We provide a counterexample to the weak Gibbs and regularity conjectures. The example is a spin-one system on the square lattice with a translation-invariant nearest-neighbor interaction. At high temperatures and high on-site field strengths, it has a unique KMS state that does not have the weak Gibbs or regularity properties.
Introduction
For the finite-volume Gibbs states of a quantum spin system, the relative entropy is equal to the free energy excess multiplied by the inverse temperature, see for example [3]. Schematically, for every density matrix σ,
D(σ∥γ)=−S(σ)+βEH(σ)+PH(β),
where S, EH, and PH denote the entropy, energy, and pressure contribution, respectively, and γ is the Gibbs state at inverse temperature β>0. A natural question to ask is whether, for KMS states of infinitely extended systems, the same relation persists for the respective volume-averaged quantities. States that satisfy this property are called regular [11 Definition 2.15]. In this note, we construct a translation-invariant nearest-neighbor quantum spin system whose unique KMS state is nonregular.
More precisely, consider two translation-invariant states ν,ω on the quasi-local algebra A of the spin system. Their relative entropy density is defined as
whenever the limit exists. Here νZ and ωZ are the local density matrices, and Z↗Zd denotes increasing cubes centered at the origin. The state ω is said to be regular if, for every translation-invariant state ν on A, the relative entropy density exists and it holds that
s(ν∥ω)=−s(ν)+βeH(ν)+pH(β),(1)
where the entropy density, energy density, and pressure are defined as
s(ν)=−Z↗Zdlim∣Z∣trZ(νZlogνZ),
eH(ν)=Z↗Zdlim∣Z∣trZ(νZHZ),
pH(β)=Z↗Zdlim∣Z∣logtrZexp(−βHZ).
A regular state is always KMS [11 Theorem 2.2 and the remark after Definition 2.15]. Whether the converse of this statement is true is a longstanding open problem. It is explicitly formulated as Conjecture-(R) by Jakšić, Pillet, and Szczepanek [8 Section 2.2.6]. The corresponding upper bound on the left-hand side of (1) was established by Jakšić, Pillet, and Tauber [11] using previous results of Hiai and Petz [6], and the full identity has been proven for one-dimensional finite-range models and at sufficiently high temperatures [6, 5]. In our example, we establish for a KMS state ω and an explicit translation-invariant state ν the strict inequality
Quantum weak Gibbsianity is a stronger property that implies regularity [11 Propositions 2.10 and 2.14] and is of independent interest. A translation-invariant state ω is said to be weak Gibbs at inverse temperature β>0 if there are constants cZ≥0 for finite lattice subsets Z such that
The quantum definition was formalized in [11], following earlier quantum Gibbs comparison estimates [1, 6] and the classical notion of weak Gibbs measures [16]. Since weak Gibbsianity implies regularity, a weak Gibbs state is always a KMS state, and it was conjectured in [10] that the converse is also true. We refer to this as the weak Gibbs conjecture. Weak Gibbsianity of KMS states has been established in finite-range one-dimensional systems and at sufficiently high temperature [11], and for gauge-invariant free lattice fermions [9]. Since the KMS state in our example is nonregular, it is also not weak Gibbs.
Pettinari [13] recently disproved the weak Gibbs conjecture for a translation-invariant nearest-neighbor qubit interaction on Z2. The present work was developed independently and establishes failure of regularity in a model with a unique KMS state. Both proofs build on the multiscale branching idea of Bouch [2].
Our example is a spin-one system on the square lattice with nearest-neighbor interaction terms and a uniform on-site field. It is closely related to models studied in the context of quantum breakdown [12, 7]. Its precise definition can be found in Section 2. It has three important features: First, the Hamiltonian preserves an exponentially weighted occupation number, which we use to express the relative entropy of the empty product state in terms of an empty-region probability. Second, the interaction terms can amplify the occupation number in a cascading manner, allowing for configurations that are entirely localized within a boundary region to have volume-order effects. Third, the on-site field provides a parameter with which we can obtain nonregularity while retaining uniqueness of the equilibrium state.
Our proof proceeds as follows. We first prove uniqueness for small inverse temperature, with a threshold independent of the field. For this, we use the results of Rouzé and Stilck França [14], which provide, for each finite volume, an auxiliary dissipative dynamics that converges to the Gibbs state, called a Gibbs sampler. We construct an infinite-volume version of this Gibbs sampler that sends all observables to scalar multiples of the identity and show that all KMS states are invariant under this evolution, from which we can then conclude uniqueness. This gives us a β0>0 such that, for all 0<β<β0 and all field strengths h, there is a unique KMS state ω. Note that choosing h large moves us outside the usual high-temperature regime, where regularity is known to hold. We test regularity against the empty product state ν. Conservation of the exponentially weighted occupation makes the empty vector an eigenvector of each local density matrix of ω, so the relative entropy is minus the logarithm of the empty-region probability. On each box ΛL of side length 2L around the origin, we have
D(νΛL∥ωΛL)=−logω(νΛL).
We then use positivity of matrix entries and bounds on the partition function to obtain a bound on the above expression in terms of the local partition function ZΛL, an imaginary-time return amplitude GΛL+=⟨uΛL+,exp(−βHΛL)uΛL+⟩, and a comparison error term. The vector uΛL+ in this return amplitude is defined on ΛL+, an enlarged box with one additional row and column to the bottom and left. It is empty on ΛL with only the additional boundary occupied. This allows us to obtain a comparison error of boundary order:
D(νΛL∥ωΛL)≤logZΛL−logGΛL++O(L).
Thus, we want to show that −logGΛL+≤(−s(ν)+βeH(ν)−δ)∣ΛL∣+O(L) for some δ>0. We show this along a subsequence of boxes with L=2k−1, which yields (2). For the empty product state ν, we have s(ν)=0 and eH(ν)=2h. Indeed, we can show that −logGΛL+≤(23h+2log(1+h)+14log2)∣ΛL∣+O(L). Choosing h large enough completes the proof. The final estimate is obtained by recursively bounding the return amplitude from below along shrinking dyadic boxes. This is where the special form of the interactions and the multiscale branching idea of Bouch enter.
Section 2 gives the precise description of the model and states the main result. Section 3 proves the theorem from three lemmas: uniqueness of the KMS state, a comparison between the empty-region probability and the return amplitude, and the recursive return-amplitude estimate. Their proofs are given in Sections 4, 5, and 6, respectively.
Model and main result
We define a quantum spin system on Z2. The local Hilbert space is C3 with basis ∣0⟩,∣1⟩,∣2⟩. For a finite subset Z⊆Z2, the local Hilbert space and the algebra of local observables are
HZ=x∈Z⨂C3,AZ=x∈Z⨂B(C3).
We use the canonical embeddings obtained by tensoring with the identity and write A for the usual quasi-local C∗-algebra obtained by norm completion of the union of all local algebras. For a state ρ on A, we denote by ρZ∈AZ the unique density matrix with ρ(A)=trZ(ρZA) for A∈AZ. The relative entropy is defined as
D(ρZ∥σZ)=trZ(ρZ(logρZ−logσZ)),
with value +∞ if suppρZ⊆suppσZ. Here log denotes the natural logarithm.
On the local Hilbert space, define the particle number and annihilation operators
n=∣1⟩⟨1∣+2∣2⟩⟨2∣,c=∣0⟩⟨1∣+∣1⟩⟨2∣
and denote by nx and cx their embeddings into A acting on site x∈Z2. We write x→y if y−x is one of the two positive coordinate vectors. For a field h≥0, the Hamiltonian on a finite lattice subset Z is
On each directed bond, Fxy is a hopping term that turns one particle into two along the bond and two into one in the opposite direction. It is of norm one and, on any local Hilbert space HZ with x,y∈Z, its occupation-basis matrix entries are nonnegative. For β>0, the local partition function and Gibbs density matrix are
ZZ=trZexp(−βHZ),γZ=ZZexp(−βHZ).
The Hamiltonian conserves the dyadically weighted particle number
CZ=x∈Z∑2−(x1+x2)nx.(4)
Indeed, for every x∈Z, we have [nx,CZ]=0, and for x→y with x,y∈Z, using [n,c]=−c and [n,c∗]=c∗, we also obtain
Theorem 2.1. There exists β0>0 such that, for all 0<β<β0 and h≥0, the translation-invariant nearest-neighbor Hamiltonian (3) has exactly one β-KMS state on A. There exists an h0, dependent on β, such that, for all h>h0, this state is not regular and hence not weak Gibbs.
Proof of Theorem 2.1
We carry out the proof as outlined in the introduction using three lemmas whose proofs are postponed to the following sections. We first obtain a unique KMS state from the following uniqueness lemma. Since it is a limit of local Gibbs states, we will be able to estimate its local density matrices by finite-volume comparisons.
Lemma 3.1. There exists β0>0 such that, for all 0<β<β0 and h≥0, the Hamiltonian (3) has a unique β-KMS state. This state is translation invariant and is the local limit of the finite-volume Gibbs states along growing boxes.
[Proof]
Fix 0<β<β0. For now, let h≥0 be arbitrary and let ω be the state supplied by Lemma 3.1. This setup allows us to choose the field after deriving our relative-entropy bound while retaining uniqueness. We will bound the relative entropy between ω and the empty product state ν, whose local density matrices are
νZ=x∈Z⨂∣0⟩⟨0∣x.
Every bond interaction term annihilates the empty vector, so
s(ν)=0,trZ(νZHZ)=2h∣Z∣,eH(ν)=2h.
Regularity of ω would require
Z↗Z2lim∣Z∣D(νZ∥ωZ)=2βh+pH(β).
We will obtain a strictly smaller upper bound along a sequence of squares.
Here we used cyclicity of the trace and local convergence from Lemma 3.1. For K>L, the commutators [CΛL,A] and [CΛK,A] agree because the terms of CΛK outside ΛL commute with A. The last equality follows from (5). Because this holds for every A∈AΛL, we have [ωΛL,CΛL]=0. All weights in (4) are positive, so the zero eigenspace of CΛL consists precisely of the multiples of the empty vector. Thus νΛL is its spectral projection at zero and also commutes with ωΛL. Since νΛL has rank one, it follows that
ωΛLνΛL=ω(νΛL)νΛL.
If ω(νΛL)>0, the rank-one projection νΛL has zero entropy, and functional calculus gives (logωΛL)νΛL=logω(νΛL)νΛL. Hence
If ω(νΛL)=0, the support of νΛL is not contained in that of ωΛL, so the relative entropy is +∞. Thus (6) also holds with the convention −log0=+∞.
It therefore suffices to bound the empty-region probability ω(νΛL) from below. We compare it with an imaginary-time return amplitude that allows particles to enter the square from its bottom and left edges. Define the enlarged square and the added boundary by
ΛL+={−L−1,…,L}2,BL=ΛL+∖ΛL.
Thus BL consists of the added bottom and left edges, and ∣BL∣=4L+3; see Figure 1.
Figure 1. Starting with ΛL centered at the origin, add the bottom row and left column BL to obtain ΛL+=ΛL∪BL.
Let uΛL+ be the product vector with occupation zero on ΛL and occupation two on BL,
uΛL+=(x∈ΛL⨂∣0⟩x)⊗(x∈BL⨂∣2⟩x).
Define the return amplitude by
GΛL+=⟨uΛL+,exp(−βHΛL+)uΛL+⟩.
The boundary spins participate in the evolution, with their occupations prescribed only in the initial and final vector uΛL+. The following lemma bounds the empty-region probability in terms of this return amplitude, with a correction proportional to ∣BL∣.
Lemma 3.2. For the unique β-KMS state ω fixed above and every L≥1,
ω(νΛL)≥e−(4β+log3)∣BL∣ZΛLGΛL+.(7)
[Proof]
Since GΛL+>0, this also shows that ω(νΛL)>0. Taking minus the logarithm and using (6) gives
Since ∣BL∣/∣ΛL∣→0 and logZΛL/∣ΛL∣→pH(β), it remains to bound −logGΛL+/∣ΛL∣ strictly below 2βh along a sequence of squares. The next lemma provides the required return bound by repeatedly subdividing a square into four smaller ones.
Lemma 3.3. Set Lk=2k−1 for integers k≥0. For every β>0, h≥0, and integer k≥1,
−logGΛLk+≤(23βh+2log(1+h)+14log2)∣ΛLk+∣.(9)
[Proof]
We can now choose the field. With β fixed as above, choose an h>0 so large that
The same limit would hold along L=Lk, contradicting (11). Hence ω is not regular.
Proof of Lemma 3.1
The proof uses an auxiliary evolution of observables, called a Gibbs sampler. In a finite volume, this evolution preserves the Gibbs expectation and sends each observable to that expectation times the identity in the limit t→∞. We construct its infinite-volume limit and show that, as t→∞, it sends every local observable to a scalar multiple of the identity given by the expectation in a KMS-state ω. We then prove that it preserves every KMS state. These properties force all KMS states to agree with ω. Indeed, if Tt(A) tends to ω(A)1 and φ is invariant under Tt, then φ(A)=φ(Tt(A))→ω(A).
We write Tt for this auxiliary evolution and τt for the physical time evolution entering the KMS condition.
The finite-volume Gibbs sampler
We use the following consequence of [14 Theorems 6.1 and 4.4]. For spins with local Hilbert space C3 on the square lattice, there is an ε0∈(0,1] such that, at inverse temperature one, an arbitrary sum of on-site Hamiltonians may be perturbed by nearest-neighbor terms of total norm less than ε0 at each site, and the resulting Hamiltonian admits the construction and estimates stated below. The threshold and convergence rate are independent of the on-site Hamiltonians and the volume. We apply this result to βHZ, whose Gibbs matrix at inverse temperature one is γZ. Its on-site terms are βh(2−nx), while
xsupy∼zx∈{y,z}∑∥βFyz∥=4β.
Thus we may take β0=ε0/4, independently of h. Fix 0<β<β0 and h≥0 for the rest of the proof.
For every finite Z and x∈Z, the construction supplies a unital completely positive map Φx,Z on AZ satisfying
trZ(γZΦx,Z(A))=trZ(γZA).
We extend these maps to A by the identity map outside Z and use the operator norm for maps on A. Set
TZ,t=exp[tx∈Z∑(Φx,Z−id)].
The cited theorems give, for self-adjoint A∈AZ and t≥0,
∥TZ,t(A)−trZ(γZA)1∥≤2∣Z∣e−κt∥A∥,κ=21−e−1.(12)
By [14 Lemma B.1], there are b,μ>0, independent of h and Z, such that
∥Φx,Z−Φx,Z∩(x+Λℓ)∥≤be−μℓ,ℓ≥1.(13)
For the KMS argument, we also need the coefficients defining these updates. By [14 Proposition 3.4],
Φx,Z=exp(Kx,Z),
Kx,Z(A)=a=1∑9Lx,a,Z∗ALx,a,Z+Dx,ZA+ADx,Z∗,
where Lx,a,Z,Dx,Z∈AZ. With QZ=exp(−βHZ/2), they satisfy
Lx,a,ZQZ=QZLx,a,Z∗,QZDx,Z=Dx,Z∗QZ,
a=1∑9Lx,a,Z∗Lx,a,Z+Dx,Z+Dx,Z∗=0.(14)
Finally, the estimates in the proof of [14 Lemma 6.3] give, with Zx,ℓ=Z∩(x+Λℓ),
where the n=0 term is the identity map. This is a convex combination of compositions of unital completely positive maps.
We next take limits of the individual updates. Telescoping (15) gives
∥Dx,Z−Dx,Zx,ℓ∥≤C0r=ℓ∑∞(r+1)3e−μr,
and the same bound holds for each Lx,a,Z. The tail tends to zero uniformly in Z, and once Z contains x+Λℓ, the restricted coefficients depend only on that square. The coefficients are therefore Cauchy, with norm limits Lx,a and Dx. Hence the generators converge in norm to
Kx(A)=a=1∑9Lx,a∗ALx,a+DxA+ADx∗.
Continuity of the exponential gives the norm limit Φx=exp(Kx) of Φx,Z. Passing to the limit in the last identity of (14) and in (13) gives
Fix L≥1 and A∈AΛL. To prove that the finite-volume evolutions are Cauchy, we compare them with a common restricted evolution. For integers K≥4L and 1≤ℓ≤K/4, set
SΛK,ℓ,t=exp[tx∈ΛK∑(Φx,x+Λℓ−id)].
These updates act inside Λ⌊3K/2⌋, so SΛK,ℓ,s(A) remains supported there for every s≥0.
Let J≥2K. Adding and subtracting the updates of radius ℓ centered in Λ2K, Duhamel’s formula gives
In the first sum, each difference has norm at most be−μℓ by [13]. Contractivity therefore bounds its contribution by b∣Λ2K∣e−μℓ∥A∥.
For the second sum, we use that SΛK,ℓ,s is generated by uniformly bounded maps of range 2ℓ in the ℓ∞-metric, each generating contractions. The maps in the sum annihilate 1 and are supported at distance at least K−L−ℓ≥K/2 from ΛL. A finite-range Lieb–Robinson bound therefore controls their action on the evolved observable SΛK,ℓ,s(A). Applying [15], Theorem 3.5, using their norm bound of two, and summing over the centers gives
2∣Λ2K∣∣ΛL∣exp(Cℓ4s−4ℓK)∥A∥.
Here C is independent of h, J, K, and ℓ.
In the third sum, the centers lie outside Λ2K, while SΛK,ℓ,s(A) stays supported in the smaller box Λ⌊3K/2⌋. We restrict each update to a square that misses this support. The restricted update leaves the observable unchanged, so only the restriction error contributes. For a center with ∣x∣∞=m>2K, choose the restriction radius ⌊m/4⌋. Every site y in this square satisfies
∣y∣∞≥m−⌊m/4⌋≥43m>23K,
so the square is disjoint from the support of SΛK,ℓ,s(A). Thus [13] and contractivity give
This error decreases exponentially with the distance m of the center. There are only 8m sites with ∣x∣∞=m, so the errors are summable even as J→∞. Extending the sum to all m>2K and using ⌊m/4⌋≥m/4−1 gives
8b∥A∥m=2K+1∑∞me−μ⌊m/4⌋≤CKe−μK/2∥A∥.
Using contractivity of TΛJ,t−s and ∣Λ2K∣≤25K2, integration gives
Contractivity extends the limit to every A∈A, still uniformly on compact time intervals. The maps Tt are unital and completely positive, since these properties pass to the limit. To pass the semigroup identity to the limit, use
Hence Tt+s=TtTs and T0=id. Uniform convergence on compact time intervals also gives norm continuity of t↦Tt(A) for every A∈A.
We now show that the sampler converges to the expectation in a KMS state. For each integer n≥1, extend the Gibbs state on Λ2n8 by the normalized trace outside that box. By weak∗ compactness, choose a subsequence nj→∞ along which these states converge to a state ω. Since the interaction is finite range, ω is a β-KMS state by [3 Proposition 6.2.15].
Fix a self-adjoint A∈AΛL. For sufficiently large j, the triangle inequality, (19) with K=nj8, and (12) at time t=nj give
The right-hand side tends to zero: the first two terms vanish since ∣Λ2nj8∣≤25nj8, and the last term vanishes by the choice of ω. For every t≥nj, the semigroup property, unitality, and contractivity imply
This proves convergence as t→∞. Linearity extends it to all local observables, and contractivity and norm density extend it to every A∈A. Thus
t→∞lim∥Tt(A)−ω(A)1∥=0,A∈A.
Invariance of KMS states
Let φ be any β-KMS state. We first prove that it is invariant under each limiting update Φx. Write τZ,t(A)=exp(itHZ)Aexp(−itHZ) for the physical evolution, with infinite-volume limit τt. The balance identities (14) give
To pass to the limit at imaginary times, consider the bounded analytic function z↦τZ,z(Lx,a,Z) on the strip −β/2≤Imz≤0. By (20), its values on the upper and lower boundary lines are τZ,t(Lx,a,Z) and τZ,t(Lx,a,Z∗), respectively. Both converge in norm on bounded real-time intervals and are uniformly bounded.
The three-lines theorem bounds the norm of a bounded analytic function on a strip by its boundary norms [3 Proposition 5.3.5]. Apply it to e−z2 times the difference between two volumes. This factor decays as ∣t∣→∞ on both boundary lines, making convergence uniform there. The theorem then gives uniform norm convergence on bounded parts of the closed strip, extending τt(Lx,a) to the lower boundary. For Dx,Z, the same argument uses the strip 0≤Imz≤β/2. Hence
τ−iβ/2(Lx,a)=Lx,a∗,τiβ/2(Dx)=Dx∗.(21)
We use the halfway form of the KMS identity [3 Propositions 5.3.3 and 5.3.7],
φ(BC)=φ(τ−iβ/2(C)τiβ/2(B)).
It applies when the evolutions of B and C are analytic inside the strips 0<Imz<β/2 and −β/2<Imz<0, respectively, and extend continuously in norm to their boundary lines. We have just shown this for Dx on the upper strip and Lx,a on the lower strip. Taking adjoints gives the corresponding extensions for Lx,a∗ and Dx∗. Choose A whose evolution is analytic for all complex times. Using (21) gives
The middle line follows from invariance of φ under real-time evolution and analytic continuation to the upper half-strip. Summing the first identity over a, adding the last two, and using the balance identity in (16) gives
Contractivity and the locality estimate in (16) imply
∣φ(SΛK,ℓ,t(A))−φ(A)∣≤bt∣ΛK∣e−μℓ∥A∥.
Fix t and let K→∞, with ℓ=⌊K1/8⌋ as above. The right-hand side tends to zero. Letting J→∞ in (18) also shows that SΛK,ℓ,t(A)→Tt(A) in norm. Thus, first for local A and then by norm density,
φ(Tt(A))=φ(A),A∈A,t≥0.
4. Conclusion. For every β-KMS state φ, parts 2 and 3 give
φ(A)=φ(Tt(A))⟶ω(A),A∈A.
Thus ω is the unique KMS state. Every translate of ω is again KMS, since the Hamiltonian is translation invariant. Uniqueness therefore makes ω translation invariant. Every weak* limit point of the finite-volume Gibbs states is KMS by [3 Proposition 6.2.15], and hence equals ω. Weak* compactness then gives local convergence of the full sequence to ω.
Proof of Lemma 3.2
For A∈AZ, we write Aη,ξ for its matrix entries in the occupation basis, indexed by η,ξ∈{0,1,2}Z.
Fix L≥1 and take K>L, so that ΛL+⊂ΛK. By Lemma 3.1, the finite-volume Gibbs states converge locally to the fixed state ω. In particular,
We estimate the numerator and denominator separately, choosing the comparisons so that the same exterior partition function ZΛK∖ΛL+ appears in both.
Lower bound on the numerator
For the numerator, expand the trace in the occupation basis. Every diagonal entry of exp(−βHΛK) is positive, so retaining only the configurations with occupation two on BL gives
The right-hand side is independent of K. Thus we may take the local limit in (22), keeping L fixed, to obtain (7).
Proof of Lemma 3.3
The proof prepares four smaller squares by filling one row and two column segments. A comparison of return amplitudes then reduces the estimate to a scalar recursion. This repeated subdivision is based on the multiscale branching idea of [2 Section 6].
Setup of the recursive estimate
Fix β>0 and h≥0. For a finite lattice subset Z, set
AZ=2h∣Z∣1−HZ=hx∈Z∑nx+x→yx,y∈Z∑Fxy.
This matrix is self-adjoint and has nonnegative entries in the occupation basis. Moreover,
exp(−βHZ)=e−2βh∣Z∣exp(βAZ).(29)
We will use operations on the configuration that can lower boundary occupations, so we need a class of initial vectors that is preserved by our subdivision procedure. Let UΛL+ be the set of occupation-basis vectors on ΛL+ that are empty on ΛL, have occupation one or two on BL∖{(−L−1,−L−1)}, and have arbitrary occupation at the remaining corner. For L≥0, set
gΛL+=u∈UΛL+min⟨u,exp(βAΛL+)u⟩.(30)
The exponential series gives gΛL+≥1. Since uΛL+∈UΛL+, (29) implies
GΛL+≥e−2βh∣ΛL+∣gΛL+.(31)
Comparison along a sequence of conversions
Let
u=z∈Z⨂∣az⟩,v=z∈Z⨂∣bz⟩,az,bz∈{0,1,2},
be two occupation-basis vectors on Z. We say that u and v are connected by an allowed move, called a conversion, if there is a directed bond x→y in Z such that az=bz for every z∈Z∖{x,y} and one of the following four cases holds:
A conversion corresponds to one particle hopping along the bond from x to y and splitting into two, or two particles hopping in the opposite direction and merging into one. For the bond that realizes this conversion, Fxyu=v, so the corresponding matrix entry is one. Now suppose that u and v are joined by a sequence u=u0,u1,…,uq=v of q≥1 allowed conversions. For each j=1,…,q, choose a bond x→y with Fxyuj−1=uj. The matrix AZ−Fxy is the sum of the remaining field and bond terms, all of which have nonnegative entries. Its action on the basis vector uj−1 therefore has nonnegative coefficients, giving
AZuj−1=uj+(AZ−Fxy)uj−1≥entuj.
where ≥ent compares the coefficients in the occupation basis. Multiplication by AZ preserves these comparisons, so iteration gives
AZqu≥entAZq−1u1≥ent⋯≥entuq=v.
The exponential series shows that exp(βAZ/2) also has nonnegative entries and therefore preserves this comparison. Taking norms and commuting AZq with the exponential gives
and the same identity holds for v. Squaring the preceding norm inequality and dividing by $|A_Z|^{2q} therefore proves that any two occupation-basis vectors u and v on Z connected by a sequence of q≥1 allowed conversions satisfy
⟨u,exp(βAZ)u⟩≥∥AZ∥−2q⟨v,exp(βAZ)v⟩.(32)
Subdivision into four squares
Recall that Lk=2k−1. Fix k≥1 and u∈UΛLk+. We construct an occupation-basis vector v from u using the allowed conversions (1,0)↦(0,2) and (2,0)↦(1,2). In each of the following three stages, apply these conversions successively along the indicated row or column, starting at the specified source; see Figure 2.
Figure 2. The three sweeps for Lk=3. Each shaded square is a translate of ΛLk−1+, with its added bottom and left edges marked. The circled sources become corners of the smaller squares, where arbitrary occupation is allowed. The central circled site is the origin, where sweep (iii) starts.
Fill the row (x1,0), x1=−Lk,…,Lk, from left to right, starting with the source (−Lk−1,0).
(ii) Fill the lower column (0,x2), x2=−Lk,…,−1, from bottom to top, starting with the source (0,−Lk−1).
(iii) Fill the upper column (0,x2), x2=1,…,Lk, from bottom to top, starting with the source (0,0).
The sources in (i) and (ii) have occupation one or two by the definition of UΛLk+. Each target is empty when first reached and receives occupation two, so it can supply the next site. The source (0,0) in (iii) has occupation one after the horizontal sweep and can therefore supply the upper column segment. Thus every step is an allowed conversion. Their total number is
qk=(2Lk+1)+Lk+Lk=2k+2−3.
The enlarged square is the disjoint union
ΛLk+=z∈{−2k−1,2k−1}2⋃(z+ΛLk−1+).
On each of these four squares, v is empty on the translated core and has occupation one or two on its added bottom and left edges, except possibly at their common corner. The three sources (−Lk−1,0), (0,−Lk−1), and (0,0) that may become empty are the bottom-left corners of three smaller squares; the fourth inherits the original bottom-left corner. Thus the tensor factor vz of v on z+ΛLk−1+ is a translate of a vector in UΛLk−1+.
Deleting the bonds between the four squares gives
0≤entz∈{−2k−1,2k−1}2∑Az+ΛLk−1+≤entAΛLk+.
The exponential of the sum factors over the four squares. Comparing exponential series therefore yields
The last inequality uses translation invariance and (30).
4. Iteration of the return bound. Counting the bonds in ΛL+ and using ∥Fxy∥=1 and ∥nx∥=2 gives
∥AΛL+∥≤2h∣ΛL+∣+2(2L+2)(2L+1)≤2(1+h)∣ΛL+∣.
For each u∈UΛLk+, the vector v constructed in Part 3 is connected to u by qk allowed conversions within ΛLk+. Applying (32) to these two vectors gives
Here we used ∑j≥12−j=1 and ∑j≥1j2−j=2. Equation (31) now proves (9).
Acknowledgments
This work was funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) – 470903074.
AI declaration
Large language models were used to assist with the exploration of models and proof strategies, the development and checking of mathematical arguments, literature searches, and the drafting and revision of the manuscript. The author takes full responsibility for its mathematical content and references.
References
[1]Huzihiro Araki and P. D. F. Ion. “On the equivalence of KMS and Gibbs conditions for states of quantum lattice systems”. In: Communications in Mathematical Physics 35.1, pp. 1–12 (1974). DOI: 10.1007/BF01646450.DOI
[2]Gabriel Bouch. “Complex-time singularity and locality estimates for quantum lattice systems”. In: Journal of Mathematical Physics 56.12, 123303 (2015). DOI: 10.1063/1.4936209.DOI
[3]Ola Bratteli and Derek W. Robinson. Operator Algebras and Quantum Statistical Mechanics. Equilibrium States. Models in Quantum Statistical Mechanics. 2nd ed. Vol. 2. Texts and Monographs in Physics. Berlin and Heidelberg: Springer, 1997. DOI: 10.1007/978-3-662-03444-6.DOI
[4]Peter J. Forrester and Colin J. Thompson. “The Golden–Thompson inequality—historical aspects and random matrix applications”. In: Journal of Mathematical Physics 55.2, 023503 (2014). DOI: 10.1063/1.4863477.DOI
[5]Fumio Hiai, Milán Mosonyi, Hiromichi Ohno, and Dénes Petz. “Free energy density for mean field perturbation of states of a one-dimensional spin chain”. In: Reviews in Mathematical Physics 20.3, pp. 335–365 (2008). DOI: 10.1142/S0129055X08003298.DOI
[6]Fumio Hiai and Dénes Petz. “Entropy densities for Gibbs states of quantum spin systems”. In: Reviews in Mathematical Physics 5.4, pp. 693–712 (1993). DOI: 10.1142/S0129055X93000218.DOI
[7]Yu-Min Hu, Zhaoyu Han, and Biao Lian. “Exponential U(1) Symmetry-Breaking Phase as a Disorder-Free Quantum Glass”. In: Physical Review Letters 137, 096002 (2026). DOI: 10.1103/ryb5–8ntp.DOI
[8]Vojkan Jakšić, Claude-Alain Pillet, and Anna Szczepanek. Approach to Equilibrium in Translation-Invariant Quantum Systems: Structural Results II. 2026. arXiv: 2610.21475 [math-ph].arxiv.org/abs/2610.21475
[9]Vojkan Jakšić, Claude-Alain Pillet, and Anna Szczepanek. “Entropy Maximization and Weak Gibbsianity of Quasi-Free Fermionic States”. In: Letters in Mathematical Physics 116, 103 (2026). DOI: 10.1007/s11005–026–02100–9.DOI
[10]Vojkan Jakšić, Claude-Alain Pillet, and Clément Tauber. “A note on adiabatic time evolution and quasi-static processes in translation-invariant quantum systems”. In: Annales Henri Poincaré 25.1, pp. 751–771 (2024). DOI: 10.1007/s00023–023–01282–5.arxiv.org/abs/2204.02177
[11]Vojkan Jakšić, Claude-Alain Pillet, and Clément Tauber. “Approach to equilibrium in translation-invariant quantum systems: some structural results”. In: Annales Henri Poincaré 25.1, pp. 715–749 (2024). DOI: 10.1007/s00023–023–01281–6.arxiv.org/abs/2204.00440
[12]Biao Lian. “Quantum breakdown model: From many-body localization to chaos with scars”. In: Physical Review B 107, 115171 (2023). DOI: 10.1103/PhysRevB.107.115171.DOI
[13]Lorenzo Pettinari. A Finite-Range, Translation-Invariant Counterexample to Weak Gibbsianity. 2026. arXiv: 2610.41871 [math-ph].arxiv.org/abs/2610.41871
[14]Cambyse Rouzé and Daniel Stilck França. Rapid mixing of Gibbs samplers via quantum Dobrushin–Shlosman conditions. 2026. arXiv: 2610.03578v1 [quant-ph].arxiv.org/abs/2610.03578
[15]Carla Rubliani, Marius Wesle, and Tom Wessel. Lieb–Robinson bounds for irreversible dynamics via adaptive Dyson expansion. 2026. arXiv: 2610.12731v1 [math-ph].
[16]Michiko Yuri. “Weak Gibbs measures and the local product structure”. In: Ergodic Theory and Dynamical Systems 22.6, pp. 1933–1955 (2002). DOI: 10.1017/S0143385702000780.DOI
Léa M. Feuillet, Steve Kraemer, Anna Trindade Falcão, Travis C. Fischer, Mitchell Revalski, Lee Armus, Riccardo Middei, James N. Reeves, Valentina Braito ·