Extensions of the Haldane gap proof: higher dimensions, near degeneracy, fermions and tensor-network states

The paper in detail

Write \(Z(L,\beta)=\operatorname{tr}e^{-\beta H_L}\) on a periodic chain. The proof assumes a Hermitian spatial transfer matrix \(X_\beta\), independent of \(L\), with \(Z(L,\beta)=\operatorname{tr}X_\beta^L\) for even \(L\). Two probability distributions then share the same normaliser: Gibbs weights \(w_b\propto e^{-\beta E_b}\), and transfer weights \(v_a\propto|\lambda_a|^L\). Their purities are ratios of partition functions, \(T=Z(L,2\beta)/Z(L,\beta)^2\) and \(S=Z(2L,\beta)/Z(L,\beta)^2\). A purity of the Gibbs weights near one forces the ground-state weight near one, and hence a gap.

Section III doubles each spatial length and \(\beta\) in turn. In \(d\) dimensions the test is one number, \(Q(L,\beta)=\mathcal Z(2L,2\beta)/\mathcal Z(L,\beta)^{2^{d+1}}\) on the hypercube. If \(Q(L_0,\beta_0)\ge1-1/(2a^2)\) with \(a=2^{d+1}-1\), then every even \(\ell\ge L_0\) has gap above \(\log2/\beta_0\). The Heisenberg bilayer with \(J/J_\perp\le1/10\) is certified this way.

Section IV treats an Ising-symmetric Hamiltonian, where two nearly degenerate states make the plain test fail. The fix uses the fact that inserting the symmetry generator \(U\) in the thermal trace becomes a spatial twist after space and time are exchanged. Summing each symmetry sector over both boundary twists, \(\mathcal Z_k=\operatorname{tr}A_0^L+k\operatorname{tr}A_1^L\), restores a valid transfer purity in the \(k=+1\) sector. The paper then derives a gap \(\Gamma_\ell>\log8/\beta_0\) above the pair and a splitting \(W_\ell\le\frac{12}{5\beta_0}e^{-(\log20)\ell/(2L_0)}\). In two dimensions the splitting improves to the area law \(e^{-cL^2}\).

For fermions, antiperiodic spatial boundaries make the rotated trace an ordinary trace, and a fermion-parity twist then reaches periodic boundaries. For PEPS, the squared norm plays the partition function and the inverse correlation length plays the gap. Bounds on the \(12\times12\) and \(24\times24\) norms of the Jian–Zaletel state give \(|\langle O_1O_2\rangle-\langle O_1\rangle\langle O_2\rangle|\le\|O_1\|\|O_2\|20^{-r/24}\). This excludes Néel, quadrupolar and dimer order.

Transferable methods

1. The purity bootstrap

   Doubling \(\beta\) squares the Gibbs weights, and doubling \(L\) squares the transfer weights. Squaring a nearly pure distribution makes it quadratically purer, so each round of alternating doublings squares the deficit while only doubling \(\beta\). The deficit therefore falls doubly exponentially, faster than \(1/\beta\) grows.

   If \(w\) has purity \(1-u\), its normalised square \(w^{(2)}=w^2/\operatorname{tr}w^2\) has purity \(1-u'\) with \[u'=\frac{2\sum_{a<b}w_a^2w_b^2}{(1-u)^2}\le\frac{u^2}{2(1-u)^2},\] because \(u=2\sum_{a<b}w_aw_b\) and \(\sum_{a<b}w_a^2w_b^2\le(\sum_{a<b}w_aw_b)^2\). The partition-function identity \(S(L,2\beta)=S(L,\beta)^2T(2L,\beta)/T(L,\beta)^2\) moves a bound across from one distribution to the other. Together they give the rule \[S(L,\beta),\,T(2L,\beta)\ge1-u\ \Longrightarrow\ S(2L,2\beta),\,T(4L,2\beta)\ge1-5u^2\qquad(u\le0.01).\]

   Iterating gives \(u_j=\tfrac15(5u_0)^{2^j}\le\tfrac15\,20^{-2^j}\) at \(\beta_j=2^j\beta_0\). Since \(T\le w_0\), the gap obeys \[\Delta\ge\frac1{\beta_j}\log\frac{1-u_j}{u_j}>\frac{2^j\log20}{2^j\beta_0}=\frac{\log20}{\beta_0},\] which is independent of size. Editorial check: on 50,000 random distributions the ratio \(u'/[u^2/2(1-u)^2]\) never exceeded one (equality for two outcomes), and the recursion values \(u_j\) match the closed form exactly from \(u_0=0.01\).

   How to reuse it

   Whenever a quantity is a ratio of partition functions in two directions of a lattice, look for a squaring inequality that improves a bound faster than the scale grows. The same pattern converts a finite computation into a statement about the infinite system, but it needs a reflection-positive, Hermitian transfer in every direction used.

2. Symmetry twists replace symmetry projections

   Projecting onto a symmetry sector breaks the purity interpretation, because the projected transfer trace can exceed its square. Summing over boundary twists implements the projection on the rotated side, where it is harmless. The two sectors' ratio then bounds the splitting of the nearly degenerate pair.

   For even \(L\), \(\mathcal Z_{\pm1}=\operatorname{tr}A_0^L\pm\operatorname{tr}A_1^L\) with Hermitian \(A_0,A_1\), so \(\mathcal Z_{+1}=\operatorname{tr}(A_0\oplus A_1)^L\) is an honest transfer trace. Since \((\operatorname{tr}A^L)^2\ge\operatorname{tr}A^{2L}\) for even \(L\), \[S_{+1}\le\frac{(\operatorname{tr}A_0^L)^2+(\operatorname{tr}A_1^L)^2}{(\operatorname{tr}A_0^L+\operatorname{tr}A_1^L)^2}=\frac12\Bigl[1+\Bigl(\frac{\mathcal Z_{-1}}{\mathcal Z_{+1}}\Bigr)^2\Bigr],\] so \(\sqrt{2S_{+1}-1}\le\mathcal Z_{-1}/\mathcal Z_{+1}\le1\). As the bootstrap drives \(S_{+1}\to1\), the two sectors carry equal weight, and the lowest energies in them converge.

   Editorial check of why projection alone fails: for the classical Ising chain the projected untwisted partition function is \(\tfrac12[(2\cosh\beta J)^L+(2\sinh\beta J)^L]\). At \(\beta J=1\) and \(L=8\) the would-be purity \(Z(2L)/Z(L)^2\) evaluates to 1.63, which is impossible for a purity.

   How to reuse it

   When a symmetry is spontaneously broken, work with the sum over its twisted boundary conditions rather than its projector: under space–time exchange the two are the same object, and the twisted sum keeps positivity. The approach should apply to any finite Abelian symmetry, including higher-form ones, whenever each twisted Hamiltonian's ground energy is no lower than the untwisted one.

Source: Sections II, III and IV.1; Appendix B.2. Based on the linked paper version. Derivations marked as editorial, and suggested uses in other problems, are editorial guidance rather than results of the paper.

Context & limitations

Why it matters. Rigorous gaps for non-frustration-free interacting models are rare. Here a computation on a \(4\times4\) and an \(8\times8\) lattice controls every larger system. The symmetry-twist step turns the method from a test for a unique gapped ground state into one for topological and symmetry-broken phases, and recovers the expected \(e^{-cL}\) versus \(e^{-cL^2}\) splittings for 1-form versus 0-form symmetry breaking. The PEPS result settles a question previously supported only numerically.

Limitations. Every application needs Hermitian spatial transfer matrices in each direction, which restricts the Hamiltonians and bases that can be used; the paper gives sufficient local conditions but not a general criterion. The certified parameter windows are far from the known phase boundaries, and the examples are, as the author says, models whose behaviour is already understood. Whether the parent Hamiltonian of the Jian–Zaletel state is gapped is left open. The finite partition-function bounds in Appendix E were not rechecked here.

Description based on Sections I–VI of the linked arXiv version, read 2026-11-07, including the one-dimensional and two-dimensional update rules, the twist argument of Section IV.1 and the statements of the four applications. Appendices A–E, which hold the finite-size bounds, were not checked line by line.