betaFeedback

WorksPhysics and Astronomy

Establishing Néel order in the square-lattice antiferromagnetic Heisenberg model: a bootstrap proof

Abstract

Rigorously establishing spontaneous symmetry breaking (SSB) in the thermodynamic limit is generally challenging. This is exemplified by the two-dimensional (2D) S=1/2S = 1/2 antiferromagnetic Heisenberg model where the existence of Néel order is not rigorously established despite overwhelming evidence from a variety of methods. We show that a bootstrap approach can be used to rigorously establish SSB and use it to prove the existence of Néel order in the 2D S=1/2S = 1/2 XXZ model by deriving a strictly positive lower bound on the Néel order parameter for a set of discrete values of the anisotropy parameter ΔΔ in the range [0,2][0,2] including the isotropic (Heisenberg) point Δ=1Δ= 1. Our approach employs a generalization of the Kennedy–Lieb–Shastry inequalities, which use an infrared bound to reduce the existence of long-range order to bounds on certain finite-distance correlators. We compute these bounds for all sufficiently large systems with periodic boundary conditions using the bootstrap. An important ingredient in our proof is a novel method for obtaining exactly certified variational upper bounds on the energy density in the thermodynamic limit from matrix product states. Our approach can be readily generalized to other local lattice models for which infrared susceptibility bounds can be obtained.

Introduction—Determining phase diagrams for microscopic quantum Hamiltonians is a central goal in many-body physics. For spontaneous symmetry breaking (SSB), the central challenge is to establish order in the thermodynamic limit [1]. Unbiased numerical methods such as quantum Monte Carlo and exact diagonalization provide powerful evidence for SSB but rely on finite-size scaling [2] and cannot directly access the thermodynamic limit. On the other hand, variational methods can access much larger or even infinite systems [3, 4] but cannot by themselves certify long-range order.

The absence of a method to certify long-range order means that even for models where the existence of such order is physically well established based on experimental, theoretical, and numerical evidence, a rigorous proof can be absent. This is exemplified by the square-lattice antiferromagnetic Heisenberg model. Pioneering work by Dyson, Lieb, and Simon [5] and subsequent developments [6], particularly by Kennedy, Lieb, and Shastry (KLS) [7], reduced the problem of the existence of long-range Néel order in the model to the violation of certain inequalities involving local correlators. Together with variational energy bounds, the violation of these inequalities rigorously established Néel order for spin S≥1S \ge1 [6, 7] and for spin-1/2 with sufficiently strong easy-axis or easy-plane anisotropy [8, 9, 10, 11, 12, 13]. However, the most interesting case, the isotropic square-lattice Heisenberg model with S=1/2S = 1/2, has remained an open problem [14]. Numerical evidence that the KLS inequalities are violated for the isotropic spin-1/2 model, once sufficiently long-range correlators are included, was obtained by Löw using finite-size quantum Monte Carlo [15].

In this Letter, we show that bootstrap methods [16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41] can be used to establish SSB through finite computations yielding rigorous lower bounds on the order parameter directly in the thermodynamic limit. We use this method to prove the existence of Néel order in the spin-1/2 square-lattice Heisenberg antiferromagnet at zero temperature by obtaining a strictly positive lower bound on the Néel order parameter. This implies the existence of ground states with spontaneous staggered magnetization in the thermodynamic limit [42]. All numerical bounds are converted into exact rational certificates, which makes them rigorous and yields a full mathematical proof.

Our proof builds on the KLS inequalities, which use infrared bounds to analytically control long-wavelength fluctuations and reduce the existence of long-range order to bounds on finite-distance correlators. Our key advances are (i) the derivation of a tighter family of KLS-type inequalities with optimizable parameters, (ii) the use of the bootstrap to obtain rigorous bounds on the correlators that enter such inequalities, and (iii) a new approach to obtain certified bounds on the energy density in the thermodynamic limit from matrix product states (MPSs). The latter is used to tighten the bootstrap bounds, leading to a strictly positive lower bound on the Néel order parameter for a set of discrete values of the anisotropy parameter Δ\Delta in the range [0,2][0, 2], including the isotropic point Δ=1\Delta= 1 (see Figure 1). In principle, a positive lower bound at a given Δ\Delta also extends to a finite interval around Δ\Delta (see End Matter).

Figure 1. Exactly certified lower bounds on the order parameter. The figure shows lower bounds on mx2m_x^2 for Δ≤1\Delta\le1 and mz2m_z^2 for Δ≥1\Delta\ge1 as a function of Δ\Delta. Red circles denote the bounds using a smaller operator set with exactly certified tensor-network energy ceilings. The teal diamonds show the lower bounds for Δ\Delta close to 1 using larger operator sets and tighter variational ceilings using the method described in the text. The shaded columns mark the ranges 0≤Δ≤0.220 \le\Delta\le0.22 and Δ≥1.47\Delta\ge1.47 where Néel order was previously proven [12]. The purple square shows the KLS bound at Δ=0\Delta= 0 [9].

We emphasize that the main structural property of the model that was needed, beyond being a local lattice model, is the infrared bound which is usually a property of models with reflection positivity [5, 43]. More broadly, this framework opens a route to the rigorous and systematic determination of phase diagrams in interacting quantum systems.

The XXZ Model—We consider the spin-1/2 XXZ model on a square-lattice torus Λ\Lambda with unit lattice spacing:

HΔ=∑⟨ij⟩hij=∑⟨ij⟩(XiXj+YiYj+ΔZiZj),(1)H_{\Delta}=\sum_{\langle ij\rangle}h_{ij}=\sum_{\langle ij\rangle}(X_iX_j+Y_iY_j+\Delta Z_iZ_j), \tag*{(1)}

where S=X,Y,ZS=X,Y,Z are the spin-1/2 operators given by Si=σiS/2S_i=\sigma_i^S/2 in terms of the Pauli matrices. HΔH_{\Delta} has a U(1)U(1) internal symmetry with respect to rotations about the zz-axis, and a Z2\mathbb{Z}_2 Ising spin flip symmetry that flips ZZ and YY while leaving XX invariant. The model reduces to the XYXY model when Δ=0\Delta=0, and to the Ising model when Δ→∞\Delta\to\infty. Δ=1\Delta=1 corresponds to the SU(2)SU(2)-symmetric Heisenberg model.

We define the thermodynamic ground state as the limit of the finite-volume ground states on Λ\Lambda as ∣Λ∣→∞|\Lambda|\to\infty. For every finite Λ\Lambda, there is a unique, fully symmetric ground state [44] with correlations denoted as ⟨⋅⟩Λ\langle\cdot\rangle_{\Lambda}. Although the Mermin–Wagner theorem [45] precludes order at any finite temperature, the thermodynamic ground state can spontaneously break symmetries [6], evidenced by a singularity in the thermodynamic structure factor S(q)=lim⁡∣Λ∣→∞SΛ(q)S(\mathbf{q})=\lim_{|\Lambda|\to\infty}S_{\Lambda}(\mathbf{q}), where

SΛ(q)=⟨X~qX~−q⟩Λ,(2)S_{\Lambda}(\mathbf{q})=\langle\widetilde{X}_{\mathbf{q}}\widetilde{X}_{-\mathbf{q}}\rangle_{\Lambda}, \tag*{(2)}

and the Fourier transform is defined as X~q=1∣Λ∣∑reiq⋅rXr\widetilde{X}_{\mathbf{q}}=\frac{1}{\sqrt{|\Lambda|}}\sum_{\mathbf{r}}e^{i\mathbf{q}\cdot\mathbf{r}}X_{\mathbf{r}}.

The Néel order parameter is given by the staggered magnetization Mx=∑rηrXrM_x=\sum_{\mathbf{r}}\eta_{\mathbf{r}}X_{\mathbf{r}}, where ηr=±1\eta_{\mathbf{r}}=\pm1 on the A and B sublattices respectively. The squared order parameter density is defined by

mΛ,x2=⟨M2⟩Λ∣Λ∣2=SΛ(Q)∣Λ∣(3)m^2_{\Lambda,x}=\frac{\langle M^2\rangle_{\Lambda}}{|\Lambda|^2}=\frac{S_{\Lambda}(\mathbf{Q})}{|\Lambda|} \tag*{(3)}

where Q=(π,π)\mathbf{Q}=(\pi,\pi).

In the thermodynamic limit, mx2=lim⁡∣Λ∣→∞mΛ,x2m_x^2=\lim_{|\Lambda|\to\infty}m^2_{\Lambda,x} measures the coefficient of a δ\delta-function singularity in the structure factor at q=Q\mathbf{q}=\mathbf{Q}.

The KLS construction—We start by reviewing the construction of Kennedy-Lieb-Shastry (KLS) [9], which establishes a lower bound on the magnetization in terms of local correlators using an upper bound on S(q)S(\mathbf{q}) introduced by Refs. [5, 43] along with the sum rules obeyed by S(q)S(\mathbf{q}). The KLS inequality, combined with variational energy upper bounds, was used previously to establish Néel order in the 2D XYXY model and in the Heisenberg model in more than two spatial dimensions. However, it was insufficient to establish Néel order in 2D Heisenberg model.

The KLS construction begins with the sum rule

∑q1∣Λ∣γn(q)SΛ(q)=⟨X0Xn⟩=CnX,n∈Z2,(4)\sum_{\mathbf{q}}\frac{1}{|\Lambda|}\gamma_{\mathbf{n}}(\mathbf{q})S_{\Lambda}(\mathbf{q})=\langle X_0X_{\mathbf{n}}\rangle=C_{\mathbf{n}}^X,\qquad\mathbf{n}\in\mathbb{Z}^2, \tag*{(4)}

where γn(q)=cos⁡(nxqx+nyqy)\gamma_{\mathbf{n}}(\mathbf{q})=\cos(n_xq_x+n_yq_y). For n=0=(0,0)\mathbf{n}=0=(0,0), C0X=1/4C_0^X=1/4 for any state, whereas other choices of n\mathbf{n} relate various harmonics of the structure factor to local correlators. Although KLS only considered the cases of n=0\mathbf{n}=0 or (1,0)(1,0), Eq. (4) represents a straightforward generalization.

correlators. Although KLS only considered the cases of n=0\mathbf{n}=0 or (1,0)(1,0), Eq. (4) represents a straightforward generalization.

Taking linear combinations of Eq. (4) for various n\mathbf{n} and isolating the q=Q\mathbf{q}=\mathbf{Q} term gives the generalized sum rule:

w(Q)mΛ,x2=a04+∑nanCnX−1∣Λ∣∑q≠Qw(q)SΛ(q)(5)w(\mathbf{Q})m^2_{\Lambda,x}=\frac{a_0}{4}+\sum_{\mathbf{n}}a_{\mathbf{n}}C_{\mathbf{n}}^X-\frac{1}{|\Lambda|}\sum_{\mathbf{q}\ne\mathbf{Q}}w(\mathbf{q})S_{\Lambda}(\mathbf{q}) \tag*{(5)}

where w(q)w(\mathbf{q}) is defined as w(q)=a0+∑nanγn(q)w(\mathbf{q})=a_0+\sum_{\mathbf{n}}a_{\mathbf{n}}\gamma_{\mathbf{n}}(\mathbf{q}) and is normalized such that w(Q)=1w(\mathbf{Q})=1. Since SΛ(q)≥0S_{\Lambda}(\mathbf{q})\geq0, we obtain a lower bound on the order parameter by dropping the regions where w(q)<0w(\mathbf{q})<0:

mΛ,x2≥a04+∑nanCnX−1∣Λ∣∑q≠Q{w(q)}+SΛ(q),(6)m^2_{\Lambda,x}\geq\frac{a_0}{4}+\sum_{\mathbf{n}}a_{\mathbf{n}}C_{\mathbf{n}}^X-\frac{1}{|\Lambda|}\sum_{\mathbf{q}\ne\mathbf{Q}}\{w(\mathbf{q})\}_{+}S_{\Lambda}(\mathbf{q}), \tag*{(6)}

where {w(q)}+=max⁡(0,w(q))\{w(\mathbf{q})\}_{+}=\max(0,w(\mathbf{q})). The first two terms on the right-hand side depend only on local correlators, but the last one involves a sum over all momenta, which cannot be expressed in terms of finitely many local correlators.

To deal with the last term, KLS derived an upper bound on SΛ(q)S_{\Lambda}(\mathbf{q}) using an infrared bound on the susceptibility as we review below. The dynamical structure factor is

SΛ(q,ω)=∑n∣⟨n∣X~q∣0⟩∣2δ(ω−ωn)≥0,(7)S_{\Lambda}(\mathbf{q},\omega)=\sum_n|\langle n|\widetilde{X}_{\mathbf{q}}|0\rangle|^2\delta(\omega-\omega_n)\geq0, \tag*{(7)}

where ωn=En−E0\omega_n=E_n-E_0. Some of its moments mp(q)=∫0∞dω ωpSΛ(q,ω)m_p(\mathbf{q})=\int_0^\infty d\omega\,\omega^pS_{\Lambda}(\mathbf{q},\omega) are familiar objects: m0=SΛ(q)m_0=S_{\Lambda}(\mathbf{q}), m1=12Dx(q)m_1=\frac{1}{2}D_x(\mathbf{q}) and m−1=12χx(q)m_{-1}=\frac{1}{2}\chi_x(\mathbf{q}), where Dx(q)=⟨[X~q†,[H,X~q]]⟩D_x(\mathbf{q})=\langle[\widetilde{X}_{\mathbf{q}}^\dagger,[H,\widetilde{X}_{\mathbf{q}}]]\rangle is the double commutator and χx(q)\chi_x(\mathbf{q}) is the static susceptibility. The Cauchy–Schwarz inequality in ω\omega gives

S(q)≤m1m−1=12Dx(q)χx(q).(8)S(\mathbf{q})\leq\sqrt{m_1m_{-1}}=\frac{1}{2}\sqrt{D_x(\mathbf{q})\chi_x(\mathbf{q})}. \tag*{(8)}

The double commutator only retains local terms. In particular, for the nearest-neighbor Hamiltonian, it only involves nearest-neighbor correlations,

Dx(q)=−4[(1−Δγ(q))C1,0X+(Δ−γ(q))C1,0Z],(9)D_x(\mathbf{q})=-4\left[(1-\Delta\gamma(\mathbf{q}))C_{1,0}^X+(\Delta-\gamma(\mathbf{q}))C_{1,0}^Z\right], \tag*{(9)}

where we have defined γ(q)=12(γ1,0(q)+γ0,1(q))\gamma(\mathbf{q})=\frac{1}{2}(\gamma_{1,0}(\mathbf{q})+\gamma_{0,1}(\mathbf{q})). On the other hand, χx(q)=2⟨X~q†(H−E0)−1X~q⟩\chi_x(\mathbf{q})=2\langle\widetilde{X}_{\mathbf{q}}^\dagger(H-E_0)^{-1}\widetilde{X}_{\mathbf{q}}\rangle depends on the whole low-energy spectrum. A small amount of spectral weight at very low frequency changes every local expectation value negligibly but changes χ\chi by an arbitrary amount. Thus, while Eq. (8) gives us an upper bound, it simply trades one unknown, S(q)S(\mathbf{q}), for another, χx(q)\chi_x(\mathbf{q}).

So far no special property of the Hamiltonian has been used apart from locality. Using the reflection positivity of the Hamiltonian, Refs. [5, 43] derived the infrared bound:

χx(q)≤χ‾x(q)=14(1+γ(q)),q≠Q,(10)\chi_x(\mathbf{q})\leq\overline{\chi}_x(\mathbf{q})=\frac{1}{4(1+\gamma(\mathbf{q}))},\qquad\mathbf{q}\ne\mathbf{Q}, \tag*{(10)}

which holds in the Gibbs state at any temperature, and hence in the ground state, on any even torus and for every Δ≥0\Delta\ge0 (see Ref. [46] for the precise hypotheses, and the End Matter for a derivation).

In two dimensions at zero temperature, the singularity of this bound at q=Q\mathbf{q} = \mathbf{Q} is integrable. Together, Eqs. (8) and (10) give an upper bound on the structure factor that only involves local correlators:

S(q)≤U(q)=12−[(1−Δγ(q))C1,0X+(Δ−γ(q))C1,0Z]1+γ(q)(11)S(\mathbf{q}) \le U(\mathbf{q}) = \frac{1}{2}\sqrt{-\frac{[(1-\Delta\gamma(\mathbf{q}))C^X_{1,0}+(\Delta-\gamma(\mathbf{q}))C^Z_{1,0}]}{1+\gamma(\mathbf{q})}} \tag*{(11)}

Inserting Eq. (11) into Eq. (6) produces the KLS inequality [9]

mΛ,x2≥a04+∑nanCnX−1∣Λ∣∑q≠Q{w(q)}+U(q).(12)m^2_{\Lambda,x} \ge\frac{a_0}{4}+\sum_{\mathbf{n}}a_{\mathbf{n}}C^X_{\mathbf{n}}-\frac{1}{|\Lambda|}\sum_{\mathbf{q}\ne\mathbf{Q}}\{w(\mathbf{q})\}_+U(\mathbf{q}). \tag*{(12)}

In the thermodynamic limit, the sum becomes an integral over the Brillouin zone. KLS used variational upper bounds and the Kubo inequality ∣C1,0Z∣≤−C1,0X|C^Z_{1,0}| \le-C^X_{1,0} to establish Néel order in the ground-state of the XY model.

Generalizing the KLS inequality—The only non-local input entering the KLS argument is the infrared bound in Eq. (10); everything else is a statement about the moments of the non-negative function S(q,ω)S(\mathbf{q},\omega). Thus, including higher moments may help us tighten the bound. Indeed, for inversion symmetric states, the odd moments

m2p+1(q)=m2p+1(−q)=12⟨[X~q†,ad⁡H2p+1X~q]⟩m_{2p+1}(\mathbf{q})=m_{2p+1}(-\mathbf{q})=\frac{1}{2}\langle[\widetilde{X}_{\mathbf{q}}^\dagger,\operatorname{ad}^{2p+1}_H\widetilde{X}_{\mathbf{q}}]\rangle

are equal to linear combinations of local correlators. Each commutator with the Hamiltonian enlarges the support of an operator by at most one site, so ad⁡H2p+1X0\operatorname{ad}^{2p+1}_H X_0 is a sum of Pauli strings on at most 2p+22p+2 sites within a taxicab distance 2p+12p+1 of the origin, and ⟨[Xr,ad⁡H2p+1X0]⟩=0\langle[X_{\mathbf{r}},\operatorname{ad}^{2p+1}_H X_0]\rangle=0 unless ∣r∣≤2p+1|\mathbf{r}| \le2p+1. Therefore, each m2p+1(q)m_{2p+1}(\mathbf{q}) is a finite trigonometric polynomial in q\mathbf{q} whose coefficients are linear combinations of local correlators:

m2p+1(q)=∑∣r∣≤2p+1γr(q)G2p+1(r),(13)m_{2p+1}(\mathbf{q})=\sum_{|\mathbf{r}|\le2p+1}\gamma_{\mathbf{r}}(\mathbf{q})G_{2p+1}(\mathbf{r}), \tag*{(13)}

where Gp(r)=12⟨[Xr,ad⁡HpX0]⟩G_p(\mathbf{r})=\frac{1}{2}\langle[X_{\mathbf{r}},\operatorname{ad}^p_HX_0]\rangle. The even moments, including m0(q)=S(q)m_0(\mathbf{q})=S(\mathbf{q}), are genuinely non-local and remain unknown.

Bounding S(q)S(\mathbf{q}) thus becomes a moment problem: how large can m0=S(q)m_0=S(\mathbf{q}) be for a nonnegative function S(q,ω)S(\mathbf{q},\omega) on ω>0\omega>0 whose odd moments m1,m3,…,m2J−1m_1,m_3,\ldots,m_{2J-1} are given and whose moment m−1m_{-1} is at most χˉ(q)/2\bar{\chi}(\mathbf{q})/2? Every polynomial

P(x)=λ−x+∑p=1Jα2p−1x2p≥0for all x≥0(14)P(x)=\lambda-x+\sum_{p=1}^{J}\alpha_{2p-1}x^{2p}\ge0\quad\text{for all }x\ge0 \tag*{(14)}

gives such an upper bound: since ∫dω P(ω)ωS(q,ω)≥0\int d\omega\,\frac{P(\omega)}{\omega}S(\mathbf{q},\omega)\ge0, we get

S(q)≤fP(q)=λχˉ(q)2+∑p=1Jα2p−1m2p−1(q).(15)S(\mathbf{q})\le f_P(\mathbf{q})=\lambda\frac{\bar{\chi}(\mathbf{q})}{2}+\sum_{p=1}^{J}\alpha_{2p-1}m_{2p-1}(\mathbf{q}). \tag*{(15)}

Apart from the −x-x term, PP contains only even powers, so dividing by ω\omega and integrating against S(q,ω)S(\mathbf{q},\omega) produces only m−1m_{-1}, m0m_0 and the local odd moments. Choosing the best PP gives the tightest cap, which equals the largest m0m_0 compatible with the infrared bound and the given moments, i.e. the solution of a truncated Stieltjes moment problem [47, 48]. For J=1J=1, P≥0P\ge0 if and only if 4λα1≥14\lambda\alpha_1\ge1, and the optimum gives the KLS cap Eq. (11).

A different polynomial PP may be used at each momentum. For any family P={Pq}P=\{P_{\mathbf{q}}\} satisfying Eq. (14), inserting the cap into the generalized sum rule gives

mx2≥a04+∑nanCnX−∫d2q(2π)2{w(q)}+fP(q).(16)m_x^2\ge\frac{a_0}{4}+\sum_{\mathbf{n}}a_{\mathbf{n}}C^X_{\mathbf{n}}-\int\frac{d^2\mathbf{q}}{(2\pi)^2}\{w(\mathbf{q})\}_+f_P(\mathbf{q}). \tag*{(16)}

The tightest bound uses, at each q\mathbf{q}, the polynomial that minimizes fP(q)f_P(\mathbf{q}); any other choice gives a larger cap and hence a weaker but valid bound.

Easy-axis order for Δ≥1\Delta\ge1—For Δ≥1\Delta\ge1 the spins order along zz, and the same construction bounds mz2=lim⁡Λ→∞1∣Λ∣2⟨(∑rnrZr)2⟩m_z^2=\lim_{\Lambda\to\infty}\frac{1}{|\Lambda|^2}\langle(\sum_{\mathbf{r}}n_{\mathbf{r}}Z_{\mathbf{r}})^2\rangle, with Z~q\widetilde{Z}_{\mathbf{q}} instead of X~q\widetilde{X}_{\mathbf{q}}, the correlators CnZC^Z_{\mathbf{n}} appearing in the sum rule. The corresponding infrared bound changes to χZ(q)≤1/[4Δ(1+γ(q))]\chi_Z(\mathbf{q})\le1/[4\Delta(1+\gamma(\mathbf{q}))] [5, 46] (End Matter), and the double commutator becomes

Dz(q)=−8C1,0X(1−γ(q)).(17)D_z(\mathbf{q})=-8C^X_{1,0}(1-\gamma(\mathbf{q})). \tag*{(17)}

The higher moments follow from ad⁡H2p+1Z0\operatorname{ad}_{H}^{2p+1}Z_0 exactly as before.

Bootstrapping ground-state correlations—Eq. (16) uses a variety of local correlators beyond nearest-neighbor XXXX and ZZZZ correlations both through the inclusion of higher nn in the sum rule and the consideration of higher moments in the S(q)S(\mathbf{q}) upper bound. Previous approaches to this problem so far could only get rigorous one-sided bounds on nearest-neighbour correlations coming from the variational ceiling, with no information about the longer-range and multi-site correlators. The full freedom in Eq. (16) therefore could not be exploited, and earlier bounds were restricted to the nearest-neighbor sum rule and the KLS cap, allowing the proof of easy-plane order for 0≤Δ≤0.220 \le\Delta\le0.22 and easy-axis order for Δ≥1.47\Delta\ge1.47 [8, 10, 11, 12, 49]. However, the proof of order around the Heisenberg point Δ=1\Delta= 1 has remained an open problem [50].

We get rigorous control of arbitrary local ground-state correlations using the Hamiltonian bootstrap framework [33]. Instead of explicitly computing the ground-state, the bootstrap looks for values of correlators that are compatible with a set of consistency conditions defined for the ground-state. The ground-state expectation value ⟨⋅⟩\langle\cdot\rangle is a linear map from local operators to the complex numbers, so it is fixed by its values on a basis of Pauli strings. We choose a finite set of local operators OaO_a and let y\mathbf{y} collect the expectation values of the distinct Pauli strings that appear in the products Oa†ObO_a^\dagger O_b and Oa†[H,Ob]O_a^\dagger[H,O_b]; every expectation value we need is then a linear function of y\mathbf{y}. We use the symmetries of the Hamiltonian to identify Pauli strings related by lattice and internal symmetries. Each consistency condition is either a linear equality or inequality in y\mathbf{y}, or a linear matrix inequality M(y)⪰0M(\mathbf{y}) \succeq0, where the entries of MM depend linearly on y\mathbf{y}. Because positive semidefinite (PSD) matrices form a convex cone, these conditions define a convex feasible set K\mathcal{K}, which is guaranteed to contain the true ground-state values.

The conditions we impose are

⟨I⟩=1,⟨Oa†Ob⟩⪰0,∣⟨P⟩∣≤1,\langle I\rangle= 1,\qquad\langle O_a^\dagger O_b\rangle\succeq0,\qquad|\langle P\rangle| \le1,
⟨[H,O]⟩=0,⟨Oa†[H,Ob]⟩⪰0,(18)\langle[H,O]\rangle= 0,\qquad\langle O_a^\dagger[H,O_b]\rangle\succeq0, \tag*{(18)}

where ⪰0\succeq0 means that the matrix with indices a,ba,b is positive semidefinite. The first line holds in any physical state: normalization, positivity, and the norm bound ∣⟨P⟩∣≤1|\langle P\rangle| \le1 on every Pauli string PP.

The second line characterizes ground states of HH: stationarity holds for any eigenstate, and perturbative positivity, ⟨O†[H,O]⟩=⟨O†(H−Egs)O⟩≥0\langle O^\dagger[H,O]\rangle= \langle O^\dagger(H-E_{\mathrm{gs}})O\rangle\ge0, states that no operator can lower the energy. These conditions involve only local operators and commutators with HH, so they remain meaningful in the thermodynamic limit [28].

We also impose the lattice and spin symmetries of the ground state, which is unique on every even torus, as equalities between expectation values of symmetry-related operators, and use them to block-diagonalize the PSD constraints [33]. All these conditions hold on every large enough even torus, and therefore in the infinite-volume limit.

While these relations are, in principle, sufficient to exactly define the ground states when applied to all operators in the quasi-local algebra, in practice, we can only apply them for a small subset of operators supported on a finite region. Thus, it is useful to supplement these with other relations that incorporate additional information or rely on a specific property of the Hamiltonian. For the XXZ model, translation and U(1)U(1) symmetry make the two bonds per site contribute equally, so the ground-state energy density is

ϵ0=lim⁡∣Λ∣→∞E0(Λ)∣Λ∣=4C1,0X+2ΔC1,0Z.(19)\epsilon_0 = \lim_{|\Lambda|\to\infty}\frac{E_0(\Lambda)}{|\Lambda|} = 4C_{1,0}^{X} + 2\Delta C_{1,0}^{Z}. \tag*{(19)}

A certified variational upper bound ϵ0≤ϵvar\epsilon_0 \le\epsilon_{\mathrm{var}}, which we obtain with tensor networks below, therefore becomes the linear constraint 4C1,0X+2ΔC1,0Z≤ϵvar4C_{1,0}^{X}+2\Delta C_{1,0}^{Z}\le\epsilon_{\mathrm{var}} on y\mathbf{y}. Unlike the conditions above, this constraint need not hold on a finite torus, where E0(Λ)/∣Λ∣E_0(\Lambda)/|\Lambda| can lie below ϵ0\epsilon_0, but it holds in the infinite-volume limit, which is all we use. We also use reflection positivity (RP), ⟨AaΘ(Ab)⟩≥0\langle A_a\Theta(A_b)\rangle\ge0, and the Marshall sign rule, ηW⟨W⟩≥0\eta_W\langle W\rangle\ge0. RP follows from the structure of the Hamiltonian: Θ\Theta maps operators AaA_a on one side of a bond-centered mirror plane to their reflections Θ(Aa)\Theta(A_a) on the other side (see End Matter). The Marshall sign rule [44, 51] states that the ground-state expectation value of any product WW of raising and lowering operators S±S^\pm has the sign ηW=(−1)#(sites of W on the B sublattice)\eta_W=(-1)^{\#\text{(sites of }W\text{ on the }B\text{ sublattice})}.

For a fixed choice of ω\omega and PP, the cap FP(q)F_{P}(\mathbf{q}) is linear in y\mathbf{y}, making the right-hand side of Eq. (16) an affine function Fω,P(y)F_{\omega,P}(\mathbf{y}). Since the true ground-state values y∗\mathbf{y}^* lie in K\mathcal{K}, a single semidefinite program bounds all the correlators in Fω,PF_{\omega,P} jointly:

mx2≥Fω,P(y∗)≥min⁡y∈KFω,P(y).(20)m_x^2 \ge F_{\omega,P}(\mathbf{y}^*) \ge\min_{\mathbf{y}\in\mathcal{K}}F_{\omega,P}(\mathbf{y}). \tag*{(20)}

Since m2p+1m_{2p+1} already includes a number of strings on up to 2p+22p+2 sites, getting tight bounds for each of these strings requires progressively larger operator sets. For instance, m7m_7 alone requires about 1.9×1051.9\times10^5 distinct strings, with most of them supported on 8 sites. In practice, we restrict ourselves to J=3J=3, which includes odd moments up to m5m_5.

The optimal bound for J=1J=1 is the KLS cap, which includes a square root over nearest-neighbour correlators. However, with fixed coefficients, the J=1J=1 polynomial in Eq. (14) is a tangent line to this square root with α1=1/(4λ)\alpha_1=1/(4\lambda), λχˉ/2+D/(8λ)≥12Dχˉ\lambda\bar{\chi}/2+D/(8\lambda)\ge\frac{1}{2}\sqrt{D\bar{\chi}} for every λ>0\lambda>0, with equality at one value of λ\lambda. Choosing λ\lambda at a reference point therefore gives a valid cap that is affine in y\mathbf{y} and tight at that point. We use the KLS cap close to Q0Q_0, where χˉ\bar{\chi} diverges, and the J=3J=3 cap elsewhere (End Matter).

Optimizing the weights and the cap—Every weight ww with w(Q)=1w(\mathbf{Q})=1, and every set of cap polynomials P=P= {Pq}\{P_q\} nonnegative on [0,∞)[0,\infty), gives a valid bound, so

mx2≥max⁡w,Pmin⁡y∈KFw,P(y).(21)m_x^2 \ge\max_{w,P}\min_{\mathbf{y}\in\mathcal{K}} F_{w,P}(\mathbf{y}). \tag*{(21)}

For fixed PP, the maximization over ww can be combined with the bootstrap into a single semidefinite program (End Matter). The polynomials are chosen from the moment problem at each momentum. For given m1m_1, m3m_3, m5m_5 and χˉ\bar{\chi}, the largest S(q)S(q) is attained when S(q,ω)S(q,\omega) consists of two delta functions, a low-frequency magnon-like peak and a high-frequency continuum peak. The optimal polynomial is the one that vanishes, with zero slope, at exactly these two frequencies, so that the cap is saturated. We compute these frequencies from the moments at the current optimum, round the polynomial’s coefficients to exact rationals, and verify exactly that it is nonnegative on [0,∞)[0,\infty). We then re-optimize ww with the new polynomials and repeat; a few rounds suffice. A final semidefinite program at fixed ww and PP gives the reported bound. Its validity does not depend on how ww and PP were found.

**Certified variational upper bound—**Variational energy bounds have long played an important role in rigorous studies of antiferromagnetic order [7, 9]. Through Eq. (19), a rigorous upper bound on ϵ0\epsilon_0 directly constrains the nearest-neighbor correlations entering the Néel bound. In fact, our bootstrap lower bound on the Néel order parameter at Δ=1\Delta=1 becomes positive only after imposing the improved energy bound derived below.

The Heisenberg ground-state energy is known numerically to very high precision, with quantum Monte Carlo giving ϵ0=−0.669441857(7)\epsilon_0=-0.669441857(7) [53] and consistent estimates from variational Monte Carlo [56], tensor networks [57], and coupled-cluster calculations [55]. These calculations, however, do not provide certified upper bounds. By the variational principle, any explicit ansatz gives a rigorous upper bound provided its energy can be evaluated with controlled accuracy. The best previous rigorous bound, ϵ0≤−0.66068\epsilon_0\leq-0.66068, was obtained from a fermionic mean-field ansatz [52]. Another strategy is to solve finite patches and tile the lattice with the resulting patch states [58, 59, 60]. Such product states can be contracted exactly in the thermodynamic limit, but lose quantum correlations across patch boundaries.

Building on this idea, we develop a new certified variational construction that replaces finite patches with infinite strips and restores entanglement across the remaining seams with a shallow quantum circuit, as shown in Figure 2(a). Writing ψ\psi for the width-WW infinite-strip MPS and UU for the seam circuit,

Figure 2. Certified variational upper bounds. (a) Infinite width-WW MPS strips are coupled across their seams by a shallow circuit of four-spin RR and eight-spin QQ gates. (b) Certified hierarchy at Δ=1\Delta=1 for W=12W=12 and χ=3072\chi=3072: bare strips, classical edge mean field, a 1×21\times2 seam gate, and the full R+QR+Q circuit. Dashed line: previous rigorous bound [52]; red line: QMC [53]. (c) Certified bounds across the XXZ model (blue), compared with QMC [54], coupled-cluster (CCM) results [55], and our evaluation of Ovchinnikov’s ansatz at each Δ\Delta [52].

∣Ψ(ψ,U)⟩=U⨂s∣ψ⟩s.(22)|\Psi(\psi,U)\rangle= U \bigotimes_s |\psi\rangle_s. \tag*{(22)}

Ordering the spins column by column reduces every contraction within a strip to a one-dimensional sequential process, so the reduced density matrix ρA(s)\rho_A^{(s)} of any finite region

region AA on strip ss can be obtained directly in the thermodynamic limit. We use four-spin gates RR and eight-spin gates QQ to restore correlations across the seams, as shown in Figure 2(a). Because the circuit has finite depth, the backward causal cone of a seam bond hijh_{ij} has finite support A∪BA\cup B, with A⊂sA\subset s and B⊂s+1B\subset s+1 on the two neighboring strips. Let UijU_{ij} denote the gates within this causal cone. Since the strips are independent before the circuit is applied,

⟨hij⟩Ψ=Tr⁡[(ρA(s)⊗ρB(s+1))Uij†hijUij].(23)\langle h_{ij}\rangle_{\Psi}=\operatorname{Tr}\left[\left(\rho_A^{(s)}\otimes\rho_B^{(s+1)}\right)U_{ij}^{\dagger}h_{ij}U_{ij}\right]. \tag*{(23)}

Thus each bond energy reduces to one-dimensional strip contractions for ρA(s)\rho_A^{(s)} and ρB(s+1)\rho_B^{(s+1)}, followed by a finite contraction over the circuit causal cone.

We optimize the strip state and circuit alternately,

ψn,Un→DMRG / Heff(Un)ψn+1,Un→optimize U / E(U)ψn+1,Un+1.(24)\psi_n,U_n \xrightarrow{\mathrm{DMRG}\,/\,H_{\mathrm{eff}}(U_n)} \psi_{n+1},U_n \xrightarrow{\mathrm{optimize}\ U\,/\,E(U)} \psi_{n+1},U_{n+1}. \tag*{(24)}

For fixed UU, the circuit defines an effective one-dimensional MPO Heff(U)H_{\mathrm{eff}}(U) optimized by DMRG. For fixed ψ\psi, the circuit gates are optimized using the finite energy functional E(U)E(U). The resulting state is then certified independently. A finite-patch inequality converts its local energy evaluation into a thermodynamic upper bound, with all numerical errors enclosed using verified floating-point bounds and outward-rounded interval arithmetic. The optimization itself does not enter the certificate. Details are given in the End Matter.

Figure 2(b) shows the resulting hierarchy at the Heisenberg point for W=12W = 12 and χ=3072\chi= 3072. Bare decoupled strips already give ϵ0≤−0.661161\epsilon_0 \le-0.661161, surpassing the previous rigorous bound. Adding a classical edge mean field [59] gives essentially no further improvement, showing that classical boundary polarization alone does not recover the remaining seam energy. In contrast, a 1×21 \times2 gate acting across each seam bond improves the certified bound to −0.667203-0.667203. The full R+QR + Q circuit gives

ϵ0≤−0.668047,(25)\epsilon_0 \le-0.668047, \tag*{(25)}

only 0.21%0.21\% above the quantum Monte Carlo energy.

The same construction gives certified upper bounds throughout the XXZ regime used below, as shown in Figure 2(c). Our variational ansatz preserves the SzS^z symmetry and is therefore naturally suited to the easy-axis regime Δ≥1\Delta\ge1. For Δ<1\Delta< 1, where the easy-plane ground state breaks this symmetry, an exact spin rotation maps the energy bound to an easy-axis problem at Δ′=2/(1+Δ)>1\Delta' = 2/(1 + \Delta) > 1, with the energy rescaled by (1+Δ)/2(1 + \Delta)/2. The rotation and its certification are described in the End Matter. The resulting bounds closely track quantum Monte Carlo and coupled-cluster benchmarks [54, 55] and provide the variational energy ceilings used in the bootstrap below.

Results— Our main results are shown in Figure 1. We evaluate the lower bound in Eq. (16) on a grid of Δ∈[0,2]\Delta\in[0, 2], bounding mx2m_x^2 for Δ≤1\Delta\le1 and mz2m_z^2 for Δ≥1\Delta\ge1. Away from Δ=1\Delta= 1, a small operator set (∼2.3×104\sim2.3 \times10^4 free moments) with a moderately tight energy ceiling is sufficient to establish order[61]. Close to Δ=1\Delta= 1, these bounds are negative, requiring larger operator sets. With operator sets as many as 1.2×1051.2 \times10^5 free moments at Δ=1\Delta= 1, and the tighter variational ceilings described above, we recover a positive bound at every Δ\Delta on the grid.

At Δ=0\Delta= 0, the smaller operator set already gives mx2≥0.068m_x^2 \ge0.068, which is about six times larger than the KLS bound mx2≥0.0106m_x^2 \ge0.0106[9]. At the Heisenberg point, our best bound using the largest operator set and the best certified ceiling ϵvar=−0.668047\epsilon_{\mathrm{var}} = -0.668047 is:

mx2≥7.63×10−4.(26)m_x^2 \ge7.63 \times10^{-4}. \tag*{(26)}

The generalized sum rule uses every correlator CnXC_n^X that the operator set can bound: all symmetry-inequivalent separations with taxicab distance ∣n∣≤6|\mathbf{n}| \le6 (15 shells) for the smaller operator set shown in Figure 1, and taxicab distance ∣n∣≤16|\mathbf{n}| \le16 (80 shells) for the largest set at Δ=1\Delta= 1. All the bounds are exact certificates, with the certification method described in the End Matter.

We emphasize that a straightforward generalization of the construction also allows us to bootstrap upper bounds on the order parameter without needing the infrared bound[62]. Moreover, we also discuss how to extend the bound certificate at a given Δ\Delta to a finite interval around Δ\Delta.

Acknowledgments— NC is grateful to Subir Sachdev for fruitful discussions and for pointing out Ref. [15]. This work is supported by NSF CAREER grant DMR award No. 2441781 and Alfred P. Sloan Research Fellowship to EK. TW acknowledges additional support from the Harvard Quantum Initiative Fellowship and DOE (USA) office of Basic Sciences Grant No. DE-FG02-03ER46076. MGS is supported by a postdoctoral fellowship from the Julian Schwinger Foundation. PN is supported by the Gordon and Betty Moore Foundation through Grant GBMF14251 to the American Chemical Society for the ACS-Moore Postdoctoral Bridge Award in Quantum Materials. This work uses the FASRC Cannon cluster supported by the FAS Division of Science Research Computing Group at Harvard University, the MIT Office of Research Computing and Data (ORCD). This work used Anvil [63] at Purdue University through allocation PHY260501 from the Advanced Cyberinfrastructure Coordination Ecosystem: Services & Support (ACCESS) program [64], which is supported by U.S. National Science Foundation grants #2138259, #2138286, #2138307, #2137603, and #2138296. The authors acknowledge the use of AI tools Claude Opus 5.5, Fable 5.1 and ChatGPT 6.1 for searching the literature. In addition Claude Opus 5.5 and Fable 5.1 were used for assistance with coding, and algorithmic improvements, particularly the certification procedure.

References

End Matter

Reflection positivity—Consider the reflection r(x,y)=(1−x,y)r(x,y)=(1-x,y) across the vertical plane x=1/2x=1/2, which divides the even torus into left (LL) and right (RR) halves with operator algebras AL\mathcal{A}_L and AR\mathcal{A}_R. If there is an antilinear isomorphism Θ:AL→AR\Theta:\mathcal{A}_L\to\mathcal{A}_R,

Θ(cA)=cˉΘ(A),Θ(A†)=Θ(A)†,Θ(AB)=Θ(A)Θ(B),(27)\Theta(cA)=\bar{c}\Theta(A),\qquad\Theta(A^\dagger)=\Theta(A)^\dagger,\qquad\Theta(AB)=\Theta(A)\Theta(B), \tag*{(27)}

such that

H=HL+Θ(HL)−∑k∈∂LckOkΘ(Ok),ck≥0,(28)H=H_L+\Theta(H_L)-\sum_{k\in\partial L}c_kO_k\Theta(O_k),\qquad c_k\geq0, \tag*{(28)}

with HL,Ok∈ALH_L,O_k\in\mathcal{A}_L. Then ⟨OΘ(O)⟩β≥0\langle O\Theta(O)\rangle_\beta\geq0 for all O∈ALO\in\mathcal{A}_L in every Gibbs state of HH, and, taking β→∞\beta\to\infty, in the ground state [5, 65]. Equivalently, the matrix ⟨AaΘ(Ab)⟩\langle A_a\Theta(A_b)\rangle is positive semidefinite, which is the RP constraint we impose.

For the XXZ model, Θ\Theta is the mirror reflection combined with time reversal, Θ(Si)=−Sr(i)\Theta(S_i)=-S_{r(i)}, and each bond crossing the plane gives XiXr(i)+YiYr(i)+ΔZiZr(i)=−[XiΘ(Xi)+YiΘ(Yi)+ΔZiΘ(Zi)]X_iX_{r(i)}+Y_iY_{r(i)}+\Delta Z_iZ_{r(i)}=-[X_i\Theta(X_i)+Y_i\Theta(Y_i)+\Delta Z_i\Theta(Z_i)], so the OkO_k are the spin components on the sites next to the plane, with weights (1,1,Δ)≥0(1,1,\Delta)\geq0 for every Δ≥0\Delta\geq0.

Deriving the infrared bound.—We investigate ordering along the spin component SS (XX for in-plane and ZZ for easy-axis order), and g=1g=1, Δ\Delta is the respective coupling in HH. Since neighbours lie on opposite sublattices and Si2=14S_i^2=\frac{1}{4}, each SS-bond coupling is gSiSj=g2(τi−τj)2−g4gS_iS_j=\frac{g}{2}(\tau_i-\tau_j)^2-\frac{g}{4}, where τi=ηiSi\tau_i=\eta_iS_i is the staggered spin. We probe the staggered fluctuations with a real site field h=(hi)h=(h_i) which shifts (τi−τj)→(τi−τj−hi+hj)(\tau_i-\tau_j)\to(\tau_i-\tau_j-h_i+h_j) in the SS-coupling, with the other two spin components unchanged. This gives the perturbed Hamiltonian:

H(h)=H−g(τ,Kh)+g2(h,Kh),(Kh)i=∑j∼i(hi−hj),(29)H(h)=H-g(\tau,Kh)+\frac{g}{2}(h,Kh),\qquad(Kh)_i=\sum_{j\sim i}(h_i-h_j), \tag*{(29)}

where (u,v)=∑iuivi(u,v)=\sum_i u_iv_i, j∼ij\sim i, are the sites coupled to a given ii. Let Z(h)=Tr⁡e−βH(h)Z(h)=\operatorname{Tr}e^{-\beta H(h)}. For an unconstrained classical Gaussian field, the shift τ→τ+h\tau\to\tau+h would make Z(h)Z(h) independent of hh. For spins one gets the weaker statement of Gaussian domination, Z(h)≤Z(0)Z(h)\leq Z(0) [5, 43], which rests on the following consequence of reflection positivity.

The Hilbert space is a tensor product of the left and right halves, for any F,G∈ALF,G\in\mathcal{A}_L, Tr⁡[FΘ(G)]=Tr⁡L(F)Tr⁡L(G)‾\operatorname{Tr}[F\Theta(G)]=\operatorname{Tr}_L(F)\overline{\operatorname{Tr}_L(G)}, where Tr⁡L\operatorname{Tr}_L is the trace over the left half and the complex conjugate comes from the antilinearity of Θ\Theta. For operators A,B,Ck,Dk∈ALA,B,C_k,D_k\in\mathcal{A}_L and coefficients ck≥0c_k\geq0, let F(A,C;B,D)=Tr⁡exp⁡[A+Θ(B)+∑kckCkΘ(Dk)]\mathcal{F}(A,C;B,D)=\operatorname{Tr}\exp[A+\Theta(B)+\sum_k c_kC_k\Theta(D_k)]. We can expand the exponential as a Dyson series in powers of the crossing terms to express F=∑ακαTr⁡L(Fα)Tr⁡L(Gα)‾\mathcal{F}=\sum_\alpha\kappa_\alpha\operatorname{Tr}_L(F_\alpha)\overline{\operatorname{Tr}_L(G_\alpha)} with non-negative coefficients κα\kappa_\alpha. FαF_\alpha is a product of AA and the CkC_k, and GαG_\alpha is the same product built from BB and the DkD_k. The Cauchy–Schwarz inequality then gives:

∣F(A,C;B,D)∣2≤F(A,C;A,C)F(B,D;B,D).(30)|\mathcal{F}(A,C;B,D)|^2\leq\mathcal{F}(A,C;A,C)\mathcal{F}(B,D;B,D). \tag*{(30)}

For Z(h)Z(h), AA is −β-\beta times the couplings within the left half, B∈ALB\in\mathcal{A}_L is such that Θ(B)\Theta(B) is −β-\beta times the couplings within the right half. A bond (i,r(i))(i,r(i)) crossing the plane, with ii on the left, has g2(ai−bi)2\frac{g}{2}(a_i-b_i)^2 with ai=τi−hia_i=\tau_i-h_i and bi=τi−hr(i)b_i=\tau_i-h_{r(i)}, since the reflection exchanges the sublattices, so Θ(τi)=τr(i)\Theta(\tau_i)=\tau_{r(i)}, and hh is real. Its squares g2ai2\frac{g}{2}a_i^2 and Θ(g2bi2)\Theta(\frac{g}{2}b_i^2) belong to AA and Θ(B)\Theta(B), and its cross term appears in −βH(h)-\beta H(h) as βgaiΘ(bi)\beta g a_i\Theta(b_i), of the form of Eq. (A4) with coefficient βg≥0\beta g\geq0. The other two components of the bond cross as −cOΘ(O)-cO\Theta(O) with c∈{1,Δ}c\in\{1,\Delta\}, as above. Setting (B,D)=(A,C)(B,D)=(A,C) amounts to replacing hh by h(L)h^{(L)}, defined by hi(L)=hr(i)(L)=hih_i^{(L)}=h_{r(i)}^{(L)}=h_i for ii on the left, and similarly for h(R)h^{(R)}. Thus, we get:

Z(h)2≤Z(h(L))Z(h(R)).(31)Z(h)^2\leq Z(h^{(L)})Z(h^{(R)}). \tag*{(31)}

Using translation and rotation symmetry this holds for every mirror plane between adjacent columns or rows. Since constant fields drop out, we may set h=0h=0 at one site; on the remaining fields (h,Kh)(h,Kh) grows like ∣h∣2|h|^2 while ∥(τ,Kh)∥\|(\tau,Kh)\| grows only like ∣h∣|h|, so Z(h)→0Z(h)\to0 as ∣h∣→∞|h|\to\infty and ZZ attains its maximum for some hh. But if hh is a maximizer, h(L)h^{(L)} and h(R)h^{(R)} must be maximizer. By reflecting across different planes in each direction successively eventually leads to the uniform hh configuration [46]. Thus, Z(0)Z(0) must also be a maximizer and Z(h)≤Z(0)Z(h)\leq Z(0) for every real field configuration hh.

Now we replace hh by ϵh\epsilon h. Since f(ϵ)=ln⁡Z(ϵh)−ln⁡Z(0)≤0f(\epsilon)=\ln Z(\epsilon h)-\ln Z(0)\leq0 and f(0)=0f(0)=0, f′(0)=βg⟨(τ,Kh)⟩=0f'(0)=\beta g\langle(\tau,Kh)\rangle=0 and f′′(0)≤0f''(0)\leq0. Because HH and (τ,Kh)(\tau,Kh) do not commute, the second-order term is a Duhamel (Kubo) two-point function:

f′′(0)=β2g2⟨(τ,Kh),(τ,Kh)⟩D−βg(h,Kh)≤0,(32)f''(0)=\beta^2g^2\langle(\tau,Kh),(\tau,Kh)\rangle_{\mathrm{D}}-\beta g(h,Kh)\leq0, \tag*{(32)}

with (U,V)D=Z−1∫01dsTr⁡(e−(1−s)βHU†e−sβHV)(U,V)_{\mathrm{D}}=Z^{-1}\int_0^1\mathrm{d}s\operatorname{Tr}(e^{-(1-s)\beta H}U^\dagger e^{-s\beta H}V). For a momentum k≠0k\ne0 of the torus, the fields hj=cos⁡(k⋅j)h_j=\cos(k\cdot j) and sin⁡(k⋅j)\sin(k\cdot j) satisfy Kh=4(1−γk)hKh=4(1-\gamma_k)h, where γk=γ(k)\gamma_k=\gamma(k). Applying the inequality to both and adding gives

β(τk,τk)D≤14g(1−γk),τk=1∣Λ∣∑jeik⋅jτj=Sk+Q.(33)\beta(\tau_k,\tau_k)_{\mathrm{D}}\leq\frac{1}{4g(1-\gamma_k)},\qquad\tau_k=\frac{1}{\sqrt{|\Lambda|}}\sum_j e^{ik\cdot j}\tau_j=S_{k+Q}. \tag*{(33)}

Finally, let β→∞\beta\to\infty on the fixed torus. The ground state ∣0⟩|0\rangle is unique [44, 51], and ⟨0∣τk∣0⟩=0\langle0|\tau_k|0\rangle=0 by spin-rotation symmetry (a π\pi rotation about ZZ for S=XS=X, the spin flip for S=ZS=Z). Hence β(τk,τk)D\beta(\tau_k,\tau_k)_{\mathrm{D}} tends to the static susceptibility χS(q)=∑n≠0(∣⟨n∣Sq∣0⟩∣2+∣⟨n∣Sq†∣0⟩∣2)/(En−E0)\chi_S(q)=\sum_{n\ne0}(|\langle n|S_q|0\rangle|^2+|\langle n|S^\dagger_q|0\rangle|^2)/(E_n-E_0) at q=k+Q\mathbf{q}=\mathbf{k}+\mathbf{Q}. Since γk=−γq\gamma_{\mathbf{k}}=-\gamma_{\mathbf{q}},

χX(q)≤14(1+γq),χZ(q)≤14Δ(1+γq),q≠Q,(34)\chi_X(\mathbf{q}) \le\frac{1}{4(1+\gamma_{\mathbf{q}})},\qquad\chi_Z(\mathbf{q}) \le\frac{1}{4\Delta(1+\gamma_{\mathbf{q}})},\qquad\mathbf{q}\ne\mathbf{Q}, \tag*{(34)}

for all Δ≥0\Delta\ge0 (Δ>0\Delta>0 for χZ\chi_Z). We use the first bound for Δ≤1\Delta\le1 and the second for Δ≥1\Delta\ge1.

Optimizing the weights—For fixed cap polynomials PP, the bound min⁡y∈KFw,P(y)\min_{\mathbf{y}\in\mathcal{K}}F_{w,P}(\mathbf{y}) is not linear in the weight coefficients a=(a0,an)\mathbf{a}=(a_0,a_n), because of the positive part {w(q)}+\{w(\mathbf{q})\}_+. We replace {w(q)}+\{w(\mathbf{q})\}_+ on the momentum grid by an auxiliary variable p(q)p(\mathbf{q}) with p(q)≥w(q)p(\mathbf{q})\ge w(\mathbf{q}) and p(q)≥0p(\mathbf{q})\ge0. Since S(q)≥0S(\mathbf{q})\ge0 and S(q)≤fP(q)S(\mathbf{q})\le f_P(\mathbf{q}), any such pp still gives a valid bound, and the tightest choice is p={w}+p=\{w\}_+. The objective is then linear in (a,p)(\mathbf{a},p) for fixed y\mathbf{y}. Writing the inner minimization through its dual certificate, Eq. (35), every coefficient of that identity is linear in (a,p)(\mathbf{a},p) and in the dual variables, so maximizing over all of them together, subject to w(Q)=1w(\mathbf{Q})=1, is a single semidefinite program. Near Q\mathbf{Q}, where we keep the square root of the KLS cap, each grid point adds one second-order cone constraint. Because this program uses a quadrature grid, we use its optimum only to propose weights; the reported bound comes from the final program at fixed ww and PP, with the rigorous momentum integrals described below.

Rigorously bounding the momentum integrals—For a fixed ww and PP, Fw,PF_{w,P} is affine in y\mathbf{y}, with Brillouin zone integrals appearing in the constant and the coefficient for each correlator. For instance, the coefficient of G2p+1(r)G_{2p+1}(\mathbf{r}) is −∫d2q(2π)2{w(q)}+α2p+1(q)γr(q)-\int\frac{\mathrm{d}^2\mathbf{q}}{(2\pi)^2}\{w(\mathbf{q})\}_+\alpha_{2p+1}(\mathbf{q})\gamma_{\mathbf{r}}(\mathbf{q}).

During the optimization these integrals are evaluated by quadrature on a momentum grid. However, for the final bound this is not enough. Quadrature is inaccurate near Q\mathbf{Q}, where χˉ\bar{\chi} diverges. More importantly, to turn a bound for the quadrature functional into one for the true functional we must bound their difference for every feasible y\mathbf{y}. Using only the norm bounds ∣⟨P⟩∣≤1|\langle P\rangle|\le1 this is expensive: the coefficients of m5m_5 multiply long combinations of Pauli expectation values (the expansion of a5dHX0a_5^dH_{X_0}), so errors of the order of 10−910^{-9} in them cost ∼10−4\sim10^{-4} in the bound. We therefore build the final functional directly from rigorous enclosures of the integrals.

First, the cap must hold at every momentum, not only at grid points. On each cell where the higher-moment cap is used, the cell’s polynomial is applied at every q\mathbf{q} in the cell. This is valid because the polynomial is nonnegative on [0,∞)[0,\infty) and χˉ(q)\bar{\chi}(\mathbf{q}) and m2p+1(q)m_{2p+1}(\mathbf{q}) are evaluated at the actual q\mathbf{q}, not just at the cell’s representative. Near Q\mathbf{Q} we use the Cauchy–Schwarz cap, Eq. (8), with a tangent point that varies with q\mathbf{q} so that the 1/∣q−Q∣1/|\mathbf{q}-\mathbf{Q}| singularity stays integrable.

Second, each integral is enclosed in an interval. On cells where the integrand is analytic we use tensor Gauss–Legendre quadrature with a rigorous remainder bound from Cauchy estimates. Cells on which ww changes sign are bisected repeatedly and bounded crudely at the finest level; they dominate the final width. All arithmetic is rounded outward, and every value of sin⁡\sin and cos⁡\cos used, about 101010^{10} per functional, is checked against correctly rounded multiple-precision evaluation (MPFR), with no failures.

Finally, the coefficients of the functional are set to the midpoints of their intervals, which are exact binary fractions. The half-widths are absorbed through the norm bounds ∣⟨P⟩∣≤1|\langle P\rangle|\le1 on the correlators they multiply. For the functionals behind our certified bounds, this enclosure loss, which we denote ℓ\ell, is 55–7×10−107\times10^{-10}, and the enclosure takes minutes on a single node.

Dependence on the bootstrap program—Figure 3 separates the three choices that enter the bootstrap program at Δ=1\Delta=1: the operator set, the weights ww, and the cap. We built the operator sets by enlarging smaller ones in several independent directions, each aimed at the correlations entering Eq. (16). A larger set therefore need not always contain a smaller one, and the bound is not monotonic in dim⁡y\dim y.

Figure 3. Variation in lower bounds with the bootstrap program. Exact lower bounds on m2Zm_2^Z at Δ=1\Delta=1 against the number of free moments dim⁡y\dim y of the operator set, all with the energy ceiling −0.668047-0.668047. Red circles show the results after optimizing weights and using the J=3J=3 cap. We also plot the respective bounds for a fixed set of weights, with the KLS cap (purple triangles) or the J=3J=3 cap (teal squares). The red points need not increase with dim⁡y\dim y, because a larger set need not contain a smaller one.

Enlarging the operator set alone improves the bound only slowly, by about 3×10−33\times10^{-3} over the full range shown. With fixed weights and the KLS cap it stays near −9×10−3-9\times10^{-3}. Replacing the KLS cap by the J=3J=3 cap raises the bound by about 5×10−35\times10^{-3} for every set considered. Optimizing the weights together with the cap polynomials adds a further 11–4×10−34\times10^{-3}, growing with the size of the set, since larger sets bound the longer-range correlations that the optimized weights use. This last step is what makes the bound positive.

Exactly certifying the bootstrap bound—The semidefinite program is solved using MOSEK [66] which only finds an approximate optimum up to numerical tolerances. A rigorous bound F≥LF \ge L on K\mathcal{K} requires a certificate of the form

F(y)−L=∑jtr⁡[ZjAj(y)]+∑kμkgk(y)+∑aνaea(y),(35)F(\mathbf{y})-L=\sum_j \operatorname{tr}[Z_j A_j(\mathbf{y})]+\sum_k \mu_k g_k(\mathbf{y})+\sum_a \nu_a e_a(\mathbf{y}), \tag*{(35)}

holding identically in y\mathbf{y}, i.e. coefficient by coefficient, with Zj⪰0Z_j \succeq0 and μk≥0\mu_k \ge0. Aj(y)A_j(\mathbf{y}) are the positivity, perturbative-positivity and reflection-positivity matrices, which depend affinely on y\mathbf{y} and are required to be positive semidefinite; gk(y)≥0g_k(\mathbf{y}) \ge0 are the scalar inequalities (Marshall sign rule, norm bounds, energy ceiling); and ea(y)=0e_a(\mathbf{y})=0 are the equalities (normalization, eigenstate conditions, symmetries). The first two sums are nonnegative and the last one vanishes on the feasible set, which certifies F≥LF \ge L.

We note that exact certificates of SDPs have been obtained before using a variety of techniques [34, 67, 68, 69, 70]. Here, we use a simple strategy that exploits one feature of our SDP: every variable yαy_\alpha is the expectation value of a Pauli string, so ∣yα∣≤1\lvert y_\alpha\rvert\le1, and any small residual can therefore be absorbed exactly.

The entire construction of the SDP, including the elimination of redundant moments and the block diagonalization of the Aj(y)A_j(\mathbf{y}) by symmetries, is carried out in exact rational arithmetic; the solver works with a floating-point copy of these data. Together with the binary-fraction coefficients of FF (see Rigorously bounding the momentum integrals) and the energy ceiling entered as an exact decimal, every coefficient in Eq. (35) is an exact rational number.

We factor each approximate Zj≈BjBjTZ_j \approx B_j B_j^{T} by setting any slightly negative eigenvalues to 0, round Bj→Bj′B_j \to B'_j to binary fractions, and set Zj=Bj′Bj′TZ_j=B'_j B_j^{\prime T}, which is positive semidefinite by construction irrespective of its rank. The multipliers μk\mu_k and νa\nu_a are rounded in the same way, with negative μk\mu_k set to zero. After this rounding, Eq. (35) fails only by an affine residual r0+∑αrαyαr_0+\sum_\alpha r_\alpha y_\alpha, which we compute exactly. The residual is near machine precision and is removed by absorbing r0r_0 into LL and each rαyαr_\alpha y_\alpha into the norm-bound terms ∣rα∣(1±yα)\lvert r_\alpha\rvert(1\pm y_\alpha) already present in Eq. (35). This gives an exact certificate with a negligible loss, Lexact=L+r0−∑α∣rα∣L_{\mathrm{exact}}=L+r_0-\sum_\alpha\lvert r_\alpha\rvert, without another linear solve or a strictly feasible point. For our problems, with up to 1.2×1051.2\times10^5 coordinates, this takes minutes on a single node and lowers LL by 10−1310^{-13} to 10−1210^{-12}. A separate program then re-reads the exact data and the certificate and re-checks every coefficient of Eq. (35) in rational arithmetic.

Combining this with the enclosure loss ℓ\ell of the momentum integrals, the ground-state moments y⋆∈K\mathbf{y}^\star\in\mathcal{K} give w(Q)mx2≥Ftrue(y⋆)≥Lexact−ℓw(\mathbf{Q})m_x^2\ge F_{\mathrm{true}}(\mathbf{y}^\star)\ge L_{\mathrm{exact}}-\ell, where FtrueF_{\mathrm{true}} is the functional with the exact integrals and w(Q)w(\mathbf{Q}) is an exact rational within 10−1010^{-10} of 1 after rounding. The resulting lower bound (Lexact−ℓ)/w(Q)(L_{\mathrm{exact}}-\ell)/w(\mathbf{Q}) on mx2m_x^2 is an exactly certified rational number.

Extending the lower bound to an interval around Δ\Delta—Since each certificate at a given Δ\Delta is an exact algebraic identity in which Δ\Delta enters only through explicit rational functions (the stationarity rows, the double commutator and, for Δ≥1\Delta\ge1, the infrared bound), the same Gram matrices and multipliers remain a valid certificate at a nearby Δ′\Delta', up to a residual that is an explicit rational function of Δ′\Delta'. Bounding that residual with the norm constraints ∣yα∣≤1\lvert y_\alpha\rvert\le1 gives L(Δ′)≥L(Δ)−K∣Δ′−Δ∣L(\Delta')\ge L(\Delta)-K\lvert\Delta'-\Delta\rvert with an explicit constant KK, so a single positive certificate in principle establishes order on the whole interval ∣Δ′−Δ∣<L(Δ)/K\lvert\Delta'-\Delta\rvert<L(\Delta)/K. In practice, our certificates have large KK, leading to narrow intervals of width ∼0.002\sim0.002, which would require a denser sampling of the Δ\Delta grid to certify order for the entire range. In this work, we only report bounds at the computed values of Δ\Delta.

Certifiable strip construction—The purpose of this step is to turn the numerically optimized strip MPS into an exact variational quantum state. We choose a sequential ordering of the spins along the strip. At step ii, the real MPS tensor ViV_i takes an incoming auxiliary state, emits physical spin ii, and produces the outgoing auxiliary degree of freedom. For an exactly normalized MPS tensor one would have ViTVi=IV_i^T V_i=I. The finite-precision tensor is only approximately isometric, so we rigorously certify ∥ViTVi−I∥op≤βi\lVert V_i^T V_i-I\rVert_{\mathrm{op}}\le\beta_i. We choose ai=(1+βi)−1a_i=(1+\beta_i)^{-1}, which guarantees aiViTVi≤Ia_iV_i^T V_i\le I, and define Di=I−aiViTVi⪰0D_i=I-a_iV_i^T V_i\succeq0. The certified one-site preparation map is

Fi(σi)=aiViσiViT+Tr⁡(Diσi)∣ri⟩⟨ri∣,(36)\mathcal{F}_i(\sigma_i)=a_iV_i\sigma_iV_i^T+\operatorname{Tr}(D_i\sigma_i)\lvert r_i\rangle\langle r_i\rvert, \tag*{(36)}

where σi\sigma_i is the incoming auxiliary density matrix and ∣ri⟩\lvert r_i\rangle is a fixed normalized state of the emitted spin and outgoing auxiliary space. The first term follows the optimized MPS tensor, while the second restores the probability removed by the rescaling. Since Di⪰0D_i\succeq0 and aiViTVi+Di=Ia_iV_i^T V_i+D_i=I, Fi\mathcal{F}_i is completely positive and trace preserving (CPTP). The outgoing auxiliary state is σi+1=Tr⁡iFi(σi)\sigma_{i+1}=\operatorname{Tr}_i\mathcal{F}_i(\sigma_i), where the trace is over the emitted spin. Repeated application of the Fi\mathcal{F}_i, with each subsequent map acting only on the outgoing auxiliary degree of freedom, therefore defines an exact positive normalized many-body state on the strip. No fixed-point assumption is required.

Seam circuit—A product of certified strips is already a valid two-dimensional variational state, but it contains no entanglement across neighboring strips. We restore these correlations with a finite-depth unitary circuit across each seam. The simplest circuit in Figure 2(b) applies one two-spin gate to each bond crossing a seam. Reality and conservation of total SzS^z leave the generator A=θ(Si+Sj−−Si−Sj+)A=\theta(S_i^+S_j^- - S_i^-S_j^+), where S±=X±iYS^\pm=X\pm iY. We parameterize the gate by

U(A)=(I+A)−1(I−A).(37)U(A) = (I + A)^{-1}(I - A). \tag*{(37)}

Since AA is real and antisymmetric, U(A)U(A) is exactly orthogonal. Since AA preserves total SzS^z, so does the gate. The full circuit in Figure 2(a) uses the same construction with larger gates. A four-spin gate RR acts on a 2×22 \times2 region crossing the seam, followed by an eight-spin gate QQ on a 2×42 \times4 region, with the pattern repeated every two columns. Their generators are real, antisymmetric, and block diagonal in total SzS^z. The resulting circuit therefore preserves the U(1)U(1) symmetry of the strip source. Because the circuit has finite depth, propagating any bond Hamiltonian backward through it produces a finite causal cone involving at most two neighboring strips.

*Thermodynamic energy certificate—*The remaining issue is that the numerical infinite-MPS environment need not be an exact fixed point of the certified strip channel. Let TT be the auxiliary CPTP channel obtained by composing the certified preparation maps over one two-column cell. We define an exact reference auxiliary state directly from the stored Schmidt coefficients,

σ0=diag⁡(Si2)∑jSj2.(38)\sigma_0 = \frac{\operatorname{diag}(S_i^2)}{\sum_j S_j^2}. \tag*{(38)}

The stored SiS_i are treated as exact dyadic numbers, so σ0⪰0\sigma_0 \succeq0 and Tr⁡σ0=1\operatorname{Tr}\sigma_0 = 1 hold algebraically. To certify stationarity without subtracting two nearly equal O(1)O(1) density matrices, we propagate the stationarity defect itself. Define

D0=T(σ0)−σ0,Dk=Tk(D0).(39)D_0 = T(\sigma_0) - \sigma_0,\qquad D_k = T^k(D_0). \tag*{(39)}

Then

Dk=Tk+1(σ0)−Tk(σ0).D_k = T^{k+1}(\sigma_0) - T^k(\sigma_0).

For a chosen burn-in depth nbn_b, we set σ=Tnb(σ0)\sigma= T^{n_b}(\sigma_0), so

T(σ)−σ=Dnb,σ−σ0=∑k=0nb−1Dk.(40)T(\sigma) - \sigma= D_{n_b},\qquad\sigma- \sigma_0 = \sum_{k=0}^{n_b-1} D_k. \tag*{(40)}

Verified propagation therefore gives

δ≥∥Dnb∥1,d0≥∑k=0nb−1∥Dk∥1≥∥σ−σ0∥1.(41)\delta\ge\lVert D_{n_b}\rVert_1,\qquad d_0 \ge\sum_{k=0}^{n_b-1}\lVert D_k\rVert_1 \ge\lVert\sigma- \sigma_0\rVert_1. \tag*{(41)}

The small defect DkD_k, rather than an O(1)O(1) density matrix, is propagated in round-to-nearest binary64 with explicit roundoff majorants, and all resulting bounds are enclosed outward using Arb ball arithmetic. Thus the certified residual is not limited by the O(10−16)O(10^{-16}) absolute scale associated with subtracting two normalized density matrices. For any nonnegative integer kk, trace-norm contractivity of the CPTP channel and a telescoping sum give

∥Tk(σ)−σ∥1≤kδ.(42)\lVert T^k(\sigma) - \sigma\rVert_1 \le k\delta. \tag*{(42)}

The exact variational state used for the upper bound is obtained by initializing each strip patch with σ\sigma, applying the certified channel for LL two-column cells, periodically repeating the resulting patches, and then applying the seam circuit. This is an exact positive normalized many-body state for every finite LL. The patch can be extremely long, but Eq. (42) allows its bulk energy to be bounded without contracting all LL cells individually.

We evaluate the reference energy using the exact diagonal state σ0\sigma_0. For every physical bond in one two-column circuit unit cell, the bond Hamiltonian is propagated backward through the seam circuit to its finite causal cone and contracted using σ0\sigma_0 and the certified strip maps. Every physical bond is included exactly once. Each bond expectation value is evaluated with rigorous floating-point error bounds, and the resulting intervals are accumulated using outward-rounded arithmetic. Normalizing the total bond-energy upper bound by the 2W2W sites in the unit cell defines the certified reference energy density erefe_{\mathrm{ref}}.

It remains to compare this reference contraction with the exact length-LL patch. For one strip, an auxiliary state entering a bulk causal cone after kk cells differs from σ0\sigma_0 by at most

∥Tk(σ)−σ0∥1≤d0+kδ.(43)\lVert T^k(\sigma) - \sigma_0\rVert_1 \le d_0 + k\delta. \tag*{(43)}

Every bond causal cone involves at most two strips, so the standard tensor-product trace-distance bound gives an additional factor of two. Longitudinally, the R+QR+Q causal cone extends at most two two-column cells beyond the cell containing the bond. Including one additional cell for full-cell coverage, together with the anchoring and closing applications of TT, every auxiliary state entering a bulk causal cone occurs within at most L+5L+5 applications of TT. The source density matrix of any bulk causal cone therefore differs from the reference source by at most

2[d0+(L+5)δ](44)2\left[d_0 + (L+5)\delta\right] \tag*{(44)}

in trace norm. For a spin-1/2 Heisenberg bond h=Si⋅Sjh = \mathbf{S}_i \cdot\mathbf{S}_j, the spectrum is {1/4,−3/4}\{1/4,-3/4\}, so ∥h∥op=3/4\lVert h\rVert_{\mathrm{op}} = 3/4. Unitary evolution through the seam circuit preserves trace distance, partial trace cannot increase it, and Hölder’s inequality gives

∣Tr⁡(h δρ)∣≤∥h∥op∥δρ∥1.|\operatorname{Tr}(h\,\delta\rho)| \le\lVert h\rVert_{\mathrm{op}}\lVert\delta\rho\rVert_1.

Since there are two bonds per site on the square lattice, Eq. (44) changes the bulk energy density by at most

3[d0+(L+5)δ].(45)3\left[d_0 + (L+5)\delta\right]. \tag*{(45)}

The bulk comparison excludes at most four cells at each end of a patch, giving an uncontrolled fraction 8/L8/L. The Heisenberg bond spectrum bounds the energy per site to [−3/2,1/2][-3/2,1/2], whose width is 2, so these boundary regions contribute at most 16/L16/L. Applying the variational principle to the exact periodic patch state therefore gives

ϵ0≤eref+3d0+3(L+5)δ+16L.(46)\epsilon_0 \le e_{\rm ref} + 3d_0 + 3(L+5)\delta+ \frac{16}{L}. \tag*{(46)}

Thus erefe_{\rm ref} is evaluated entirely at the exact reference state σ0\sigma_0, d0d_0 connects this reference contraction to the propagated state σ\sigma used to initialize the patch, and δ\delta controls the subsequent drift across the patch.

After erefe_{\rm ref}, δ\delta, and d0d_0 have been certified, LL is chosen to minimize Eq. (46). Short patches incur a boundary cost proportional to 1/L1/L, while long patches accumulate a stationarity error proportional to LδL\delta. Minimizing the LL-dependent terms gives

L⋆=163δ.(47)L^\star= \sqrt{\frac{16}{3\delta}}. \tag*{(47)}

For the isotropic certificate used in Eq. (25), nb=11n_b=11 gives

δ≤1.684×10−19,d0≤9.01×10−12.(48)\delta\le1.684 \times10^{-19}, \qquad d_0 \le9.01 \times10^{-12}. \tag*{(48)}

The corresponding optimal patch length is L≃5.63×109L \simeq5.63 \times10^9 cells. The patch correction,

3(L+5)δ+16L,3(L+5)\delta+ \frac{16}{L},

is at most 5.69×10−95.69 \times10^{-9}, while the reference-state shift contributes at most 3d0≤2.71×10−113d_0 \le2.71 \times10^{-11}. The total correction in Eq. (46) is therefore at most 5.72×10−95.72 \times10^{-9}. Related bootstrap approaches instead bound tensor-network contractions by constraining the space of admissible MPS environments [71]. For the XXZ calculations, the corresponding norm allowances are multiplied by the conservative factor max⁡(1,Δ)\max(1,\Delta) used in the certificate.

Easy-plane mapping—The strip source and seam circuit preserve total SzS^z, which is naturally suited to the easy-axis regime. For Δ<1\Delta< 1, rather than introduce a separate symmetry-breaking variational family, we map the easy-plane energy problem exactly to an easy-axis one. Define the global π/2\pi/2 rotations Rx=exp⁡[−i(π/2)∑jXj]R_x=\exp[-i(\pi/2)\sum_j X_j] and Ry=exp⁡[−i(π/2)∑jYj]R_y=\exp[-i(\pi/2)\sum_j Y_j]. They obey

12[Rx†H(Δ)Rx+Ry†H(Δ)Ry]=1+Δ2H(21+Δ).(49)\frac{1}{2}\left[R_x^\dagger H(\Delta)R_x + R_y^\dagger H(\Delta)R_y\right] = \frac{1+\Delta}{2}H\left(\frac{2}{1+\Delta}\right). \tag*{(49)}

For Δ<1\Delta< 1, the transformed coupling Δ′=2/(1+Δ)\Delta' = 2/(1+\Delta) lies in the easy-axis regime. If ρ′\rho' is any certified variational state for H(Δ′)H(\Delta'), then

ρ=12(Rxρ′Rx†+Ryρ′Ry†)(50)\rho= \frac{1}{2}\left(R_x\rho'R_x^\dagger+ R_y\rho'R_y^\dagger\right) \tag*{(50)}

is again a positive normalized state, and Eq. (49) gives

Tr⁡[ρH(Δ)]=1+Δ2Tr⁡[ρ′H(Δ′)].(51)\operatorname{Tr}[\rho H(\Delta)] = \frac{1+\Delta}{2}\operatorname{Tr}[\rho' H(\Delta')]. \tag*{(51)}

The variational principle therefore gives

ϵ0(Δ)≤1+Δ2 ϵvar(Δ′),(52)\epsilon_0(\Delta) \le\frac{1+\Delta}{2}\,\epsilon_{\rm var}(\Delta'), \tag*{(52)}

which gives the easy-plane bounds shown in Figure 2(c).

References

  1. [1]H. Tasaki, Physics and Mathematics of Quantum Many-Body Systems, Graduate Texts in Physics (Springer, Cham, 2020).DOI
  2. [2]A. W. Sandvik, Computational studies of quantum spin systems, AIP Conf. Proc. 1297, 135 (2010).DOI
  3. [3]J. Jordan, R. Orús, G. Vidal, F. Verstraete, and J. I. Cirac, Classical simulation of infinite-size quantum lattice systems in two spatial dimensions, Phys. Rev. Lett. 101, 250602 (2008).pubmed.ncbi.nlm.nih.gov/19113687
  4. [4]R. Orús, A practical introduction to tensor networks: Matrix product states and projected entangled pair states, Ann. Phys. 349, 117 (2014), arXiv:1306.2164.DOI
  5. [5]F. J. Dyson, E. H. Lieb, and B. Simon, Phase transitions in quantum spin systems with isotropic and nonisotropic interactions, J. Stat. Phys. 18, 335 (1978).
  6. [6]E. J. Neves and J. F. Perez, Long range order in the ground state of two-dimensional antiferromagnets, Phys. Lett. A 114, 331 (1986).DOI
  7. [7]T. Kennedy, E. H. Lieb, and B. S. Shastry, Existence of Néel order in some spin-1/2 Heisenberg antiferromagnets, J. Stat. Phys. 53, 1019 (1988).DOI
  8. [8]K. Kubo and T. Kishi, Existence of long-range order in the XYZ model, Phys. Rev. Lett. 61, 2585 (1988)
  9. [9]T. Kennedy, E. H. Lieb, and B. S. Shastry, The XY model has long-range order for all spins and all dimensions greater than one, Phys. Rev. Lett. 61, 2582 (1988).pubmed.ncbi.nlm.nih.gov/10039162
  10. [10]H. Nishimori and Y. Ozeki, Ground-state long-range order in the two-dimensional XXZ model, J. Phys. Soc. Jpn. 58, 1027 (1989).
  11. [11]Y. Ozeki, H. Nishimori, and Y. Tomita, Long-range order in antiferromagnetic quantum spin systems, J. Phys. Soc. Jpn. 58, 82 (1989).
  12. [12]H.-A. Wischmann and E. Müller-Hartmann, Extended proof of long-range order in the two-dimensional quantum spin-1/2 XXZ-model at T = 0, J. Phys. I France 1, 647 (1991).DOI
  13. [13]W.-L. You, Long-range order in two-dimensional XXZ model, Int. J. Mod. Phys. B 23, 2195 (2009).DOI
  14. [14]E. H. Lieb, Long range order for the quantum Heisenberg model (1999), open Problems in Mathematical Physics, problem B.
  15. [15]U. Löw, Néel order in the two-dimensional S = 1/2 Heisenberg model, Phys. Rev. B 76, 220409(R) (2007).DOI
  16. [16]M. Navascués, S. Pironio, and A. Acín, Bounding the set of quantum correlations, Phys. Rev. Lett. 98, 010401 (2007).pubmed.ncbi.nlm.nih.gov/17358458
  17. [17]M. Navascués, S. Pironio, and A. Acín, A convergent hierarchy of semidefinite programs characterizing the set of quantum correlations, New J. Phys. 10, 073013 (2008).DOI
  18. [18]T. Barthel and R. Hüebener, Solving condensed-matter ground-state problems by semidefinite relaxations, Phys. Rev. Lett. 108, 200404 (2012).pubmed.ncbi.nlm.nih.gov/23003130
  19. [19]T. Baumgratz and M. B. Plenio, Lower bounds for ground states of condensed matter systems, New J. Phys. 14, 023027 (2012).DOI
  20. [20]X. Han, S. A. Hartnoll, and J. Kruthoff, Bootstrapping matrix quantum mechanics, Phys. Rev. Lett. 125, 041601 (2020).
  21. [21]H. W. Lin, Bootstraps to strings: solving random matrix models with positivity, J. High Energy Phys. 2020 (6), 90.DOI
  22. [22]M. Cho, B. Gabai, Y.-H. Lin, V. A. Rodriguez, J. Sandor, and X. Yin, Bootstrapping the Ising model on the lattice (2022), arXiv:2206.12538 [hep-th].arxiv.org/abs/2206.12538
  23. [23]M. B. Hastings, Perturbation theory and the sum of squares (2022), arXiv:2205.12325.
  24. [24]V. Kazakov and Z. Zheng, Bootstrap for lattice Yang-Mills theory, Phys. Rev. D 107, L051501 (2023).DOI
  25. [25]D. A. Mazziotti, Quantum many-body theory from a solution of the N-representability problem, Phys. Rev. Lett. 130, 153001 (2023).DOI
  26. [26]D. Berenstein, G. Hulsey, and P. N. T. Lloyd, Numerical exploration of the bootstrap in spin chain models (2024), arXiv:2406.17844 [hep-th].arxiv.org/abs/2406.17844
  27. [27]J. Wang, J. Sura, I. Frérot, B. Legat, M.-O. Renou, V. Magron, and A. Acín, Certifying ground-state properties of many-body systems, Phys. Rev. X 14, 031006 (2024).DOI
  28. [28]H. Fawzi, O. Fawzi, and S. O. Scalet, Certified algorithms for equilibrium states of local quantum Hamiltonians, Nat. Commun. 15, 7394 (2024).
  29. [29]Q. Gao, R. A. Lanzetta, P. Ledwith, J. Wang, and E. Khalaf, Bootstrapping the quantum Hall problem, Phys. Rev. X 15, 031034 (2025).DOI
  30. [30]M. Cho, B. Gabai, J. Sandor, and X. Yin, Thermal bootstrap of matrix quantum mechanics, J. High Energy Phys. 04, 186, arXiv:2410.04262 [hep-th].arxiv.org/abs/2410.04262
  31. [31]A. O. Schouten, S. Ewing, and D. A. Mazziotti, Bootstrapping the electronic structure of quantum materials (2025), arXiv:2504.02861 [cond-mat.str-el].arxiv.org/abs/2504.02861
  32. [32]M. G. Scheer, N. Chadha, D.-C. Lu, and E. Khalaf, Defect bootstrap: Tight ground state bounds in spontaneous symmetry breaking phases (2025), arXiv:2511.20860 [cond-mat.str-el].arxiv.org/abs/2511.20860
  33. [33]M. G. Scheer, Bootstrapping quantum Hamiltonians with symmetry, Phys. Rev. Lett. 136, 246501 (2026).DOI
  34. [34]M. G. Scheer and E. Khalaf, A hierarchy of exact lower bounds on the Laughlin spectral gap (2026), arXiv:2610.18854.arxiv.org/abs/2610.18854
  35. [35]M. Cho, C. O. Nancarrow, P. Tadić, Y. Xin, and Z. Zheng, Coarse-grained bootstrap of quantum many-body systems, J. High Energy Phys. 02, 222, arXiv:2412.07837 [hep-th].arxiv.org/abs/2412.07837
  36. [36]M. Cho, B. Gabai, H. W. Lin, J. Yeh, and Z. Zheng, Bootstrapping Euclidean two-point correlators, J. High Energy Phys. 08, 147, arXiv:2511.08560 [hep-th].arxiv.org/abs/2511.08560
  37. [37]Q. Gao, Z. Han, and E. Khalaf, Bootstrapping flatband superconductors: Rigorous lower bounds on superfluid stiffness, Phys. Rev. Lett. 136, 076505 (2026).DOI
  38. [38]N. Paul and G. Refuel, Bootstrapping ground state properties of classical frustrated magnets (2026), arXiv:2605.06784 [cond-mat.stat-mech].arxiv.org/abs/2605.06784
  39. [39]N. Chadha, M. G. Scheer, and E. Khalaf, Bootstrap bounds for quantum spin systems using string operators (2026), arXiv:2606.06584 [cond-mat.str-el].arxiv.org/abs/2606.06584
  40. [40]Y. Önder, M. G. Scheer, M. Cho, and E. Khalaf, Bootstrapping disordered quantum systems (2026), arXiv:2609.20916.arxiv.org/abs/2609.20916
  41. [41]M. Cho, Y. Önder, E. Khalaf, and M. G. Scheer, Bootstrapping classical systems with quenched disorder (2026), arXiv:2609.20927.arxiv.org/abs/2609.20927
  42. [42]T. Koma and H. Tasaki, Symmetry breaking in Heisenberg antiferromagnets, Commun. Math. Phys. 158, 191 (1993).DOI
  43. [43]J. Fröhlich, B. Simon, and T. Spencer, Infrared bounds, phase transitions and continuous symmetry breaking, Commun. Math. Phys. 50, 79 (1976).DOI
  44. [44]E. Lieb and D. Mattis, Ordering energy levels of interacting spin systems, J. Math. Phys. 3, 749 (1962).
  45. [45]N. D. Mermin and H. Wagner, Absence of ferromagnetism or antiferromagnetism in one- or two-dimensional isotropic Heisenberg models, Phys. Rev. Lett. 17, 1133 (1966).
  46. [46]J. E. Björnberg and D. Ueltschi, Reflection positivity and infrared bounds for quantum spin systems, in The Physics and Mathematics of Elliott Lieb, Vol. I (EMS Press, 2022) pp. 77–108, arXiv:2204.12896.arxiv.org/abs/2204.12896
  47. [47]M. G. Krein and A. A. Nudel’man, The Markov Moment Problem and Extremal Problems, Translations of Mathematical Monographs, Vol. 50 (American Mathematical Society, Providence, RI, 1977).
  48. [48]K. Schmüdgen, The Moment Problem, Graduate Texts in Mathematics, Vol. 277 (Springer, Cham, 2017).DOI
  49. [49]H. Nishimori, K. Kubo, Y. Ozeki, Y. Tomita, and T. Kishi, Long-range order in the XXZ model, J. Stat. Phys. 55, 259 (1989).
  50. [50]H. Tasaki, Long-range order, “tower” of states, and symmetry breaking in lattice quantum systems, J. Stat. Phys. 174, 735 (2019), arXiv:1807.05847.DOI
  51. [51]W. Marshall, Antiferromagnetism, Proc. R. Soc. Lond. A 232, 48 (1955).
  52. [52]A. A. Ovchinnikov, Variational bound for the energy of two-dimensional quantum antiferromagnet, Pis’ma Zh. Eksp. Teor. Fiz. 57, 796 (1993), arXiv:cond-mat/9307067.DOI
  53. [53]A. W. Sandvik, High-precision ground state parameters of the two-dimensional spin-1/2 Heisenberg model on the square lattice, J. Stat. Mech. 2026, 043101 (2026), arXiv:2601.20189.DOI
  54. [54]N. Caci, D.-B. Hering, M. R. Walther, K. P. Schmidt, S. Wessel, and G. S. Uhrig, Quantitative description of long-range order in the XXZ antiferromagnet on the square lattice, Phys. Rev. B 110, 054411 (2024).
  55. [55]R. F. Bishop, P. H. Y. Li, R. Zinke, R. Darradi, J. Richter, D. J. J. Farnell, and J. Schulenburg, The spin-half XXZ antiferromagnet on the square lattice revisited: A high-order coupled cluster treatment, J. Magn. Magn. Mater. 428, 178 (2017), arXiv:1608.06157.DOI
  56. [56]Z. Liu and E. Manousakis, Variational calculations for the square-lattice quantum antiferromagnet, Phys. Rev. B 40, 11437 (1989).pubmed.ncbi.nlm.nih.gov/9991739
  57. [57]P. Corboz, Variational optimization with infinite projected entangled-pair states, Phys. Rev. B 94, 035133 (2016), arXiv:1605.03006.DOI
  58. [58]T. Oguchi, A theory of antiferromagnetism, II, Prog. Theor. Phys. 13, 148 (1955).DOI
  59. [59]G. J. Mata and G. B. Arnold, Short-range correlations in magnetic systems: Effective-cluster approach, Phys. Rev. B 38, 11582 (1988).pubmed.ncbi.nlm.nih.gov/9946041
  60. [60]D. Yamamoto, Correlated cluster mean-field theory for spin systems, Phys. Rev. B 79, 144427 (2009).DOI
  61. [61]The energy ceilings are computed using simpler certified ceilings: finite patches (8 × 8 for Δ ≤ 1 and 12 × 6 for Δ > 1), with two-site gates across each seam and energies evaluated in exact rational arithmetic.
  62. [62]For any trigonometric polynomial w(q) ≥ 0 with w(Q) = 1, positivity of S(q) gives m²ₓ ≤ ∫ d²q/(2π)² w(q)S(q) = a₀/4 + ∑ₙ aₙCⁿ₂, and maximizing the right-hand side over K bounds m²ₓ from above. With an optimized nonnegative w of range 16 and 4.4 × 10⁴ free moments we find m²ₓ ≤ 0.059 (mₛ ≤ 0.42), a numerically checked bound.
  63. [63]X. C. Song, P. Smith, R. Kalyanam, X. Zhu, E. Adams, K. Colby, P. Finnegan, E. Gough, E. Hillery, R. Irvine, A. Maji, and J. St. John, Anvil – system architecture and experiences from deployment and early user operations, in Proc. PEARC ’22 (ACM, 2022).DOI
  64. [64]T. J. Boerner, S. Deems, T. R. Furlani, S. L. Knuth, and J. Towns, ACCESS: Advancing innovation: NSF’s advanced cyberinfrastructure coordination ecosystem: Services & support, in Proc. PEARC ’23 (ACM, 2023) pp. 173–176.
  65. [65]J. Fröhlich, R. Israel, E. H. Lieb, and B. Simon, Phase transitions and reflection positivity. I. General theory and long range lattice models, Commun. Math. Phys. 62, 1 (1978).
  66. [66]MOSEK ApS, MOSEK Optimizer API for Julia (2025).
  67. [67]H. Peyrl and P. A. Parrilo, Computing sum of squares decompositions with rational coefficients, Theor. Comput. Sci. 409, 269 (2008), symbolic-Numerical Computations.
  68. [68]C. Jansson, D. Chaykin, and C. Keil, Rigorous error bounds for the optimal value in semidefinite programming, SIAM J. Numer. Anal. 46, 180 (2008).DOI
  69. [69]E. L. Kaltofen, B. Li, Z. Yang, and L. Zhi, Exact certification in global polynomial optimization via sums-of-squares of rational functions with rational coefficients, J. Symb. Comput. 47, 1 (2012).DOI
  70. [70]Y. Naceur, J. Wang, V. Magron, and A. Acín, Certified bounds on optimization problems in quantum theory, Quantum 10, 2254 (2026), arXiv:2512.1773.arxiv.org/abs/2512.17713
  71. [71]S. Ono, Y. Zhang, and H. C. Po, Toward bootstrapping tensor-network contractions (2026), arXiv:2603.17856 [cond-mat.str-el].arxiv.org/abs/2603.17856

Paper details

vlnc:W_vhwbhhdj
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 9

downloads · 2

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents