ Λr−rν and b ∈ / Λr+rν with {a, b} ∈ Γ(ni′ ) \ Γ(ni′ ) being the earliest such edge. We
consider the map ni′ 7→ mi′ where (mi′ )a,b = (ni′ )a,b −1 and mi′ coincides with ni′ everywhere
else. This maps ni′ to a current mi′ with sources {ui , yi }∆{a, b} (b might coincide with
ui or yi ) such that the backbone Γ(mi′ ) always connects ui and b (by definition of the
exploration). Hence, using the chain rule for backbones,

⟨σ{ui ,yi }∆{a,b} ⟩β {ui ,yi }∆{a,b}

Puy [Γ(mi′ ) connects ui and b]
X
β [Ki ] ≤ βJa,b Pβ
a∈Λr−rν
⟨σui σyi ⟩β
b∈Λ
/ r+rν
X ⟨σui σb ⟩β ⟨σa σyi ⟩β
≤ βJa,b .
a∈Λr−rν
⟨σui σyi ⟩β
b∈Λ
/ r+rν

Since ui ∈ Ayi (2ki ), ⟨σa σyi ⟩β ≤ C⟨σui σyi ⟩β . Hence,

X ⟨σui σb ⟩β ⟨σa σyi ⟩β X
β Ja,b ≤ Cβ Ja,b ⟨σui σb ⟩β .
a∈Λr−rν
⟨σui σyi ⟩β a∈Λ r−r ν
b∈Λ
/ r+rν b∈Λ
/ r+rν

We distinguish two-cases according to whether b is close to ui or not. By (6.1) and (IRB),

X X Cr4
β Ja,b ⟨σui σb ⟩β ≤ Cr4 J0,x ≤ .
a∈Λr−rν |x|≥m/4
m2+ε
b∈Λ|ui |/2 (ui )

Again, using (6.1) and (IRB),

X Cr4 X Cr4−ν(2+ε)
Ja,b ⟨σui σb ⟩β ≤ J0,x ≤ .
a∈Λr−rν
m2 |x|≥rν m2
b∈Λ
/ r+rν ∪Λ|ui |/2 (ui )

By definition of r and m/n, we get the existence of ζ ′ > 0 such that

C
Puy
β [Ki ] ≤ .
mζ ′
As a result, if we choose µ and 1/γ sufficiently small, we get that for some ϵ′ > 0,
ϵ′
n

Pxu,uy
β [Fic ] ≤C .
N
□

We now turn to the proof of Theorem 6.20.

Proof of Theorem 6.20. Introduce the coefficients δ(u, x, y) defined by,

t
axi ,yi (ui )
δ(u, x, y) := 1[∃(k1 , . . . , kt ) ∈ Kt , u ∈ Ay1 (2k1 ) × . . . × Ayt (2kt )]
Y
.
i=1
|K|Axi ,yi (2ki )

Note that
X
δ(u, x, y) = 1.
(k1 ,...,kt )∈Kt
u∈Ay1 (2k1 )×...×Ayt (2kt )
48 R. PANIS

yi

ui

Λn

∂Λm

Figure 6. An illustration of the event Ki . We represented the current ni′

in black, with a bold line representing its backbone Γ(ni′ ). The red/green
dashed lines correspond to the open edges of Γ(ni′ )\Γ(ni′ ), they all carry an
non-zero even weight. The grey region corresponds to the annulus Ann(r −
rν , r + rν ). In this picture, Γ(ni′ ) does not cross Ann(r + rν , m), ni′ \ Γ(ni′ )
does not cross Ann(n, r − rν ), yet the long green open edge of Γ(ni′ ) \ Γ(ni′ )
creates a cluster of ni′ crossing Ann(n, m).

Equation (6.8) together with Lemma 6.23 yield,

Pxy uy
X
(6.10) β [E ∩ F ] − δ(u, x, y)Pxu
β [E]Pβ [F ]
u
ϵ
C1 n C3

≤p + C2 ≤p ,
log(N/n) N log(N/n)

as long as N ≥ nγ where γ > 2 is given by Lemma 6.23. We begin by proving (6.6)

when yi , yi′ are in regular scales (but not necessarily the same ones). Applying the above
inequality once for y and once for y′ with the event E and F = ΩZd ,

′ 2C3
Pxy xy
(δ(u, x, y) − δ(u, x, y′ ))Pxu
X
β [E] − Pβ [E] ≤ β [E] + p . (6.11)
u log(N/n)

Since all the yi , yi′ are in regular scales, one has Ayi (2ki ) = Ayi′ (2ki ) = Ann(2ki , 2ki +1 ).
Moreover, using Property (P2) of regular scales,
1−µ
M n
  
′
δ(u, x, y) − δ(u, x, y ) ≤ C4 δ(u, x, y) ≤ C5 δ(u, x, y),
N N
where µ is also given by Lemma 6.23. Indeed, in that case δ(u, x, y) and δ(u, x, y′ ) are
both close to
Y ⟨σxi σui ⟩
.
|K| vi ∈Ann(2ki ,2ki +1 ) ⟨σxi σvi ⟩
P
i≤t

This gives (6.6) in that case. Now, assume that N ≥ n2(γ/µ)+1 (so that m ≥ nγ ). Consider
z = (z1 , . . . , zt ) with zi ∈ Ann(m, M ) in a regular scale. Also, pick y on which we do not
 TRIVIALITY OF CRITICAL ISING AND φ4 MODELS WITH deff ≥ 4 49

assume anything. We have,

Pxy xz
Pxy
X
β [E] − Pβ [E] = β [E] − δ(u, x, y)Pxz
β [E]
u
C6
Pxy
X
≤ β [E] − δ(u, x, y)Pxu
β [E] + p
u log(m/n)
C7
≤ p ,
log(N/n)
where in the second line we used (6.11) with (y, y′ ) = (u, z) together with the fact that
m/n = (N/n)µ/2 ≥ nγ , and in the third line we used (6.10) with F = ΩZd . This gives
(6.6).
The same argument works for (6.7) for every x, x′ , y, noticing that for every regular u
for which δ(u, x, y) ̸= 0, using once again (P2),
µ/2
n n
  
′
δ(u, x, y) − δ(u, x , y) ≤ C8 δ(u, x, y) ≤ C9 δ(u, x, y).
m N
To get (6.5) we repeat the same line of reasoning. We start by applying (6.10). Then,
xy uy
we replace each Pxu
β [E] by Pβ [E] using the above reasoning. Finally, replace Pβ [F ] by
xy
Pβ [F ] using the same methods.
□

6.5. Proof of the clustering bound. With the intersection property and the mixing
statement, we are now in a position to conclude. We will use the following (natural)
monotonicity property of the currents measures.

Proposition 6.24 (Monotonicity in the number of sources, [ADC21, Corollary A.2]). For
every β > 0, for every x, y, z, t ∈ Zd and every set S ⊂ Zd , one has,

P0x,0z,0y,0t
β [Cn1 +n3 (0) ∩ Cn2 +n4 (0) ∩ S = ∅]
≤ Pβ0x,0z,∅,∅ [Cn1 +n3 (0) ∩ Cn2 +n4 (0) ∩ S = ∅].

Proof of Proposition 6.2. Fix γ > 2 sufficiently large so that Theorem 6.20 holds. Re-
member by Remark 6.17 that we may choose D = D(γ) sufficiently large in the definition
of L such that ℓk+1 ≥ ℓγk .
We may assume u = 0. Since x, y are at distance at least 2ℓK of each other, one of
them must be at distance at least ℓK of u. Without loss of generality we assume that this
is the case of x and make the same assumption about z. Let δ > 0 to be fixed below.
(δ)
Let SK denote the set of subsets of {2δK, . . . , K − 3} which contain only even integers.
(δ)
If {Mu (I; L, K) < δK} occurs, then there must be S ∈ SK with |S| ≥ (1/2 − 2δ)K such
that BS occurs, where BS is the event that the clusters of 0 in n1 + n3 and n2 + n4 do
not intersect in any of the annuli Ann(ℓi , ℓi+1 ) for i ∈ S. Using the monotonicity property
recalled above,

P0x,0z,0y,0t Pβ0x,0z,0y,0t [BS ]

X
(6.12) β [Mu (I; L, K) < δK] ≤
(δ)
S∈SK
|S|≥(1/2−2δ)K

Pβ0x,0z,∅,∅ [BS ].
X
≤
(δ)
S∈SK
|S|≥(1/2−2δ)K
50 R. PANIS

Let J be the event defined by

K−3
[
J := {n1 + n3 ∈ Jump(ℓk , ℓk+kν )} ∪ {n2 + n4 ∈ Jump(ℓk , ℓk+kν )}.
k=2δK

Using Lemma 6.5 and Remark 6.17, if D is large enough, there exist C0 , C1 , η > 0 such
that
C0 K δK
P0x,0z,∅,∅
β [J] ≤ η ≤ C1 e−η2 . (6.13)
ℓδK
(δ)
Fix some S ∈ SK . Let AS be the event that none of the events Ik (defined in Definition
6.18) occur for k ∈ S.
We now make a crucial observation: if k ∈ S and Jc occurs, the events Ik and BS are
incompatible. Indeed, The occurrence of Jc ∩ Ik ensures that the only cluster
of n1 + n3
(resp. n2 + n4 ) crossing Ann(ℓk , ℓk+1 ) is the cluster of 0. Hence, BS ∩ Jc ⊂ AS . Using
(6.13),
!
−η2δK (1/2 − 2δ)K
Pβ0x,0z,0y,0t [Mu (I; L, K) Pβ0x,0z,∅,∅ [AS ]
X
< δK] ≤ + C1 e .
(δ)
2δK
S∈SK
|S|≥(1/2−2δ)K

Standard binomial estimates give that the second term on the right-hand side above is
bounded by C2 2−δK . We are left with the study of Pβ0x,0z,∅,∅ [AS ].
Since ℓK ≤ L(β), we know by Proposition 3.28 that there exists y ∈ Ann(ℓK−1 , ℓK ) in
a regular scale. The event AS depends on edges with both endpoints in ΛℓK−2 . Using the
mixing property together with Remark 6.17, we get that
−c(K−1)
C3 D

P0x,0z,∅,∅
β [AS ] ≤ P0y,0y,∅,∅
β [AS ] + ≤ P0y,0y,∅,∅
β [AS ] + C4 .
(log ℓK−1 )c log 2
Let s = max S. Using once again the mixing property between n = ℓs−1 and N = ℓs , we
get,

C3
P0y,0y,∅,∅
β [AS ] ≤ P0y,0y,∅,∅
β [Isc ]P0y,0y,∅,∅
β [AS\{s} ] +
(log ℓs /ℓs−1 )c
≤ (1 − κ)P0y,0y,∅,∅
β [AS\{s} ] + C5 (1 − κ)|S|−1 ,

where κ is given by Lemma 6.19 and where we chose D large enough. Iterating the above
yields,

Pβ0y,0y,∅,∅ [AS ] ≤ C6 (1 − κ)|S| .

Going back to (6.12), we get that

!
(1/2 − 2δ)K
Pβ0x,0z,0y,0t [Mu (I; L, K) < δK] ≤ C7 (1 − κ)(1/2−2δ)K + C2 2−δK ≤ C8 2−δK ,
2δK

for δ > 0 sufficiently small. □

6.6. Proof of Corollary 1.8. Now that we were able to obtain the improved the tree
diagram bound, the proof of Corollary 1.8 follows the strategy used in Section 5, although
some additional technical difficulties appear due to the lack of knowledge on the growth
of BL (β). We refer to [ADC21] for the details of the proof.
 TRIVIALITY OF CRITICAL ISING AND φ4 MODELS WITH deff ≥ 4 51

Proof of Corollary 1.8. We import notations from the proof of Theorem 5.5. Applying
the improved tree diagram bound we obtain
X ⟨σx σx1 ⟩β ⟨σx σx2 ⟩β ⟨σx σx3 ⟩β ⟨σx σx4 ⟩β
S(β, L, f ) ≤ C , (6.14)
BL(x1 ,x2 ,x3 ,x4 ) (β)c ΣL (β)2
x∈Zd
x1 ,x2 ,x3 ,x4 ∈Λrf L

where L(x1 , x2 , x3 , x4 ) is the minimal distance between the xi . We now fix a ∈ (0, 1). The
strategy consists in splitting the right-hand side of (6.14) according to the following four
possibilities: x ∈ Λdrf L and L(x1 , x2 , x3 , x4 ) ≤ La , x ∈
/ Λdrf L and L(x1 , x2 , x3 , x4 ) ≤ La ,
x ∈ Λdrf L and L(x1 , x2 , x3 , x4 ) > La , x ∈
/ Λdrf L and L(x1 , x2 , x3 , x4 ) > La .
From there the conclusion builds on the same tools as the ones used in Section 5, and
also on Lemma 6.16. □
Remark 6.25. In [ADC21], the proof of this result crucially relies on a sharp sliding-scale
infrared bound which leads to a sharp bound in Lemma 6.16. This shows how remarkable
this bound is since it yields a quantitative decay even though we do not even know that
BL (βc ) explodes. This argument breaks down in the next section (for d ∈ {2, 3}) due to
the lack of such a sharp result.

7. Reflection positive Ising models in dimension 1 ≤ d ≤ 3 satisfying deff = 4

We now explain how to adapt the above strategy to the remaining “marginal case”:
deff = 4 > d. As seen in Section 5 (and more precisely in Remark 5.3), for algebraically
decaying RP interactions Jx,y = C|x−y|−d−α1 , this corresponds to choosing d−2(α∧2) = 0.
We will assume that J satisfies (A1)-(A5) with the complementary assumption18 that
there exist c, C > 0 such that, for all x ∈ Zd \ {0},
c C
≤ J0,x ≤ , (7.1)
|x|d+α(d) |x|d+α(d)
where α(d) = d/2.
Our goal is to prove the following result, which is a slightly stronger version of Theorem
1.9.
Theorem 7.1 (Improved tree diagram bound). Let 1 ≤ d ≤ 3. Assume that J satisfies
(A1)-(A5) together with (7.1). There exists C > 0 such that the following holds: for all
β ≤ βc , there exists an increasing function ϕβ : R → R> such that for all x, y, z, t ∈ Zd at
mutual distance at least L of each other with 1 ≤ L ≤ L(β),
C
|U4β (x, y, z, t)| ≤
X
⟨σx σu ⟩β ⟨σy σu ⟩β ⟨σz σu ⟩β ⟨σt σu ⟩β .
ϕβ (BL (β)) d u∈Z
If B(βc ) = ∞, one has ϕβc (t) → ∞ as t → ∞.
Remark 7.2. We could replace ϕβ (BL (β)) by BL (β)c (as in Theorem 1.8) provided we
could prove the following (under the assumptions of Theorem 7.1): there exists C > 0 such
that, if 1 ≤ ℓ ≤ L ≤ L(β),
χN (β) χn (β)
≤ C γ(d) ,
N γ(d) n
where γ(2) = 1 and γ(3) = 3/2. This will become more transparent below.
When d = 1, such a sharp result is obtained thanks to the exact knowledge on the decay
of the two-point function.
18This assumption is a little more restrictive than before since we now require some regularity property
on the interaction. This restriction is essentially technical and we believe that the proof which follows
should hold under more general assumptions.
52 R. PANIS

From this result, we will obtain

Corollary 7.3. We keep the assumptions of Theorem 7.1. Then, for σ ∈ (0, d/2),
lim gσ (β) = 0.
β↗βc

As a consequence, for β = βc , every sub-sequential scaling limit of the model is Gaussian.

To prove the improved tree diagram bound, we will extend the strategy of Section 6
and prove a mixing statement together with the intersection property. The proofs heavily
relies on a finer analysis of the geometry of the clusters compared to what was done in
Section 6.2 (see Remark 6.10), that is in fact valid in a wider generality (i.e. also when
deff > 4). Hence, we begin by proving the mixing statement in its most general form.
In Sections 7.1, 7.2, and 7.3 we consider an interaction J on Zd (d ≥ 1) satisfying
(A1)-(A5) together with the following assumption: there exist c1 , C1 > 0 such that, for
all x ∈ Zd \ {0},
c1 C1
d+α
≤ J0,x ≤ , (Assumptionα )
|x| |x|d+α
where α > 0 will be specified below. By the results of Section 3, it has for consequence
the existence of C2 > 0 such that: for all β ≤ βc , for all x ∈ Zd \ {0},
C2
⟨σ0 σx ⟩β ≤ . (IRBα )
βc |x|d−α∧2 (log |x|)δ2,α
Moreover, using Proposition 3.25, we find that there exists C3 > 0 such that: for all
β ≤ βc , for all x ∈ Zd \ {0} with 1 ≤ |x| ≤ L(β),
C3
⟨σ0 σx ⟩β ≥ , (LBα )
β|x|d−1 fα (|x| + 1)
where if t > 1 fα (t) := 1 if α > 1, f1 (t) := log t, and fα (t) := t1−α for α ∈ (0, 1).

7.1. Existence of weak regular scales. In the case d ≥ 3, the existence of regular
scales was proved in Section 3. For d = 2, the proof of (P4) failed. The reason behind this
is purely technical: the sliding-scale infrared bound is not optimal in dimension 2 since
we expect the growth of χn (βc ) to be smaller than n2 . We can circumvent this technical
difficulty by allowing ourselves a “weaker” property (P4′ ) in the definition of a regular
scale.
Definition 7.4 (Weak regular scales). Fix c, C > 0. An annular region Ann(n/2, 8n) is
said to be (c, C)-weak regular if it satisfies the properties (P1)-(P3) and
(P4′ ) For every x ∈ Λn and y ∈
/ ΛC(log n)2 n , Sβ (y) ≤ 21 Sβ (x).
A scale k is said to be weak regular if n = 2k is such that Ann(n/2, 8n) is (c, C)- weak
regular, a vertex x ∈ Zd will be said to be in a weak regular scale if it belongs to an annulus
Ann(n, 2n) with n = 2k and k a weak regular scale.
Proposition 7.5 (Existence of weak regular scales). Let d ≥ 1. Assume that J satisfies
(A1)-(A5) and (Assumptionα ) where α > 0 if d ≥ 3, α ∈ (0, 1] if d = 2, and α ∈ (0, 1)
if d = 1. Let γ > 2. There exist c0 , c1 , C0 > 0 such that for every β ≤ βc , and every
1 ≤ nγ ≤ N ≤ L(β), there are at least c1 log2 N n (c0 , C0 )-weak regular scales between n
and N .
Proof. The cases d ≥ 3, and d ∈ {1, 2} with α ∈ (0, 1) were already settled in Propositions
3.28 and 3.29 since (P4′ ) is weaker than (P4). We only need to take care of the case
 TRIVIALITY OF CRITICAL ISING AND φ4 MODELS WITH deff ≥ 4 53

d = 2 and α = 1. Using (LBα ) together with (IRBα ), we get that for x ∈ Zd with
2 ≤ |x| ≤ L(β),
c1 C1
≤ ⟨σ0 σx ⟩β ≤ . (7.2)
β|x|(log |x|) β|x|
Using (7.2) and the assumption19 1 ≤ nγ ≤ N , we get the existence of c2 , c3 , c4 > 0 such
that,
c2 N c2 N n N 1 N 1/2
     
χN (β) ≥ = ≥ c3 χn (β) ≥ c4 χn (β).
β log N β n log N n log N n
Using Theorem 3.18, we find r, c5 > 0 and independent of n, N , such that there are at
least c5 log2 (N/n) scales m = 2k between n and N such that
χrm (β) ≥ χ16dm (β) + χm (β). (7.3)
We prove that such an m is a (c0 , C0 )-weak regular scale for a good choice of c0 , C0 . The
proof of (P1)-(P3) follows the same line as before. Now, using (7.2), we get that for
1 ≤ ℓ ≤ L with ℓ ≤ L(β),
χL (β) log ℓ
≤ C2 χℓ (β). (7.4)
L ℓ
Let R ≥ 1. Using the same strategy we used to obtain (3.12) and replacing the sliding-scale
infrared bound by (7.4), we get for y ∈
/ ΛdR(log m)2 m and x ∈ Λm ,
|ΛR(log m)2 m |Sβ (y) ≤ χR(log m)2 m (β) ≤ C3 R(log m)3 χm (β) ≤ C4 R(log m)3 m2 Sβ (x),
which implies that
C5 1
Sβ (y) ≤ Sβ (x) ≤ Sβ (x),
R log m 2
if R is large enough. This concludes the proof. □
7.2. Properties of the current. In this subsection, we let d ≥ 1 and assume that J
satisfies (A1)-(A5) together with (Assumptionα ) with d − 2(α ∧ 2) ≥ 0. As explained
in Section 5, this choice corresponds to deff ≥ 4.
We import the notations from Section 6.2. The main difficulty below will come from
the fact that Jump(k, k + k ν ) (for ν < 1) now occurs with high probability (if α ∈ (0, 2]).
However, to the price of considering thicker annuli we can keep a similar statement.
Many of the computations done here are very similar to what was done in Section 6.2
so we only present the main changes and omit the trivial modifications.
Lemma 7.6. Let d ≥ 1. Assume that J satisfies (A1)-(A5) and (Assumptionα ) with
α > 0. Let ϵ > 0. There exist C, η > 0 such that for all β ≤ βc , y ∈ Zd in a weak regular
scale, with 1 ≤ |y| ≤ L(β), and for all k ≥ 1 such that k 4 ≤ |y|,
C
P0y,∅
β [Jump(k, k
1+ϵ
)] ≤
.
kη
Proof. We repeat the strategy of proof of Lemma 6.5. Lemma 6.3 yields

P0y,∅
X
1+ϵ
β [Jump(k, k )] ≤ 2β Ju,v ⟨σu σv ⟩β
u∈Λk , v ∈Λ
/ k1+ϵ
!
⟨σ0 σu ⟩β ⟨σv σy ⟩β ⟨σ0 σv ⟩β ⟨σu σy ⟩β
+ + =: A1 + A2 + A3 .
⟨σ0 σy ⟩β ⟨σ0 σy ⟩β

19In fact n ≤ N/ log N is enough.

54 R. PANIS

Using (IRBα ) and (Assumptionα ),

X 1 kd
A1 ≤ C 1 ≤ C2 .
u∈Λk , v ∈Λ
/ k1+ϵ
|u − v|d+α+d−α∧2 k (1+ϵ)(d+α−α∧2)

Using (P1) of weak regular scales, together with (IRBα ) and (Assumptionα ), we simi-
larly obtain
kd
A2 ≤ C3 .
k (1+ϵ)(d+α−α∧2)
Finally, proceeding as in the proof of Lemma 6.5, using additionally (LBα ),
X ⟨σ0 σu ⟩β ⟨σv σy ⟩β k α∧2 |y|d−1 fα (|y|)|y|α∧2
β Ju,v ≤ C5 ,
u∈Λk , v∈Λ|y|/2 (y)
⟨σ0 σy ⟩β |y|d+α

where the right-hand side is bounded by C5 k α /|y|α for α ∈ (0, 1), by C5 k log(|y|)/|y| for
α = 1, and by C5 k α∧2 /|y| for α > 1. Hence, if k 4 ≤ |y|, we always have that it is bounded
by C4 /|y|δ for some δ = δ(α) > 0.
Using again (P1) as in Lemma 6.5,
X ⟨σ0 σu ⟩β ⟨σv σy ⟩β k α∧2
β Ju,v ≤ C6 (1+ϵ)α .
/ k1+ϵ ∪Λ|y|/2 (y)
u∈Λk , v ∈Λ
⟨σ0 σy ⟩β k

This concludes the proof.

□
Remark 7.7. The main contribution to the above probability comes from the term A2
pictured in Figure 4.
Despite being equipped with a very weak version of Lemma 6.5, we can still obtain a
version of Corollary 6.9 in our context.
Corollary 7.8 (No zigzag for the backbone). Let d ≥ 1. Assume that J satisfies (A1)-
(A5) and (Assumptionα ) with d − 2(α ∧ 2) ≥ 0. Fix ν ∈ (0, 1) and ϵ > 0. There exist
C, η > 0 such that, for all β ≤ βc , for all k, ℓ ≥ 1 and y ∈ Zd in a weak regular scale with
k [2d/(1−ν)]∨[2d/(να)] ≤ ℓ and ℓ4 ≤ |y|,
C
P0y
β [Γ(n1 ) ∈ ZZ(0, y; k, ℓ, ∞)] ≤ .
ℓη
Proof. Notice that
ZZ(0, y; k, ℓ, ∞) ⊂ ZZ(0, y; k, ℓ, ℓ1+ϵ ) ∪ Jump(ℓ, ℓ1+ϵ ).
Using Lemma 7.6 we find C1 , η > 0 such that
C1
P0y
β [Jump(ℓ, ℓ
1+ϵ
)] ≤
.
ℓη
If ZZ(0, y; k, ℓ, ℓ1+ϵ ) occurs, there are two possibilities: either the backbone actually visits
Ann(ℓ, ℓ + ℓν ) before hitting Λk , an event we denote by B1 ; or it does not in which case
there must be an open edge which jumps from Λℓ to Ann(ℓ + ℓν , ℓ1+ϵ ), an event we denote
by B2 . By the chain rule for the backbone, we find that,
⟨σ0 σu ⟩β ⟨σu σv ⟩β ⟨σv σy ⟩β k d ℓd−1+ν C2
P0y
X
β [B1 ] ≤ ≤ C2 2d−2(α∧2) ≤ (1−ν)/2 ,
u∈Ann(ℓ,ℓ+ℓν )
⟨σ0 σy ⟩β ℓ ℓ
v∈Λk

where we used (IRBα ), the property (P2) of weak regular scales to compare ⟨σv σy ⟩β and
⟨σ0 σy ⟩β , and the hypothesis d − 2(α ∧ 2) ≥ 0. Using (B.2), we see that for a one-step walk
 TRIVIALITY OF CRITICAL ISING AND φ4 MODELS WITH deff ≥ 4 55

γ : a → b, one has ρ(γ) ≤ ρ{a,b} (γ) = tanh(βJa,b ). Combining this observation with the
chain rule,
⟨σ0 σa ⟩β tanh(βJa,b )⟨σb σc ⟩β ⟨σc σy ⟩β C3 k d ℓα∧2
P0y
X
β [B2 ] ≤ ≤ d−α∧2 να ,
a∈Λℓ
⟨σ0 σy ⟩β ℓ ℓ
b∈Ann(ℓ+ℓν ,ℓ1+ϵ )
c∈Λk

where we used (P2) to compare ⟨σc σy ⟩β and ⟨σ0 σy ⟩β , (IRBα ) to argue that
X kd
⟨σb σc ⟩β ≤ C4 ,
c∈Λk
ℓd−α∧2
and (IRBα ) once again with (Assumptionα ) to get
X ℓα∧2
⟨σ0 σa ⟩β tanh(βJa,b ) ≤ C5 .
a∈Λℓ
ℓαν
b∈Λ
/ ℓ+ℓν

This concludes the proof. □

We can also obtain the corresponding modification of Lemma 6.11.
Lemma 7.9. Let d ≥ 1. Assume that J satisfies (A1)-(A5) and (Assumptionα ) with
d − 2(α ∧ 2) ≥ 0. Let ϵ > 0. There exist C, η > 0 such that, for all β ≤ βc , for
n < m ≤ M ≤ k with 1 ≤ M 2/ϵ ≤ k ≤ L(β), for all x ∈ Λn , and all u ∈ Ann(m, M ),
C
Pxu,∅
β [Jump(k, k 1+ϵ )] ≤ η .
k
Proof. We repeat the proof of Lemma 6.11. Using Lemma 6.3,

Pxu,∅
X
β [nw,v ≥ 1]
w∈Λk , v ∈Λ
/ k1+ϵ
!
X ⟨σx σw ⟩β ⟨σv σu ⟩β ⟨σx σv ⟩β ⟨σw σu ⟩β
≤ 2β Jw,v ⟨σw σv ⟩β + + .
w∈Λk , v ∈Λ
/ k1+ϵ
⟨σx σu ⟩β ⟨σx σu ⟩β

Using (Assumptionα ) and (IRBα ),

X X pd−1 C2
β Jw,v ⟨σw σv ⟩β ≤ C1 k d ≤ . (7.5)
pd+α+d−(α∧2) k dϵ
w∈Λk , v ∈Λ
/ k1+ϵ p≥k1+ϵ

Then, using (LBα ) (which is licit since 1 ≤ |x−u| ≤ L(β)) together with (Assumptionα )
and (IRBα ), we get
X ⟨σx σv ⟩β ⟨σw σu ⟩β X
β Jw,v ≤ β 2 C3 M d−1 fα (M ) Jw,v ⟨σx σv ⟩β ⟨σw σu ⟩β
w∈Λk , v ∈Λ
/ k1+ϵ
⟨σx σu ⟩β w∈Λk
v ∈Λ
/ k1+ϵ

|v|−(d−α∧2) J0,v
X
≤ C4 M d−1 fα (M )k α∧2
v ∈Λ
/ k1+ϵ
d α∧2−d(1+ϵ)
≤ C5 M k .
Finally, with the same reasoning we also get
X ⟨σx σw ⟩β ⟨σv σu ⟩β
Jw,v ≤ C6 M d k α∧2−d(1+ϵ) .
w∈Λ , v ∈Λ
/
⟨σ σ ⟩
x u β
k k1+ϵ

Now, clearly, if M 2/ϵ

≤ k, using that d = 2(α ∧ 2) ≥ 0, we can choose η = dϵ/2 and C a
sufficiently large constant. □
56 R. PANIS

x
Λn a b

∂ΛM Ann(k 1/(1+ϵ) , k) Ann(k + k ν , k 1+ϵ )

Figure 7. An illustration of the occurrence of the event B2 defined in the

proof of Corollary 7.10. The ”exclusion zone“ Ann(k, k + k ν ) is represented
in red. The green dashed line illustrates the long open edge which jumps
above it.

Corollary 7.10. Let d ≥ 1. Assume that J satisfies (A1)-(A5) and (Assumptionα )

with d − 2(α ∧ 2) ≥ 0. There exist ν ∈ (0, 1) and C, ϵ, η > 0 such that for all β ≤ βc , for
all n < m ≤ M ≤ k with 1 ≤ M [2(1+ϵ)/ϵ]∨[2d/(1−ν)] ≤ k ≤ L(β), for all x ∈ Λn and all
u ∈ Ann(m, M ),
C
Pxu
β [ZZ(x, u; M, k, ∞)] ≤ .
kη
Proof. We repeat the strategy used to get Corollary 7.8. Let ν ∈ (0, 1) and ϵ > 0 to be
fixed below. Notice that,

ZZ(x, u; M, k, ∞) ⊂ ZZ(x, u; M, k, k 1+ϵ ) ∩ (Jump(k 1/(1+ϵ) , k))c

∪ Jump(k 1/(1+ϵ) , k) ∪ Jump(k, k 1+ϵ ).
Using Lemma 7.9 together with the hypothesis on M and k, we find C1 , η > 0 such that
C1
Pxu
β [Jump(k
1/(1+ϵ)
, k) ∪ Jump(k, k 1+ϵ )] ≤ .
kη
We handle ZZ(x, u; M, k, k 1+ϵ ) ∩ (Jump(k 1/(1+ϵ) , k))c as we did in the proof of Corollary
7.8 by splitting it into two events B1 and B2 according to whether or not the backbone
reaches Ann(k, k + k ν ). Using the chain rule together with (LBα ) and (IRBα ),
X ⟨σx σv ⟩β ⟨σv σu ⟩β
Pxu
β [B1 ] ≤
v∈Ann(k,k+kν )
⟨σx σu ⟩β

C2 M d−1 fα (M )k d−1+ν
≤
k 2d−2(α∧2)
C2 M d C2
≤ ≤ (1−ν)/2 ,
k 1−ν k
where we used the assumption that d ≥ 2(α ∧ 2).
 TRIVIALITY OF CRITICAL ISING AND φ4 MODELS WITH deff ≥ 4 57

It remains to analyse B2 . In that case, the backbone has to jump above Ann(k, k +
kν ): it goes from x to a point in Ann(k 1/(1+ϵ) , k) (recall that we excluded jumps above
Ann(k 1/(1+ϵ) , k)), then jumps in Ann(k + k ν , k 1+ϵ ), before finally hitting u (see Figure 7).
Using the chain rule for the backbone as we did in the proof of Corollary 7.8 together
with (Assumptionα ), (LBα ), and (IRBα ), we get
X ⟨σx σa ⟩β tanh(βJa,b )⟨σb σu ⟩β
Pxu
β [B2 ] ≤
⟨σx σu ⟩β
a∈Ann(k1/(1+ϵ) ,k)
b∈Ann(k+kν ,k1+ϵ )

C3 M d−1 fα (M )k d k d(1+ϵ)
≤ d−α∧2 .
k 1+ϵ k ν(d+α) k d−α∧2
Now, recall that M d ≤ k dϵ , so that we may find ζ = ζ(ϵ, ν) > 0 such that
C3 k 2d+3dϵ C3
Pxu
β [B2 ] ≤ d−α∧2 ≤ ,
k k ν(d+α) k d−α∧2
1+ϵ kζ
provided ϵ > 0 is small enough, and ν is sufficiently close to 1. This concludes the
proof. □

Finally, as we did in Section 6.2, we conclude this subsection with some properties
concerning the current n \ Γ(n).
Lemma 7.11. Let d ≥ 1. Assume that J satisfies (A1)-(A5) and (Assumptionα ) with
α > 0. Let ϵ > 0. There exist C, η > 0 such that for all β ≤ βc , for all k ≥ 1, for all
x, y ∈ Zd ,
C
Pxy
β [n \ Γ(n) ∈ Jump(k, k
1+ϵ
)] ≤ Pxy,∅
β [(n1 + n2 ) \ Γ(n1 ) ∈ Jump(k, k
1+ϵ
)] ≤ .
kη
Proof. Proceeding exactly as in the proof of Corollary 6.15 we find that
Pxy,∅
X
1+ϵ
β [(n1 + n2 ) \ Γ(n1 ) ∈ Jump(k, k )] ≤ 2 βJu,v ⟨σu σv ⟩β .
u∈Λk , v ∈Λ
/ k1+ϵ

This last sum is then smaller than Ck −dϵ as shown in (7.5). □

Recall that the event Cross was defined in Definition 6.14.

Corollary 7.12. Let d ≥ 1. Assume that J satisfies (A1)-(A5) and (Assumptionα )
with d − 2(α ∧ 2) ≥ 0. There exist ν ∈ (0, 1) and C, ϵ, η > 0 such that for all β ≤ βc , for
all k, ℓ ≥ 1 with k 2d/(1−ν) ≤ ℓ, for all x, u ∈ Zd ,
xu,∅ C
Pxu
β [n \ Γ(n) ∈ Cross(k, ℓ)] ≤ Pβ [(n1 + n2 ) \ Γ(n1 ) ∈ Cross(k, ℓ)] ≤ .
ℓη
Proof. We use the ideas developed in the proofs of Lemma 6.13 and Corollary 7.10. Let
ν ∈ (0, 1) and ϵ > 0 to be fixed below. Start by writing,
{(n1 + n2 ) \ Γ(n1 ) ∈ Cross(k, ℓ)} = [(B1 ∪ B2 ) ∩ Jc ] ∪ J,
where B1 is the event that (n1 +n2 )\Γ(n1 ) crosses Ann(k, ℓ) by passing through Ann(ℓ, ℓ+
ℓν ), B2 is the complement of B1 in {(n1 + n2 ) \ Γ(n1 ) ∈ Cross(k, ℓ)}, and J = {(n1 + n2 ) \
Γ(n1 ) ∈ Jump(ℓ1/(1+ϵ) , ℓ)} ∪ {(n1 + n2 ) \ Γ(n1 ) ∈ Jump(ℓ, ℓ1+ϵ )}.
Using Lemma 7.11, we get the existence of C1 , η1 > 0 such that
C1
Pxu,∅
β [J] ≤ .
ℓη1
 58 R. PANIS

Using (IRBα ) together with the what was done in the proof of Corollary 6.15,
C2 ℓd−1+ν k d C2
Pxu,∅
X
β [B1 , Jc ] ≤ ⟨σw σv ⟩2β ≤ 2d−2(α∧2)
≤ (1−ν)/2 ,
v∈Ann(ℓ,ℓ+ℓν )
ℓ ℓ
w∈Λk

where we used that d − 2(α ∧ 2) ≥ 0. Finally, using a similar strategy as in the proof of
Lemma 6.13,
∅,∅
Pxu,∅
X X
β [B2 , Jc ] ≤ Pxu c
β [Γ(n) = γ]Pγ c ,Zd ,β [a ↔ b in γ , (n1 + n2 )b,c ≥ 1].
γ:x→u a∈Λk
consistent
b∈Ann(ℓ1/(1+ϵ) ,ℓ)
c∈Ann(ℓ+ℓν ,ℓ1+ϵ )

Using the generalisation of the switching lemma mentioned in the proof of Corollary 6.15
together with Griffith’s inequality,

Pγ∅,∅ c
c ,Zd ,β [a ↔ b in γ , (n1 + n2 )b,c ≥ 1] = ⟨σa σb ⟩γ c ,β ⟨σa σb ⟩β Pab,ab
γ c ,Zd ,β
[(n1 + n2 )b,c ≥ 1]
≤ βJb,c ⟨σa σc ⟩γ c ,β ⟨σa σb ⟩β + βJb,c ⟨σa σb ⟩γ c ,β ⟨σa σc ⟩β
≤ 2⟨σa σb ⟩β βJb,c ⟨σc σa ⟩β .
We obtained,

Pxu,∅
X
β [B2 , Jc ] ≤ 2 ⟨σa σb ⟩β βJb,c ⟨σc σa ⟩β .
a∈Λk
b∈Ann(ℓ1/(1+ϵ) ,ℓ)
c∈Ann(ℓ+ℓν ,ℓ1+ϵ )

Now, using (Assumptionα ) and (IRBα ),

C3 k d ℓd ℓd(1+ϵ)
Pxu,∅
β [B2 , Jc ] ≤ d−α∧2 .
ℓ(d−α∧2) ℓ 1+ϵ ℓν(d+α)
We can now use that k d ≤ ℓ(1−ν) and choose ϵ > 0 sufficiently small, and ν sufficiently
close to 1 to conclude. □

7.3. Mixing property for deff ≥ 4. The goal of this subsection is to prove the following
result.
Theorem 7.13 (Mixing property for deff ≥ 4). Let d ≥ 1 and s ≥ 1. Assume that J
satisfies (A1)-(A5) and (Assumptionα ) with d − 2(α ∧ 2) ≥ 0. There exist γ, c, C > 0,
such that for every 1 ≤ t ≤ s, every β ≤ βc , every 1 ≤ nγ ≤ N ≤ L(β), every xi ∈ Λn and
yi ∈
/ ΛN (i ≤ t), and every events E and F depending on the restriction of (n1 , . . . , ns ) to
edges with endpoints within Λn and outside ΛN respectively,

(7.6) Pxβ1 y1 ,...,xt yt ,∅,...,∅ [E ∩ F ] − Pβx1 y1 ,...,xt yt ,∅,...,∅ [E]Pxβ1 y1 ,...,xt yt ,∅,...,∅ [F ]
−1/2
log(N/n)

≤C .
log log(N/n)
Furthermore, for every x′1 , . . . , x′t ∈ Λn and y1′ , . . . , yt′ ∈
/ ΛN , we have that
−1/2
log(N/n)

x y ′ ,...,xt yt′ ,∅,...,∅
Pxβ1 y1 ,...,xt yt ,∅,...,∅ [E] − Pβ1 1 [E] ≤C , (7.7)
log log(N/n)

−1/2
log(N/n)

x′ y1 ,...,x′t yt ,∅,...,∅
Pxβ1 y1 ,...,xt yt ,∅,...,∅ [F ] − Pβ1 [F ] ≤ C . (7.8)
log log(N/n)
 TRIVIALITY OF CRITICAL ISING AND φ4 MODELS WITH deff ≥ 4 59

We follow the strategy employed above and import all the notations from Section 6.4.
Fix β ≤ βc . Fix two integers t, s satisfying 1 ≤ t ≤ s. Introduce integers m, M such that
n ≤ m ≤ M ≤ N , m/n = (N/n)µ/2 , and N/M = (N/n)1−µ for µ small to be fixed.
We fix ν ∈ (0, 1) and ϵ > 0 such that Corollaries 7.10 and 7.12 hold.
Introduce the set K of (c0 , C0 )-weak regular scales k between m and M/2 with every
k
2 for k ∈ K differing by a multiplicative factor at least 2C0 log(N/n)2 . By Proposition
7.5, we may assume that |K| ≥ c1 loglog(N/n)
log(N/n) for a sufficiently small c1 = c1 (µ) > 0. Recall
that U was defined in Section 6.4.
The property (P4′ ) of weak regular scales allows us to prove,
Lemma 7.14 (Concentration of U). For all γ > 2, there exists C1 = C1 (d, t, γ) > 0 such
that for all n sufficiently large satisfying nγ ≤ N ≤ L(β),
log log(N/n)
Exy,∅
β [(U − 1)2 ] ≤ C1 .
log(N/n)
The only place where the argument needs to be adapted is located in the proof of
Lemma 6.23, which bounds the occurrence of G(u)c (where G(u) was defined in Definition
6.22). The remainder of the subsection concerns the extension of this lemma to our setup.
Lemma 7.15. We keep the assumptions of Theorem 7.13. There exist C, δ > 0, γ =
γ(δ) > 0 large enough and µ = µ(δ) > 0 small enough such that for every nγ ≤ N ≤ L(β),
and every u with ui ∈ Ayi (2ki ) with m ≤ 2ki ≤ M/2 for every 1 ≤ i ≤ t,
δ
n

Pxu,uy
β [G(u)c ] ≤C .
N
Proof. Recall that we have fixed the values of ϵ and ν. We follow the notations used in
the proof of Lemma 6.23.
Recall that G(u) = ∩1≤i≤s Gi , and Hi ∩ Fi ⊂ Gi where
Hi = {ni ∈
/ Cross(M, N )}, Fi = {ni′ ∈
/ Cross(n, m)}.
Introduce intermediate scales n ≤ r ≤ m ≤ M ≤ R ≤ N with r, R chosen below.
• Bound on Hi . Define,
[
Ji := {ni ∈ Jump(p, p1+ϵ )}.
p∈{R,N 1/(1+ϵ) ,N }

Notice that,
Pxu,uy
β [Hic ] ≤ Pxu
β [Γ(ni ) ∈ ZZ(xi , ui ; M, R, ∞)]

+ Pxu
β [ni \ Γ(ni ) ∈ Cross(R
1+ϵ
, N )] + Pxu
β [Ji ].

Assume R = N ι where ι > µ will be fixed below, and recall that M ≤ N µ+1/γ . We
might decrease µ and increase γ to ensure that (2/ϵ)(µ + 1/γ) ≤ ι. As a result, we may
use Lemma 7.9 to obtain C1 , η1 > 0 such that
C1
Pxu
β [Ji ] ≤ .
Rη1
Moreover, thanks to Corollaries 7.10 and 7.12, if we additionally require that20:
M [2(1+ϵ)/ϵ]∨[2d/(1−ν)] ≤ R, R2d(1+ϵ)/(1−ν) ≤ N,
we find C2 , η2 > 0 such that,
C2 C2
Pxu
β [Γ(ni ) ∈ ZZ(xi , ui ; M, R, ∞)] ≤ , Pxu
β [ni \ Γ(ni ) ∈ Cross(R
1+ϵ
, N )] ≤ .
R η2 R η2
20This might decrease µ and increase γ.
60 R. PANIS

• Bound on Fi . We follow the exact same strategy as in the proof of Lemma 6.23. The
modifications are similar to what was done for the bound on Hi . Again, we replace
Corollary 6.15 by Corollary 7.12 and choose accordingly the values of µ and γ.
We set r = mκ with 2κd(1 + ϵ) ≤ α ∧ 2. Recall that m ≥ N µ/2 . We find that

Pxu,uy
β [Fic ] ≤ Puy ′
β [Γ(ni ) ∈ ZZ(ui , yi ; r
1+ϵ
, m, ∞)]
+ Puy ′ ′ uy f
β [ni \ Γ(ni ) ∈ Cross(n, r)] + Pβ [Ki ],

where Kfi is the event that there exists a ∈ Λr and b ∈ / Λr1+ϵ such that (ni′ )a,b ≥ 2 and
{a, b} ∈ Γ(ni′ ) \ Γ(ni′ ).
Using (IRBα ) and the assumption that ui ∈ Ayi (2ki ) to get that ⟨σv σyi ⟩β ≤ C3 ⟨σui σyi ⟩β ,
we obtain
X ⟨σu σv ⟩β ⟨σv σy ⟩β
Puy ′
β [Γ(ni ) ∈ ZZ(ui , yi ; r
1+ϵ
, m, ∞)] ≤ i i

v∈Λ
⟨σ ui σy i ⟩ β
r 1+ϵ

rd(1+ϵ) C4
≤ C4≤ (α∧2)/2 ,
md−α∧2 m
where we used that d − α ∧ 2 ≥ α ∧ 2. Moreover, using Corollary 7.12 (which requires that
n2d/(1−ν) ≤ r and hence decreases the values of µ and 1/γ), there exist ζ > 0 such that
C5
Puy ′ ′
β [ni \ Γ(ni ) ∈ Cross(n, r)] ≤ .
rζ
We conclude the proof with the bound on K
fi . Proceeding as in the proof of Lemma
6.23,
Puy
X
β [Ki ] ≤ C6 βJa,b ⟨σui σb ⟩β .
f
a∈Λr
b∈Λ
/ r1+ϵ

Using (Assumptionα ) and (IRBα ) we obtain ζ ′ > 0 such that

C7
Puy
β [Ki ] ≤ .
mζ ′
This concludes the proof □
We are now in a position to conclude.
Proof of Theorem 7.13. The proof follows the exact same lines as the proof of Theorem
6.20 except that we replace Lemma 6.23 by Lemma 7.15. □

7.4. Proof of Theorem 7.1. The lack of precision of the sliding-scale infrared bound
in comparison to the situation in d = 4 slightly weakens the result. The three cases of
interest are a little different and thus are treated in different sections.

7.4.1. The case d = 3. We assume that d = 3 and that J satisfies (A1)-(A5) and
(Assumptionα ) with α = 3/2.
Let us first observe that the sliding-scale infrared bound of Theorem 3.18 is not sharp
in this setup. Indeed, we expect the finite-volume susceptibility to grow like n3/2 (below
L(β)). (LBα ) and (IRBα ) can (almost) make up for this lack of precision as they yield
the existence of C > 0 such that for 1 ≤ n ≤ N ≤ L(β),
χN (β) √ χn (β)
3/2
≤ C n 3/2 . (7.9)
N n
Recall that (IRBα ) still gives BL (β) − Bℓ (β) ≤ C0 log(L/ℓ) in our setup. However, (7.9)
not being sharp, we modify Lemma 6.16 accordingly.
 TRIVIALITY OF CRITICAL ISING AND φ4 MODELS WITH deff ≥ 4 61

Lemma 7.16. There exists C > 0 such that for every β ≤ βc , and for every 1 ≤ ℓ ≤ L ≤
L(β),
log2 (L/ℓ)
 
BL (β) ≤ 1 + C ℓ Bℓ (β).
log2 (ℓ)
Proof. We repeat the proof of Lemma 6.16 except that we replace the use of Theorem 3.18
by (7.9). □

We define a (possibly finite) sequence L3 = L3 (β, D) by ℓ0 = 0 and

ℓk+1 = inf {ℓ ≥ ℓk , Bℓ (β) ≥ D · (ℓk + 1) · Bℓk (β)} .
We also define a sequence U3 = U3 (β, D) by uk = ℓ3k for k ≥ 0.
Remark 7.17. Note that the sequence L3 grows much faster than L introduced in Sec-
tion 6. The reason why we need an additional sequence U3 is technical and will become
transparent later.
We begin with a technical result on L3 .
Lemma 7.18 (Growth of L3 ). There exist c, C > 0 such that, for all k ≥ 1,
k−1
Y k−1
Y
[D · (ℓi + 1)] ≤ Bℓk (β) ≤ C [D · (ℓi + 1)],
i=0 i=0

and, as long as ℓk+1 ≤ L(β),

ℓk+1 ≥ ℓcD
k .

Proof. We repeat the argument used to study L in Section 6. The lower bound is imme-
diate and for the upper bound, using that BL (β) − Bℓ (β) ≤ C0 log(L/ℓ), for k ≥ 1,
1
  
Bℓk −1 (β) ≤ D(ℓk−1 + 1)Bℓk−1 (β) ≤ D(ℓk−1 + 1) Bℓk−1 −1 (β) − C0 log 1 −
ℓk−1
k−1
! k−1
Qk−1
Y X j=k−i [D · (ℓj + 1)]
≤ [D · (ℓi + 1)] Bℓ1 −1 (β) + C0
i=1 i=1
ℓk−i
k−1
!
Y
≤ C [D · (ℓi + 1)] .
i=0

for C large enough (independent of D and k). We conclude by noticing that

Bℓk (β) ≤ Bℓk −1 (β) + C0 log 2.
As for the growth of L3 , we use Lemma 7.18 and proceed as in Remark 6.17 to get that
Bℓk (β) 1
log2 (ℓk ) ≤ Cℓk log2 (ℓk+1 /ℓk ) ≤ Cℓk log2 (ℓk+1 /ℓk ) ,
Bℓk+1 (β) − Bℓk (β) (Dℓk − 1)

which yields, for some c1 > 0, ℓk+1 ≥ ℓck1 D . This concludes the proof. □
Remark 7.19. The second part of the above statement is in some sense the most important
one since it ensures that there is “room” between successive scales. This has been used
before to apply the results of Section 6.2 and Theorem 6.20. It will also be useful in
our case, and it explains the introduction of the additional multiplicative factor ℓk in the
definition of L3 . This choice backfires when we try to estimate Bℓk (β).
Our goal now is to prove,
62 R. PANIS

Proposition 7.20 (Clustering bound for d = 3 and α = 3/2). For D large enough, there
exists δ = δ(D) > 0 such that for all β ≤ βc , for all K > 3 with uK+1 ≤ L(β), and for all
v, x, y, z, t ∈ Z3 with mutual distance between x, y, z, t larger than 2uK ,
Pvx,vz,vy,vt
β [Mu (I; U3 , K) < δK] ≤ 2−δK .
We first see how this result implies Theorem 7.1.
Proof of Theorem 7.1 for d = 3. We follow the proof of Section 6. The only change occurs
in the connection between L and BL (β) when L = 2uK . Using Lemma 7.18, we find that
3K
!
Y
BL (β) ≤ Bℓ3K+1 (β) ≤ C [D · (ℓi + 1)] =: Πβ,D (3K),
i=0
so if Φ = Φβ,D is defined for t ≥ 1 by:
Φ(t) := inf{k ≥ 0, Πβ,D (3k) ≥ t} ∧ (⌊N0 (L3 )/3⌋ + 1) ,
where N0 (L3 ) is the index of the last element of L3 (possibly equal to ∞), we find that
K ≥ Φ(BL (β)). This gives the result setting ϕβ (t) := 2−δΦβ,D (t)/5 . □
As before, Proposition 7.20 will follow from a combination of Theorem 7.13 and of an
intersection property. We begin by modifying the definition of the intersection event.
Below, we fix ν ∈ (0, 1) and ϵ > 0 such that the results of Section 7.2 hold.
Definition 7.21 (Intersection event for d = 3 and α = 3/2). Let k ≥ 1 and y ∈ / Λuk+2 . A
pair of currents (n, m) with (∂n, ∂m) = ({0, y}, {0, y}) realises the event Ik if the following
e
properties are satisfied:
(i) The restrictions of n and m to edges with both endpoints in Ann(ℓ3k , ℓ1+ϵ
3k+3 ) contain
1+ϵ
a unique cluster “strongly crossing” Ann(ℓ3k , ℓ3k+3 ), in the sense that it contains
a vertex in Ann(ℓ3k , ℓ1+ϵ 1+ϵ
3k ) and a vertex in Ann(ℓ3k+3 , ℓ3k+3) ).
(ii) The two clusters described in (i) intersect.
Lemma 7.22 (Intersection property for d = 3 and α = 3/2). For D large enough, there
exists κ > 0 such that for every β ≤ βc , every k ≥ 2, and every y ∈
/ Λuk+2 in a weak
regular scale with 1 ≤ |y| ≤ L(β),
P0y,0y,∅,∅
β [(n1 + n3 , n2 + n4 ) ∈ Iek ] ≥ κ.
Proof. We repeat the two-step proof done in the preceding section.
p Introduce intermediate
p
scales uk = ℓ3k ≤ n ≤ m ≤ M ≤ N ≤ ℓ3k+3 = uk+1 with n = ℓk ℓk+1 , N = ℓk+2 ℓk+3 ,
m = ℓ3k+1 , and M = ℓ3k+2 . Keeping the same notations as in the proof of Lemma 6.19,
one has for some c1 > 0,
E0y,0y,∅,∅
β [|M|] ≥ c1 (BM (β) − Bm−1 (β)),
and for some c2 > 0
E0y,0y,∅,∅
β [|M|2 ] ≤ c2 (BM (β) − Bm−1 (β))B2M (β).
Now, by definition of L3 , one has BM (β) ≥ D(ℓ3k+1 +1)Bm (β) so that BM (β)−Bm−1 (β) ≥
BM (β)
2 for D large enough. Moreover, by (IRBα ), B2M (β) ≤ BM (β) + C log 2. As a result,
we may find c3 > 0 such that,
Pβ0y,0y,∅,∅ [|M| > 0] ≥ c3 .
The conclusion of the proof follows the same lines as in Lemma 6.19: we replace Lemma
6.5 by Lemma 7.6, and Corollaries 6.9 and 6.15 by Corollaries 7.8 and 7.12. The proof is
enabled by Lemma 7.18 which ensures that the different scales are sufficiently “distanced”
when D is large enough. □
 TRIVIALITY OF CRITICAL ISING AND φ4 MODELS WITH deff ≥ 4 63

We are now equipped to prove Proposition 7.20.

Proof of Proposition 7.20. Fix γ > 2 sufficiently large so that Theorem 7.13 holds. Re-
member by Lemma 7.18 that we may choose D = D(γ) sufficiently large in the definition
of L3 such that ℓk+1 ≥ ℓγk . We may assume v = 0. Since x, y are at distance at least 2uK
of each other, one of them must be at distance at least uK of u. Without loss of generality
we assume that this is the case of x and make the same assumption about z. Let δ > 0
(δ)
to be fixed below. Recall that SK denote the set of subsets of {2δK, . . . , K − 3} which
contain only even integers.
As in the proof of Proposition 6.2,
P0x,0z,0y,0t Pβ0x,0z,∅,∅ [BS ].
X
β [Mu (I; U3 , K) < δK] ≤
(δ)
S∈SK
|S|≥(1/2−2δ)K

Let J be the event defined by

K−3
[
J := {n1 + n3 ∈ Jump(uk , u1+ϵ 1+ϵ
k )} ∪ {n2 + n4 ∈ Jump(uk , uk )}.
k=2δK
Using Lemmas 7.6 and 7.18, if D is large enough, there exists C0 , C1 , η > 0 such that
C0 K δK
P0x,0z,∅,∅
β [J] ≤ η ≤ C1 e−η2 . (7.10)
ℓ3δK
(δ)
Fix some S ∈ SK . Let A
e S be the event that none of the events Ie (defined in Definition
k
7.21) occur for k ∈ S. As above, if k ∈ S and Jc occurs, the events Iek and BS are
incompatible. Using (7.10),
!
−η2δK (1/2 − 2δ)K
P0x,0z,0y,0t Pβ0x,0z,∅,∅ [A
X
β [Mu (I; U3 , K) < δK] ≤ eS] + C1 e .
(δ)
2δK
S∈SK
|S|≥(1/2−2δ)K

From this point, the analysis follows the exact same lines as before and we refer to the
proof of Proposition 6.2 for the rest of the argument. □

7.4.2. The case d = 2. We now assume that d = 2 and that J satisfies (A1)-(A5) and
(Assumptionα ) with α = 1.
As before, the sliding-scale infrared bound of Theorem 3.18 is not sharp since we expect
the finite-volume susceptibility to grow like n (below L(β)). Proposition 3.25 and (IRBα )
yield the existence of C > 0 such that for 1 ≤ n ≤ N ≤ L(β),
χN (β) χn (β)
≤ C log n . (7.11)
N n
Lemma 7.23. There exists C > 0 such that for every β ≤ βc , and for every 1 ≤ ℓ ≤ L ≤
L(β),
log (L/ℓ)
 
BL (β) ≤ 1 + C 2 log ℓ Bℓ (β).
log2 (ℓ)
Proof. We repeat the proof of Lemma 7.16 using this time (7.11). □
We define a (possibly finite) sequence L2 = L2 (β, D) by ℓ0 = 1 and
ℓk+1 = inf {ℓ ≥ ℓk , Bℓ (β) ≥ D · (log(ℓk ) + 1) · Bℓk (β)} .
We also define a sequence U2 = U2 (β, D) by uk = ℓ3k for k ≥ 0. Adapting the proof of
Lemma 7.18 to our setup, we obtain,
64 R. PANIS

Lemma 7.24 (Growth of L2 ). There exists c, C1 , C2 > 0 such that, for all k ≥ 1,
k−1
Y k−1
Y
[D · (log(ℓi ) + 1)] ≤ Bℓk (β) ≤ C [D · (log(ℓi ) + 1)],
i=0 i=0

and as long as ℓk+1 ≤ L(β),

ℓk+1 ≥ ℓcD
k .

The second part of Theorem 1.9 will follow from the following proposition.
Proposition 7.25 (Clustering bound for d = 2 and α = 1). For D large enough, there
exists δ = δ(D) > 0 such that for all β ≤ βc , for all K > 3 with uK+1 ≤ L(β), and for all
v, x, y, z, t ∈ Z2 with mutual distance between x, y, z, t larger than 2uK ,
Pvx,vz,vy,vt
β [Mu (I; U2 , K) < δK] ≤ 2−δK .

Proof. The proof follows the exact same lines as the proof of Proposition 7.20 (in particular,
we keep the same intersection event Iek ). □

As above, this result easily implies 7.1 for d = 2.

Proof of Theorem 7.1 for d = 2. We follow the proof of Section 7.4.1. Using Lemma 7.24,
we find that
3K
[D · (log(ℓi ) + 1)] =: Π′β,D (3K),
Y
BL (β) ≤ Bℓ3K+1 (β) ≤ C
i=0

so that K ≥ Φ′ (B L (β)) where Φ′ = Φ′β,D is defined for t ≥ 1 by:

Φ′ (t) := inf{k ≥ 0, Π′ (3k) ≥ t} ∧ (⌊N0 (L2 )/3⌋ + 1) .
This concludes the proof. □

7.4.3. The case d = 1. Finally, we treat the case d = 1. Assume that J satisfies (A1)-
(A5) and (Assumptionα ) with α = 1/2. The results of Section 3 give us the good rate
of decay for Sβ : there exist c, C > 0 such that for all x ∈ Zd with 1 ≤ |x| ≤ L(β),
c C
1/2
≤ ⟨σ0 σx ⟩β ≤ . (7.12)
|x| |x|1/2
This observation greatly simplifies the proof21and allows to proceed like in Section 6.
As before, define a (possibly finite) sequence L1 = L1 (β, D) by ℓ0 = 0 and
ℓk+1 = inf {ℓ ≥ ℓk , Bℓ (β) ≥ DBℓk (β)} .
With the above work, it is easy to obtain,
Proposition 7.26 (Clustering bound for d = 1 and α = 1/2). For D large enough, there
exists δ = δ(D) > 0 such that for all β ≤ βc , for all K > 3 with ℓK+1 ≤ L(β), and for all
u, x, y, z, t ∈ Z with mutual distance between x, y, z, t larger than 2ℓK ,
Pux,uz,uy,ut
β [Mu (I; L, K) < δK] ≤ 2−δK .

From this result and (7.12), we can obtain Theorem 7.1 with ϕβ (BL (β)) = (log L)c for
some constant c > 0.
21To avoid writing yet another proof of triviality we simply import the results of Section 6. However,
note that the knowledge of the critical exponent η yields a shorter proof of the improved tree diagram
bound, see [ADC21, Section 4]
 TRIVIALITY OF CRITICAL ISING AND φ4 MODELS WITH deff ≥ 4 65

7.5. Proof Corollary 7.3.

Proof of Corollary 7.3. We use the same strategy as in Appendix D. In particular, we will
use Proposition D.1. Recall that α = d/2. Let σ ∈ (0, d/2) so that ξσ (β) is well defined
for all β < βc . We begin by noticing that there exists C > 0 such that, for β < βc ,
χ(β) ≤ Cξσ (β)d/2 . (7.13)
Indeed, if K > 0, (IRBα ) implies that for some C1 = C1 (d) > 0,
χKξσ (β) (β) ≤ C1 (Kξσ (β))d/2 .
Moreover, using Proposition D.1,
χ(β)
χ(β) − χKξσ (β) (β) ≤ C2
.
Kσ
Combining the two last inequalities, and choosing K large enough, we obtain (7.13).
Now, assume that we are given 1 ≤ L ≤ L(β). Write,
0 ≤ gσ (β) ≤ A1 + A2 ,
where
1 1
U4β (0, x, y, z), U4β (0, x, y, z).
X X
A1 := − A2 := −
χ(β) ξσ (β)d
2 χ(β) ξσ (β)d
2
x,y,z∈Zd x,y,z∈Zd
L(0,x,y,z)≤L L(0,x,y,z)>L

Using the (standard) tree diagram bound (4.7), we get that

χ(β)
A1 ≤ C3 Ld ≤ C4 Ld ξσ (β)−d/2 .
ξσ (β)d
Moreover, using this time Theorem 7.1,
C5 χ(β)2 C6
A2 ≤ d
≤ .
ϕβ (BL (β)) ξσ (β) ϕβ (BL (β))
As a result, for any L ≥ 1,
C6
lim sup gσ (β) ≤ .
β↗βc ϕβc (BL (βc ))
Hence, if B(βc ) = ∞, one obtains the result taking L to infinity. If B(βc ) < ∞, we may
conclude using Theorem D.2. □

8. Extension of the results to models in the Griffiths-Simon class

In this section, we extend the results of Sections 6 and 7 to the case of single-site
measures in the GS class. Let us mention that using Remark 5.7 we can adapt the proof
of Theorem 1.3 to the case of measures in the GS class.
We focus on the results of Section 6 and briefly explains in Section 8.4 how similar
considerations permit to extend the results of Section 7. Let J be an interaction satisfying
(A1)-(A6), and ρ be a measure in the GS class.
Since the measure ρ might be of unbounded support, we will have to be careful in the
derivation of the diagrammatic bounds. More precisely, to be able to take weak limits in
ρ, we will need to write them in a spin-dimension balanced way. Let us give a concrete
example. The tree diagram bound (4.7) has four spins on the left and four pairs of
spins on the right. As such, it is not spin balanced. We can obtain a balanced version
of this inequality by “site-splitting” each term where an Ising spin is repeated by using
Lemma 3.6. The resulting bound is given in (5.4). Though they are more complicated,
66 R. PANIS

the diagrammatic bounds obtained via this procedure have the advantage of being spin-
balanced. An alternative route22 would be to divide by ⟨τ02 ⟩ρ,β .
Below, we let U4ρ,β be the corresponding four-point Ursell function for the field variable
τ , and for n ≥ 1,
X
Bn (ρ, β) := ⟨τ0 τx ⟩2ρ,β .
x∈Λn

Also, we will use the definitions of βc (ρ) and L(ρ, β) that were introduced in Section 3.
We will prove the following result, which in particular covers the case of the φ4 lattice
models by Proposition 2.2.
Theorem 8.1. Let d = 4. Assume that J satisfies (A1)-(A6). Let κ > 0. There exist
c, C > 0 such that, for all ρ in the GS class satisfying βc (ρ) ≥ κ, for all β ≤ βc (ρ), for all
x, y, z, t ∈ Z4 at mutual distance at least L with 1 ≤ L ≤ L(ρ, β),

(8.1) |U4ρ,β (x, y, z, t)|

c X
B0 (ρ, β)
 X
≤C ⟨τx τu ⟩ρ,β βJu,u′ ⟨τu′ τy ⟩ρ,β ⟨τz τu ⟩ρ,β βJu,u′′ ⟨τu′′ τt ⟩ρ,β .
BL (ρ, β)
u∈Z4 u′ ,u′′ ∈Z4

As for the Ising model, we can deduce from Theorem 1.13 and Proposition 4.6 the
following triviality statement for measures in the GS class.
Corollary 8.2. Let d = 4. Assume that J satisfies (A1)-(A6). Let κ > 0. There exist
C, c, γ > 0 such that, for any ρ in the GS class satisfying βc (ρ) ≥ κ, for all β ≤ βc (ρ),
1 ≤ L ≤ L(ρ, β), f ∈ C0 (Rd ), and z ∈ R,
!
z2
⟨exp (zTf,L,β (τ ))⟩ρ,β − exp ⟨Tf,L,β (τ )2 ⟩ρ,β
2
C∥f ∥4∞ rfγ z 4
!
z2
≤ exp ⟨T (τ )2 ⟩ρ,β .
2 |f |,L,β (log L)c
We will extend the results to the GS class using the following strategy.
Step 1 Fix a measure ρ0 of the Ising-type in the GS class, i.e. a measure that falls into
(i) of Definition 2.1. Prove that ρ0 satisfies Theorem 1.13 with constants c, C > 0
which only depend on βc (ρ0 ). To do so, we prove an analogous version of the
intersection clustering bound that was derived in Proposition 6.2. We proceed as
above by first proving that big jumps occur with small probability, and then by
obtaining a version of Proposition 6.19, together with a mixing statement as in
Theorem 6.20.
Step 2 Take any ρ is the GS class that is obtained as a weak limit of measures (ρk )k≥1 of
the type (i) in Definition 2.1. Prove that the statement available for each k ≥ 1
passes to the limit k → ∞. This requires a control of (L(ρk , β))k≥1 and (βc (ρk ))k≥1 ,
together with “infinite volume” version of the GS approximation in the sense that:
for all β < βc (ρ), for all x, y ∈ Zd ,
lim ⟨τx τy ⟩ρk ,β = ⟨τx τy ⟩ρ,β .
k→∞
In Section 8.1, we prove Theorem 1.13 for measures of the Ising type in the GS class modulo
an intermediate result (Proposition 8.3) that is similar to Proposition 6.2. In Section 8.2,
we implement Step 2 of the above strategy and extend the result to all measures in the
GS class. In Section 8.3, we prove Proposition 8.3. Finally, in Section 8.4 we explain how
this strategy can also be used to extend the results of Section 7.
22The two methods are comparable through (3.4).
 TRIVIALITY OF CRITICAL ISING AND φ4 MODELS WITH deff ≥ 4 67

8.1. Proof of the improved tree diagram bound for measures of the Ising type
in the GS class. Fix ρ in the GS class of the Ising-type, and β < βc (ρ). The measure
⟨·⟩ρ,β can be represented as an Ising measure on Zd × KN that we still denote by ⟨·⟩ρ,β .
In that case, we can identify τx with averages of the form
N
X
Qi σ(x,i) ,
i=1

where Qi ≥ 0 for 1 ≤ i ≤ N . For x ∈ Zd , we will denote Bx := {(x, i), 1 ≤ i ≤ N }. This

point of view allows to use the random current representation. We introduce a measure
Pxy
Λ,ρ,β on ΩΛ×KN which we define in the following two steps procedure:
- first, we sample two integers 1 ≤ i, j ≤ N with probability
Qi Qj ⟨σ(x,i) σ(y,j) ⟩ρ,β
,
⟨τx τy ⟩Λ,ρ,β
{(x,i),(y,j)}
- then, sample a current according to the “usual” current measure Pρ,β in-
troduced in Section 4.
It is also possible to define the infinite volume version of the above measure that we will
denote Pxy xy
ρ,β . Samples of Pρ,β are random currents with random sources in Bx and By . The
interest of this measure lies in the fact that it is better suited for the derivation of bounds
on connection probabilities in terms of the correlation functions of the field variable τ .
These bounds can be directly imported from [ADC21] and are recalled in Appendix C.
Define a sequence L similarly as in (6.4), using this time Bℓ (ρ, β). Call Iu the set of
vertices in v ∈ Zd such that Bv is connected in n1 + n3 and n2 + n4 to Bu . With this
definition, we now consider coarse intersections instead of proper intersections. This point
of view is better for the analysis that follows.
For models in the GS class of the Ising type, the clustering bound takes the following
form.
Proposition 8.3 (Clustering bound for models in the GS class). Let d = 4. Assume that
J satisfies (A1)-(A6). Let κ > 0. For D large enough, there exists δ = δ(D, κ) > 0 such
that for every ρ in the GS class of the Ising-type with βc (ρ) ≥ κ, for every β < βc (ρ),
every K > 3 with ℓK+1 ≤ L(ρ, β), every x, y, z, t ∈ Zd with mutual distance between
x, y, z, t larger than 2ℓK , every u, u′ , u′′ ∈ Zd with u′ , u′′ J-neighbours23 of u satisfying
|u − x|, |u − z| ≥ ℓK ,
′ ′′
Pux,uz,u
ρ,β
y,u t
[Mu (Iu′ ; L, K) < δK] ≤ 2−δK ,
where Iu′ is the set of vertices in v ∈ Zd such that Bv is connected in24 n1 + n3 +
δ(∂n1 ∩Bu ,∂n3 ∩Bu′ ) and n2 + n4 + δ(∂n2 ∩Bu ,∂n4 ∩Bu′′ ) to Bu .
Remark 8.4. The reason why we have Iu′ instead of Iu in the bound above is technical.
This is a consequence of the form the switching lemma takes in that context, as seen in
[ADC21, Lemma A.7].
Remark 8.5. In fact, the same result holds with L(α) (ρ, β), α ∈ (0, 1), instead of L(ρ, β)
(with a change of parameter25 δ). This will be useful below.
We postpone the derivation of this bound to the next section and now explain how one
can derive Theorem 1.13 for measures ρ of the Ising type.
23In the sense that J ′ , J ′′ > 0.
u,u u,u
24Here, for u, v ∈ Zd such that J
u,v > 0, δ(u,v) denotes the current identically equal to 0 except on the
pair {u, v} where it is equal to 1.
25One way to see this is to observe that the constant c in Proposition 3.23 depends on α.
68 R. PANIS

Proof of Theorem 1.13 for a measure ρ of the Ising type in the GS class. As for the case
of the Ising model, one can show (by summing (4.7) over points in Bx , By , Bz and Bt ) that
|U4ρ,β (x, y, z, t)| ≤ 2⟨τx τy ⟩ρ,β ⟨τz τt ⟩ρ,β Pρ,β
xy,zt,∅,∅
[Cn1 +n3 (∂n1 ) ∩ Cn2 +n4 (∂n2 ) ̸= ∅],
where Cn1 +n3 (∂n1 ) and Cn2 +n4 (∂n2 ) refer to the clusters in n1 + n2 and n3 + n4 of
the (random) sources ∂n1 and ∂n2 respectively. As above, we may find c0 > 0 such
that if x, y, z, t are at mutual distance at least L, there exists K = K(L) such that
K ≥ c0 log(BL (ρ, β)/B0 (ρ, β)) and 2ℓK ≤ L. The rest of the proof is conceptually identical
to what was done before, except that now we look at coarse intersections. Let D be large
enough so that Proposition 8.3 holds for some δ = δ(D, κ) > 0. Using Markov’s inequality
together with (C.1),

⟨τx τy ⟩ρ,β ⟨τz τt ⟩ρ,β Pxy,zt,∅,∅

ρ,β [|Cn1 +n3 (∂n1 ) ∩ Cn2 +n4 (∂n2 )| ≥ 2δK/5 ]
≤ 2−δK/5
X
⟨τx τu ⟩ρ,β βJu,u′ ⟨τu′ τy ⟩ρ,β ⟨τz τu ⟩ρ,β βJu,u′′ ⟨τu′′ τt ⟩ρ,β .
u,u′ ,u′′ ∈Zd

Now, using Lemma 6.1, write

⟨τx τy ⟩ρ,β ⟨τz τt ⟩ρ,β Pxy,zt,∅,∅

ρ,β [0 < |Cn1 +n3 (∂n1 ) ∩ Cn2 +n4 (∂n2 )| < 2δK/5 ]
n +n n +n
Pxy,zt,∅,∅
X 1 3 2 4
≤ ρ,β [∂n1 ←→ Bu , ∂n2 ←→ Bu , Mu (Iu ; L, K) < δK].
u∈Zd

Notice that by hypothesis u must satisfy |u − x| ∨ |u − y| ≥ ℓK and |u − z| ∨ |u − t| ≥ ℓK .

Hence, the upper bound above can be rewritten into the sum of four terms which represent
each case. We assume without loss of generality that |u − x| ≥ ℓK and |u − z| ≥ ℓK .
Using a proper formulation of the switching lemma in that context [ADC21, Lemma A.7],
we get
n +n n +n
Pxy,zt,∅,∅
ρ,β
1
[∂n1 ←→3 2
Bu , ∂n2 ←→4
Bu , Mu (Iu ; L, K) < δK]
′ ′′
⟨τx τu ⟩ρ,β βJu,u′ ⟨τu′ τy ⟩ρ,β ⟨τz τu ⟩ρ,β βJu,u′′ ⟨τu′′ τt ⟩ρ,β Pux,uz,u y,u t
[Mu (Iu′ ; L, K) < δK].
X
≤ ρ,β
u′ ,u′′ ̸=u

We then conclude using Proposition 8.3. We obtained the existence of c, C > 0 such
that: for all ρ in the GS class of the Ising type such that βc (ρ) ≥ κ, for all β < βc (ρ), for
all x, y, z, t ∈ Z4 at mutual distance at least L with L ≤ L(ρ, β), (8.1) holds. We can then
extend the result to βc (ρ) by a continuity argument26 together with the observation that
L(ρ, β) → ∞ as β → βc (ρ).
□
Remark 8.6. Using Remark 8.5 we see that we also obtained the same result replacing
L(ρ, β) by L(α) (ρ, β), for some α ∈ (0, 1). Note that this affects the constant c, C in
Theorem 1.13.
8.2. Extension of the tree diagram bound to weak limits of Ising type measures.
The goal of this section is to extend the improved tree diagram bound to the entire GS
class of measures. The result will be a consequence of the following proposition.
Proposition 8.7. Let ρ be a measure in the GS class. There exists a sequence of measures
(ρk )k≥1 of the Ising type in the GS class such that:
(1) (ρk )k≥1 converges weakly to ρ,
(2) lim inf βc (ρk ) ≥ βc (ρ),
26Here we use the left-continuity of the two-point and four-point correlation functions together with
(IRB) which uniformly bounds the two-point function for β ≤ βc (ρ).
 TRIVIALITY OF CRITICAL ISING AND φ4 MODELS WITH deff ≥ 4 69

(3) for every β < βc (ρ), for every x, y, z, t ∈ Zd ,

lim ⟨τx τy ⟩ρk ,β = ⟨τx τy ⟩ρ,β , lim ⟨τx τy τz τt ⟩ρk ,β = ⟨τx τy τz τt ⟩ρ,β ,
k→∞ k→∞

(4) for every β > 0, lim inf L(1/4) (ρk , β) ≥ L(ρ, β).
Assuming this result, and knowing Theorem 1.13 for Ising type measures, we can easily
extend Theorem 1.13 to all models in the GS class.
Proof of Theorem 1.13. Fix ρ in the GS class which is a weak limit of Ising type measure
in the GS class of measures, and which satisfies βc (ρ) ≥ κ. Let also (ρk )k≥1 be given
by Proposition 8.7. By the property (2) of the same proposition, the exists k0 ≥ 0 such
that for k ≥ k0 , βc (ρk ) ≥ κ/2. Since the tree diagram bound holds uniformly over Ising
type measures ρ′ satisfying βc (ρ′ ) ≥ κ/2, and using Remark 8.6, there exist C, c > 0 such
that for all k ≥ k0 , for all β ≤ βc (ρk ), for all x, y, z, t at mutual distance at least L with
1 ≤ L ≤ L(1/4) (ρk , β),

(8.2) |U4ρk ,β (x, y, z, t)|

c
B0 (ρk , β)
 X
≤C ⟨τx τu ⟩ρk ,β βJu,u′ ⟨τu′ τy ⟩ρk ,β ⟨τz τu ⟩ρk ,β βJu,u′′ ⟨τu′′ τt ⟩ρk ,β .
BL (ρk , β)
u,u′ ,u′′ ∈Zd

Fix β < βc (ρ) and 1 ≤ L ≤ L(ρ, β). By properties (2) and (4) of Proposition 8.7, there
exists k1 ≥ k0 such that β ≤ βc (ρk ) and L ≤ L(1/4) (ρk , β) for k ≥ k1 . As a result, (8.2)
holds with β and L for k ≥ k1 . We now use (3) to pass the inequality to the limit. Using
(IRB), we know that there exists C = C(d) > 0 such that for all u, v ∈ Zd , for k ≥ k1
C
⟨τu τv ⟩ρk ,β ≤ .
βc (ρ)|J||u − v|d−2
This justifies passing to the limit in (8.2) and yields the result. We extend the result to
β = βc (ρ) by a continuity argument as above. □
We now prove Proposition 8.7. We split the statement into lemmas.
Lemma 8.8. Assume that (ρk )k≥1 converges weakly to ρ. Then,
lim inf βc (ρk ) ≥ βc (ρ),
Proof. Assume β > lim inf βc (ρk ). If S ⊂ Zd is a finite set containing 0,
lim φρk ,β (S) = φρ,β (S). (8.3)
k→∞
Since β > lim inf βc (ρk ), one has that for k large enough β ≥ βc (ρk ) and hence27 φρk ,β (S) ≥
1 (using the same argument as in Remark 3.22) so that
φρ,β (S) ≥ 1.
Since this holds for any finite set S containing 0, one has β ≥ βc (ρ). □
Lemma 8.9. Assume that (ρk )k≥1 converges weakly to ρ. Then, for all β > 0, for all
α ∈ (0, 1/2)
L(ρ, β) ≤ lim inf L(α) (ρk , β).
Proof. If lim inf L(α) (ρk , β) = ∞ then the upper bound is trivial. Otherwise, fix n >
(2d) lim inf L(α) (ρk , β). There exists S ⊂ Zd finite and containing 0 with rad(S) ≤ 2n such
that for all k sufficiently large,
φρk ,β (S) < α.
Using 8.3, we get that φρ,β (S) ≤ α < 1/2 so that n ≥ (2d)L(ρ, β). □
27Otherwise the susceptibility would be finite at β.
70 R. PANIS

Lemma 8.10. Let d = 4. Assume that (ρk )k≥1 converges weakly to ρ. Let β < βc (ρ). For
every x, y, z, t ∈ Z4 ,
lim ⟨τx τy ⟩ρk ,β = ⟨τx τy ⟩ρ,β , lim ⟨τx τy τz τt ⟩ρk ,β = ⟨τx τy τz τt ⟩ρ,β .
k→∞ k→∞

Proof. We only prove the first part of the statement, the second part follows by a similar
argument. Let x, y ∈ Zd . It is sufficient to show that


lim sup ⟨τx τy ⟩ρk ,β − ⟨τx τy ⟩Λn ,ρk ,β = 0. (8.4)
n→∞ k≥1

Fix n large enough such that x, y ∈ Λn . Recall that ρk is defined by averages on KNk for
some Nk ≥ 1. Using the switching lemma,
" #
(n1 +n2 )|Λn ×KN
⟨τx τy ⟩ρk ,β − ⟨τx τy ⟩Λn ,ρk ,β = ⟨τx τy ⟩ρk ,β Pxy,∅
Zd ,Λn ,ρk ,β
∂n1 ∩ Bx ↮ k
∂n1 ∩ By
 n|Λn ×KN 
≤ ⟨τx τy ⟩ρ,β Pxy
ρk ,β ∂n ∩ Bx ↮ k
∂n ∩ By

Let ℓ := |x| + |y| and introduce the event ZZGSk (x, y; ℓ, n, ∞) that the backbone of n goes
from ∂n ∩ Bx to ∂n ∩ By by exiting Λn × KNk . As explained below, it is possible to extend
the proof of Corollary 6.12 to this setup to get that there exist η, C1 > 0 such that for all
n large enough, for all k ≥ 1,
C1
Pxy
ρk ,β [ZZGSk (x, y; ℓ, n, ∞)] ≤ .
nη
 n|Λn ×KN 
k
The observation that ∂n ∩ Bx ↮ ∂n ∩ By ⊂ ZZGSk (x, y; ℓ, n, ∞) gives (8.4). □

Proof of Proposition 8.7. If ρ falls into (i) of Definition 2.1 the statement is trivial. Oth-
erwise, fix any sequence (ρk )k≥1 of Ising type measures in the GS class that converges
weakly to ρ. Using the three above lemmas we verify that (ρk )k≥1 satisfies all the desired
properties.
□

8.3. Derivation of the intersection clustering bound for models in the GS class.
We now turn to the proof of Proposition 8.3. The proof follows the exact same lines as
for the Ising case and is reduced to the adaptation of the results of Section 6.2, together
with an extension of the intersection property of Lemma 6.19, and the mixing statement
of Theorem 6.20. We fix d = 4 and an interaction J which satisfies (A1)-(A6).
We start by excluding the existence of “big jumps” in our context. As it turns out, the
results of Section 6.2 directly follow from the following adaptation of Lemma 6.3.
Lemma 8.11. Let β > 0. Let ρ be of the Ising type in the GS class. For x, y, u, v ∈ Zd ,

Pxy,∅
ρ,β [∃i, j, n(u,i),(v,j) ≥ 1]
!
⟨τx τu ⟩ρ,β ⟨τv τy ⟩ρ,β ⟨τx τv ⟩ρ,β ⟨τu τy ⟩ρ,β
≤ βJu,v 2⟨τu τv ⟩ρ,β + + .
⟨τx τy ⟩ρ,β ⟨τx τy ⟩ρ,β
At this stage of the proof, the arguments essentially build on what was done in the Ising
case together with a proper adaptation of the proofs, as already explained in [ADC21].
Below we explain the main changes in the proofs and refer to [ADC21] for more details.
We first state the intersection property for Ising-type models in the GS class.
Lemma 8.12 (Intersection property for models in the GS class). Let κ > 0. For D =
D(κ) > 0 large enough, there exists δ = δ(κ) > 0 such that for every ρ of the Ising type in
 TRIVIALITY OF CRITICAL ISING AND φ4 MODELS WITH deff ≥ 4 71

the GS class satisfying βc (ρ) ≥ κ, every β ≤ βc (ρ), every k ≥ 2, and every y ∈

/ Λℓk+2 in a
regular scale with 1 ≤ |y| ≤ L(ρ, β),
P0y,0y,∅,∅
ρ,β [(n1 + n3 , n2 + n4 ) ∈ Ik (0)] ≥ δ,
where Ik (0) is defined similarly to the intersection event of Definition 6.18, except that we
now ask that the clusters of n1 + n3 and n2 + n4 coarse-intersect in the sense that there
exists v ∈ Ann(ℓk , ℓk+1 ) such that Bv is connected to B0 in n1 + n3 and n2 + n4 .
Proof. We keep the notations introduced in the proof of Lemma 6.19. Define,
Q2i 1[∂n1 ←→ (v, i)]Q2i′ 1[∂n2 ←→
n +n n +n
(v, i′ )].
X X 1 3 2 4
M :=
v∈Ann(m,M ) i,i′

The extra Q2i , Q2i′ terms allow to rewrite moments of M in terms of correlation functions
of the field variables (τz )z∈Zd . Using a similar computation as for the case of the Ising
model, together with the results of Proposition C.1, we get c1 , C1 > 0 such that
′
y,∅,∅
E0y,0
ρ,β [|M|] ≥ c1 (BM (ρ, β) − Bm−1 (ρ, β)),
′
y,∅,∅
E0y,0
ρ,β [|M|2 ] ≤ C1 Bℓk+1 (ρ, β)2 .
Similarly as above, we deduce, for some c2 > 0,
0y,0y,∅,∅
Pρ,β ̸ ∅] ≥ c2 .
[M =
The second part of the proof consists in making the intersection event local. We proceed
exactly as we did for the Ising case by first excluding the possibility of jumping any of
the intermediate scales, and by then repeating the analysis that lead to the bounds on the
events F1 , . . . , F5 . At this stage one needs to be careful in the use of the infrared bound
and it is required to have bounds involving β|J|. This will ensure that the bound on the
intersection probability we end up with does not depend on ρ. □
Theorem 8.13 (Mixing property for models in the GS class). Let d = 4. Let κ > 0
and s ≥ 1 There exist γ, c, C > 0, such that for every ρ of the Ising-type in the GS class
satisfying βc (ρ) ≥ κ, for every 1 ≤ t ≤ s, every β ≤ βc (ρ), every nγ ≤ N ≤ L(ρ, β), every
xi ∈ Λn and yi ∈ / ΛN (i ≤ t), and every events E and F depending on the restriction of
(n1 , . . . , ns ) to edges with endpoints within Λn and outside ΛN respectively,
x1 y1 ,...,xt yt ,∅,...,∅ x1 y1 ,...,xt yt ,∅,...,∅ x1 y1 ,...,xt yt ,∅,...,∅
(8.5) Pρ,β [E ∩ F ] − Pρ,β [E]Pρ,β [F ]
−1/2
N

≤ C log .
n
Furthermore, for every x′1 , . . . , x′t ∈ Λn and y1′ , . . . , yt′ ∈
/ ΛN , we have that
−1/2
N

x1 y1 ,...,xt yt ,∅,...,∅ x1 y1′ ,...,xt yt′ ,∅,...,∅
Pρ,β [E] − Pρ,β [E] ≤ C log , (8.6)
n
−1/2
N

x′1 y1 ,...,x′t yt ,∅,...,∅
Pxρ,β
1 y1 ,...,xt yt ,∅,...,∅
[F ] − Pρ,β [F ] ≤ C log . (8.7)
n
Proof. The main modification in the proof comes in the definition of Ui :
N
1 X 1 ni +n′
Q2j 1[(u, j) ←→i ∂ni ],
X X
Ui :=
|K| k∈K Axi ,yi (2k )
u∈Ayi (2k ) j=1

where
⟨τx τu ⟩ρ,β ⟨τu τy ⟩ρ,β X
ax,y (u) := , Ax,y (k) := ax,y (u).
⟨τx τy ⟩ρ,β
u∈Ayi (2k )
72 R. PANIS

Note that, as above, the extra term Q2j allows to express the moments in terms of the field
variables (τz )z∈Zd . Also, in the derivation of an analogue of Lemma 6.23, one will have to
be careful to use infrared bounds involving β|J|. □
We are now in a position to prove Proposition 8.3.
Proof of Proposition 8.3. The proof follows the exact same lines as for the Ising case,
except that we need to slightly take care of the monotonicity property we want to use.
We keep the notations introduced in the proof of Proposition 6.2. Let δ > 0 to be fixed
(δ)
later. Let S ∈ SK . Let BS (resp. B′S ) be the event that the clusters of Bu in n1 + n3
and n2 + n4 (resp. n1 + n3 + δ(∂n1 ∩Bu ,∂n3 ∩Bu′ ) and n2 + n4 + δ(∂n2 ∩Bu ,∂n4 ∩Bu′′ ) ), do not
coarse intersect in any of the annuli Ann(ℓi , ℓi+1 ) for i ∈ S. Then, using an adaptation of
the monotonicity argument of Proposition 6.24 to our context (see Proposition C.2),
′ ′′ ′ ′′
(8.8) Pux,uz,u y,u t
[Mu (Iu′ ; L, K) < δK] ≤ ux,uz,u y,u t
[B′S ]
X
ρ,β Pρ,β
(δ)
S∈SK
|S|≥(1/2−2δ)K

Pux,uz,∅,∅
X
≤ ρ,β [BS ].
(δ)
S∈SK
|S|≥(1/2−2δ)K

The rest of the proof is identical to what was done in Section 6. □

8.4. Extension of the results of Section 7. We now briefly explain how to extend to
results of Section 7 to models in the GS class. The strategy is very similar to what was
done above so we only present the main modifications in the proof. We begin by discussing
the modifications involved in the proofs of the results obtained in Sections 7.1 and 7.2
Let d ≥ 1. We fix an interaction J on Zd satisfying (A1)-(A5) and (Assumptionα )
with d − 2(α ∧ 2) ≥ 0. In that setup, we get that for any ρ in the GS class: if β ≤ βc (ρ)
and x ∈ Zd \ {0},
C
⟨τ0 τx ⟩ρ,β ≤ . (8.9)
βc (ρ)|x|d−α∧2 (log |x|)δ2,α
The first important observation is to notice that, although stated for the Ising model,
the results of Section 7.1 extend mutatis mutandis to every single-site measure ρ in the
GS class thanks to Proposition 3.25.
Similarly, we may extend the results of Section 7.2 to all measures ρ of the Ising type
in the GS class by using Lemma 8.11 and (8.9). With these tools, it is possible to extend
the results of Section 7 to measures of the Ising-type in the GS class by using the same
strategy as in Section 8.3.
The extension to all measures in the GS class uses again the approximation step of
Section 8.2. The only non-trivial modification concerns Proposition 8.7, and more precisely
Lemma 8.10. We will prove the following result.
Lemma 8.14. Let 1 ≤ d ≤ 3. Assume that J satisfies (A1)-(A5) and (Assumptionα )
with d − 2(α ∧ 2) ≥ 0. Let ρ be a measure in the GS class, and let (ρk )k≥1 be a sequence
of measures of the Ising type which converges weakly to ρ. Let β < βc (ρ). For every
x, y, z, t ∈ Zd ,
lim ⟨τx τy ⟩ρk ,β = ⟨τx τy ⟩ρ,β , lim ⟨τx τy τz τt ⟩ρk ,β = ⟨τx τy τz τt ⟩ρ,β .
k→∞ k→∞

Proof. Again we only prove the first part of the statement.We follow the proof of Lemma
 n|Λn ×KN
k
8.10. As before, if ℓ := |x|+|y|, the key observation is that ∂n ∩ Bx ↮ ∂n ∩ By is
included in the event ZZGSk (x, y; ℓ, n, ∞). However, as explained above, we can extend the
 TRIVIALITY OF CRITICAL ISING AND φ4 MODELS WITH deff ≥ 4 73

results of Section 7.2, and in particular Corollary 7.10, to obtain a bound the probability
of the latter event. This is enough to conclude.
□

Appendix A. Spectral representation of reflection positive Ising models

The aim of this appendix is to prove Theorem 3.11. We use the notations of Section 3.
In what follows ρ is a measure in the GS class. We assume that J satisfies (A1)-(A5).
We will make good use of the spectral theorem (see [Hal13]) which will be applied to
diagonalize the shift operator T given by,
T : x ∈ Zd 7→ x + (1, 0, . . . , 0).
Before that, we introduce some notations and a proper Hilbert space.
Let β > 0. Let e1 = (1, 0, . . . , 0). Let Σ be the hyperplane orthogonal to e1 passing
through 0. Let Θ be the reflection through Σ. Notice that Σ cuts Zd in two half-planes
Λ+ and Λ− with Λ+ ∩ Λ− = Σ. Let A+ be the algebra generated by local functions with
support in Λ+ . Reflection positivity with respect to Θ implies that for all f ∈ A+ ,
⟨Θ(f )f ⟩ρ,β ≥ 0.
We define a positive semi-definite bi-linear form on A+ by : for all f, g ∈ A+ ,
(f, g) := ⟨Θ(f )g⟩ρ,β .
Quotienting A+ by the kernel of (·, ·) and completing the resulting space one obtains a
Hilbert space (H, (·, ·)). We denote by ∥.∥ the norm on this Hilbert space and ∥.∥op the
associated operator norm. The shift T in the e1 direction defines an operator on H whose
properties are described in the next proposition whose proof can be found in [BC96, Ott19].
Proposition A.1 (Properties of T ). The shift operator T : H → H has the following
properties,
(i) T is self-adjoint,
(ii) T is positive,
(iii) T is bounded.
In what follows we introduce many classical objects in the study of bounded self-adjoint
operators in a Hilbert space. For all the definitions we refer to [Hal13]. We are now in a
position to apply the spectral theorem [Hal13, Theorem 7.12].
Proposition A.2. There exists a unique projection valued measure µT such that
Z
T = λdµT (λ).
σ(T )

Remark A.3. One has that σ(T ) ⊂ [0, 1].

We also state two propositions (which can be found in chapter 7 of [Hal13]) that will
allow us to make good use of the preceding proposition.
Proposition A.4. Let f : σ(T ) → C be a bounded measurable function. Then,
Z
f (T ) = f (λ)dµT (λ).
σ(T )

Proposition A.5. If f : [0, 1] → C is a bounded measurable function, and ψ ∈ H, there

exists a (positive) real-valued measure µψ such that
 Z 1   Z 1
T
ψ, f (λ)dµ (λ) ψ = f dµψ .
0 0
Remark A.6. The measure µψ is given for E ∈ Ω, by
µψ (E) = (ψ, µ(E)ψ).
74 R. PANIS

Recall that for f, g ∈ H, the truncated correlation of f and g is given by

⟨f ; g⟩ρ,β := ⟨f g⟩ρ,β − ⟨f ⟩ρ,β ⟨g⟩ρ,β .
Proposition A.7 (Representation of truncated correlation functions). For all f ∈ H,
and all n ≥ 0, there exists f⊥ ∈ H such that,
⟨Θ(f ); T n f ⟩ρ,β = (f⊥ , T n f⊥ ).
Proof. Let 1 ∈ H be such that T 1 = 1. By definition, for all f ∈ H, ⟨f ⟩ρ,β = (1, f ). Thus,
⟨Θ(f ); T n f ⟩ρ,β = (f, T n f ) − (1, f )(1, f ).
Write P⊥ the orthogonal projection on Vect(1)⊥ . Letting
f⊥ := P⊥ f = f − (1, f )1,
we find that
⟨Θ(f ); T n f ⟩ρ,β = (f⊥ , T n f⊥ ).
□
We are now in a position to prove the main result of this section.
Proof of Theorem 3.11. Let β ≤ β (ρ). Apply Proposition A.5 to f : x ∈ [0, 1] 7→ xn for
P c
n ≥ 0, and ψ = V⊥ where V = x⊥ ∈Zd−1 vx⊥ τ(0,x⊥ ) ∈ H to get
Z 1
(V⊥ , T n V⊥ ) = λn dµV⊥ (λ).
0
Using Proposition A.7, we obtain
Z 1
n
⟨Θ(V )T V ⟩ρ,β = λn dµV⊥ (λ).
0

Now, notice that ⟨Θ(V )T n V ⟩ρ,β is exactly the left-hand side of (3.6). Moreover, consid-
ering the push-forward of µV⊥ under the map a ∈ [0, 1] 7→ − log a ∈ R+ ∪ {∞}, that we
denote µv,β ,
Z 1 Z ∞
n
λ dµV⊥ (λ) = e−an dµv,β (a),
0 0
and the result follows for all n ∈ Z using that ⟨·⟩ρ,β is invariant under T . □
Remark A.8. Note that one may have
µv,β ({∞}) > 0,
which is exactly equivalent to the fact that ξ(ρ, β) = ∞.
We now present the proof of the monotonicity property of the two-point function’s
Fourier transform.
Proof of Proposition 3.16. First, notice that
(mod) X
Sbρ,β (p) = 2 eip·x Sρ,β (x).
x∈Zd
x1 +x2 =0[2]

We follow the proof of Theorem 3.11 and keep the same notations. This time we introduce
the operator T ′ : x 7→ x+(1, 1, 0, . . .). Let R be the reflection with respect to the hyperplane
Σ′ orthogonal to e1 + e2 passing through 0. Σ′ cuts Zd in two half-planes Λ′+ and Λ′− with
Λ′+ ∩ Λ′− = Σ′ . Let A′+ be the algebra generated by local functions with support in Λ′+ .
Reflection positivity with respect to R implies that for all f ∈ A′+ ,
⟨R(f )f ⟩ρ,β ≥ 0.
 TRIVIALITY OF CRITICAL ISING AND φ4 MODELS WITH deff ≥ 4 75

We define a positive semi-definite bilinear form on A′+ by : for all f, g ∈ A′+ ,

(f, g) := ⟨R(f )g⟩β .
Quotienting A′+ by the kernel of (·, ·) and completing the obtained space one obtains a
Hilbert space (H′ , (·, ·)). Then, T ′ can be seen as an operator of H′ . Using the same
arguments as in Proposition A.1, we also have that T ′ is a self-adjoint, bounded and
positive operator of H′ . Just as in Theorem 3.11, we obtain that for all v : Zd−1 → C in
ℓ2 (Zd−1 ), there exists a positive measure µ′v,β such that, for all n ∈ Z,
Z ∞
′ ′
e−a|n| dµ′v,β (a).
X
v(e,x♭ ) v (e′ ,y ♭)
Sρ,β (((e − e ) + n, −(e − e ) + n, x♭ − y♭ )) =
0
(e,x♭ ),(e′ ,y♭ )∈Zd−1

Fix p♭ = (p3 , . . . , pd ). Let q ∈ R which will be fixed later. Considering the sequence of ℓ2
functions given for L ≥ 1 by
eiqe eip♭ ·x♭
(L)
v(e,x ) = q 1 (d−1) ,
♭ (d−1) (e,x♭ )∈ΛL
|ΛL |
we get, just like in the proof of Corollary 3.12, that there exists a positive measure µ′q,p♭ ,β
such that for r ∈ R,
ea − e−a
Z ∞
eirn+iqe+ip♭ ·z♭ Sρ,β (e + n, −e + n, z♭ ) = ′
X

2 dµq,p♭ ,β (a).
(n,e,z♭ )∈Zd
0 E1 (r) + ea/2 − e−a/2
Taking r = p1 + p2 and q = p1 − p2 , we get that
ea − e−a
Z ∞
′
X
eip·x Sρ,β (x) =
2 dµp1 −p2 ,p♭ ,β (a).
x∈Zd
0 E1 (p1 + p2 ) + ea/2 − e−a/2
x1 +x2 =0[2]
(mod)
Notice that in the formula above, one can use the symmetries of Sbρ,β (p) to change p2
into −p2 . As a result, we obtain that
ea − e−a
Z ∞
(mod) ′
Sbρ,β (p) = 2
2 dµp1 +p2 ,p♭ ,β (a).
0 E1 (p1 − p2 ) + ea/2 − e −a/2

This yields the result using the monotonicity of u ∈ [0, π] 7→ E1 (u), as in the proof of
Corollary 3.15. □

Appendix B. The backbone representation of the Ising model

We recall the definition and some of the main properties of the backbone represen-
tation of the Ising model. This representation is closely related to the random current
representation and was first introduced in [AF86, ABF87], and later used to capture fine
properties of the Ising model [DCT16, ADCTW19, ADC21]. We refer to these papers for
more details about this object. We will use the notations of Section 4.
We fix a finite subset Λ of Zd and fix any ordering of ({u, v})u,v∈Λ that we denote ≺.
Definition B.1. Let n ∈ ΩΛ . Assume that ∂n = {x, y}. The backbone of n, denoted
Γ(n), is the unique oriented and edge self-avoiding path from x to y supported on pairs
{u, v} with nu,v odd which is minimal for ≺.
The backbone Γ(n) can be obtained via the following exploration process:
(1) Let x0 = x. The first edge {x, x1 } of Γ(n) is the earliest one of all the edges
emerging from x with nx,x1 odd.
(2) Each edge {xi , xi+1 } is the earliest of all edges emerging from xi that have not been
cancelled previously, and for which the flux number is odd.
76 R. PANIS

(3) The path stops when it reaches a site from which there are no more non-cancelled
edges with odd flux number available. This always happen at a source of n (in that
case y).
We let Γ(n) be the set of explored edges (this set is made of the {xi , xi+1 } together with
all cancelled edges).
A path γ : x → y (viewed as a sequence of steps) is said to be consistent if no step of
the sequence uses a bond cancelled by a previous step.
One can write
X
⟨σx σy ⟩Λ,β = ρΛ (γ), (B.1)
γ:x→y consistent
where for a consistent path γ : x → y,
wβ (n)1[Γ(n) = γ]
P
∂n=∂γ
ρΛ (γ) := P .
∂n=∅ wβ (n)
The backbone representation has the following useful properties:
(1) If γ is a consistent path and E is a subset of edges of Λ such that γ ∩ E c = ∅, then
ρΛ (γ) ≤ ρE (γ). (B.2)
(2) If a consistent path γ is the concatenation of γ1 and γ2 (which we denote by
γ = γ1 ◦ γ2 ),
ρΛ (γ) = ρΛ (γ1 )ρΛ\γ1 (γ2 ). (B.3)
This last property has the following consequence.
Proposition B.2 (Chain rule for the backbone). Let x, y, u, v ∈ Λ. Then,
⟨σx σu ⟩Λ,β ⟨σu σv ⟩Λ,β ⟨σv σy ⟩Λ,β
Pxy
Λ,β [Γ(n) passes through u first and then through v] ≤ .
⟨σx σy ⟩Λ,β
Remark B.3. The backbone expansion is very close to the Brydges-Fröhlich-Spencer ran-
dom walk expansion introduced in [BFS82], although it appears to be less canonical.

Appendix C. Properties of currents for models of the Ising-type in the GS

class
We recall a few classical bounds that can be found in [ADC21, Appendix A.4]. We keep
the notations introduced in Section 8. Fix a measure ρ of the Ising type in the GS class,
and β > 0.
Proposition C.1. For every distinct x, y, u, v ∈ Zd ,
n1 +n2 X ⟨τx τu ⟩ρ,β (βJu,u′ )⟨τu′ τy ⟩ρ,β
Pxy,∅
ρ,β [∂n1 ←→ Bu ] ≤ , (C.1)
′ d
⟨τx τy ⟩ρ,β
u ∈Z
and
n +n
P∅,∅ 1 2
X
ρ,β [Bx ←→ By ] ≤ ⟨τx τy ⟩ρ,β (βJy,y′ )⟨τy′ τx′ ⟩ρ,β βJx′ ,x . (C.2)
x′ ,y ′ ∈Zd
Moreover,
0x,∅ 1 2 n +n X ⟨τx τu ⟩ρ,β (βJu,u′ )⟨τu′ τv ⟩ρ,β (βJv,v′ )⟨τv′ τy ⟩ρ,β
(C.3) Pρ,β [∂n1 ←→ Bu , Bv ] ≤
⟨τx τy ⟩ρ,β
u′ ,v ′ ∈Zd
⟨τx τv ⟩ρ,β (βJv,v′ )⟨τv′ τu ⟩ρ,β (βJu,u′ )⟨τu′ τy ⟩ρ,β
+
⟨τx τy ⟩ρ,β
In the spirit of Proposition 6.24 we also have the following result.
 TRIVIALITY OF CRITICAL ISING AND φ4 MODELS WITH deff ≥ 4 77

Proposition C.2 (Monotonicity in the number of sources for the GS class). For every
x, y, z, t ∈ Zd , every u, u′ , u′′ with u′ , u′′ J-neighbours of u, and every S ⊂ Zd × KN ,
′ ′′
Pux,uz,u
ρ,β
y,u t
[Cn1 +n3 +δ(∂n1 ∩Bu ,∂n3 ∩B (∂n1 ) ∩ Cn2 +n4 +δ(∂n2 ∩Bu ,∂n4 ∩B (∂n2 ) ∩ S = ∅]
u′ ) u′′ )
ux,uz,∅,∅
≤ Pρ,β [Cn1 +n3 (∂n1 ) ∩ Cn2 +n4 (∂n2 ) ∩ S = ∅].

Appendix D. Triviality and finiteness of the Bubble diagram

In this appendix, we prove that models in the GS class for which the bubble diagram
is finite at criticality behave trivially. This provides an alternative proof to the results
of Section 5 but it also captures more cases (for instance we may apply it to the case of
algebraically decaying RP interactions for d = 4 and α = 2).
Below we fix a measure ρ in the GS class and, as in Section 3, we denote the spin-field
by τ . The correlation length or order σ > 0, mentioned in the introduction is given by
1/σ
|x|σ ⟨τ0 τx ⟩ρ,β
P
x∈Zd
ξσ (ρ, β) := .
χ(ρ, β)
We assume that we are given an interaction J on Zd satisfying (A1)-(A5) and such that
the above quantity can be defined for σ small enough throughout the critical phase.
The following result can be found in [Sok82] and is a direct consequence of the Messager-
Miracle-Solé inequality.
Proposition D.1. Let β < βc (ρ). There exists a constant C > 0 such that for all x ∈ Zd ,
χ(ρ, β)ξσ (ρ, β)σ
⟨τ0 τx ⟩ρ,β ≤ C .
(1 + |x|)d+σ
Proof. Using (MMS2),
/ Λ|x|/(2d) , Sρ,β (y) ≥ Sρ,β (x)}| ≥ C1 (1 + |x|)d ,
|{y ∈
for some C1 > 0. As a consequence,
X
χ(ρ, β)ξσ (ρ, β)σ = |y|σ Sρ,β (y) ≥ C2 (1 + |x|)d+σ Sρ,β (x),
y∈Zd

which concludes the proof. □

Recall that the renormalised coupling constant of order σ is defined by,

1
U4ρ,β (0, x, y, z).
X
gσ (ρ, β) := −
χ(ρ, β)2 ξσ (ρ, β)d
x,y,z∈Zd

Theorem D.2 (The bubble condition implies triviality). Let d ≥ 2. For a reflection
positive model in Zd with an interaction J satisfying the above conditions and such that
X
B(ρ, βc (ρ)) = ⟨τ0 τx ⟩2ρ,βc (ρ) < ∞,
x∈Zd

one has,
lim gσ (ρ, β) = 0.
β↗βc (ρ)

Proof. Using the tree diagram bound (4.7), we get

χ(ρ, β)2
0 ≤ gσ (ρ, β) ≤ 2 . (D.1)
ξσ (ρ, β)d
78 R. PANIS

Now, take L ≫ ε > 0 to be fixed later. Write

   
χ(ρ, β) = χεξσ (ρ,β) (ρ, β) + χLξσ (ρ,β) (ρ, β) − χεξσ (ρ,β) (ρ, β) + χ(ρ, β) − χLξσ (ρ,β) (ρ, β) .
| {z } | {z } | {z }
(1) (2) (3)

Using Cauchy–Schwarz inequality one gets for C1 > 0,

q
(1) ≤ C1 εd/2 ξσ (ρ, β)d/2 B(ρ, βc (ρ)),
and for C2 = C2 (L, ε) > 0,
q
(2) ≤ C2 ξσ (ρ, β)d/2 BLξσ (ρ,β) (ρ, β) − Bεξσ (ρ,β) (ρ, β).
Moreover, using Proposition D.1, we get that for C3 > 0,
ξσ (ρ, β)d/2 χ(ρ, β)
(3) ≤ C3 .
Lσ ξσ (ρ, β)d/2
Putting all the pieces together we get that,
χ(ρ, β) q
d/2
q C3 χ(ρ, β)
≤ C 1 B(ρ, βc )ε + C 2 BLξσ (ρ,β) (ρ, β) − Bεξσ (ρ,β) (ρ, β) + σ .
ξσ (ρ, β)d/2 L ξσ (ρ, β)d/2
C3
Fix L > 0 large enough so that Lσ < 1. Using the left-continuity of the two-point function,
and the fact that ξσ (ρ, β) → ∞ as β ↗ βc (ρ) together with the monotone convergence
theorem, we get that
BLξσ (ρ,β) (ρ, β), Bεξσ (ρ,β) (ρ, β) −→ B(ρ, βc (ρ)),
β↗βc (ρ)

so that for all ε > 0,

p
χ(ρ, β) C1 B(ρ, βc (ρ)) d/2
lim sup d/2
≤ C3
ε ,
β→βc (ρ) ξσ (ρ, β) 1− L σ

which yields the result using (D.1). □

Remark D.3. The above result could be extended to more general models in the GS class:
if J satisfies (A1)-(A4), and if we consider an interaction J for which both the bubble
condition and the MMS inequalities hold (or more precisely (MMS2)), then the renor-
malised coupling constant vanishes at criticality. Using the proof of the MMS inequality of
[ADCTW19], we may then extend our result to finite-range interactions28.

References
[ABF87] Michael Aizenman, David J. Barsky, and Roberto Fernández. The phase transition in a
general class of Ising-type models is sharp. Journal of Statistical Physics, 47:343–374, 1987.
[ADC21] Michael Aizenman and Hugo Duminil-Copin. Marginal triviality of the scaling limits of crit-
ical 4d Ising and φ44 models. Annals of Mathematics, 194(1):163–235, 2021.
[ADCS15] Michael Aizenman, Hugo Duminil-Copin, and Vladas Sidoravicius. Random currents and
continuity of Ising model’s spontaneous magnetization. Communications in Mathematical
Physics, 334(2):719–742, 2015.
[ADCTW19] Michael Aizenman, Hugo Duminil-Copin, Vincent Tassion, and Simone Warzel. Emergent
planarity in two-dimensional Ising models with finite-range interactions. Inventiones Math-
ematicae, 216(3):661–743, 2019.
[AF86] Michael Aizenman and Roberto Fernández. On the critical behavior of the magnetization in
high-dimensional Ising models. Journal of Statistical Physics, 44(3-4):393–454, 1986.
[AF88] Michael Aizenman and Roberto Fernández. Critical exponents for long-range interactions.
letters in mathematical physics, 16(1):39–49, 1988.
[AG83] Michael Aizenman and Ross Graham. On the renormalized coupling constant and the sus-
ceptibility in φ44 field theory and the Ising model in four dimensions. Nuclear Physics B,
225(2):261–288, 1983.

28Recall that these interactions are not reflection positive in general.

 TRIVIALITY OF CRITICAL ISING AND φ4 MODELS WITH deff ≥ 4 79

[AIOV21] Yacine Aoun, Dmitry Ioffe, Sébastien Ott, and Yvan Velenik. Non-analyticity of the correla-
tion length in systems with exponentially decaying interactions. Communications in Mathe-
matical Physics, 386(1):433–467, 2021.
[Aiz82] Michael Aizenman. Geometric analysis of φ4 fields and Ising models. parts I and II. Com-
munications in Mathematical Physics, 86(1):1–48, 1982.
[Aiz21] Michael Aizenman. A geometric perspective on the scaling limits of critical Ising and φ4d
models. arXiv preprint arXiv:2112.04248, 2021.
[Aou21] Yacine Aoun. Sharp asymptotics of correlation functions in the subcritical long-range
random-cluster and Potts models. Electronic Communications in Probability, 26:1–9, 2021.
[BBS14] Roland Bauerschmidt, David C. Brydges, and Gordon Slade. Scaling limits and critical be-
haviour of the 4-dimensional n-component |φ|4 spin model. Journal of Statistical Physics,
157:692–742, 2014.
[BBS19] R. Bauerschmidt, D.C. Brydges, and G. Slade. Introduction to a Renormalisation Group
Method, volume 2242. Springer Nature, 2019.
[BC96] C. Borgs and JT. Chayes. The covariance matrix of the Potts model: a random cluster
analysis. Journal of Statistical Physics, 82(5):1235–1297, 1996.
[BDH98] D. Brydges, J. Dimock, and TR. Hurd. A non-Gaussian fixed point for φ4 in 4−ε dimensions.
Communications in mathematical physics, 198:111–156, 1998.
[BFS82] David Brydges, Jürg Fröhlich, and Thomas Spencer. The random walk representation of
classical spin systems and correlation inequalities. Communications in Mathematical Physics,
83(1):123–150, 1982.
[Bis09] Marek Biskup. Reflection positivity and phase transitions in lattice spin models. In Methods
of contemporary mathematical statistical physics, pages 1–86. Springer, 2009.
[CS15] Lung-Chi Chen and Akira Sakai. Critical two-point functions for long-range statistical-
mechanical models in high dimensions. The Annals of Probability, 43(2):639–681, 2015.
[CS19] Lung-Chi Chen and Akira Sakai. Critical two-point function for long-range models with
power-law couplings: The marginal case for d ≥ dc . Communications in Mathematical
Physics, 372(2):543–572, 2019.
[DCCF83] C. Aragao De Carvalho, S. Caracciolo, and J. Fröhlich. Polymers and g|φ|4 theory in four
dimensions. Nuclear Physics B, 215(2):209–248, 1983.
[DCT16] Hugo Duminil-Copin and Vincent Tassion. A new proof of the sharpness of the phase transi-
tion for Bernoulli percolation and the Ising model. Communications in Mathematical Physics,
343(2):725–745, 2016.
[Dys69] Freeman J. Dyson. Existence of a phase-transition in a one-dimensional Ising ferromagnet.
Comm. Math. Phys., 12(1):91–107, 1969.
[FILS78] Jürg Fröhlich, Robert Israel, Elliot H. Lieb, and Barry Simon. Phase transitions and reflection
positivity. I. General theory and long range lattice models. In Statistical Mechanics, pages
213–246. Springer, 1978.
[Fis67] Michael E. Fisher. Critical temperatures of anisotropic Ising lattices. II. General upper
bounds. Physical Review, 162(2):480, 1967.
[FMN72] Michael E. Fisher, Shang-keng Ma, and BG. Nickel. Critical exponents for long-range inter-
actions. Physical Review Letters, 29(14):917, 1972.
[FMRS87] J. Feldman, J. Magnen, V. Rivasseau, and R Sénéor. Construction and borel summability of
infrared φ44 by a phase space expansion. Communications in mathematical physics, 109:437–
480, 1987.
[Frö82] Jürg Fröhlich. On the triviality of λϕ4d theories and the approach to the critical point in
d > 4 dimensions. Nuclear Physics B, 200(2):281–296, 1982.
(−)
[FSS76] Jürg Fröhlich, Barry Simon, and Thomas Spencer. Infrared bounds, phase transitions and
continuous symmetry breaking. Communications in Mathematical Physics, 50(1):79–95,
1976.
[GHS70] Robert B. Griffiths, Charles A. Hurst, and Seymour Sherman. Concavity of magnetization of
an Ising ferromagnet in a positive external field. Journal of Mathematical Physics, 11(3):790–
795, 1970.
[GJ73] James Glimm and Arthur Jaffe. Positivity of the φ43 Hamiltonian. Fortschritte der Physik,
21(7):327–376, 1973.
[GJ77] J. Glimm and A. Jaffe. Critical exponents and elementary particles. Communications in
Mathematical Physics, 52(3):203–209, 1977.
[GJ12] James Glimm and Arthur Jaffe. Quantum physics: a functional integral point of view.
Springer Science & Business Media, 2012.
80 R. PANIS

[GK85] K. Gawedzki and A. Kupiainen. Massless lattice φ44 theory: rigorous control of a renor-
malizable asymptotically free model. Communications in Mathematical Physics, 99:197–252,
1985.
[GPPS22] Trishen S. Gunaratnam, Christoforos Panagiotis, Romain Panis, and Franco Severo. Random
tangled currents for φ4 : translation invariant Gibbs measures and continuity of the phase
transition. arXiv preprint arXiv:2211.00319, 2022.
[Gri67] Robert B. Griffiths. Correlations in Ising ferromagnets. I. Journal of Mathematical Physics,
8(3):478–483, 1967.
[Gri69] Robert B. Griffiths. Rigorous results for Ising ferromagnets of arbitrary spin. Journal of
Mathematical Physics, 10(9):1559–1565, 1969.
[Gri70] R.B. Griffiths. Dependence of critical indices on a parameter. Physical Review Letters,
24(26):1479, 1970.
[Hal13] Brian C. Hall. Quantum theory for mathematicians, volume 267. Springer, 2013.
[Heg77] Gerhard C. Hegerfeldt. Correlation inequalities for Ising ferromagnets with symmetries. Com-
munications in Mathematical Physics, 57(3):259–266, 1977.
[HT87] Takashi Hara and Hal Tasaki. A rigorous control of logarithmic corrections in four-
dimensional φ4 spin systems: II. critical behavior of susceptibility and correlation length.
Journal of statistical physics, 47:99–121, 1987.
[HvdHS08] Markus Heydenreich, Remco van der Hofstad, and Akira Sakai. Mean-field behavior for
long-and finite range Ising model, percolation and self-avoiding walk. Journal of Statistical
Physics, 132:1001–1049, 2008.
[Kad93] L.P. Kadanoff. Critical behavior. universality and scaling. In From Order To Chaos: Essays:
Critical, Chaotic and Otherwise, pages 222–239. World Scientific, 1993.
[Leb74] Joel L. Lebowitz. GHS and other inequalities. Communications in Mathematical Physics,
35(2):87–92, 1974.
[Lie04] Elliott H. Lieb. A refinement of Simon’s correlation inequality. Statistical Mechanics: Selecta
of Elliott H. Lieb, pages 447–455, 2004.
[LL10] Gregory F. Lawler and Vlada Limic. Random walk: a modern introduction, volume 123.
Cambridge University Press, 2010.
[MMS77] A. Messager and S. Miracle-Sole. Correlation functions and boundary conditions in the Ising
ferromagnet. Journal of Statistical Physics, 17(4):245–262, 1977.
[MPS23] Emmanuel Michta, Jiwoon Park, and Gordon Slade. Boundary conditions and universal
finite-size scaling for the hierarchical |φ|4 model in dimensions 4 and higher. arXiv preprint
arXiv:2306.00896, 2023.
[MS76] J. Magnen and R. Seneor. The infinite volume limit of the φ43 model. Annales de l’IHP
Physique théorique, 24(2):95–159, 1976.
[Nel66] E. Nelson. A quartic interaction in two dimensions. In Mathematical Theory of Elementary
Particles, (Dedham, Mass., 1965), pages 69–73, 1966.
[New75] Charles M. Newman. Inequalities for Ising models and field theories which obey the Lee-Yang
theorem. Communications in Mathematical Physics, 41(1):1–9, 1975.
[New80] CM. Newman. Normal fluctuations and the FKG inequalities. Communications in Mathe-
matical Physics, 74(2):119–128, 1980.
[NS98] CM. Newman and H. Spohn. The Shiba relation for the spin-boson model and asymptotic
decay in ferromagnetic Ising models. preprint, 1998.
[OS73] Konrad Osterwalder and Robert Schrader. Axioms for Euclidean Green’s functions. Com-
munications in Mathematical Physics, 31:83–112, 1973.
[OS75] Konrad Osterwalder and Robert Schrader. Axioms for Euclidean Green’s functions II. Com-
munications in Mathematical Physics, 42(3):281–305, 1975.
[Ott19] Sébastien Ott. On Ornstein-Zernike theory and some applications. PhD thesis, University of
Geneva, 2019.
[Pei36] Rudolf Peierls. On Ising’s model of ferromagnetism. Mathematical Proceedings of the Cam-
bridge Philosophical Society, 32(3):477–481, 1936.
[RSDZ17] Slava Rychkov, David Simmons-Duffin, and Bernardo Zan. Non-gaussianity of the critical 3d
Ising model. SciPost Physics, 2(1):001, 2017.
[Sch77] Robert Schrader. New correlation inequalities for the Ising model and P (φ) theories. Physical
Review B, 15(5):2798, 1977.
[SG73] Barry Simon and Robert B. Griffiths. The φ42 field theory as a classical Ising model. Com-
munications in Mathematical Physics, 33:145–164, 1973.
[Sim80] Barry Simon. Correlation inequalities and the decay of correlations in ferromagnets. Com-
munications in Mathematical Physics, 77(2):111–126, 1980.
 TRIVIALITY OF CRITICAL ISING AND φ4 MODELS WITH deff ≥ 4 81

[Sla18] Gordon Slade. Critical exponents for long-range O(n) models below the upper critical di-
mension. Communications in Mathematical Physics, 358:343–436, 2018.
[Sok82] Alan D. Sokal. An alternate constructive approach to the φ43 quantum field theory, and a
possible destructive approach to φ44 . Annales de l’IHP Physique Théorique, 37(4):317–398,
1982.
[SYI72] M. Suzuki, Y. Yamazaki, and G. Igarashi. Wilson-type expansions of critical exponents for
long-range interactions. Physics Letters A, 42(4):313–314, 1972.
[Uch98] Kohei Uchiyama. Green’s functions for random walks on Zn . Proceedings of the London
Mathematical Society, 77(1):215–240, 1998.
[WF72] Kenneth G. Wilson and Michael E. Fisher. Critical exponents in 3.99 dimensions. Physical
Review Letters, 28(4):240, 1972.
[Wig56] Arthur S. Wightman. Quantum field theory in terms of vacuum expectation values. Physical
Review, 101(2):860, 1956.

Section de mathématique,
Université de Genève,
Genève, Switzerland.
Email address: [email protected]