betaFeedback

WorksPhysics and Astronomy

Free fermions from integer arithmetic: a Galois-theoretic test for free-fermionic spectra

Abstract

The spectrum of a free-fermionic operator is built from a few single-particle quantities. For operators with rational entries this confines, as is known, the Galois group of the characteristic polynomial of each sector to a hyperoctahedral group, up to elementary abelian 22-groups. We turn this into a test for free-fermionic spectra that uses only integer arithmetic: Frobenius cycle types modulo primes can prove that a spectrum is not free-fermionic, and an exact recovery of the single-particle data, certified by power sums, confirms it when it is. As a worked example, the test recovers Onsager's dispersion relation from the characteristic polynomial of the Ising transfer matrix. As applications, we prove that the periodic Fendley chain with generic couplings admits no free-fermion solution of the Chapman–Elman–Mann form for a range of sizes, bearing on whether a simplicial clique is necessary for free-fermion solvability, and we detect hidden tensor-product and matchgate structure in any basis. The Galois group grades spectral structure: hyperoctahedral for free fermions, symmetric on subsets with particle-number conservation, apparently imprimitive for the partially decoupled Bethe equations of the exclusion process, and alternating within sectors for the interacting models tested.

Introduction

A free-fermionic model is exactly solvable because its many-body spectrum is built from a few single-particle energies. For the two-dimensional Ising model this is the content of Kaufman’s spinor solution [36] of Onsager’s model [46]: every eigenvalue of the row transfer matrix on a ring of LL sites has the form

λs=κ∏k=1Leskγk/2,s∈{±1}L(1)\lambda_s = \kappa\prod_{k=1}^{L} e^{s_k\gamma_k/2}, \qquad s \in\{\pm1\}^{L} \tag*{(1)}

with a fixed parity of ss in each spin-flip sector. The same structure governs matchgate circuits, which are classically simulable for this reason [57, 56, 31], and Hamiltonians that are free only after a nonlocal change of variables, the free fermions in disguise of Fendley [16], characterized graph-theoretically by Chapman, Flammia, Elman, and Mann [10, 12, 9].

Arranged by ss, the spectrum (1) is the rank-one tensor κ(u1,u1−1)⊗⋯⊗(uL,uL−1)\kappa(u_1,u_1^{-1}) \otimes\cdots\otimes(u_L,u_L^{-1}), uk=eγk/2u_k=e^{\gamma_k/2}, a point of the Segre variety (P1)L(\mathbb{P}^{1})^{L}, and its eigenvalues satisfy λsλt=λs′λt′\lambda_s\lambda_t=\lambda_{s'}\lambda_{t'} whenever s+t=s′+t′s+t=s'+t'. For an operator with rational entries every automorphism σ\sigma of Q\mathbb{Q} preserves these relations, so the Galois group of the characteristic polynomial is confined to the symmetries of the hypercube: for the square of the transfer matrix in a spin-flip sector it is a subgroup of the hyperoctahedral group W(BL)W(B_L), of order 2LL!2^L L!, and for the transfer matrix itself, whose eigenvalues involve the half-angles γk/2\gamma_k/2 and a possibly irrational κ\kappa, an extension of such a subgroup by an elementary abelian 2-group (Proposition 2.4); a generic matrix of size 2L2^L has the symmetric group S2LS_{2^L}. This confinement is a special case of a theorem of Jouve, Kowalski, and Zywina [30]: for an element gg of a reductive group, the Galois group of the characteristic polynomial of ρ(g)\rho(g) lies in the Weyl group extended by the Galois action on characters, and acts through the weights of ρ\rho. Free fermions are the case of Spin⁡(2L)\operatorname{Spin}(2L) in its spin representation, whose weights are the sign vectors (±12,…,±12)(\pm\frac{1}{2},\ldots,\pm\frac{1}{2}). With particle-number conservation the hidden group is U(L)U(L) acting on the NN-particle space ΛNCL\Lambda^{N}\mathbb{C}^{L}, and its Weyl group SLS_L acts on NN-element subsets.

This paper uses the confinement as a test that reads off the hidden group through its Weyl group, with free fermions as the main case. The Galois group of an integer polynomial is accessible through its factorizations modulo primes: by Chebotarev’s density theorem [53], the degrees of the irreducible factors modulo pp are cycle types of Frobenius elements, which equidistribute in the group. A cycle length incompatible with the allowed group orders, or a certificate that the group contains an alternating group AdA_d with d≥5d \ge5 and d>Ld > L, rules out free-fermionic structure. The test needs no fermionic representation, works in any basis, and costs a few factorizations of integer polynomials. It concerns the spectrum only: it detects free-fermionic spectral structure, not fermionic dynamics, and operators whose spectra are trivially of free form, such as commuting stabilizer Hamiltonians, pass it.

Results

Section 2 proves the hyperoctahedral lemma and derives rigorous criteria for non-freeness: for transfer matrices and their squares, for complex and orthogonal operators, for number-conserving free fermions and free parafermions, and for Pauli Hamiltonians with symmetry sectors. Section 3 gives a three-step procedure (screen, recover, certify) and works it out for the Ising transfer matrix, recovering Onsager’s dispersion relation from the integer characteristic polynomial and certifying the Kaufman form exactly for strips of width 3 to 7. Section 4 contains the applications. The first concerns claw-free frustration graphs without a simplicial clique, which lie outside the free-fermion solutions of Chapman, Elman, and Mann [9]: we prove that for generic couplings the periodic Fendley chain and further cycle powers admit no free-fermion solution of their form at a range of sizes (Theorem 4.1), and we certify 23 graphs on at most 9 vertices, 176 on 10, and 29 on 11. This complements exact free solutions reported for other graphs of this kind, including the periodic Fendley chain on 8 and 9 sites [2]. The second recovers hidden group structure, tensor factors and free-fermion modes, from the multiplicative relations among eigenvalues. The third concerns quantum circuits (§ 4.3). Section 5 shows how the Galois group grades spectral structure: interacting integrable models have alternating groups inside their symmetry sectors, in agreement with recent work on Bethe equations [6, 40], while the totally asymmetric exclusion process, whose Bethe equations partially decouple, appears to have imprimitive groups of intermediate size. Appendix A validates the test on free-fermion models from the literature: in every family the points flagged free are known free-fermion points or have roots-of-unity spectra, except for three-dimensional Ising bars with a Clifford background, which Appendix A explains through the Lie algebra of a Pauli orbit. At one of these points the hidden group is symplectic, which the Galois group cannot distinguish from free fermions; Table 6 summarizes the families. Appendix § 8 recovers the single-particle spectra of further known models (Table 7), and § 6 lists open questions.

settingflagged freeknown free or roots-of-unity points
Ising strips, real and imaginary fieldszero field; h=iπ/2h=i\pi/2planar zero field [46, 36]; iπ/2i\pi/2
eight-vertexa2+b2=c2+d2a^2+b^2=c^2+d^2Fan–Wu [14]
crossing and four-spin couplings(2)Fan–Wu via the vertex map
three-dimensional IsingWang–Hazzard locus[58]
braid representationsAlexander; Jones at k=3,4,6k=3,4,6free [35]; roots of unity [20]
circuitsmatchgates[57, 31]
number-conserving chainsSLS_L on subsetstight-binding chains

Table 6. Points flagged free in the main families tested; further models are in Table 5 and the text. Points with roots-of-unity spectra pass the screen without being free-fermionic in any useful sense; the bars with a Clifford background and generic bonds of Appendix A are explained there by the Lie algebra of a Pauli orbit, and one of them is symplectic, not free.

modelrecoveredagreement with the literaturecertificate
dimers, strip width 3≤L≤83 \le L \le8∏k(y−ηk2)\prod_k(y-\eta_k^2)ηk=cos⁡πkL+1\eta_k=\cos\frac{\pi k}{L+1}, exact [33, 55, 41]exact
transverse-field Ising chain, open, 4≤L≤64 \le L \le6∏k(y−εk2)\prod_k(y-\varepsilon_k^2)exact [42, 47]exact
XY chain, open, 4≤L≤64 \le L \le6∏k(y−εk2)\prod_k(y-\varepsilon_k^2)exact [42]exact
Kitaev honeycomb, flux-free sectorεq\varepsilon_q∣2(Jxeiq⋅n1+Jyeiq⋅n2+Jz)∣\lvert2(J_x e^{i\mathbf{q}\cdot\mathbf{n}_1}+J_y e^{i\mathbf{q}\cdot\mathbf{n}_2}+J_z)\rvert [38]numerical
chain in disguise, 8≤N≤158 \le N \le15∏k(y−εk2)\prod_k(y-\varepsilon_k^2)exact [16]exact
free parafermions, n=3n=3, L=3,4L=3,4∏k(y−εk3)\prod_k(y-\varepsilon_k^3)exact [15]exact
Fendley–Pozsgay chains, 3≤L≤73 \le L \le7∏k(y−εk2)\prod_k(y-\varepsilon_k^2)exact [17]none
Ising on the Möbius strip, 4≤L≤74 \le L \le7signed hypercubenumerical [43]numerical

Table 7. Recoveries of known single-particle spectra. “Exact” agreement means a rational identity between the recovered polynomial and the published one; an exact certificate means agreement of all power sums with the exact traces.

The arithmetic of spectra has been used before for particular models, for example the Galois action on the spectra of small Heisenberg rings [44] and integer factorization as a signature of Hilbert-space fragmentation [48]. The use of hyperoctahedral Galois groups as a test for free-fermionic structure, the certificates of non-freeness, and the three-step procedure appear to be new.

The hyperoctahedral lemma and certificates of non-freeness

Definition 2.1. Let χ∈Q[μ]\chi\in\mathbb{Q}[\mu]. We say χ\chi has free form of rank LL if there are κ∈Q×\kappa\in\mathbb{Q}^{\times}, a squarefree reciprocal polynomial P∈Q[u]P \in\mathbb{Q}[u] of degree 2L2L with roots u1±1,…,uL±1u_1^{\pm1},\ldots,u_L^{\pm1}, none equal to ±1\pm1, and a set SS of sign vectors, either all of {±1}L\{\pm1\}^{L} or its even-parity half, such that

χ(μ)=∏s∈S(μ−κ∏k=1Luksk).\chi(\mu)=\prod_{s\in S}\left(\mu-\kappa\prod_{k=1}^{L}u_k^{s_k}\right).

The arguments below use only the splitting field NN of PP. For uniform couplings PP need not be squarefree (γq=γ−q\gamma_q=\gamma_{-q} below), and χ\chi is then not literally of free form of lower rank; but NN is the splitting field of the squarefree part of PP with the factors u≠1u\ne1 removed, so parts (1) and (2) of Lemma 2.2 and Corollary 2.3 hold with the rank r≤Lr\le L of that part. For the Ising transfer matrix VV, χ\chi is the characteristic polynomial of V2V^2 on a spin-flip sector, PP is the characteristic polynomial of the single-particle matrix, and uk=eγku_k=e^{\gamma_k} [36]. For a free-fermion Hamiltonian the form is additive, Es=e0+∑kskϵkE_s=e_0+\sum_k s_k\epsilon_k, and the arguments below apply with sums in place of products.

Let W(BL)=C2≀SLW(B_L)=C_2\wr S_L act on {±1}L\{\pm1\}^L by (π,ϵ)⋅s=(ϵπ−1(k)sπ−1(k))k(\pi,\epsilon)\cdot s=(\epsilon_{\pi^{-1}(k)}s_{\pi^{-1}(k)})_k, and let W(DL)W(D_L) be the subgroup with ∏kϵk=1\prod_k\epsilon_k=1.

Lemma 2.2. Let χ\chi have free form of rank LL (Definition 2.1), with the roots μs=κ∏kuksk\mu_s=\kappa\prod_k u_k^{s_k}, s∈Ss\in S, pairwise distinct, and let NN be the splitting field of PP. Then:

(1) the splitting field of χ\chi is contained in NN, so Gal⁡(χ)\operatorname{Gal}(\chi) is a quotient of Gal⁡(N/Q)\operatorname{Gal}(N/\mathbb{Q});

(2) Gal⁡(N/Q)\operatorname{Gal}(N/\mathbb{Q}) embeds in W(BL)W(B_L) via σ(uk)=uπ(k)ϵk\sigma(u_k)=u_{\pi(k)}^{\epsilon_k};

(3) Gal⁡(N/Q)\operatorname{Gal}(N/\mathbb{Q}) acts on the roots of χ\chi through sign vectors: σ(μs)=μ(π,ϵ)⋅s\sigma(\mu_s)=\mu_{(\pi,\epsilon)\cdot s};

(4) if SS is the even half and no μs\mu_s with ss even equals κ∏kuktk\kappa\prod_k u_k^{t_k} with tt odd, the image lies in W(DL)W(D_L).

In particular every cycle of a Frobenius element at a prime of good reduction has length at most the largest order maxord⁡(BL)\operatorname{maxord}(B_L) of an element of W(BL)W(B_L); if S={±1}LS=\{\pm1\}^L, or under the hypothesis of (4), its cycle type is that of an element of W(BL)W(B_L), respectively W(DL)W(D_L), acting on $S.

Proof. Each μs\mu_s is a monomial in the roots of PP, which gives (1). An automorphism σ\sigma permutes the roots of PP and commutes with u↦u−1u\mapsto u^{-1}, so it permutes the LL pairs {uk,uk−1}\{u_k,u_k^{-1}\}, which are distinct because PP is squarefree with no root ±1\pm1. Recording the permutation π\pi of the pairs and whether uku_k goes to uπ(k)u_{\pi(k)} or to its inverse gives an injective homomorphism into W(BL)W(B_L), which is (2). Then σ(μs)=κ∏kuπ(k)ϵksk=κ∏jujϵπ−1(j)sπ−1(j)=μ(π,ϵ)⋅s\sigma(\mu_s)=\kappa\prod_k u_{\pi(k)}^{\epsilon_k s_k}=\kappa\prod_j u_j^{\epsilon_{\pi^{-1}(j)}s_{\pi^{-1}(j)}}=\mu_{(\pi,\epsilon)\cdot s}, which is (3). For (4), σ\sigma maps roots of χ\chi to roots of χ\chi, and (π,ϵ)(\pi,\epsilon) changes the parity of ss by ∏kϵk\prod_k\epsilon_k. □

Lemma 2.2 is the elementary core of a general statement. For a reductive group GG over Q\mathbb{Q} with a faithful representation ρ\rho, Jouve, Kowalski, and Zywina show that the Galois group of the splitting field of det⁡(T−ρ(g))\det(T-\rho(g)) is a subquotient of the group generated by the Weyl group and the Galois action on the characters of a maximal torus, that the eigenvalues are the values of the weights of ρ\rho at gg, and that Galois acts through its action on the weights [30 Prop. 2.1, Lemma 2.4]; random elements attain the whole group [30 Thm. 1.1] (see also [24]). For Spin⁡(2L)×Gm\operatorname{Spin}(2L)\times G_m in the spin representation this gives Lemma 2.2. Galois groups and multiplicative relations among eigenvalues are related in [39, 52].

Corollary 2.3 (non-freeness certificate). If an irreducible factor of χ\chi of degree d≥5d\ge5 has Galois group containing AdA_d and d>Ld>L, then χ\chi divides no polynomial of free form of rank at most $L.

Proof. Parts (1) and (2) of Lemma 2.2 do not use that the μs\mu_s are distinct: the splitting field of every divisor of a polynomial of free form of rank LL lies in NN, so the Galois group of every factor is a quotient of a subgroup of W(BL)=C2L⋊SLW(B_L)=C_2^L\rtimes S_L. Since the factor is irreducible of degree dd, its Galois group is a subgroup of SdS_d; containing AdA_d, it is AdA_d or SdS_d, and for d≥5d\geq5 it has the simple group AdA_d as a composition factor, so AdA_d is a composition factor of a subgroup of W(BL)W(B_L). A subgroup HH has the abelian normal subgroup H∩C2LH\cap C_2^L with quotient in SLS_L, so AdA_d is a composition factor of a subgroup of SLS_L, and d!/2≤L!d!/2\leq L!. This forces d≤Ld\leq L. □

The same argument applies to every bound below: in each case the nonabelian composition factors of the relevant Galois group are those of subgroups of SLS_L (of SαS_\alpha or StS_t in Proposition 2.9), so the order bounds stated there are only needed for element orders, and an alternating group AdA_d, d≥5d\geq5, is excluded as soon as d>Ld>L (respectively d>αd>\alpha and d>td>t). In each application the rank is fixed in advance by the size of the system: for a transfer matrix on LL sites, or a circuit on LL qubits, it is LL, the number of Majorana pairs, and for a Hamiltonian it is the number of modes; certificates exclude free form of this rank. The bounds below apply in the same way to sectors and factors: they constrain every polynomial whose roots are of the given form, not only the full characteristic polynomial. The operator itself, not its square, is usually given, and its entries may be complex.

Proposition 2.4. Suppose every eigenvalue of a rational matrix is of the form λs=κ∏kvksk\lambda_s=\kappa\prod_k v_k^{s_k} for some s∈{±1}Ls\in\{\pm1\}^L, with κ2∈Q\kappa^2\in\mathbb{Q} and vk±2v_k^{\pm2} the roots of a rational reciprocal polynomial of degree 2L2L. Then all eigenvalues lie in a Galois extension of Q\mathbb{Q} of degree at most GL=22L+1L!G_L=2^{2L+1}L!, in whose Galois group every element has order dividing 2a2a for the order aa of some element of W(BL)W(B_L). If instead the matrix is the realification of a matrix over Q(i)\mathbb{Q}(i) whose eigenvalues have this form with κ2∈Q(i)\kappa^2\in\mathbb{Q}(i) and vk±2v_k^{\pm2} the roots of a reciprocal polynomial over Q(i)\mathbb{Q}(i), then all eigenvalues lie in a Galois extension of Q\mathbb{Q} of degree at most 2GL22G_L^2 in whose Galois group every element has order dividing 4lcm⁡(a,b)4\operatorname{lcm}(a,b) for the orders a,ba,b of some elements of W(BL)W(B_L). Consequently a factor of degree d≥5d\geq5 with Galois group containing AdA_d and d>Ld>L, or a Frobenius cycle whose length divides no such 2a2a (respectively no such 4lcm⁡(a,b)4\operatorname{lcm}(a,b)), for instance, for a single set, a cycle longer than 2maxord⁡(BL)2\operatorname{maxord}(B_L) in the first case, proves that the eigenvalues are not all of this form.

Proof. Let NN be the splitting field of the reciprocal polynomial, of degree at most 2LL!2^L L!. The set {vk±2}\{v_k^{\pm2}\} is Galois stable, so N(v1,…,vL,κ)N(v_1,\ldots,v_L,\kappa) is Galois over Q\mathbb{Q} of degree at most 2L+1⋅2LL!2^{L+1}\cdot2^L L!, and the kernel of its Galois group over Gal⁡(N/Q)\operatorname{Gal}(N/\mathbb{Q}) is an elementary abelian 22-group; if the image of gg has order aa, then g2a=1g^{2a}=1. In the second case the same construction over Q(i)\mathbb{Q}(i) gives a Galois extension KK of Q(i)\mathbb{Q}(i) of degree at most GLG_L containing the eigenvalues of the complex matrix. The eigenvalues of its realification are these and their complex conjugates, which lie in KK‾K\overline{K}, a Galois extension of Q\mathbb{Q} of degree at most 2GL22G_L^2. The group Gal⁡(KK‾/Q(i))\operatorname{Gal}(K\overline{K}/\mathbb{Q}(i)) embeds in Gal⁡(K/Q(i))×Gal⁡(K‾/Q(i))\operatorname{Gal}(K/\mathbb{Q}(i))\times\operatorname{Gal}(\overline{K}/\mathbb{Q}(i)), so its elements have orders dividing lcm⁡(2a,2b)\operatorname{lcm}(2a,2b), and every element of Gal⁡(KK‾/Q)\operatorname{Gal}(K\overline{K}/\mathbb{Q}) has its square in this subgroup. In both cases the nonabelian composition factors are those of subgroups of SLS_L, and AdA_d with d≥5d\geq5 is one only if d≤Ld\leq L, as in Corollary 2.3. □

Remark 2.5 (two sectors). Suppose the eigenvalues form two sets as in Proposition 2.4, each with its own rational data κi2∈Q\kappa_i^2\in\mathbb{Q} and Pi∈Q[u]P_i\in\mathbb{Q}[u] (respectively over Q(i)\mathbb{Q}(i)), as for the two parity sectors of a transfer matrix, with no assumption that the sectors are defined over Q\mathbb{Q}. The argument of the proof, applied to the compositum of the two fields, gives element orders dividing 2lcm⁡(a,b)2\operatorname{lcm}(a,b), respectively 4lcm⁡(a,b,c,d)4\operatorname{lcm}(a,b,c,d), for orders a,b,c,da,b,c,d of elements of W(BL)W(B_L), and the composition-factor criterion is unchanged. Every cycle-length certificate in this paper also satisfies this stronger condition.

Remark 2.6 (orthogonal operators). If M=cUM=cU with UU real orthogonal and c2∈Qc^2\in\mathbb{Q}, then after removal of the factor x2−c2x^2-c^2 (of the factors x∓cx\mp c if c∈Qc\in\mathbb{Q}) the characteristic polynomial is f(x)=xmh(x+c2/x)f(x)=x^m h(x+c^2/x), and the Galois group of ff is imprimitive with blocks {x,c2/x}\{x,c^2/x\}, so it never contains Adeg⁡fA_{\deg f}. Proposition 2.4 applies instead to the half-degree polynomial hh, whose splitting field lies in that of ff: a factor of hh of degree d≥5d\geq5 containing AdA_d with d>Ld>L certifies non-freeness.

Proposition 2.7 (number-conserving free fermions). Suppose every eigenvalue of a rational matrix is EA=∑k∈AεkE_A=\sum_{k\in A}\varepsilon_k for an NN-element subset AA of {1,…,L}\{1,\ldots,L\}, where the εk\varepsilon_k are the distinct roots of a rational polynomial QQ of degree LL. Then the Galois group of the characteristic polynomial is a quotient of Gal⁡(Q)≤SL\operatorname{Gal}(Q)\leq S_L. If the eigenvalues are the EAE_A for all AA and are pairwise distinct, it acts on them through the action of SLS_L on NN-element subsets.

Proof. Each EAE_A lies in the splitting field of QQ. An automorphism σ\sigma permutes the roots of QQ by a permutation π\pi, and σ(EA)=Eπ(A)\sigma(E_A)=E_{\pi(A)}. □

This is the case of U(L)U(L) acting on ΛNCL\Lambda^N\mathbb{C}^L in [30 Lemma 2.4], whose Weyl group is SLS_L. As in Corollary 2.3, a factor of degree d≥5d\geq5 whose Galois group contains AdA_d with d>Ld>L excludes this form; § 5 confirms the action on subsets for tight-binding chains.

Remark 2.8 (free parafermions). For free parafermions with Z3\mathbb{Z}_3 symmetry [15] the energies are E=e0+∑kωakεkE=e_0+\sum_k\omega^{a_k}\varepsilon_k, a∈{0,1,2}La\in\{0,1,2\}^L, ω=e2πi/3\omega=e^{2\pi i/3}. If e0∈Q(ω)e_0\in\mathbb{Q}(\omega) and the εk3\varepsilon_k^3 are the roots of a polynomial of degree LL over Q(ω)\mathbb{Q}(\omega), then, as in Proposition 2.4, every energy lies in the field generated over Q(ω)\mathbb{Q}(\omega) by the cube roots of the roots of this polynomial and of its complex conjugate, a Galois extension of Q\mathbb{Q} of degree at most 2(3LL!)22(3^L L!)^2 whose nonabelian composition factors are those of subgroups of SLS_L; a factor of degree d≥5d\geq5 whose Galois group contains AdA_d with d>Ld>L excludes this form.

Pauli Hamiltonians

Let H=∑jbjPjH=\sum_j b_jP_j be a sum of Pauli strings with frustration graph GG (an edge between anticommuting strings). In the free-fermion solutions of Chapman, Elman, and Mann [9 Sec. I.A, Result 1], each symmetry sector, a joint eigenspace of commuting conserved charges, carries an additive free spectrum ∑j=1αsjεj\sum_{j=1}^{\alpha}s_j\varepsilon_j with α=α(G)\alpha=\alpha(G) modes, the independence number of GG. Every Pauli realization of GG is a representation of the twisted group algebra of GG; its irreducible modules, which we call central sectors, are labeled by the eigenvalues of central Pauli products, are defined over Q(i)\mathbb{Q}(i), and have dimension D=2r/2D=2^{r/2} with rr the rank of the anticommutation form. For rational couplings the restriction of HH to a central sector is a Hermitian matrix over Q(i)\mathbb{Q}(i), so its characteristic polynomial lies in Q[E]\mathbb{Q}[E]. We say that a central sector has the Chapman–Elman–Mann form if it is the direct sum of the joint eigenspaces JJ of commuting operators with entries in Q(i)\mathbb{Q}(i) that lie in the algebra generated by the terms of HH and commute with HH, as the conserved charges of [9] do, and if on each JJ the spectrum is the multiset of the numbers ∑jsjεJ,j\sum_j s_j\varepsilon_{J,j}, with no constant term, over all s∈{±1}αs\in\{\pm1\}^{\alpha} or over a parity half, each ss counted at least once, where ±εJ,j\pm\varepsilon_{J,j} are the roots of a polynomial whose coefficients lie in the field generated over Q(i)\mathbb{Q}(i) by the eigenvalues labeling JJ. Allowing parity halves only weakens the criterion below; each JJ has at least 2α−12^{\alpha-1} states.

Proposition 2.9. Suppose a central sector of dimension DD has the Chapman–Elman–Mann form. If a factor of degree d≥5d\geq5 of its characteristic polynomial, in EE or in E2E^2, has Galois group containing AdA_d, then d≤αd\leq\alpha or d2α−1≤Dd2^{\alpha-1}\leq D. If α\alpha is even, the set of eigenvalues of the sector is symmetric under E↦−EE\mapsto-E.

Proof. Work over Q(i)\mathbb{Q}(i). The Galois group of the factor over Q(i)\mathbb{Q}(i) is a subgroup of index at most 22 of its group over Q\mathbb{Q}, so it still contains AdA_d, and the factor remains irreducible over Q(i)\mathbb{Q}(i). Let EE be an eigenvalue in the subspace JJ such that EE or E2E^2 is a root of the factor, let FF be the field generated over Q(i)\mathbb{Q}(i) by the labels of JJ, and let t=[F:Q(i)]t=[F:\mathbb{Q}(i)]. An automorphism of Q‾\overline{\mathbb{Q}} over Q(i)\mathbb{Q}(i) fixes the central sector and maps JJ to the joint eigenspace with the conjugate labels; the tt embeddings of FF give tt distinct joint eigenspaces, each of dimension at least 2α−12^{\alpha-1}, so t2α−1≤Dt2^{\alpha-1}\le D. The splitting field over FF of the polynomial with roots ±εJ,j\pm\varepsilon_{J,j} contains EE and E2E^2, and as in Lemma 2.2 its Galois group is isomorphic to a subgroup of W(Bα)W(B_\alpha). (Repeated or vanishing εJ,j\varepsilon_{J,j} only shrink this group.) Its Galois closure Ω\Omega over Q(i)\mathbb{Q}(i) is the compositum of its images τ(M)\tau(M) under the tt embeddings τ\tau of FF, where MM denotes the splitting field. The group Gal⁡(Ω/Q(i))\operatorname{Gal}(\Omega/\mathbb{Q}(i)) permutes the embeddings by composition, with image in StS_t; its kernel KK fixes every τ(F)\tau(F) pointwise, so it maps each τ(M)\tau(M), which is Galois over τ(F)\tau(F), into itself, and KK embeds in ∏τGal⁡(τ(M)/τ(F))≅Gal⁡(M/F)t\prod_\tau\operatorname{Gal}(\tau(M)/\tau(F))\cong\operatorname{Gal}(M/F)^t. The Galois group of the factor over Q(i)\mathbb{Q}(i) is a quotient of Gal⁡(Ω/Q(i))\operatorname{Gal}(\Omega/\mathbb{Q}(i)) and contains the simple group AdA_d, so AdA_d is a composition factor of KK or of a subgroup of StS_t, hence of a subgroup of W(Bα)W(B_\alpha) or of StS_t. As in Corollary 2.3, d≤αd\le\alpha or d≤t≤D/2α−1d\le t\le D/2^{\alpha-1}. For even α\alpha the map s↦−ss\mapsto-s preserves parity, so the set of eigenvalues on each JJ, and hence on the sector, is symmetric. □

We therefore certify non-freeness of the Chapman–Elman–Mann form when

d>αandd2α−1>D,d>\alpha\qquad\text{and}\qquad d2^{\alpha-1}>D,

or, for even α\alpha, when the set of eigenvalues of a central sector is not symmetric under E↦−EE\mapsto-E, that is, when its squarefree part χred\chi_{\mathrm{red}} does not satisfy χred(−E)=±χred(E)\chi_{\mathrm{red}}(-E)=\pm\chi_{\mathrm{red}}(E).

Certifying alternating groups

A primitive permutation group of degree dd containing a cycle of prime length p≤d−3p\le d-3 contains AdA_d, by a theorem of Jordan [11 Thm. 3.3E]. A transitive group containing a pp-cycle cc with p>d/2p>d/2 is primitive. Indeed, cc permutes the blocks of a block system in orbits of size 11 or pp. A block fixed by cc that contains a point moved by cc contains all pp such points, so it has more than d/2d/2 points and is everything; otherwise the moved points lie in at least p>d/2p>d/2 distinct blocks, which are then singletons. A Frobenius element with a single cycle of prime length pp, d/2<p<d−2d/2<p<d-2, and all other cycle lengths prime to pp has a power that is a pp-cycle, so its observation certifies AdA_d in the Galois group of an irreducible factor of degree dd. The range d/2<p<d−2d/2<p<d-2 contains a prime only for d≥8d\ge8; smaller factors are certified only through cycle lengths.

Screen, recover, certify

Let MM be a square integer matrix and χ\chi its characteristic polynomial, restricted to a symmetry sector when one is known.

  1. Screen. Remove zero roots and repeated factors, replace f(x)=g(xk)f(x)=g(x^k) by gg when all exponents are divisible by kk, and factor over Q\mathbb{Q}. For each factor of degree at least 44, factor modulo 8080 to 400400 primes of good reduction, record the cycle types, and test the cycle lengths and Jordan’s criterion. If a criterion of Section 2 applies, stop: the spectrum is not free-fermionic, with proof. Otherwise the point is a free candidate.

  1. Recover. Compute the roots to high precision. Within a sector, the ratios μs/μt\mu_s/\mu_t for s,ts,t differing in two positions i,ji,j are the most frequent nontrivial ratios, of the form ui±2uj±2u_i^{\pm2}u_j^{\pm2}; for S={±1}LS=\{\pm1\}^L the ratios ui±2u_i^{\pm2} for one position are more frequent still and give the uiu_i directly. Choosing rij=ui2uj2r_{ij}=u_i^2u_j^2 consistently among them, ui4=rijrik/rjku_i^4=r_{ij}r_{ik}/r_{jk}, and the remaining uku_k are peeled off one at a time. Accept the labeling only if it regenerates the whole spectrum, then recognize the coefficients of PP and κ\kappa as rationals.

(3) Certify. Compare exact power sums. For S={±1}LS=\{\pm1\}^{L}, ∑sμsm=κm∏k(ukm+uk−m)=κmdet⁡(2Tm(C))\sum_s \mu_s^m=\kappa^m\prod_k(u_k^m+u_k^{-m})=\kappa^m\det(2T_m(C)), where CC is the companion matrix of the polynomial with roots ck=(uk+uk−1)/2c_k=(u_k+u_k^{-1})/2 and TmT_m is the Chebyshev polynomial. For the even half,

∑s evenμsm=κm2(det⁡(2Tm(C))+σdet⁡(2Um−1(C))),σ2=det⁡(C2−I),\sum_{s\ \mathrm{even}}\mu_s^m=\frac{\kappa^m}{2}\left(\det(2T_m(C))+\sigma\det(2U_{m-1}(C))\right),\qquad\sigma^2=\det(C^2-I),

with Um−1U_{m-1} the Chebyshev polynomial of the second kind; the sign of σ\sigma is fixed by the case m=1m=1. Equality with the exact traces Tr⁡Mm\operatorname{Tr}M^m for m=1,…,deg⁡χm=1,\ldots,\deg\chi gives χ\chi exactly, by Newton’s identities.

A certificate in step 1 proves that the spectrum at a given rational point is not of free form; steps 2 and 3, when they succeed, prove that it is. If neither applies, the procedure is inconclusive at that point. Step 1 alone has false positives: rational or cyclotomic spectra (finite-order and Clifford operators), accidental factorizations at special points, and sectors too small for any group to be excluded. Commuting stabilizer Hamiltonians such as the toric code [37] have integer spectra that are sums of independent signed couplings with parity constraints, and are therefore trivially of free form: the procedure concerns free-fermionic structure of the spectrum, not whether the dynamics is that of nontrivially coupled fermions. Step 2 is unambiguous for a simple spectrum; with repeated pairs, as for the uniform couplings of § 3.1, ratios coincide, the labeling is accepted only if it regenerates the spectrum, and step 3 certifies the result. Free form for generic parameters means free form over the function field Q(b)\mathbb{Q}(b) of the parameters bb, with all data, κ\kappa, PP, and the operators defining sectors, in Q(b)\mathbb{Q}(b) or Q(i)(b)\mathbb{Q}(i)(b); since Q(i)(b∗)≅Q(i)(b)\mathbb{Q}(i)(b^*)\cong\mathbb{Q}(i)(b) for algebraically independent values b∗b^*, it is equivalent to free form at any such values, with data in Q(i)(b∗)\mathbb{Q}(i)(b^*). If a family has free form for generic parameters, then at a rational parameter point the Galois group of each factor is a quotient of a subgroup of the generic Galois group (of a decomposition group at a prime of the integral closure above the maximal ideal of the point), and every bound above passes to such subquotients, since it bounds a composition factor or the order of an element [7 Ch. V, §2, Thm. 2]. Symmetry of the set of eigenvalues under E↦−EE\mapsto-E also passes to specializations, since the eigenvalues at the point are the specializations of the generic ones. A certificate at one rational point therefore excludes free form for generic parameters.

Worked example: the Ising transfer matrix

Take the zero-field Ising model on a ring of LL sites with bond weight x1+ss′x^{1+ss'} for an integer xx, so that eK=xe^K=x and the row transfer matrix V=VLV=V_L has integer entries. The global spin flip commutes with VV; in the basis of configurations with s1=+1s_1=+1, the sector matrices V±(r,r′)=V(r,r′)±V(r,−r′)V_{\pm}(r,r')=V(r,r')\pm V(r,-r') are integer matrices of size 2L−12^{L-1}.

Recovery. Step 2, applied to the roots of the sector polynomials at 1200 bits, returns for x=2x=2 and L=4,…,7L=4,\ldots,7

cosh⁡γq=289120−cos⁡q,\cosh\gamma_q=\frac{289}{120}-\cos q,

with q=(2k+1)π/Lq=(2k+1)\pi/L in the flip-even sector and q=2πk/Lq=2\pi k/L in the flip-odd sector, and κ2=x4L(x2−x−2)L\kappa^2=x^{4L}(x^2-x^{-2})^L. For L=5,6,7L=5,6,7 the recovered single-particle polynomials agree exactly with (−1)L21−L(TL(289120−c)±1)(-1)^L2^{1-L}(T_L(\frac{289}{120}-c)\pm1), with the upper sign in the flip-even sector; for L=7L=7 in the flip-even sector the polynomial in c=cosh⁡γc=\cosh\gamma is

(120c−409)(1728000c3−11620800c2+25041960c−17261509)23583180800000000.\frac{(120c-409)(1728000c^3-11620800c^2+25041960c-17261509)^2}{3583180800000000}.

For L=4L=4 the labeling of the flip-odd sector is ambiguous: γi↦12∑jγj−γi\gamma_i\mapsto\frac{1}{2}\sum_j\gamma_j-\gamma_i preserves the multiset of sums ∑kskγk\sum_k s_k\gamma_k over even ss, a reflection of triality for D4D_4. The result is Onsager’s dispersion relation, cosh⁡γq=cosh⁡2K∗cosh⁡2K−sinh⁡2K∗sinh⁡2Kcos⁡q\cosh\gamma_q = \cosh2K^* \cosh2K - \sinh2K^* \sinh2K \cos q with tanh⁡K∗=e−2K\tanh K^* = e^{-2K} [46, 36], obtained without a fermionic or spinor construction. For a disordered and frustrated strip (L=6L = 6, couplings random in sign and size across the row and constant along it), the single-particle polynomial in cc of each sector is an irreducible rational sextic; the two sextics differ only in their constant terms, and the Galois group of each is S6S_6, so the Galois group of each sector polynomial lies in C26⋊S6=W(B6)C_2^6 \rtimes S_6 = W(B_6).

Certificates. For x=2,3x = 2, 3 and L=3,…,7L = 3, \ldots, 7, step 3 certifies in both sectors

det⁡(μ−V±2)=∏s even(μ−κ2∏qesqγq)\det(\mu- V_{\pm}^{2}) = \prod_{s\ \mathrm{even}} \left(\mu- \kappa^{2}\prod_q e^{s q\gamma_q}\right)

with cosh⁡γq=A−cos⁡q\cosh\gamma_q = A - \cos q, A=cosh⁡22K/sinh⁡2KA = \cosh^{2} 2K / \sinh2K, antiperiodic momenta in the flip-even sector and periodic momenta in the flip-odd sector. All products over momenta are determinants of Chebyshev polynomials in the companion matrix of P(c)=(−1)L21−L(TL(A−c)±1)P(c) = (-1)^L 2^{1-L}(T_L(A-c) \pm1), upper sign in the flip-even sector, so no algebraic numbers appear. The quantity σ=∏qsinh⁡γq\sigma= \prod_q \sinh\gamma_q is rational: with A±1=cosh⁡θ±A \pm1 = \cosh\theta_{\pm} one has eθ±=e2K±2K∗e^{\theta_{\pm}} = e^{2K \pm2K^*} (for x=2,20/3x = 2, 20/3 and 12/512/5), and ∏q(cq2−1)\prod_q(c_q^2 - 1) is a square; for L=3L = 3 and x=2x = 2, σ=14874031/1728000\sigma= 14874031/1728000 and 12780719/172800012780719/1728000 in the two sectors. For x>1x > 1 the matrix VLV_L is similar to a positive definite matrix, so its spectrum is determined by that of VL2V_L^2. Since the spectrum of VLV_L determines Tr⁡VLM\operatorname{Tr} V_L^M for all MM and Tr⁡VLM\operatorname{Tr} V_L^M is symmetric in LL and MM on the torus, these certificates give the exact partition function of every L×ML \times M torus with 3≤min⁡(L,M)≤73 \le\min(L,M) \le7 at these couplings, and for 3≤L≤73 \le L \le7 its restriction to each boundary sector in the direction of length LL. A proof valid for all LL requires an identity uniform in LL, supplied in the known proofs by the spinor representation [36] or a determinant formula for the partition function.

Applications

Claw-free frustration graphs without a simplicial clique

Chapman and Flammia, and Elman, Chapman, and Flammia, showed that claw-free frustration graphs satisfying further conditions give free-fermion solutions [10, 12]. Chapman, Elman, and Mann proved that a claw-free graph with a simplicial clique, a clique KK such that for every j∈Kj \in K the neighborhood of jj induces a clique in G∖KG \setminus K, admits a free-fermion solution, with energies given by the roots of independence polynomials [9 Sec. I.A, Result 1]; whether the simplicial clique is necessary is open [9 Sec. IX] (see also [50]). The spectrum depends on the realization only through the central sector [9 Sec. VIII], so the central sectors of the realization with one qubit per vertex cover all realizations. Graphs that become simplicial after twin collapse are free [49] and are excluded. Exact free solutions of the Chapman–Elman–Mann form for 107 graphs outside these classes, among them C82C_8^2 and C92C_9^2, are reported in [2].

Small graphs. There are 14, 50, 191, 881, 4944 connected claw-free graphs on 5, . . . , 9 vertices, of which 0, 0, 2, 13, 186 have no simplicial clique; all of these contain even holes, and 173 remain after twin collapse. Each was tested in every central sector for three random choices of integer couplings with Proposition 2.9 (Table 1). No control with a simplicial clique was ever certified, and none of the 106 graphs of [2] with exact free certificates that lie in our sets on at most 10 vertices is certified. Among the extensions by one vertex of the undecided graphs on 9 vertices, 176 of the 2624 graphs on 10 vertices are certified, 799 have a factor containing AdA_d to which Proposition 2.9 does not apply, 192 are consistent with free form, and 1457 give the test no power; in a sample of 800 graphs on 11 vertices with α≥3\alpha\ge3, 29 are certified, 582 are consistent, and 189 give the test no power. Here a graph is consistent if some factor has degree d≥5d \ge5 with d>αd > \alpha and d 2α−1>Dd\,2^{\alpha-1} > D, so that Proposition 2.9 could apply, but no certificate is found; otherwise the test has no power beyond the symmetry test for even α\alpha, which these graphs pass.

The classes are assigned in the order certified, AdA_d uncertified, consistent, no power. A graph counts as certified if some choice of couplings certifies it; of the 192 consistent graphs, 190 are consistent at all three choices. Numerically, none of the 190 graphs on 10 vertices consistent at all three choices, nor 120 sampled on 11, has a sector spectrum that is a full additive hypercube or a union of two full 3-mode hypercubes; the free forms not excluded for them involve parity halves, which at these sizes impose almost no constraint.

classgraphscertifiedAdA_d, uncertifiedconsistentno power
no simplicial clique, n=7n = 711000
no simplicial clique, n=8n = 892304
no simplicial clique, n=9n = 91632048689
simplicial clique (controls)3000525
simplicial after twin collapse (controls)2800028
with a claw (controls)2013007

Table 1. Claw-free frustration graphs without a simplicial clique, n≤9n \le9, and controls. “AdA_d, uncertified”: some factor has Galois group containing AdA_d, but d 2α−1>Dd\,2^{\alpha-1} > D fails, so Proposition 2.9 does not apply. “Consistent” and “no power” are defined in the text.

Cycle powers. The square CM2C_M^2 of the cycle is the frustration graph of Fendley’s chain [16] with periodic boundary conditions, whose spectrum Zheng, Pozsgay, and Chen obtain from functional relations of a commuting family of transfer matrices, with free, nested free, or interacting spectra depending on the couplings [60] (see also [21]). Cycle powers CMkC_M^k are claw-free for every kk. For each of three random choices of integer couplings, every central sector is certified by Proposition 2.9: by a factor whose Galois group contains the alternating group (for C172C_{17}^2, of degree 128), or, for CM2C_M^2 with M=11,13,14M = 11, 13, 14, by the asymmetry of the spectrum. By the remark on specialization in § 3, this proves the following.

Theorem 4.1. For generic couplings, in particular for algebraically independent couplings, no Pauli realization of the frustration graph CM2C_M^2 of the periodic Fendley chain has a free-fermion solution of the Chapman–Elman–Mann form for M=7,10,11M = 7, 10, 11 and 13≤M≤1913 \le M \le19; the same holds for CM3C_M^3 with 13≤M≤1813 \le M \le18 and for CM4C_M^4 with M∈{11,13,14,15,17,18}M \in\{11, 13, 14, 15, 17, 18\}.

For CM2C_M^2 with M=8,9,12M = 8, 9, 12 the sector spectra are degenerate and small, and the test has no power; for M=8,9M = 8, 9 exact free solutions are reported in [2], which also finds numerically that C72C_7^2 and C102C_{10}^2 have none.

The undecided graphs are mostly those with α=2\alpha= 2. For them, if the spectrum is symmetric under E↦−EE \mapsto-E, the factors of the polynomial in E2E^2 have degree at most D/2D/2, and a factor in EE is either even, so imprimitive, or paired with its reflection, so of degree at most D/2D/2. Then either d 2α−1>Dd\,2^{\alpha-1} > D fails or AdA_d cannot occur, and deciding them requires a test of free form that is not vacuous on parity halves.

Hidden group structure from multiplicative relations

Galois automorphisms preserve every multiplicative relation ∏iλini=1\prod_i \lambda_i^{n_i} = 1 among eigenvalues [39, 52]. The relation lattice and the configuration of weights it induces identify the algebraic group in which an operator lives: a hypercube or half-cube for free fermions, a grid {1,…,a}×{1,…,b}\{1,\ldots,a\} \times\{1,\ldots,b\} for a tensor product A⊗BA \otimes B, a cross-polytope for a generic orthogonal or symplectic matrix, and no relations for a generic matrix. We compute the relations by lattice reduction on the logarithms of the eigenvalues at 2000 to 4000 bits, keep relations with residual at most 2−p/22^{-p/2} at a precision of pp bits (every accepted relation has residual at most 2−19972^{-1997}), and identify the configuration against a library of these shapes; the relations are verified numerically, not exactly. Table 2 shows that the factor structure is recovered from the characteristic polynomial alone, also after conjugation by a random unimodular matrix: the number of free-fermion modes, the sizes of hidden tensor factors, and mixtures of the two. Spectral methods cannot distinguish free fermions from a product of independent qubits, which give the same cube.

operatorexpected rankfoundconfiguration
A⊗BA \otimes B: 2×42 \times4, 3×33 \times3, 2×32 \times3, 4×44 \times4, 2×2×22 \times2 \times2∑iai−r+1\sum_i a_i-r+15,5,4,7,45,5,4,7,4grids (cube)
generic, size 6,9,16,326,9,16,32fullfullno relations
Ising, zero field, L=4,…,7L=4,\ldots,7L+1L+15,6,7,85,6,7,8half-cube DLD_L
Ising, real field, L=5L=532323232no relations
matchgate circuits, n=5,6n=5,6modes +1+15,75,7cube, D6D_6
XXZ, eight-vertex (integrable), L=6L=6fullfullno relations
Ising (L=4,5L=4,5) ⊗\otimes generic 3×33 \times3L+3L+37,87,8simplex ×\times half-cube

Table 2. Rank of the multiplicative group generated by the distinct eigenvalues and the recovered configuration. For a tensor product of rr factors of sizes a1,…,ara_1,\ldots,a_r the expected rank is ∑iai−r+1\sum_i a_i-r+1; “full” means equal to the number of distinct eigenvalues. Circuit matrices are integer multiples of rational orthogonal matrices, and the scale adds one to the rank; for odd nn, matchgate circuits have a zero mode that lowers it by one. The Ising entries use random couplings.

Quantum circuits

Circuits of real orthogonal two-qubit gates with rational entries (rational points on the circle) have rational unitaries. Matchgates G(A,B)G(A,B), acting as AA on span⁡{∣00⟩,∣11⟩}\operatorname{span}\{|00\rangle,|11\rangle\} and as BB on span⁡{∣01⟩,∣10⟩}\operatorname{span}\{|01\rangle,|10\rangle\} with det⁡A=det⁡B\det A=\det B, on nearest-neighbor qubits generate free-fermionic circuits [57, 56, 31]. On 5, 6, and 7 qubits, brickworks of matchgates are free, with longest cycles 4, 12, 12, also after conjugation by random local rotations or, on 5 and 6 qubits, by a random rational orthogonal matrix of full size, since the test depends only on the spectrum; so are next-nearest-neighbor matchgates routed through fermionic swaps and Floquet matchgate circuits. The screen flags 25 circuits: on each of 5, 6, and 7 qubits, brickworks of generic gates (longest cycles 30, 62, 126), a single generic gate repeated as a Floquet circuit, and matchgate circuits with one generic gate, with one SWAP, with next-nearest-neighbor matchgates without fermionic swaps, with gates G(A,B)G(A,B) having det⁡A=−det⁡B\det A=-\det B, with a single-qubit kick, or with alternate layers conjugated by a generic single-qubit rotation; and matchgates on the edges of a 2×32 \times3 grid. By Remark 2.6 all 25 are certified not free: each has a factor of the half-degree polynomial whose Galois group contains AdA_d with d>nd>n. For periodic brickworks at even nn, each parity sector is free with its own modes, as for the Ising sectors of § 3.1, and a hypercube labeling of each sector is recovered. Clifford-type circuits have roots-of-unity spectra and pass the screen; their spectra are highly degenerate, which is why step 1 alone is not a proof of freeness.

What the Galois group sees

The Galois groups of sector factors grade the algebraic structure of a spectrum (Table 3). The bounds in the first two rows are Lemma 2.2 and Proposition 2.7, and equality with the full group is observed through Frobenius statistics; the last row is certified in every case listed; the third is observed.

spectral structureGalois group of a sector factorexamples
free fermionsin W(BL)W(B_L) or W(DL)W(D_L)Ising strips, matchgate circuits, Fendley’s chains
free fermions, particle number conservedin SLS_L on NN-element subsetstight-binding chains, Alexander braid representations
partially decoupled Bethe equationsimprimitive (observed), blocks over real cyclotomic fieldstotally asymmetric exclusion
interacting, integrable or notcontains AdA_dXXZ, eight-vertex, symmetric exclusion, M1M_1, Baxter–Wu

Table 3. Galois groups of factors of sector polynomials.

Number conservation. For tight-binding chains with random integer hoppings and on-site energies, with (L,N)=(8,3)(L,N)=(8,3), (9,4)(9,4), and (10,3)(10,3), the NN-particle sector polynomial is irreducible of degree (LN)\binom{L}{N}, the longest Frobenius cycle equals Landau’s function g(L)=15,20,30g(L)=15,20,30, and over 1500 primes the frequencies of cycle types agree with the class distribution of SLS_L acting on NN-element subsets within total variation 0.034, 0.057, 0.045, with no Frobenius element outside the group. The Alexander braid representations of Appendix A show the same signature.

Interacting integrable models. Integrable models that are not free behave like generic ones inside their symmetry sectors (Table 4). In inhomogeneous eight-vertex models whose columns share Baxter’s invariants Δ=(a2+b2−c2−d2)/(2(ab+cd))\Delta=(a^2+b^2-c^2-d^2)/(2(ab+cd)) and Γ=(ab−cd)/(ab+cd)\Gamma=(ab-cd)/(ab+cd), so that the transfer matrices commute [13], and in the six-vertex model, the sector factors have Galois groups containing alternating groups of full degree, while free-fermion columns, a2+b2=c2+d2a^2+b^2=c^2+d^2 [14], give small groups.

columnsfactor degreeslargest factor
free fermion, all different44 (×4\times4), 1212 (×4\times4)longest cycle 4
integrable, fixed (Δ,Γ)(\Delta,\Gamma)12 or 16contains A12A_{12}, A16A_{16}
six-vertex, fixed Δ\Delta1, 6, 10, 15contains A15A_{15}
generic16 (×4\times4)contains A16A_{16}

Table 4. Inhomogeneous eight-vertex transfer matrices, L=6L=6, 200 primes per factor, with, in the integrable rows, Δ≠0\Delta\ne0 and, except in the six-vertex row (Γ=1\Gamma=1), away from the degenerate values Δ=±1\Delta=\pm1, Γ=±1\Gamma=\pm1, Δ=±Γ\Delta=\pm\Gamma. Containment of A12A_{12}, A15A_{15}, A16A_{16} certifies non-freeness by Proposition 2.4, since 12,15,16>612,15,16>6.

The lattice supersymmetric M1M_1 model [45, 18], whose chain is equivalent to the XXZ chain at Δ=−1/2\Delta=-1/2, shows its supersymmetry exactly in the characteristic polynomial: every nonzero factor of the sector with ff fermions divides the polynomial of a sector with f±1f\pm1 fermions, and the weighted count of zero modes equals the Witten index. Its sectors on chains, ladders, a triangular strip, a 3×43\times4 grid, and trees nevertheless contain alternating groups, for example A44A_{44} and A76A_{76} on an open chain of 12 sites. These results agree with the large Galois groups of Bethe equations for the Heisenberg chain [6] and with the theorem that the Galois group of the Bethe equations of the GLd\mathrm{GL}_d Gaudin model with one-box weights at all sites contains the alternating group whenever the target shape is neither a hook nor self-transpose [40 Thm. C]. The symmetric exclusion process on 14 sites with 7 particles gives A32A_{32}, A27A_{27}, and A20A_{20} in the sector of momentum 0.

The totally asymmetric exclusion process. On a ring of LL sites with NN particles, the Bethe equations of this process decouple, given one collective parameter [26, 25], and its transition probabilities are determinants [51]; the generator has integer entries. Over Q\mathbb{Q} the NN-particle sector splits into norms of momentum sectors, so we use the rational sectors of momentum 0 and π\pi. In every case examined these contain factors whose Galois groups contain alternating groups far above the free bounds, for example A63A_{63} for (L,N)=(14,7)(L,N)=(14,7), A80A_{80} for (15,5)(15,5), and A131A_{131} for (13,6)(13,6), which proves that no sector spectrum is of free form. Other factors are not alternating. For (L,N)=(14,7)(L,N)=(14,7) the sector of momentum 0 has a factor of degree 84 with no Jordan certificate over 600 primes, whose Frobenius cycle types are compatible only with 3 blocks of 28; among 400 further primes, at each of the 267 inert in Q(cos⁡2π/7)\mathbb{Q}(\cos2\pi/7) every cycle length is divisible by 3, the pattern of a norm from Q(cos⁡2π/7)\mathbb{Q}(\cos2\pi/7). For (15,5)(15,5) factors of degree 90 and 30 are compatible only with 2 blocks and with norms from Q(5)\mathbb{Q}(\sqrt{5}), at the 203 inert primes among 400. In both cases the block field is Q(ζg)+\mathbb{Q}(\zeta_g)^+ with g=gcd⁡(L,N)g=\gcd(L,N). No blocks occur for (13,4)(13,4), (15,3)(15,3), (16,4)(16,4), and (12,6)(12,6), where Q(ζg)+=Q\mathbb{Q}(\zeta_g)^+=\mathbb{Q} and all large factors contain alternating groups, nor for (10,5)(10,5), where the sector of momentum 0 has factors of degree 15, containing A15A_{15}, and of degree 5. These block structures are observed, not proved. Golinelli and Mallick explain high-order degeneracies of the spectrum of this process by a hidden symmetry of its Bethe ansatz [22] (see also [23]); whether that symmetry accounts for these block structures is open.

Integrability organizes the spectrum into sectors and commuting families. Free fermions in addition make every eigenvalue a monomial in a few algebraic numbers, which confines the Galois group to a Weyl group; partial decoupling of Bethe equations appears to leave an intermediate, imprimitive trace; in the other interacting models tested no trace remains. The test is therefore a test for free-fermionic spectral structure, in any basis and in disguise, and not for integrability. Like any test through Weyl groups, it cannot distinguish types BLB_L and CLC_L, whose Weyl groups coincide: a symplectic hidden group, with eigenvalues of the form cμi±1c\mu_i^{\pm1}, can pass the screen. Appendix A gives an example, which the relation lattice of § 4.2 separates.

The analogy with zeros of LL-functions is close: there Frobenius elements equidistribute in a monodromy group whose type determines the random-matrix statistics [34]; here the Frobenius statistics of a single characteristic polynomial reflect its Galois group, and in the examples above, at random parameters, they are consistent with the full Weyl group of the hidden group: W(DL)W(D_L) for free fermions and SLS_L with particle-number conservation.

Outlook

The converse of Lemma 2.2 fails for individual matrices, since a matrix with rational eigenvalues has trivial Galois group. For families one may ask whether a Galois action at generic parameter points through a subgroup of W(BL)W(B_L) acting on sign vectors, with each sector of size 2L2^L or 2L−12^{L-1}, forces free form identically in the parameters. The symplectic example of Appendix AA shows that abstract embeddability of the Galois groups in W(BL)W(B_L) does not suffice; whether its Galois action is of this finer form is not settled by our data. More generally the method reads off a hidden reductive group through its Weyl group and weights: Spin⁡(2L)\operatorname{Spin}(2L) for free fermions, U(L)U(L) with particle-number conservation, the reflection group C3≀SLC_3 \wr S_L for free parafermions with Z3\mathbb{Z}_3 symmetry, and products of general linear groups for hidden tensor products. The imprimitive groups of the exclusion process suggest a finer invariant that measures partial decoupling of Bethe equations; proving the block structure of § 5 is open. The limiting factor is size: the screen has no power when all factors have small degree, which leaves part of § 4.1 undecided.

Appendix A. Validation on known models

Unless stated otherwise, couplings are integers drawn at random across the row and constant along the transfer direction, which removes translation symmetry. “Free” means that the screen finds no certificate under the criteria of § 2; “not free” means certified by Corollary 2.3, Proposition 2.4, Remark 2.6, or Remark 2.8. The maximal element orders are maxord⁡(BL)=8,12,12,24,30\operatorname{maxord}(B_L) = 8, 12, 12, 24, 30 for L=4,…,8L = 4, \ldots, 8.

Ising strips and fields

In zero field, with couplings random in sign and size so that plaquettes are frustrated, each sector polynomial is irreducible of degree 2L−12^{L-1}, and the longest Frobenius cycle over 150 primes equals maxord⁡(BL)\operatorname{maxord}(B_L): 12 at L=6L = 6 and 24 at L=7L = 7. Over 3000 primes the frequencies of cycle types agree with the class distribution of W(DL)W(D_L) acting on the 2L−12^{L-1} even sign vectors, of order 2L−1L!2^{L-1}L! (23040 at L=6L = 6, 322560 at L=7L = 7), within total variation 0.02 to 0.034, with no Frobenius element outside the group. A real field makes the polynomial irreducible of degree 2L2^L, with a full 2L2^L-cycle; at L=6L = 6 its Galois group contains A64A_{64}. At the imaginary field h=iπ/2h = i\pi/2 the polynomial satisfies f(x)=g(x2)f(x) = g(x^2) and gg has the same hyperoctahedral image; at h=iπrh = i\pi r with r=1/3,1/4,1/6r = 1/3, 1/4, 1/6 the spectrum is not free. The Ising model on the Möbius strip and the Klein bottle [43] is free, as is a field iπ/2i\pi/2 on alternate sites; one non-planar chord per row gives A64A_{64} at L=7L = 7.

Eight-vertex models

Deforming free-fermion columns by di↦di+td_i \mapsto d_i + t, t∈{−6,…,6}t \in\{-6, \ldots, 6\}, the test is free exactly at t=0t = 0 and certifies non-freeness at the other 12 values; in a two-parameter deformation (t,u)∈{−2,…,2}2(t,u) \in\{-2, \ldots, 2\}^2 it is free exactly at the origin.

Crossing and four-spin couplings

For the square-lattice Ising model with nearest-neighbor couplings, crossing diagonal couplings L1L_1, L2L_2, and a four-spin coupling MM on each plaquette, the map of Ising models with two- and four-spin couplings to an eight-vertex model [32, 59] and the Fan–Wu condition [14] give, after the nearest-neighbor weights cancel, the free-fermion condition

e4M=cosh⁡2(L1−L2)cosh⁡2(L1+L2).(2)e^{4M} = \frac{\cosh2(L_1-L_2)}{\cosh2(L_1+L_2)}. \tag*{(2)}

We have not found condition (2) stated in this form in [32, 59, 14]. Without four-spin coupling it reduces to sinh⁡2L1sinh⁡2L2=0\sinh2L_1 \sinh2L_2 = 0, that is, to L1L2=0L_1L_2 = 0 for real couplings. On 15 integer points of (2), including sign weights (complex couplings), the test reports free; on 18 nearby points off it, it certifies non-freeness. On grids of 2744 points at zero field and at field iπ/2i\pi/2 (L=6L = 6), the 2640 points off (2) in each grid are certified not free; the other 104 points are degenerate solutions of (2), with a diagonal coupling or the four-spin weight trivial or a pure sign, and are flagged free, so the nontrivial part of the locus is tested only on the 15 points above. With both diagonal couplings neither trivial nor a pure sign and no four-spin term, all 36 points of a random-column scan (L=6L = 6) are certified not free; all 189 such points with uniform couplings at L=7L = 7 are certified not free. These nulls are consistent with Barahona’s theorem that spin-glass problems become hard off planar lattices or in a field [3].

Three-dimensional models

The three-dimensional model of Wang and Hazzard [58] is solvable at imaginary couplings K⊥=βh=iπ/4K_{\perp} = \beta h = i\pi/4. With per-link random weights on its locus the test reports free (factors of degree 8, longest cycle 6≤maxord⁡(B4)6 \le\operatorname{maxord}(B_{4}), with 4 modes per flux sector); off the locus it certifies non-freeness, with a factor of degree 256 containing A256A_{256}. Its bond graph is non-planar, but the bonds with non-Clifford weights form a perfect matching within each layer, and all three-dimensional connectivity runs through bonds and fields at iπ/4i\pi/4. In a scan of Ising models on the non-planar bars G×ZG \times\mathbb{Z} with cross-sections G=K4G = K_{4}, 2×32 \times3, prism, and 2×42 \times4, with real and Gaussian-integer couplings, all 563 points are certified not free, by alternating groups or by cycle lengths, such as 5, 7, or 9, that divide no element order allowed by Proposition 2.4, also with two sectors (Remark 2.5). With all couplings equal to iπ/4i\pi/4 the transfer matrix CC of the bars with cross-sections 2×32 \times3 and prism is a Clifford operator, and CkC^{k} is a scalar for some kk; this is a point of the tractable set of [28]. Generic bonds give a transfer matrix T=CBT = CB with BB a product of exponentials of Pauli strings, so TkT^{k} is a scalar times a product of exponentials of the Pauli strings in the orbit of those of BB under conjugation by CC. The Lie algebra generated by these strings is, by [1], a direct sum of free-fermionic, orthogonal, symplectic, or unitary algebras or of controlled versions of these (tensored with the group algebra of a group of commuting Pauli strings); TkT^{k} lies in the corresponding complexified group, which determines the structure of the spectrum of TkT^{k}, and hence that of TT up to kk-th roots of unity. With one generic bond per period, all 12 tests are flagged free, 6 of them with no factor of degree above 4; with two bonds, 4 of 12 are flagged free, 1 of them without power, and 8 are certified not free; with three or four bonds, all 24 are certified. The dimension of this Lie algebra separates the flagged from the certified points: at every certified point it is at least 240. At the flagged points it is abelian (7 points without power and 2 others), su(2)3\mathfrak{su}(2)^{3} in each sector of the global spin flip (3 points), or su(2)\mathfrak{su}(2) tensored with the group algebra of a group of 32 commuting Pauli strings (3 points, one with further central elements), all free-fermionic, except at one point with two bonds, where it is sp(8)\mathfrak{sp}(8) in each sector of the global spin flip. There each sector of TkT^{k} is the sum of four copies of the defining representation of Sp⁡(8)\operatorname{Sp}(8), with eigenvalues cμj±1c\mu_{j}^{\pm1}, j=1,…,4j = 1,\ldots,4, which are not of free form. The Galois test cannot distinguish this case, since the Weyl groups of types B4B_{4} and C4C_{4} coincide, but the relation lattice of § 4.2, applied to the transfer matrix itself, finds on its two factors of degree 16 a cross-polytope, as for an orthogonal or symplectic matrix, not a cube.

Calibrations

Table 5 lists further models known to be free or not.

modelsizeexpectedresult
transverse-field Ising chainN=5,…,8N = 5,\ldots,8freelongest cycles 12, 12, 24, 30
same, plus longitudinal fieldN=5,…,8N = 5,\ldots,8not freefull cycles, A2NA_{2N}
Fendley’s chainS=2,…,5S = 2,\ldots,5freelongest cycles 4, 4, 6, 12
∑bmzmzm+1xm+2\sum b_m z_m z_{m+1} x_{m+2} [16]modes
free parafermions [15]L=3,4L = 3,4freef(x)=g(x3)f(x) = g(x^{3}), within C3≀SLC_{3} \wr S_{L}
parafermions plus τ†\tau^{\dagger} termsL=3,4L = 3,4not freeA9,A27A_{9}, A_{27}

Table 5. Calibrations.

Table 5. Calibrations. For Hamiltonians the energies are additive and the bound is maxord⁡(BS)\operatorname{maxord}(B_{S}) in the number SS of modes. Fendley’s chain has no Jordan–Wigner map to free fermions and its frustration graph contains no claw; the control with a longitudinal field has a claw [10, 12].

Further models

The three-spin model of Baxter and Wu [5], integrable and not free, is certified not free on uniform strips of widths 7 and 8 (cycles 102 and 66, divisible by the primes 17 and 11, which divide no allowed element order, and A15A_{15}, A24A_{24}) and with random columns at widths 6 and 7 (A16A_{16}, A128A_{128}). The Kitaev–Heisenberg model [8] on a 2×42 \times4 ladder has four factors of degree 64 containing A64A_{64}; the pure Kitaev ladder has only factors of degree 4, where the test has no power. Among non-Hermitian chains, the transverse-field Ising chain with an imaginary transverse field (N=6,7N = 6, 7, realified) satisfies f(x)=g(x2)f(x) = g(x^2) with longest cycles 12 and 24, consistent with free form, while the Ising chain with an imaginary longitudinal field gives A64A_{64} and A128A_{128}. A tight-binding chain with a quasiperiodic on-site potential of rational frequency 1/61/6 (L=9L = 9, N=4N = 4) has a factor of degree 126 with longest cycle 20=g(9)20 = g(9) and total variation 0.117 from S9S_9 on 4-element subsets over 300 primes; the chain of Su, Schrieffer, and Heeger [54] on 8 sites has factors too small for the test. For the toric code [37] on a 2×22 \times2 torus with random couplings every factor is linear, and the spectrum is a constant plus the sums over two even half-cubes of the vertex and plaquette couplings.

Knot invariants and the Tutte plane

Braid-group representations on (C2)⊗n(\mathbb{C}^2)^{\otimes n} from the free-fermion six-vertex RR-matrix, which yields the Alexander–Conway polynomial [35], decompose into exterior powers of the reduced Burau matrix with factor degrees (n−1k)\binom{n-1}{k}, and their Galois groups lie in Sn−1S_{n-1} (longest cycles 6 for n=6,7n = 6, 7). Those from the Uq(sl2)U_q(\mathfrak{sl}_2) RR-matrix (Jones polynomial) at rational qq contain A9A_9 (n=6n = 6) and A14A_{14} (n=7n = 7) in all 16 long braid words tested. In the semisimple path representations at q=eiπ/kq = e^{i\pi/k}, all eigenvalues are roots of unity exactly for k∈{3,4,6}k \in\{3, 4, 6\} (and q=±1q = \pm1), and not for k=5,7,8,10k = 5, 7, 8, 10; rational q≠±1q \ne\pm1 never collapses. With Hecke parameter q2=e2πi/kq^2 = e^{2\pi i/k}, these are points at which the Jones polynomial is computable in polynomial time [28], and for k≥3k \ge3 they are the orders at which the image of the braid group in the Jones representations is finite for every nn: it is infinite modulo its center for k≥5k \ge5, k≠6,10k \ne6, 10, n≥3n \ge3, and for k=10k = 10, n≥5n \ge5, while the finite images at k=3,4,6k = 3, 4, 6 (and k=10k = 10, n=3,4n = 3, 4) were identified earlier by Jones and others [20 Thm. 2.3 and Sec. 4]. For Potts transfer matrices on planar strips the test is free for Q=2Q = 2, the Ising hyperbola of the Tutte plane, at widths W=6,7,8W = 6, 7, 8, with longest cycles 12, 24, 30 equal to maxord⁡(BW)\operatorname{maxord}(B_W), and certifies non-freeness for Q=3,4Q = 3, 4.

Supersymmetric models and hard squares

For the hard-core lattice gas at activity z=−1z = -1, which governs the ground states of the M1M_1 model [27], the row transfer matrix of hard squares has only roots of unity as eigenvalues for every width and boundary condition tested (periodic widths 6 to 13, open widths 2, 5, 8, 11), as known [19, 29, 4]; at z=±2z = \pm2, −12-\frac{1}{2}, −3-3, −13-\frac{1}{3}, 11 (periodic width 10) it does not, and the groups contain alternating groups. Hard hexagons, the honeycomb lattice, and the king’s graph at z=−1z = -1 give alternating groups.

Appendix B. Recoveries of known single-particle spectra

Section 3.1 recovers Onsager’s dispersion relation; this appendix collects the corresponding recoveries for further free-fermion and free-parafermion models. In each case below, step 2 of the procedure (recovery of the hypercube labeling) was applied to the exact characteristic or minimal polynomial with no input from the known solution, and the recovered single-particle polynomial, recognized over Q\mathbb{Q} (or over Q(ω)\mathbb{Q}(\omega)), was then compared with the published formula. Where stated, step 3 certified the result exactly: the power sums of the free form, computed from the recovered polynomial alone, agree with the exact traces for every exponent up to the dimension, so the characteristic polynomial is the composed product. Transfer matrices use the multiplicative form; Hamiltonians use the additive form Es=e0+∑kskϵkE_s = e_0 + \sum_k s_k \epsilon_k, for which the recovered object is ∏k(y−ϵk2)\prod_k(y - \epsilon_k^2).

Dimers. For the dimer transfer matrix on strips of width LL, the recovered numbers ηk=sinh⁡wk\eta_k=\sinh w_k satisfy ∏k(y−ηk2)=∏k=1L(y−cos⁡2πkL+1)\prod_k(y-\eta_k^2)=\prod_{k=1}^{L}(y-\cos^2\frac{\pi k}{L+1}) identically, the Chebyshev form of the solution of Kasteleyn, Temperley and Fisher in Lieb’s transfer-matrix formulation [33, 55, 41]; the certificate holds with κ2=1\kappa^2=1 for all m≤2Lm\le2^L.

Quantum chains. For open transverse-field Ising and XY chains with random integer couplings, the recovered polynomial is the characteristic polynomial of the Majorana single-particle matrix of Lieb, Schultz, and Mattis [42]; for uniform unit couplings, the critical chain with J=h=1J=h=1, it gives εk=2cos⁡πk2L+1\varepsilon_k=2\cos\frac{\pi k}{2L+1} to 140 digits, as in Pfeuty’s solution [47].

Kitaev’s honeycomb model. On 2×22\times2 and 3×23\times2 tori, the flux-free sector reproduces the Majorana dispersion of [38] to 10−1410^{-14}, the four loop sectors corresponding to momentum shifts by (0,0)(0, 0), (0,12)(0, \frac{1}{2}), (12,0)(\frac{1}{2}, 0), (12,12)(\frac{1}{2}, \frac{1}{2}); only the flux sector with all plaquette operators equal to +1+1 has this form.

Free fermions in disguise. For the chain of [16] with 8≤N≤158 \le N \le15 sites, the recovered polynomial is ySPM(1/y)y^{S}P_M(1/y) with PMP_M given by Fendley’s recursion, exactly, in every case, with exact certificates. Here no Jordan–Wigner transformation is available, and the single-particle data are obtained from integer arithmetic alone. The number of levels is S=⌈M/3⌉=⌊N/3⌋S = \lceil M/3\rceil= \lfloor N/3\rfloor, where M=N−2M = N - 2 is the number of terms, consistent with the example P4P_4 in [16].

Free fermions beyond Jordan–Wigner. The chains of Fendley and Pozsgay [17] have a frustration graph that contains claws, so the criterion of Chapman, Elman, and Mann does not apply to them directly. For 3≤L≤73 \le L \le7 (5 to 9 qubits) the screen has no power, since all factors have degree at most 4, but recovery gives r=⌊(L+1)/2⌋r = \lfloor(L + 1)/2\rfloor modes and ∏k(y−εk2)=yrP(1/y)\prod_k(y-\varepsilon_k^2) = y^rP(1/y) with PP as in [17 eq. (4.4)], exactly.

Free parafermions. For the three-state chain of [15], energies are sums ∑jωajεj\sum_j \omega^{a_j}\varepsilon_j with ω=e2πi/3\omega= e^{2\pi i/3}; the recovered ∏k(y−εk3)\prod_k(y-\varepsilon_k^3) agrees exactly with Fendley’s polynomial, and the traces in Z[ω]\mathbb{Z}[\omega] certify the spectrum for all m≤3Lm \le3^L. For uniform couplings and L=3L = 3, εk3=4cos⁡2πk7\varepsilon_k^3 = 4\cos^2\frac{\pi k}{7}.

The Möbius strip. The transfer matrix composed with the site reflection has the same moduli hypercube as the open strip; its signs are reproduced to 140 digits by reflection parities (−1)k+1(-1)^{k+1} of the modes, ordered by increasing wkw_k, times the reordering sign (−1)(∣A∣2)(-1)^{\binom{|A|}{2}}. Since this labeling depends on the ordering of real numbers, it is not defined over Q\mathbb{Q}, and the identification is numerical; compare [43].

References

References

  1. [1]G. Aguilar, S. Cichy, J. Eisert, and L. Bittel, Full classification of Pauli Lie algebras, preprint (2024), arXiv:2408.00081.arxiv.org/abs/2408.00081
  2. [2]Anonymous, Which frustration graphs give free fermions? A converse to the even-hole, claw-free theorem, and exact free sectors of the periodic Fendley chain, preprint, Zenodo (2026), doi:10.5281/zenodo.22994979.DOI
  3. [3]F. Barahona, On the computational complexity of Ising spin glass models, J. Phys. A 15 (1982), no. 10, 3241–3253, doi:10.1088/0305-4470/15/10/028.DOI
  4. [4]R. J. Baxter, Hard squares for z = −1, Ann. Comb. 15 (2011), no. 2, 185–195, doi:10.1007/s00026-011-0089-2.DOI
  5. [5]R. J. Baxter and F. Y. Wu, Exact solution of an Ising model with three-spin interactions on a triangular lattice, Phys. Rev. Lett. 31 (1973), 1294–1297, doi:10.1103/PhysRevLett.31.1294.DOI
  6. [6]O. R. Bellwood and W. J. Munro, Algebraic structure of finite-size Bethe solutions in the Heisenberg chain, preprint (2026), arXiv:2605.05589.arxiv.org/abs/2605.05589
  7. [7]N. Bourbaki, Commutative Algebra. Chapters 1–7, Elements of Mathematics, Springer, Berlin, 1989.
  8. [8]J. Chaloupka, G. Jackeli, and G. Khaliullin, Kitaev–Heisenberg model on a honeycomb lattice: possible exotic phases in iridium oxides A₂IrO₃, Phys. Rev. Lett. 105 (2010), 027204, doi:10.1103/PhysRevLett.105.027204.DOI
  9. [9]A. Chapman, S. J. Elman, and R. L. Mann, A unified graph-theoretic framework for free-fermion solvability, preprint (2023), arXiv:2305.15625.arxiv.org/abs/2305.15625
  10. [10]A. Chapman and S. T. Y. Chan, Characterization of solvable spin models via graph invariants, Quantum 4 (2020), 278, doi:10.22331/q-2020-06-04-278.DOI
  11. [11]J. D. Dixon and B. Mortimer, Permutation Groups, Graduate Texts in Mathematics, vol. 163, Springer, New York, 1996, doi:10.1007/978-1-4612-0731-3.
  12. [12]S. J. Elman, A. Chapman, and S. T. F. Chapman, Free fermions behind the disguise, Comm. Math. Phys. 388 (2021), no. 2, 969–1003, doi:10.1007/s00220-021-04220-w.DOI
  13. [13]K. Fabricius and B. M. McCoy, New developments in the eight vertex model, J. Stat. Phys. 111 (2003), no. 1–2, 323–337, doi:10.1023/A:1022213209641.DOI
  14. [14]C. Fan and F. Y. Wu, General lattice model of phase transitions, Phys. Rev. B 2 (1970), no. 3, 723–733, doi:10.1103/PhysRevB.2.723.DOI
  15. [15]P. Fendley, Free parafermions, J. Phys. A 47 (2014), no. 7, 075001, doi:10.1088/1751-8113/47/7/075001.DOI
  16. [16]P. Fendley, Free fermions in disguise, J. Phys. A 52 (2019), 335002, doi:10.1088/1751-8121/ab305d.DOI
  17. [17]P. Fendley and B. Pozsgay, Free fermions beyond Jordan and Wigner, SciPost Phys. 16 (2024), 102, doi:10.21468/SciPostPhys.16.4.102.DOI
  18. [18]P. Fendley, K. Schoutens, and J. de Boer, Lattice models with N = 2 supersymmetry, Phys. Rev. Lett. 90 (2003), 120402, doi:10.1103/PhysRevLett.90.120402.DOI
  19. [19]P. Fendley, K. Schoutens, and B. van Eerten, Hard squares with negative activity, J. Phys. A 38 (2005), no. 2, 315–322, doi:10.1088/0305-4470/38/2/002.DOI
  20. [20]M. H. Freedman, M. J. Larsen, and Z. Wang, The two-eigenvalue problem and density of Jones representation of braid groups, Comm. Math. Phys. 228 (2002), no. 1, 177–199, doi:10.1007/s002200200636.DOI
  21. [21]K. Fukai, B. Pozsgay, and I. Vona, Solving models with generalized free fermions II: path-product expansion and conserved charges, preprint (2026), arXiv:2605.31453.arxiv.org/abs/2605.31453
  22. [22]O. Golinelli and K. Mallick, Spectral degeneracies in the totally asymmetric exclusion process, J. Stat. Phys. 120 (2005), 779–798, doi:10.1007/s10955-005-6972-7.DOI
  23. [23]O. Golinelli and K. Mallick, The asymmetric simple exclusion process: an integrable model for non-equilibrium statistical mechanics, J. Phys. A 39 (2006), 12679–12705, doi:10.1088/0305-4470/39/41/S03.DOI
  24. [24]A. Gorodnik and A. Nevo, Splitting fields of elements in arithmetic groups, Math. Res. Lett. 18 (2011), no. 6, 1281–1288, doi:10.4310/MRL.2011.v18.n6.a16.DOI
  25. [25]L.-H. Gwa and H. Spohn, Bethe solution for the dynamical-scaling exponent of the noisy Burgers equation, Phys. Rev. A 46 (1992), 844–854, doi:10.1103/PhysRevA.46.844.DOI
  26. [26]L.-H. Gwa and H. Spohn, Six-vertex model, roughened surfaces, and an asymmetric spin Hamiltonian, Phys. Rev. Lett. 68 (1992), 725–728, doi:10.1103/PhysRevLett.68.725.DOI
  27. [27]L. Huijse and K. Schoutens, Supersymmetry, lattice fermions, independence complexes and cohomology theory, Adv. Theor. Math. Phys. 14 (2010), no. 2, 643–694, doi:10.4310/ATMP.2010.v14.n2.a8.DOI
  28. [28]F. Jaeger, D. L. Vertigan, and D. J. A. Welsh, On the computational complexity of the Jones and Tutte polynomials, Math. Proc. Cambridge Philos. Soc. 108 (1990), no. 1, 35–53, doi:10.1017/S0305004100068936.DOI
  29. [29]J. Jonsson, Hard squares with negative activity and rhombus tilings of the plane, Electron. J. Combin. 13 (2006), no. 1, R67, doi:10.37236/1093.
  30. [30]F. Jouve, E. Kowalski, and D. Zywina, Splitting fields of characteristic polynomials of random elements in arithmetic groups, Israel J. Math. 193 (2013), 263–307, doi:10.1007/s11856-012-0117-x.DOI
  31. [31]R. Jozsa and A. Miyake, Matchgates and classical simulation of quantum circuits, Proc. R. Soc. A 464 (2008), no. 2100, 3089–3106, doi:10.1098/rspa.2008.0189.DOI
  32. [32]L. P. Kadanoff and F. J. Wegner, Some critical properties of the eight-vertex model, Phys. Rev. B 4 (1971), no. 11, 3989–3993, doi:10.1103/PhysRevB.4.3989.DOI
  33. [33]P. W. Kasteleyn, The statistics of dimers on a lattice, Physica 27 (1961), no. 12, 1209–1225, doi:10.1016/0031-8914(61)90063-5.DOI
  34. [34]N. M. Katz and P. Sarnak, Zeroes of zeta functions and symmetry, Bull. Amer. Math. Soc. 36 (1999), no. 1, 1–26, doi:10.1090/S0273-0979-99-00766-1.DOI
  35. [35]L. H. Kauffman and H. Saleur, Free fermions and the Alexander–Conway polynomial, Comm. Math. Phys. 141 (1991), 293–327, doi:10.1007/BF02101508.
  36. [36]B. Kauffman, Crystal statistics. II. Partition function evaluated by spinor analysis, Phys. Rev. 76 (1949), 1232–1243, doi:10.1103/PhysRev.76.1232.DOI
  37. [37]A. Yu. Kitaev, Fault-tolerant quantum computation by anyons, Ann. Physics 303 (2003), 2–30, doi:10.1016/S0003-4916(02)00018-0.DOI
  38. [38]A. Kitaev, Anyons in an exactly solved model and beyond, Ann. Physics 321 (2006), no. 1, 2–111, doi:10.1016/j.aop.2005.10.005.DOI
  39. [39]E. Kowalski, The large sieve, monodromy, and zeta functions of algebraic curves, 2: independence of the zeros, Int. Math. Res. Not. IMRN 2008 (2008), rnn091, 57 pp., doi:10.1093/imrn/rnn091.DOI
  40. [40]S. Liao and L. Rybińkov, Maximal transitivity of the cactus group on standard Young tableaux, preprint (2025), arXiv:2506.16561.arxiv.org/abs/2506.16561
  41. [41]E. H. Lieb, Solution of the dimer problem by the transfer matrix method, J. Math. Phys. 8 (1967), no. 12, 2339–2341, doi:10.1063/1.1705163.DOI
  42. [42]E. Lieb, T. Schultz, and D. Mattis, Two soluble models of an antiferromagnetic chain, Ann. Physics 16 (1961), no. 3, 407–466, doi:10.1016/0003-4916(61)90115-4.DOI
  43. [43]W. T. Lu and F. Y. Wu, Ising model on nonorientable surfaces: exact solution for the Möbius strip and the Klein bottle, Phys. Rev. E 63 (2001), 026107, doi:10.1103/PhysRevE.63.026107.DOI
  44. [44]J. Milewski, G. Banaszak, T. Lulek, M. Labuz, and R. Stagraczyński, Galois actions on the eigenproblem of the Heisenberg heptagon, Open Syst. Inf. Dyn. 19 (2012), no. 2, 1250012, doi:10.1142/S1230161212500126.DOI
  45. [45]H. Nicolai, Supersymmetry and spin systems, J. Phys. A 9 (1976), no. 9, 1497–1506, doi:10.1088/0305-4470/9/9/010.DOI
  46. [46]L. Onsager, Crystal statistics. I. A two-dimensional model with an order-disorder transition, Phys. Rev. 65 (1944), 117–149, doi:10.1103/PhysRev.65.117.DOI
  47. [47]P. Pfeuty, The one-dimensional Ising model with a transverse field, Ann. Physics 57 (1970), no. 1, 79–90, doi:10.1016/0003-4916(70)90270-8.DOI
  48. [48]N. Regnault, S. Liu, and B. Bernevig, Integer characteristic polynomial factorization and Hilbert space fragmentation, preprint (2022), arXiv:2210.08019.arxiv.org/abs/2210.08019
  49. [49]J. Ruh and S. J. Elman, Expanding the class of free fermions via twin-collapse methods, Phys. Rev. A 113 (2026), no. 2, 022456, doi:10.1103/sfvf-vdh5.DOI
  50. [50]J. Ruh, S. J. Elman, and A. Chapman, A graph-theoretic analysis of non-generic free-fermion solvability by Krylov decompositions, preprint (2026), arXiv:2610.03557.arxiv.org/abs/2610.03557
  51. [51]G. M. Schütz, Exact solution of the master equation for the asymmetric exclusion process, J. Stat. Phys. 88 (1997), 427–445, doi:10.1007/BF02508478.DOI
  52. [52]C. J. S. Smith, Additive and multiplicative relations connecting conjugate algebraic numbers, J. Number Theory 23 (1986), no. 2, 243–254, doi:10.1016/0022-314X(86)90094-6.
  53. [53]P. Stevenhagen and H. W. Lenstra, Jr., Chebotarëv and his density theorem, Math. Intelligencer 18 (1996), no. 2, 26–37, doi:10.1007/BF03027290.DOI
  54. [54]W. P. Su, J. R. Schrieffer, and A. J. Heeger, Solitons in polyacetylene, Phys. Rev. Lett. 42 (1979), 1698–1701, doi:10.1103/PhysRevLett.42.1698.DOI
  55. [55]H. N. V. Temperley and M. E. Fisher, Dimer problem in statistical mechanics: an exact result, Philos. Mag. 6 (1961), no. 68, 1061–1063, doi:10.1080/14786436108243366.DOI
  56. [56]B. M. Terhal and D. P. DiVincenzo, Classical simulation of noninteracting-fermion quantum circuits, Phys. Rev. A 65 (2002), 032325, doi:10.1103/PhysRevA.65.032325.DOI
  57. [57]L. G. Valiant, Quantum circuits that can be simulated classically in polynomial time, SIAM J. Comput. 31 (2002), no. 4, 1229–1254, doi:10.1137/S0097539700377025.DOI
  58. [58]Z. Wang and K. R. A. Hazzard, Topological correlations in three-dimensional classical Ising models: an exact solution with a continuous phase transition, Phys. Rev. Research 5 (2023), 013086, doi:10.1103/PhysRevResearch.5.013086.DOI
  59. [59]F. Y. Wu, Ising model with four-spin interactions, Phys. Rev. B 4 (1971), no. 7, 2312–2314, doi:10.1103/PhysRevB.4.2312.DOI
  60. [60]M. Zheng, B. Pozsgay, and S. Chen, Exact solutions of hidden free fermion models, preprint (2026), arXiv:2609.14134.arxiv.org/abs/2609.14134

Paper details

vlnc:W_y65ypmss
Version
v1 · latest
Published
Fresh — published within the last 30 days
License
CC BY 4.0
Usage

views · 9

downloads · 0

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