betaFeedback

WorksPhysics and Astronomy

Uniform Spectral Gap and Stability for the Spin-2 Square-Lattice AKLT Model with Free Boundaries

Abstract

We prove a uniform spectral gap above the full ground space of the spin-2 AKLT Hamiltonian on every square-lattice rectangle with four free sides and both lengths at least two. The ground-space dimension grows exponentially with the perimeter. Starting from the classical hard-boundary coupling of Aronow and Lopatto, we compare the boundary Gram operator of one wing with the product of its marginals on the cut side and the remaining three sides. A reference system with at most eight unsaturated sites then controls a physical cut of arbitrary length. Recursion over overlapping rectangles gives the gap. We also obtain a gap on all simple tori, including those of fixed circumference. In free squares, boundary effects on fixed local observables decay exponentially with the buffer width. By the stability theorems of Nachtergaele, Sims and Young, weak quasi-local perturbations with stretched-exponentially decaying interaction terms preserve the bulk GNS gap and the finite-torus gap at bounded aspect ratio.

The Hamiltonian and the boundary problem

Let Λ=[W]×[H]⊂Z2\Lambda= [W] \times[H] \subset\mathbb{Z}^2, where [m]={1,…,m}[m] = \{1,\ldots,m\}. At every site, including every corner, take the irreducible spin-2 space V4=Sym⁡4C2V_4 = \operatorname{Sym}^4 \mathbb{C}^2. For a nearest-neighbour edge ee let pep_e be the orthogonal projection onto total spin 4 at its endpoints. Write

HΛ=∑e⊂Λpe,GΛ=proj⁡ker⁡HΛ.H_\Lambda= \sum_{e \subset\Lambda} p_e, \qquad G_\Lambda= \operatorname{proj} \ker H_\Lambda.

Each internal edge has coefficient one. There are no exterior edges or boundary fields.

Classical coupling hypothesis (H). On each sufficiently large square [−M,M]2[-M,M]^2, consider the classical S2S^2 spin law with bond weight 1+nx⋅ny1+n_x \cdot n_y and a fixed exterior. For any two exteriors differing at one adjacent site there is a coupling for which the expected number of unequal interior spins is at most ΔM\Delta_M, with a common ΔM/(2M+1)→0\Delta_M/(2M+1) \to0 as M→∞M \to\infty.

Theorem 1.1 (Free rectangles). Assume (H). There exists g>0g > 0 such that HΛ≥g(1−GΛ)H_\Lambda\ge g(1-G_\Lambda) for all integers W,H≥2W,H \ge2. The inequality holds on the entire physical Hilbert space, uniformly over all aspect ratios.

Corollary 1.2. For the spin-2 square-lattice AKLT model, the uniform lower bound of Theorem 1.1 holds on every free rectangle of side lengths at least two.

Proof. Proposition 4.4 of Aronow and Lopatto supplies (H) from their window estimate (W), established via Theorems 5.9, 7.4 and 8.1 (assembled in Section 8.4) of [5]. □

We began studying open rectangles in October 2026. While this work was in progress, Aronow and Lopatto established an explicit spectral gap for doubly periodic rectangles when both circumferences are sufficiently large (of either parity), identified the ground vector on every simple torus, and proved uniqueness of the infinite-volume frustration-free state using open-box ground states [5 Theorem 1.1, Corollary 3.3 and Appendix F]. Taking only their classical hard-boundary coupling (H) as a modular input, the present paper establishes four results for the undecorated spin-2 square-lattice model: (i) a uniform spectral gap on every free-boundary rectangle W,H≥2W,H \ge2, uniformly over all aspect ratios (Theorem 1.1); (ii) a uniform gap on all simple tori L,N≥3L,N \ge3, including fixed short circumferences down to three with arbitrary length (Theorem 6.1); (iii) exponential local topological quantum order (LTQO) in operator norm on the entire open-box ground space (Theorem 7.1); and (iv) perturbative stability of both the infinite-volume bulk GNS gap and the bounded-aspect-ratio finite-torus gap under weak quasi-local perturbations with stretched-exponentially decaying interaction terms (Corollaries 8.1 and 8.2).

On an open rectangle, the ground-space dimension grows exponentially with the perimeter as 3422W+2H−83^4 2^{2W+2H-8}. Its boundary vectors need not be products, the physical inner product on them depends on the rectangle, and dividing a block-ground vector by a fixed valence-bond state can leave degrees of freedom at a free edge inside the block. In our gap proof, the classical coupling controls both a bounded, two-sided comparison of the boundary Gram operator for one whole side (Proposition 3.1) and the uniform gap of a pinned reference system with at most eight unsaturated sites (Appendix § 9); it is this separate gapped reference system that turns the boundary comparison into a uniform contraction at a physical seam of arbitrary length (Proposition 4.1). In Section 7, the same coupling on an intact inner rectangle controls the joint classical marginal on the inner neighbour ring and yields exponential LTQO directly.

Related work. Our analysis builds on the AKLT valence-bond framework [1], boundary operators for projected entangled pair states [7], and recursive gap criteria for frustration-free Hamiltonians [6]. Kennedy, Lieb and Tasaki described finite-volume valence-bond ground states and proved exponential decay of two-point correlations on the periodic square lattice [2]. On two-dimensional degree-3 and decorated lattices, spectral gaps were established analytically on decorated hexagonal lattices by Abdul-Rahman, Lemm, Lucia, Nachtergaele and Young [15], by finite-size criteria on the hexagonal and other degree-3 lattices by Lemm, Sandvik and Wang [16] and by Pomata and Wei [17], and via boundary-state factorization on the hexagonal lattice by Kastoryano [11]; Jackson, Nachtergaele and Young proved local indistinguishability for the hexagonal and Lieb AKLT models [10]. For the undecorated square lattice, Guo, Pomata and Wei obtained a weighted finite-size sufficient condition for a spectral gap [3].

Polynomial conventions. Give VdV_d the inner product in which xd−jyjx^{d-j}y^j are orthogonal of squared norm (dj)−1\binom{d}{j}^{-1}. Identifying two spinors gives a map J:Vk⊗Vl→Vk+lJ:V_k\otimes V_l\to V_{k+l} with JJ∗=1JJ^*=1, by Vandermonde’s identity. Thus J∗JJ^*J projects onto the largest total spin and ker⁡J\ker J comprises the multiples of [uv]=xuyv−yuxv[uv]=x_u y_v-y_u x_v (by divisibility on proportional spinors, or by dimension). For the polynomial divisibility argument, see Kirillov and Korepin, Theorem 2 in the arXiv version of [4]. Distinct edge factors are relatively prime. On a box DD, multiplication UDU_D by the product of all internal edge factors maps ⨂xV4−deg⁡D(x)\bigotimes_x V_{4-\deg_D(x)} bijectively onto its ground space. The four corners leave dimension three each, the other boundary sites dimension two each, giving the count above. An edge factor can be inserted between any two residual sites of degree at most three with squared singular values between universal positive constants. Indeed multiplication is injective on each of finitely many local spaces Vk⊗VlV_k\otimes V_l and acts as the identity on all spectator variables. We use the same observation after lowering the output degrees at endpoints of removed edges.

Two comparisons for hard classical boundaries

With unit spins nx∈S2n_x \in S^2 and normalized area measure, the classical Gibbs density in a rectangle DD is proportional to

∏{x,y}⊂D, x∼y(1+nx⋅ny)∏x∈D, y∉D, x∼y(1+nx⋅ηy).\prod_{\{x,y\}\subset D,\ x\sim y}(1+n_x\cdot n_y)\prod_{x\in D,\ y\notin D,\ x\sim y}(1+n_x\cdot\eta_y).

All exterior spins ηy\eta_y are fixed. The partition function is positive since each factor is positive almost everywhere. Throughout the following reduction we assume (H). Fix an odd bb such that Δ(ℓ−1)/2≤b/96\Delta_{(\ell-1)/2}\le b/96 for each odd ℓ∈[b,2b]\ell\in[b,2b]. Let kk be the number of these lengths, so k≥b/2k\ge b/2.

Lemma 2.1. Suppose both lengths of DD are at least 2b2b. There is a reversible square-block heat bath with gap at least 1/21/2, for every hard exterior. If two exteriors differ only on YY, then at any i∈Di\in D

∥μiη−μiη′∥TV≤b2∑y∈Yexp⁡{−∥i−y∥∞/(8b)}.\left\|\mu_i^\eta-\mu_i^{\eta'}\right\|_{\mathrm{TV}}\le\frac{b}{2}\sum_{y\in Y}\exp\{-\|i-y\|_\infty/(8b)\}.

Here total variation is the supremum of the absolute event difference.

Proof. In a coordinate of length nn, assign the start jj of an ℓ\ell-interval the weight 1+b1j=1+b1j=n−ℓ+11+b\mathbf{1}_{j=1}+b\mathbf{1}_{j=n-\ell+1}. Update each contained ℓ\ell-square by conditional resampling at the product of its two starting weights divided by kb2kb^2. The sum of one-dimensional weights containing a given site lies between bb and 4b4b: an endpoint anchor suffices for the lower bound, and away from both anchors there are ℓ\ell possible starts. For a site immediately outside the interval there are at most two starts, and endpoint bonuses occur only at the two lengths x−1,n−xx-1,n-x. The sum of these outside weights across lengths is at most 2k+2b2k+2b. Thus each site has covering rate at least one and outside-neighbour rate at most 2(4b)(2k+2b)/(kb2)≤48/b2(4b)(2k+2b)/(kb^2)\le48/b. The covering rate is at most 16 per site or 32 for blocks touching a given edge.

A single interior disagreement is erased by a block containing it. For a neighbouring block the classical input bounds expected new disagreements. The net removal rate is at least 1−(48/b)(b/96)=1/21-(48/b)(b/96)=1/2. Denoting the total clock rate by λ\lambda, a single random update contracts Hamming distance by

Diagram of comparisons for hard classical boundaries

Figure 2. The comparison at a free seam needs both the boundary norm of a wing and a gapped reference system. The weighted coupling also controls the interior of an intact square. The two stability conclusions use the perturbation bounds specified in § 8.

1−1/(2λ)1 - 1/(2\lambda), first for neighbouring configurations and then along a coordinate path. The Markov operator is a positive self-adjoint average of conditional expectations. A mean-zero bounded function has finite Hamming Lipschitz seminorm, so its iterates decay in supremum norm at this rate times a volume-dependent prefactor. The spectral theorem, followed by density of bounded functions in L2L^2, gives the stated continuous-time spectral gap.

For the marginal, put wx=e−∥x−i∥∞/(8b)w_x=e^{-\lVert x-i\rVert_\infty/(8b)} and use the sum of weights of disagreements as the distance. Within an affected block each new weight is at most e1/4wy<(4/3)wye^{1/4}w_y < (4/3)w_y for a disagreement at yy. The corresponding interior removal rate is at least 1/31/3. For a changed external spin, the normal start is forced and has weight at most 3b3b, the tangential covering weight is at most 4b4b. Thus the total affected rate is at most 12, injecting expected weight at most bwy/6bw_y/6 per unit time. Couple along a coordinate path and telescope over changed exterior spins. For stationary laws, their weighted Wasserstein distance LL satisfies

L≤(1−1/(3λ))L+(b/(6λ))∑y∈Ywy.L \le(1 - 1/(3\lambda))L + (b/(6\lambda))\sum_{y\in Y} w_y.

One may glue the conditional update couplings on the finite product of spheres; the weighted Hamming cost is bounded and lower semicontinuous. Since wi=1w_i=1, the marginal total variation is at most LL.

One whole side in the physical boundary norm

In a w×hw \times h wing DD, let SS be a side of length hh, including both corner sites, and put O=∂D∖SO=\partial D\setminus S. Define the trace-one Gram operator M=UD∗UD/Tr⁡(UD∗UD)M=U_D^*U_D/\operatorname{Tr}(U_D^*U_D) and its two marginal operators MS=Tr⁡OMM_S=\operatorname{Tr}_O M, MO=Tr⁡SMM_O=\operatorname{Tr}_S M.

Proposition 3.1. For each fixed A≥1A\ge1 there is CA<∞C_A<\infty such that for every 2≤h≤Aw2\le h\le Aw and w≥2w\ge2,

CA−1MS⊗MO≤M≤CAMS⊗MO.(1)C_A^{-1}M_S\otimes M_O \le M \le C_A M_S\otimes M_O. \tag*{(1)}

Proof. The coherent-state identity (d+1)∫∣n;d⟩⟨n;d∣ dn=1(d+1)\int|n;d\rangle\langle n;d|\,dn=1 expresses MM as the expectation of the tensor product of unit coherent projectors of residual degrees 4−deg⁡D(x)4-\deg_D(x). Its probability density on all physical spins is proportional to the product of squared edge factors. Changing nxn_x to −nx-n_x on one sublattice makes the edge weight proportional to 1+nx⋅ny1+n_x\cdot n_y and applies the same variable change to the local projectors. Positivity and their unit traces are preserved.

Remove the two perimeter bonds between SS and OO and lower the output degree at each exposed end by one per removal. The residual source at each site is unchanged. Before trace normalization, re-inserting these factors compares the Grams in both directions with a universal constant by the local insertion estimate. Trace normalization and either partial trace preserve such a comparison, at a bounded cost. For bounded hh one may instead cut all hh edges between the facing column and the next column. The separated Gram is a product across S∣OS|O, giving (1) with constant depending only on this upper bound for hh. This covers the case where either length of the inner rectangle is less than 2b2b, since then h≤A(2b+2)h\le A(2b+2). The hh insertions cost at most 2h2^h on the unnormalized Gram. Normalizing and comparing the two true marginals as below gives a factor at most 43h4^{3h}.

Otherwise condition the outer ring on its two pieces s,os,o. The ring density of the two-edge-deleted model has the form a(s)c(o)Z(s,o)a(s)c(o)Z(s,o) up to normalization. Here ZZ is the partition function of the intact inner rectangle with hard exterior given by that ring. The two corner sites of SS occur in a(s)a(s). Only its other h−2h-2 sites have an inward neighbour in the inner rectangle, one per site. Change a

single SS spin with an inward neighbour ii from vv to uu, and write fu=1+ni⋅uf_u = 1 + \mathbf{n}_i \cdot\mathbf{u}. Integrating fufvf_u f_v against the same weight with this single leg omitted gives the identity between hard laws

Z(u,o)Z(v,o)=Eμv,ofuEμu,ofv.\frac{Z(u,o)}{Z(v,o)} = \frac{\mathbb{E}_{\mu_{v,o}} f_u}{\mathbb{E}_{\mu_{u,o}} f_v}.

For either law, conditional on all other spins, the density at ii is the squared modulus of a product PP of t≤4t \le4 linear spinor factors. Multiplying a normalized degree-tt monomial by xx has minimum squared singular value 1/(t+1)1/(t+1) over the whole space, as follows by taking ratios of the binomial norms. By unitary invariance this holds for any unit linear factor. The coherent-state normalization therefore gives Efu≥2/(t+2)≥1/3\mathbb{E}f_u \ge2/(t+2) \ge1/3. Changing oo to o′o' changes each logarithmic expectation by at most six times the marginal total variation (0≤fu≤20 \le f_u \le2).

Telescope along the SS legs and apply Lemma 2.1. With ξ=8b\xi= 8b, the absolute logarithmic cross ratio of Z(s,o),Z(s,o′),Z(s′,o′),Z(s′,o)Z(s,o), Z(s,o'), Z(s',o'), Z(s',o) is bounded independently of the lengths: for each adjacent side the sum is bounded by ∑a≥0,j≥1e−max⁡(a,j)/ξ\sum_{a\ge0,j\ge1} e^{-\max(a,j)/\xi}, and for the opposite side by h2e−(w−2)/ξh^2e^{-(w-2)/\xi}, bounded for h≤Awh \le Aw. Multiply out the cross-ratio inequality, include a(s)c(o)a(s′)c(o′)a(s)c(o)a(s')c(o'), and integrate in s′,o′s',o'. This compares the normalized ring law in both directions with the product of its two marginals, also at configurations where a perimeter weight vanishes. Integrate the positive coherent projectors to obtain the operator comparison for the cut Gram. To track the original marginals, write B−,M−B^-,M^- for the unnormalized and normalized cut Grams. Each singlet insertion has squared singular values in [1/2,2][1/2,2]: on the two degrees k,l≤3k,l \le3 its eigenvalues in squared norm are (t+1)(k+l+2−t)/((k+1)(l+1))(t+1)(k+l+2-t)/((k+1)(l+1)), 0≤t≤min⁡(k,l)0 \le t \le\min(k,l). Thus B−/4≤UD∗UD≤4B−B^-/4 \le U_D^*U_D \le4B^- and M−/16≤M≤16M−M^-/16 \le M \le16M^-. Each partial trace preserves the latter order. If the cut comparison has constant KK, the joint Gram and its two marginals therefore give (1) with constant 163K16^3K.

From eight exposed sites to a seam of arbitrary length

In an n×hn \times h rectangle cut the matching FF between columns j,j+1j,j+1. Let PjP_j project onto the kernel with these edges deleted and let Tj=∏f∈F(1−pf)T_j = \prod_{f\in F}(1-p_f). Its factors have disjoint supports. Their simultaneous ground projector with PjP_j is G=GΛG = G_\Lambda.

Proposition 4.1. If j−1,n−j−1≥max⁡(2,h/5)j-1,n-j-1 \ge\max(2,h/5) then ∥(Tj−G)(Pj−G)∥≤q\lVert(T_j-G)(P_j-G)\rVert\le q for a single q<1q<1 independent of the rectangle.

Proof. Keep the two columns next to the matching as physical columns YL,YRY_L,Y_R. Call the two remaining rectangles DL,DRD_L,D_R and take their facing sides SL,SRS_L,S_R as in Figure 1. In the polynomial source of the cut ground space write L\mathcal{L} for the residual factors of the two YY columns and the facing sides after all nonmatching bonds have been accounted for, and OO for the residual factors on both complements OL,ORO_L,O_R. Multiplication VV by the vertical bonds in YY and the bonds from YY to the facing sides acts trivially on OO. Then U=(UDL⊗UDR⊗1Y)VU=(U_{D_L}\otimes U_{D_R}\otimes1_Y)V multiplies by every nonmatching bond exactly once and parametrizes ran⁡Pj\operatorname{ran}P_j injectively. The matching acts only on the retained physical columns. Consequently B(f)=∥Uf∥2B(f)=\lVert U f\rVert^2 and D(f)=∥(1−Tj)Uf∥2D(f)=\lVert(1-T_j)Uf\rVert^2 evaluate the same product Gram on VfVf and on (1−Tj)Vf(1-T_j)Vf, respectively. Trace-normalizing the wing Grams changes both by the same scalar.

A vertical matching cut with regions $D_L$, $D_R$, $O_L$, $O_R$, $S_L$, $S_R$, $Y_L$, $Y_R$, and dashed edges $F$

Figure 1. A vertical matching cut. The dashed edges form FF. The two columns YL,YRY_L,Y_R carry their original physical spin 2 throughout. Each deep wing has one facing side SS (blue), including both of its corner sites. Its three other sides, with those two corners omitted, form OO. Only the reference model pins the exterior slots on OO. After the matching is restored, its eight circled sites can still have free exterior slots. All slots on the original exterior are free in Theorem 1.1.

Fill the exterior slots on each OO by a fixed product polynomial χ=χL⊗χR\chi=\chi_L\otimes\chi_R. Impose this divisibility by one orthogonal complement projection per site. After restoring the matching, the only unsaturated sites in the reference system are the four ends of the two YY columns and the four ends of the two facing sides. By Lemmas A.1 and A.2 below, the reference Hamiltonian has gap at least g8>0g_8>0. Its cut ground space is U(L⊗χ)U(\mathcal{L}\otimes\chi). For a reference cut-ground vector vv perpendicular to the reference full kernel, TjvT_jv stays perpendicular to it. For a nonmatching edge aa, let FaF_a consist of the matching edges touching aa, and put Tfar=∏f∉Fa(1−pf)T_{\mathrm{far}}=\prod_{f\notin F_a}(1-p_f). Since pav=0p_av=0 and pap_a commutes with TfarT_{\mathrm{far}},

∥paTjv∥2≤∥(1−∏f∈Fa(1−pf))Tfarv∥2=∑∅≠B′⊆Fa∥EB′v∥2.\|p_aT_jv\|^2\leq\left\|\left(1-\prod_{f\in F_a}(1-p_f)\right)T_{\mathrm{far}}v\right\|^2=\sum_{\varnothing\ne B'\subseteq F_a}\|E_{B'}v\|^2.

where EB′E_{B'} projects onto the pattern of violations precisely B′B'. Fixing any one edge in a nonempty B′B' shows that this pattern is charged at most six times. All onsite constraints are on OO and commute with TjT_j. Thus g8∥Tjv∥2≤6∥(1−Tj)v∥2g_8\|T_jv\|^2\leq6\|(1-T_j)v\|^2. For any reference cut vector this yields the energy-distance estimate with coefficient g8/(6+g8)g_8/(6+g_8) after subtraction of its reference-ground component.

In a single wing put Mχ=⟨χD∣M∣χD⟩OM^\chi=\langle\chi_D|M|\chi_D\rangle_O and s=⟨χD,MOχD⟩>0s=\langle\chi_D,M_O\chi_D\rangle>0. Equation (1) with A=5A=5 first gives C5−1MS≤Mχ/s≤C5MSC_5^{-1}M_S\leq M^\chi/s\leq C_5M_S, and then compares MM with (Mχ/s)⊗MO(M^\chi/s)\otimes M_O up to factors C5±2C_5^{\pm2}. For both wings, BB and DD are each comparable up to C5±4C_5^{\pm4} to their reference form tensored with the same positive spectator metric, with the same normalization factor. The zero subspace is K=ker⁡[(1−Tj)V∣L]⊗O\mathcal{K}=\ker[(1-T_j)V|_{\mathcal{L}}]\otimes O, because the wing Grams are positive definite and VV leaves OO untouched. In an orthonormal basis of the spectator metric the reference estimate holds componentwise. Comparing the energy and the infimum of the squared norm over the same affine subspace gives

D(f)≥ηinf⁡k′∈KB(f−k′),η=C5−8g8/(6+g8)>0.D(f)\geq\eta\inf_{k'\in\mathcal{K}}B(f-k'),\qquad\eta=C_5^{-8}g_8/(6+g_8)>0.

This quotient comparison includes arbitrary entanglement with OO and permits the physical orthogonal complements for the two metrics to differ. The matching-pass projection TjT_j has norm at most q=1−ηq=\sqrt{1-\eta} on ran⁡(Pj−G)\operatorname{ran}(P_j-G).

Ordered cuts and recursion on the area

Proof of Theorem 1.1. Orient the longest side horizontally and let its length be nn and the height be hh. Write Rj=Pj−GR_j=P_j-G. The next split enforces the jjth matching, so Pj+1≤TjP_{j+1}\leq T_j and ∥Rj+1Rj∥≤q\|R_{j+1}R_j\|\leq q for every admissible jj. For a<j<b′a<j<b' there is an exact identity Rb′RjRa=Rb′RaR_{b'}R_jR_a=R_{b'}R_a. Indeed PjP_j is the product of the ground projectors of [1,j][1,j] and [j+1,n][j+1,n], each taken at full height. In Pb′PjPaP_{b'}P_jP_a, the shorter prefix projector can be absorbed by the longer prefix and the shorter suffix by the longer suffix, commuting past disjoint intervals. Subtract GG, which is below every projector involved. It follows by insertion of all intermediate cuts and idempotence that

∥G[1,b′]G[a+1,n]−G∥=∥Rb′Ra∥≤qb′−a.\|G_{[1,b']}G_{[a+1,n]}-G\|=\|R_{b'}R_a\|\leq q^{b'-a}.

Here Pb′Pa=G[1,b′]G[a+1,n]P_{b'}P_a=G_{[1,b']}G_{[a+1,n]} by the same nesting argument.

For sufficiently large nn put d=⌊n⌋d=\lfloor\sqrt{n}\rfloor and k=⌊n/(10d)⌋k=\lfloor n/(10d)\rfloor. Place kk disjoint dd-column overlap strips within the central interval [2n/5,3n/5][2n/5,3n/5], starting just after ⌊2n/5⌋\lfloor2n/5\rfloor. For an overlap (a,b′)(a,b') use [1,b′][1,b'] and [a+1,n][a+1,n] as its two child boxes. Both children have area at most 3nh/53nh/5 and the deep wing on each side of every intermediate cut has width at least 2n/5−1≥max⁡(2,h/5)2n/5-1\geq\max(2,h/5). The intersection of the two child kernels is the parent kernel since all edges are covered. On its orthogonal complement the sum of the two complementary projectors is at least 1−qd1-q^d, by the principal-angle identity for two subspaces. If γch\gamma_{\mathrm{ch}} is the minimum gap among all the children, then summing their energy bounds over the kk pairs gives

HΛ≥γchkk+1(1−qd)(1−G).H_\Lambda\geq\gamma_{\mathrm{ch}}\frac{k}{k+1}(1-q^d)(1-G).

Indeed each original edge occurs once in each pair, with a second copy only if both endpoints are in that pair’s overlap. The overlaps are disjoint. The multiplicative factor f(n)f(n) satisfies 1−f(n)≤c/n1-f(n)\le c/\sqrt{n}: 1/(k+1)≤10/n1/(k+1)\le10/\sqrt{n} and qnq^{\sqrt{n}} decays faster.

Choose a fixed cutoff so that f(n)≥1/2f(n)\ge1/2 beyond it and the described cuts exist. Reorient each child along its own longest axis. Along any branch, counting ss steps upwards from a leaf, the area grows at least as (5/3)s(5/3)^s and the longest side is at least (5/3)s/2(5/3)^{s/2}. The negative logarithms of the factors sum to at most 2c∑s≥0(3/5)s/4<∞2c\sum_{s\ge0}(3/5)^{s/4}<\infty, uniformly in the root. The finitely many leaf shapes have side lengths at least two and a positive least nonzero eigenvalue. Their minimum times this uniform product bound proves the theorem. Thin rectangles are covered by the same recursion using the bounded-hh branch of Proposition 3.1.

A consequence for periodic rectangles

Theorem 6.1. For simple square-lattice tori of circumferences L,N≥3L,N\ge3, the ground space is one-dimensional and the spectral gap has a common positive lower bound. This conclusion follows from the open-square gaps and finite-width quantum comparisons.

Proof. Write gng_n for the gap of the free n×nn\times n square. When L,N≥2n+1L,N\ge2n+1 and n≥3n\ge3, translating that square gives the elementary finite-size estimate

gap⁡(HL,Ntor)≥n(n−1)gn−(n+1)(n−1)2.(2)\operatorname{gap}(H^{\mathrm{tor}}_{L,N})\ge\frac{n(n-1)g_n-(n+1)}{(n-1)^2}. \tag*{(2)}

We give the counting argument because the two circumferences may differ and have either parity. This is a Knabe-type finite-size comparison [8]. For weighted criteria, see [9]. Let HxH_x be the open-square Hamiltonian at translate xx, and put a=n(n−1)a=n(n-1), c=(n−1)2c=(n-1)^2. Every edge occurs aa times, so agnH≤∑xHx2ag_nH\le\sum_x H_x^2. A perpendicular pair of incident edges occurs cc times and a collinear incident pair c−1c-1 times. A disjoint pair occurs at most cc times, by intersecting the sets of allowed starting positions in both coordinates. For parallel edges in the same line, their separation reduces the along-line count to at most n−2n-2. The circumference assumption prevents two distinct short lifts of a contributing pair. Disjoint edge projectors commute, so their anticommutators are positive. For a collinear incident pair use −{pe,pf}≤pe+pf-\{p_e,p_f\}\le p_e+p_f. Each edge has two collinear neighbours. Thus ∑xHx2≤cH2+(a−c+2)H=cH2+(n+1)H\sum_x H_x^2\le cH^2+(a-c+2)H=cH^2+(n+1)H, giving (2). Theorem 1.1 permits a fixed nn for which the right side is positive.

It remains to treat finitely many short circumferences 3≤h≤2n3\le h\le2n with unrestricted length. We record the fixed-width argument. For any transverse matching F~\widetilde{F} of hh edges, let PP be the ground projector before inserting F~\widetilde{F} and T=∏f∈F(1−pf)T=\prod_{f\in F}(1-p_f). This applies to a cut across an open-ended cylinder of circumference hh, and also to its end-to-end closing matching. There is a constant ch>0c_h>0 such that every v∈ran⁡Pv\in\operatorname{ran}P satisfies

∥(1−T)v∥2≥chdist⁡(v,ran⁡P∩ran⁡T)2.\lVert(1-T)v\rVert^2\ge c_h\operatorname{dist}(v,\operatorname{ran}P\cap\operatorname{ran}T)^2.

To see this, strip from vv all nonmatching edge determinants touching an endpoint of FF, writing v=Buv=Bu. There are at most 6h6h stripped factors. All remaining divisibility constraints avoid the matching endpoints. On the resulting lower-degree polynomial spaces, fuse the two spinors at ff by J~f\widetilde{J}_f. The projections J~f∗J~f\widetilde{J}_f^*\widetilde{J}_f have disjoint supports. If u0u_0 is the projection of uu onto their common kernel, then Bu0∈ran⁡P∩ran⁡TBu_0\in\operatorname{ran}P\cap\operatorname{ran}T and ∥u−u0∥2≤∑f∥J~fu∥2\lVert u-u_0\rVert^2\le\sum_f\lVert\widetilde{J}_fu\rVert^2. Fusing vv at ff gives B(f)J~fuB(f)\widetilde{J}_fu, where B(f)B(f) is BB with the two endpoints identified. Each factor survives because the graph is simple and ff itself was never stripped. In a three-cycle the two bonds to a common neighbour can become a squared determinant after fusion. Its multiplication map is still injective since the polynomial is nonzero. The local multiplication maps by B(f)B(f) are injective. For fixed hh their smallest singular values have a common positive lower bound, and the norms of the maps by BB have a common finite upper bound: there are finitely many incidence and degree patterns and all other sites are spectators. Thus the physical fusion energies ∑fpf∥pfv∥2\sum_f p_f\|p_fv\|^2 control dist⁡(v,ran⁡P∩ran⁡T)2\operatorname{dist}(v,\operatorname{ran} P \cap\operatorname{ran} T)^2. Finally ∑fpf≤h(1−T)\sum_f p_f \le h(1-T) gives the displayed estimate.

On the open-ended cylinder the resulting cut angle is bounded away from one uniformly in the length. Use the same insertion identity and disjoint-overlap average as in the proof of Theorem 1.1, always splitting the open direction. Each child has at most 3/53/5 of its parent’s length, and the finitely many terminal lengths give a positive seed. This proves a length-independent cylinder gap γh>0\gamma_h > 0. Closing the two ends adds the hh matching projectors. The closed Hamiltonian dominates γh(1−P)+(1−T)\gamma_h(1-P)+(1-T), which has a gap by the angle estimate above and the two-projection decomposition. Take the minimum over the finitely many short circumferences. On a torus every site has degree four, so the product of all edge determinants exhausts every degree and spans the full kernel.

Exponential local indistinguishability in open boxes

Write Bx(r)={z∈Z2:∥z−x∥∞≤r}B_x(r)=\{z\in\mathbb{Z}^2:\|z-x\|_\infty\le r\}, and let GLG_L denote the full ground projection of the open Hamiltonian on LL. Local topological quantum order is a sufficient local-order input in stability results for frustration-free Hamiltonians [12]. An infinite-volume state is frustration-free if it annihilates every bond projection. Aronow and Lopatto proved uniqueness of this state in [5 Theorem F.1]. Their equation (F.1) bounds the total variation of the classical marginal on a fixed central set by 4(2M+1)(1−δ)⌊M1/8⌋4(2M+1)(1-\delta)^{\lfloor M^{1/8}\rfloor} for sufficiently large MM. Sections F.2–F.3 then use informational completeness of the coherent measurement to identify the quantum state. Here we condition on the neighbours of the support of an observable. Those spins prepare a density matrix on the support, which gives a quantitative estimate in operator norm.

Theorem 7.1 (Local topological quantum order). Let bb be as in Lemma 2.1, and let ω\omega be the unique infinite-volume frustration-free state. For every x∈Z2x\in\mathbb{Z}^2 and integers r≥0r\ge0, ℓ≥b+1\ell\ge b+1, set B=Bx(r)B=B_x(r), R=r+ℓR=r+\ell and L=Bx(R)L=B_x(R). Every operator AA supported on BB satisfies

∥GLAGL−ω(A)GL∥≤2∥A∥min⁡{1,4bRe−(ℓ−1)/(8b)}.(3)\|G_L A G_L-\omega(A)G_L\|\le2\|A\|\min\{1,4bR e^{-(\ell-1)/(8b)}\}. \tag*{(3)}

The estimate applies to complex observables and to the entire open-box ground space.

Proof. Let O={z:∥z−x∥∞=R}O=\{z:\|z-x\|_\infty=R\}, D=L∖OD=L\setminus O, and let S={z∉B:dist⁡(z,B)=1}S=\{z\notin B:\operatorname{dist}(z,B)=1\} be the ring of nearest neighbours of BB. Then S⊂DS\subset D, including when r=0r=0. We first bound the joint marginal on SS of the hard-boundary classical law on $D ({xref}Figure 3). The side length of $D$ is $2R-1\ge 2b$. In the proof of {xref}Lemma 2.1`, take wz=exp⁡[−dist⁡∞(z,S)/(8b)]w_z=\exp[-\operatorname{dist}_\infty(z,S)/(8b)]. Distance to SS is 1-Lipschitz, so a disagreement at yy can create only weights at most e1/4wye^{1/4}w_y in an affected block. The interior removal rate is again at least $1/3$ and the exterior injection rate at most bwy/6bw_y/6. The same stationary coupling inequality therefore bounds the weighted Hamming Wasserstein distance by (b/2)∑ywy(b/2)\sum_y w_y, with the sum over the changed exterior spins. A mismatch anywhere on SS has cost at least one. All exterior neighbours of DD belong to OO, which has 8R8R sites and lies at ℓ∞\ell^\infty-distance at least ℓ−1\ell-1 from SS. Thus, for arbitrary hard configurations η,η′\eta,\eta' on OO,

∥μDη∣S−μDη′∣S∥TV≤ε,ε=min⁡{1,4bRe−(ℓ−1)/(8b)}.\|\mu_D^\eta|_S-\mu_D^{\eta'}|_S\|_{\mathrm{TV}}\le\varepsilon,\qquad\varepsilon=\min\{1,4bR e^{-(\ell-1)/(8b)}\}.
Diagram showing $O=\partial L$: fixed outer spins $\eta$; $D=L\setminus O$: intact classical box; $B$: quantum sites supporting $A$ (shaded); $S$: exterior nearest neighbours of $B$ (circles)

Figure 3. Conditioning on the outer layer gives a classical law on all of DD, including the sites of BB. Its joint marginal on SS mixes the quantum states QsQ_s on BB. The four diagonal corners just outside BB do not meet it by a bond. The drawing illustrates incidence and is not to scale.

Write a ground vector on LL as ULfU_Lf, where ff is a polynomial on the outer layer OO. Insert the coherent resolution of the identity on L∖BL\setminus B, leaving the quantum spins in BB unevaluated. Write the classical spins after the sublattice sign change of [5 Appendix F.2]. For a configuration ss on SS, evaluating the exterior ends of all bonds crossing out of BB, and retaining the singlet factors internal to BB, gives a vector vsv_s on BB. This is a product of nonzero polynomials and hence is nonzero for every ss. The choices of spinor phases cancel in Qs=∣vs⟩⟨vs∣/∥vs∥2Q_s = |v_s\rangle\langle v_s|/\lVert v_s\rVert^2. The coherent identity on BB gives ∥vs∥2=cBZB(s)\lVert v_s\rVert^2 = c_B Z_B(s), where ZB(s)Z_B(s) is the classical partition function with these hard neighbours and cB>0c_B > 0 is independent of ss.

In particular, fixing η\eta on OO prepares on BB the density matrix

ση=∫Qs d(μD∣Sη)(s).\sigma_\eta= \int Q_s\,d(\mu_{D\mid S}^{\eta})(s).

To see the measure in this formula, multiply the squared factors on D∖BD\setminus B and its bonds to OO by ∥vs∥2\lVert v_s\rVert^2. The factor ZB(s)Z_B(s) integrates exactly the missing sites and bonds of the intact rectangle DD. Its normalization is ZD(η)Z_D(\eta). Consequently there is a positive measure dτ(η)d\tau(\eta) on OO, independent of ff and AA, such that, for any two boundary polynomials,

⟨ULg,AULf⟩=∫g(η)‾f(η)Tr⁡(Aση) dτ(η).(4)\langle U_Lg,AU_Lf\rangle= \int\overline{g(\eta)}f(\eta)\operatorname{Tr}(A\sigma_\eta)\,d\tau(\eta). \tag*{(4)}

Here the boundary polynomial is evaluated at the coherent spinors before the sublattice sign change, expressed as a function of η\eta. The measure dτd\tau is the product area measure times ZD(η)Z_D(\eta), the bond weights within OO, and a common positive normalization. Taking A=1A=1 gives ∫∣f∣2 dτ=∥ULf∥2\int|f|^2\,d\tau=\lVert U_Lf\rVert^2, so entangled boundary polynomials have the same representation.

Each QsQ_s has trace norm one, so ∥ση−ση′∥1≤2ε\lVert\sigma_\eta-\sigma_{\eta'}\rVert_1\le2\varepsilon. For completeness, normalized open-box ground states on an exhaustion have a locally convergent subsequence by finite-dimensional compactness and a diagonal argument. The consistent limit annihilates every bond. If ω\omega is any such state, its restriction to LL is supported on GLG_L, by positivity of the local Hamiltonian. Its restriction to BB is therefore a convex mixture of the ση\sigma_\eta by (4) and spectral decomposition. In particular ∣Tr⁡(Aση)−ω(A)∣≤2ε∥A∥|\operatorname{Tr}(A\sigma_\eta)-\omega(A)|\le2\varepsilon\lVert A\rVert for every η\eta. Subtract the identity case of (4) multiplied by ω(A)\omega(A). Cauchy–Schwarz bounds the resulting matrix element by 2ε∥A∥∥ULg∥∥ULf∥2\varepsilon\lVert A\rVert\lVert U_Lg\rVert\lVert U_Lf\rVert. Supremizing over both ground vectors proves (3). Any second frustration-free state is also supported on GLG_L upon restriction. Taking ℓ\ell to infinity for fixed BB gives equality on all local observables, and hence uniqueness.

Stability of the gap on the infinite plane

Let Φ(x,n)\Phi(x,n) be self-adjoint, supported on Bx(n)B_x(n), for x∈Z2x \in\mathbb{Z}^2 and integers n≥1n \ge1. Assume ∥Φ(x,n)∥≤Je−anθ\lVert\Phi(x,n)\rVert\le Je^{-an^\theta}, uniformly in x,nx,n, with J,a>0J,a>0 and 0<θ≤10<\theta\le1. Onsite terms may be included at radius one. The family defines the perturbation of the infinite-volume dynamics with coupling ss as in [13 Assumption 2.6].

Corollary 8.1 (Bulk gap under local perturbations). Under (H), the frustration-free state ω\omega on Z2\mathbb{Z}^2 has a positive gap γ0\gamma_0 in its GNS Hamiltonian. For each 0<γ<γ00<\gamma<\gamma_0 there is s0>0s_0>0 such that, for real ∣s∣<s0\lvert s\rvert<s_0, the spectral-flow state ωs=ω∘αs\omega_s=\omega\circ\alpha_s [18] is a ground state of the perturbed infinite-volume dynamics and its GNS Hamiltonian has gap at least γ\gamma.

Proof. First take any local AA with ω(A)=0\omega(A)=0. For increasing open squares LL containing the support and all its neighbours, let P=GLP=G_L. Frustration freeness gives ω(P)=1\omega(P)=1. For the local energy derivation δ(A)=[HL,A]\delta(A)=[H_L,A], Theorem 1.1 yields

ω(A∗δ(A))=ω(A∗HLA)≥g(ω(A∗A)−ω(A∗PA)).\omega(A^*\delta(A))=\omega(A^*H_LA)\ge g\bigl(\omega(A^*A)-\omega(A^*PA)\bigr).

Here ω(A∗PA)\omega(A^*PA) is at most ∥PAP∥2\lVert PAP\rVert^2, since ω\omega is supported on PP. It tends to zero by Theorem 7.1. This proves the GNS gap criterion [13 (2.22)] with g>0g>0.

We check the one-term-per-site convention of the stability theorem. On Z2\mathbb{Z}^2 with the ℓ∞\ell^\infty metric, let hxh_x be half the sum of the four bond projections around the unit plaquette with lower left corner xx. Every bond belongs to two plaquettes, so this defines the same infinite interaction derivation. On a square Bx(n)B_x(n), n≥1n\ge1, formula (2.13) in [13] keeps the plaquettes contained in the square. Each boundary bond then has weight 1/21/2 and each remaining bond has weight one. The local kernel is exactly GBx(n)G_{B_x(n)} and the local gap is at least g/2g/2 by Theorem 1.1. Ball volumes grow quadratically. Colour centres by both coordinates modulo 2n+12n+1. Balls of a common colour are disjoint, and there are (2n+1)2(2n+1)^2 colours. Thus the exponents in Assumption 2.3 can be taken as ν=ζ=2\nu=\zeta=2 and α=0\alpha=0 (with ν\nu the volume exponent).

For the local-order condition (2.27) in that paper, put t=m−kt=m-k for integer outer and inner radii m≥km\ge k. If t≥b+1t\ge b+1, Theorem 7.1 bounds the norm by 8b∥A∥(k+t)e−(t−1)/(8b)8b\lVert A\rVert(k+t)e^{-(t-1)/(8b)}. Since k+t≤(1+k)2(1+t)k+t\le(1+k)^2(1+t), one may take G0(t)=Cbe−t/(16b)G_0(t)=C_be^{-t/(16b)} in (2.27), enlarging CbC_b to cover the trivial bound 2∥A∥2\lVert A\rVert for 0≤t<b+10\le t<b+1. At radius zero the full one-site space is the local kernel. For real radii one takes integer parts and enlarges CbC_b once. All the moments of G0G_0 in (2.26) are finite. The interaction radius, local gaps, local order and perturbation bound now satisfy the hypotheses of [13 Theorems 2.8–2.9]. The latter theorem gives the stated infinite-volume conclusion. □

Finite tori of bounded aspect ratio

Corollary 8.2. Fix 0<κ≤10<\kappa\le1. Consider simple tori with short circumference h≥3h\ge3 and long circumference nn satisfying h≥κnh\ge\kappa n. On each torus allow any self-adjoint perturbation V=∑x,mΦ(x,m)V=\sum_{x,m}\Phi(x,m) with 1≤m≤∣diam⁡∣1\le m\le\lvert\operatorname{diam}\rvert, where the support lies in the periodic ℓ∞\ell^\infty ball of radius mm at xx and ∥Φ(x,m)∥≤Je−amθ\lVert\Phi(x,m)\rVert\le Je^{-am^\theta} with common J,a,θJ,a,\theta as above. We use the linear perturbation path s↦Htor+sVs\mapsto H^{\mathrm{tor}}+sV. Assuming (H), there are s∗>0s_*>0 and g∗>0g_*>0 such that Htor+sVH^{\mathrm{tor}}+sV has a unique ground state and gap at least g∗g_* for all these tori and all real ∣s∣≤s∗\lvert s\rvert\le s_*.

Proof. We verify the uniform perturbation conditions in [14 Section 6]. The initial Hamiltonians are uniformly gapped by Theorem 6.1. Put r0=⌊h/4⌋r_0=\lfloor h/4\rfloor. For large hh, every ball of radius m≤r0m\le r_0 is an induced open square. The torus ground state restricts to its ground space by frustration freeness. For k≤mk\le m and AA supported in the concentric kk-ball, we may therefore replace ω(A)\omega(A) in (3) by the torus expectation at a cost of at most twice the bound there. Since m≤(1+k)(1+m−k)m \le(1+k)(1+m-k), this proves their (2.9) with Ω(t)=Cbe−t/(16b)\Omega(t)=C_b e^{-t/(16b)} and indistinguishability radius at least r0r_0 everywhere. Small tt is covered by increasing CbC_b. Their cutoff radii K=L=⌊h/12⌋K=L=\lfloor h/12\rfloor thus allow the whole torus as perturbation region in (6.6).

For their local gap condition (6.11)–(6.12), choose the comparison region at a centre and scale m≥1m\ge1 to be the ball itself for m≤r0m\le r_0 and the whole torus for m>r0m>r_0. The local gaps have a common lower bound by Theorems 1.1 and 6.1. Small balls of the same radius have an intersection graph of maximum degree O(m2)O(m^2), giving O(m2)O(m^2) colours. At a larger radius use one colour per centre: hn≤h2/κ=Oκ(m2)hn\le h^2/\kappa=O_\kappa(m^2). The volume and colouring exponents are both two. For their (6.14)–(6.15), use F(t)=e−a′tθ(1+t)−pF(t)=e^{-a't^\theta}(1+t)^{-p}, with p=4p=4 and a′=2−(1+θ)aa'=2^{-(1+\theta)}a. A radius-mm ball containing a pair at distance tt has m≥t/2m\ge t/2. The number of possible centres and the volume of each ball are both O((1+m)2)O((1+m)^2), so the spare exponential decay bounds the required weighted sum after division by F(t)F(t).

The cutoffs grow linearly in the diameter with constants depending on κ\kappa. For large diameters both cutoffs exceed its square root. The uniform local gaps and exponential Ω\Omega hence satisfy (6.38) and (6.41)–(6.43) of [14], with both cutoff exponents equal to 1/21/2. Theorem 6.8 there gives a uniformly separated low-energy band. Its rank stays one along the small perturbation path. All estimates are uniform in the chosen shapes and terms. Equivalently, an offending sequence of unbounded sizes has a subsequence with both lengths increasing to infinity. Its centred fundamental rectangles form an increasing exhaustion of the plane; on each fixed finite set the metrics and the unperturbed interactions eventually agree. Definition 6.5(v) only requires a uniform bound on the perturbation norms for this finite-volume conclusion (see Remark 6.6 of [14]). Thus the individual perturbations may be chosen independently of one another and the uniform-sequence theorem applies to this subsequence. The finitely many smaller shapes are covered by the unperturbed gap and ∥V∥≤Jhn∑m≥1e−amθ\lVert V\rVert\le Jhn\sum_{m\ge1}e^{-am^\theta}. Applying the result to both signs of VV and reducing the threshold gives the closed interval stated.

The reference system

For every missing exterior edge choose a unit linear spinor factor at its boundary site. If the retained factors at xx have product LxL_x, impose one onsite orthogonal projector with kernel LxV4−deg⁡LxL_x V_{4-\deg L_x}. With no retained factor this projector is zero. Equal factors at a corner retain their multiplicity. All original edge and onsite projections have unit coefficients. A site is saturated when every missing direction carries a factor. We prove two facts about these auxiliary Hamiltonians.

Lemma A.1. With every site saturated, the ground state is unique and the gap is bounded below by g0>0g_0>0 for every W,H≥2W,H\ge2 and every choice of spinor factors.

Proof. The full product of edge determinants and onsite factors exhausts all degrees. Its normalization Ξ\Xi spans the joint kernel. The coherent integral for ∣Ξ∣2\lvert\Xi\rvert^2 gives exactly the classical hard-boundary law after the sublattice change of variables. Every exterior neighbour of a rectangle touches just one internal site, hence arbitrary chosen exterior factors define a legal hard configuration. Coherent evaluation divided by that of Ξ\Xi defines an isometry JΞJ_\Xi from the physical space into L2(μ)L^2(\mu), with JΞΞ=1J_\Xi\Xi=1 (the zero set has measure zero).

For an updated square RR take HRH_R to consist of every physical edge touching RR and every onsite term in RR, and let PRP_R project onto its kernel. Dividing a vector in this kernel by Ξ\Xi leaves a function of the exterior of RR: the required factors exhaust each degree in RR. Writing ERE_R for conditional expectation given that exterior yields

∥(1−ER)EJϕ∥2≤∥(1−PR)ϕ∥2≤g∗−1⟨ϕ,HRϕ⟩.\lVert(1-E_R)E_J\phi\rVert^2 \le\lVert(1-P_R)\phi\rVert^2 \le g_*^{-1}\langle\phi,H_R\phi\rangle.

For our fixed bb only finitely many local block geometries occur, including edges and corners. At each geometry the kernel is the image of multiplication by the prescribed product from a fixed residual source. Its rank is constant for all unit pin parameters, including coincidences. The onsite projectors vary continuously, since the multiplication matrix LL has full column rank and its range projector is L(L∗L)−1L∗L(L^*L)^{-1}L^*. Compactness thus gives a common positive g∗g_* for the nonzero local spectra. Summing with the rates of Lemma 2.1 and using the covering bound of 32 for each term gives H≥(g∗/64)(1−∣Ξ⟩⟨Ξ∣)H \ge(g_*/64)(1-\lvert\Xi\rangle\langle\Xi\rvert) whenever both side lengths are at least 2b2b.

For a fixed short side h<2bh<2b take a column as a single classical variable. Its interaction with its neighbour is B(x,y)=∏r=1h(1+xr⋅yr)B(x,y)=\prod_{r=1}^{h}(1+x_r\cdot y_r). Let wj(y)w_j(y) contain the within-column weights and any exterior factors. The two-step kernel

Lj(x,z)=∫B(x,y)wj(y)B(y,z) dyL_j(x,z)=\int B(x,y)w_j(y)B(y,z)\,dy

is continuous and everywhere positive, since each of finitely many factors has a null zero set. Compactness and the finite column geometries give 0<mh≤Lj≤Mh0<m_h\le L_j\le M_h. Conditional on a common right boundary, the law two columns forward has density proportional to Lj(x,z)F(z)L_j(x,z)F(z) with the same nonnegative right message FF for both left conditions. Both laws therefore dominate (mh/Mh)F(z) dz/∫F(m_h/M_h)F(z)\,dz/\int F. Couple every second column until agreement, then couple synchronously and fill the intermediate columns by their conditional bridges. The expected number of changed columns is at most 2Mh/mh+32M_h/m_h+3, uniformly in the interval length. A changed right condition is handled by reversal.

Set Dh=2Mh/mh+3D_h=2M_h/m_h+3 and take an odd bh>12Dhb_h>12D_h. For intervals of odd length in [bh,2bh][b_h,2b_h], use the endpoint bonuses bhb_h above. Give each interval its starting weight divided by khbhk_hb_h as its continuous-time clock rate, where khk_h is the number of allowed lengths. Each column is covered at rate between one and four. Its neighbouring intervals have total starting weight at most 2kh+2bh≤6kh2k_h+2b_h\le6k_h, hence total rate at most 6/bh6/b_h. A single column discrepancy thus contracts at rate at least 1−6Dh/bh>1/21-6D_h/b_h>1/2. Each bond is charged at rate at most eight. The updated blocks have bounded size depending on hh, so the same quantum comparison and compactness argument applies. If the length is also below 2bh2b_h, compactness applies to the entire rectangle. Take the minimum over the finitely many short sides. This argument obtains g0g_0 without using any gap assumption for the free-boundary system.

Lemma A.2. For each fixed KK, suppose at most KK boundary sites retain only some (possibly none) of their exterior factors. The gap is at least gK>0g_K>0, uniformly in size, positions, spacings, and directions.

Proof. Induct on the number jj of exceptions, starting with Lemma A.1. Choose an exceptional xx and let r∈{1,2}r\in\{1,2\} be its number of free exterior slots. Take S∗S_* to be the sum of the edge terms incident to xx and the current onsite terms in its closed star. Write P=proj⁡ker⁡S∗P=\operatorname{proj}\ker S_*. The star constraints mean precisely divisibility by their product B∗B_*, which includes each required determinant and each onsite polynomial. The same finite-type constant-rank argument as above gives S∗≥s(1−P)S_*\ge s(1-P), uniformly over the pin parameters. Expand p=Pψ=∑a=0rpap=P\psi=\sum_{a=0}^{r}p_a with pa=B∗eauap_a=B_*e_au_a, where (ea)(e_a) are the normalized monomials in VrV_r at xx and uau_a have degree zero there. Enlarge the pin product at xx by the factors of eae_a to get a Hamiltonian HaH_a with j−1j-1 exceptions. Then pap_a satisfies its enlarged onsite term and Na=ker⁡Ha⊂N=ker⁡HN_a=\ker H_a\subset N=\ker H.

For an edge ee outside S∗S_*, its endpoints do not include xx. On identifying its endpoints, the product B∗B_* becomes a polynomial B∗(e)B_*^{(e)}. Every star edge still has distinct endpoints because it contains xx, and each pin factor remains nonzero. Fusion commutes with multiplication in the sense that fusing the product equals multiplying the fused factors. Multiplication by B∗(e)B_{\ast}^{(e)} is injective, and its largest and smallest singular values have bounded ratio by compactness on a star and at most one additional edge. All other variables are spectators. In particular ee may meet another exceptional site adjacent to xx. Fusion acts only on the two endpoints of ee, so the distinct eae_a at xx remain orthogonal before multiplication by the fused star factors. Since fusion’s squared norm equals the energy of the corresponding highest spin projector, there is a fixed c>0c > 0 with ∥pep∥2≥c∑a∥pepa∥2\lVert p_e p\rVert^2 \ge c\sum_a \lVert p_e p_a\rVert^2. For a star-exterior onsite projector the same argument applies directly because it commutes with star multiplication and the component split. All the star terms vanish on pap_a, including its new pin constraint. Consequently

⟨p,Hp⟩≥cgj−1∑adist⁡(pa,Na)2≥djdist⁡(p,N)2,dj=cgj−1r+1.\langle p,Hp\rangle\ge c g_{j-1}\sum_a \operatorname{dist}(p_a,N_a)^2 \ge d_j\operatorname{dist}(p,N)^2,\qquad d_j=\frac{c g_{j-1}}{r+1}.

For the last bound, sum nearest points in NaN_a and use Cauchy–Schwarz.

For ψ⊥N\psi\perp N, write p=Pψp=P\psi, u=ψ−pu=\psi-p and E=⟨ψ,Hψ⟩E=\langle\psi,H\psi\rangle. Since N⊂ran⁡PN\subset\operatorname{ran}P, p⊥Np\perp N and ∥u∥2≤E/s\lVert u\rVert^2\le E/s. Projectors disjoint from the closed star commute with PP and cannot have larger energy on pp than on ψ\psi. Among the remaining terms outside S∗S_\ast only a bounded number D∗D_\ast meet the star. Each such projector h′h' satisfies ∥h′p∥2≤2∥h′ψ∥2+2∥u∥2\lVert h'p\rVert^2\le2\lVert h'\psi\rVert^2+2\lVert u\rVert^2. It follows that

∥ψ∥2≤[s−1+2+2D∗/sdj]E.\lVert\psi\rVert^2\le\left[s^{-1}+\frac{2+2D_\ast/s}{d_j}\right]E.

Taking minima over the finite local types closes the induction. Only K=8K=8 is used in the proof of Theorem 1.1. □

Acknowledgements

The main results follow from the analytic estimates above together with the classical hard-boundary coupling result established in the preprint of Aronow and Lopatto [5]. As supplementary sanity checks on small volumes, exact rational calculations independently verified all 78 independent entries of each of the cut-norm and matching-energy forms in the 12-dimensional cut-source sector of the 6×26\times2 rectangle, all 81 entries of a 2×22\times2 wing partial trace, and all 25 entries of the central spin density operator against the inner-neighbour preparation on a 5×55\times5 square. A supplementary Lean 4 development (OrderedCuts.lean), prepared with AI assistance and checked locally using Lean 4.8.0, formalizes the abstract ordered-cut identities and a conditional overlap-contraction lemma used in § 5. It does not formalize the physical AKLT estimates or the main spectral-gap theorem. We plan to make the source files publicly available in a separate repository in a subsequent release.

AI-assisted tools (ChatGPT/Codex and Gemini) were used in mathematical exploration, verification code, and manuscript preparation.

We thank P. M. Aronow and Patrick Lopatto for making their preprint and supporting materials available. Their Proposition 4.4 and verification of (W) supply the classical hypothesis of Corollary 1.2, and we also draw on their description of open-box ground states in Appendix F of [5].

References

  1. [1]I. Affleck, T. Kennedy, E. H. Lieb and H. Tasaki, Valence bond ground states in isotropic quantum antiferromagnets, Commun. Math. Phys. 115, 477–528 (1988).
  2. [2]T. Kennedy, E. H. Lieb and H. Tasaki, A two-dimensional isotropic quantum antiferromagnet with unique disordered ground state, J. Stat. Phys. 53, 383–415 (1988). doi:10.1007/BF01011563.DOI
  3. [3]W. Guo, N. Pomata and T.-C. Wei, Nonzero spectral gap in several uniformly spin-2 and hybrid spin-1 and spin-2 AKLT models, Phys. Rev. Research 3, 013255 (2021). doi:10.1103/PhysRevResearch.3.013255.DOI
  4. [4]A. N. Kirillov and V. E. Korepin, A resonance valence bond in quasicrystals, Leningrad Math. J. 1, 343–377 (1990). English preprint under the title The valence bond solid in quasicrystals, https://arxiv.org/abs/0909.2211v1.
  5. [5]P. M. Aronow and P. Lopatto, Uniform Spectral Gap for the AKLT Model on the Square Lattice, Hexagon preprint 2611.00580v1, 18 November 2026. https://hexagonmath.org/2611.00580v1.hub.valency.io/works/W_ucuyhfke
  6. [6]M. J. Kastoryano and A. Lucia, Divide and conquer method for proving gaps of frustration free Hamiltonians, J. Stat. Mech. (2018), 033105. https://arxiv.org/abs/1705.09491.
  7. [7]M. J. Kastoryano, A. Lucia and D. Pérez-García, Locality at the boundary implies gap in the bulk for 2D PEPS, Commun. Math. Phys. 366, 895–926 (2019). https://arxiv.org/abs/1709.07691.
  8. [8]S. Knabe, Energy gaps and elementary excitations for certain VBS-quantum antiferromagnets, J. Stat. Phys. 52, 627–638 (1988). doi:10.1007/BF01019721.DOI
  9. [9]M. Lemm and D. Xiang, Quantitatively improved finite-size criteria for spectral gaps, J. Phys. A: Math. Theor. 55, 295203 (2022). doi:10.1088/1751-8121/ac7989.DOI
  10. [10]T. Jackson, B. Nachtergaele and A. Young, Local Topological Quantum Order and Spectral Gap Stability for the AKLT Models on the Hexagonal and Lieb Lattices, Commun. Math. Phys. 407, 261 (2026). doi:10.1007/s00220-026-05752-9.DOI
  11. [11]M. J. Kastoryano, Spectral Gap of the Hexagonal AKLT Model via Boundary-State Factorization, https://arxiv.org/abs/2609.27993v1.
  12. [12]S. Michalakis and J. P. Zwolak, Stability of Frustration-Free Hamiltonians, Commun. Math. Phys. 322, 277–302 (2013). doi:10.1007/s00220-013-1762-6.DOI
  13. [13]B. Nachtergaele, R. Sims and A. Young, Stability of the bulk gap for frustration-free topologically ordered quantum lattice systems, Lett. Math. Phys. 114, article 24 (2024). https://arxiv.org/abs/2102.07209v3.
  14. [14]B. Nachtergaele, R. Sims and A. Young, Quasi-locality bounds for quantum lattice systems. Part II. Perturbations of frustration-free spin models with gapped ground states, Ann. Henri Poincaré 23, 393–511 (2022). https://arxiv.org/abs/2010.15337v3.
  15. [15]H. Abdul-Rahman, M. Lemm, A. Lucia, B. Nachtergaele and A. Young, A class of two-dimensional AKLT models with a gap, Contemp. Math. 741, 1–21 (2020). https://arxiv.org/abs/1901.09297.
  16. [16]M. Lemm, A. W. Sandvik and L. Wang, Existence of a spectral gap in the Affleck–Kennedy–Lieb–Tasaki model on the hexagonal lattice, Phys. Rev. Lett. 124, 177204 (2020). https://arxiv.org/abs/1910.11810.
  17. [17]N. Pomata and T.-C. Wei, Demonstrating the Affleck–Kennedy–Lieb–Tasaki spectral gap on 2D degree-3 lattices, Phys. Rev. Lett. 124, 177203 (2020). https://arxiv.org/abs/1911.01410.
  18. [18]S. Bachmann, S. Michalakis, B. Nachtergaele and R. Sims, Automorphic equivalence within gapped phases of quantum lattice systems, Commun. Math. Phys. 309, 835–871 (2012). https://arxiv.org/abs/1102.0842.

Paper details

vlnc:W_qba66nce
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

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

  • Physics and Astronomy

ChatGPT, Codex, Gemini (multiple versions)

View statement
Looking for related papers…
Contents