Matrix-Product State Skeletons in Onsager-Integrable Quantum Chains Authors: Imogen Camp, Nick G Jones Published: 2026 ABSTRACT AbstractMatrix-product state (MPS) skeletons are connected networks of Hamiltonians with exact MPS ground states that underlie a phase diagram. Such skeletons have previously been found in classes of free-fermion models. For the translation-invariant BDI and AIII free-fermion classes, it has been shown that the underlying skeleton is dense, giving an analytic approach to MPS approximation of ground states anywhere in the class. In this paper, we partially expose the skeleton in certain interacting spin chains: the N-state Onsager-integrable chiral clock families. We construct MPS that form a dense MPS skeleton in the gapped regions surrounding a sequence of fixed-point Hamiltonians (the generators of the Onsager algebra). Outside these gapped regions, these MPS remain eigenstates, but no longer give the many-body ground state. Rather, they are ground states in particular sectors of the spectrum. Our methods also allow us to find further MPS eigenstates; these correspond to low-lying excited states within the aforementioned gapped regions. This set of MPS excited states goes beyond the previous analysis of ground states on the N=2 free-fermion MPS skeleton. As an application of our results, we find a closed form for the disorder parameter in a family of interacting models. Finally, we remark that many of our results use only the Onsager algebra and are not specific to the chiral clock model representation. Introduction In recent decades, significant advances have been made in understanding the ground state physics of one-dimensional quantum many-body systems. While integrability provides an analytical probe of phenomena in special families of Hamiltonians [1, 2], numerical methods related to matrix-product state (MPS) approximations allow us to understand features of the general interacting case [3, 4]. When the many-body Hamiltonian has a spectral gap, the entanglement area law guarantees an MPS approximation to the ground state [5, 6]. Beyond approximations, MPS are also a valuable analytical tool. MPS-solvable Hamiltonians have exact MPS ground states, in the sense that the ground state can be represented with a finite bond dimension in the thermodynamic limit. Such models, including the AKLT chain [7], have led to many insights [6]. For example, MPS approaches have led to a deeper understanding of symmetry fractionalisation, order parameters, and the classification of symmetry-protected topological (SPT) phases in one dimension [8–11]. It is interesting to consider the connection between integrable models and Hamiltonians that admit exact MPS ground states: neither property implies the other, yet both facilitate analytic insights into quantum many-body systems. Simple examples of MPS-solvable Hamiltonian families are the one-parameter “disorder line” of product states in the XY model [12–16], and its MPS-solvable Kramers-Wannier dual [17, 18]. Refs. [19, 20] investigated this question more broadly in families of free-fermion chains (the translation-invariant BDI and AIII classes of the tenfold way [21]). Despite the simple exact-solvability of free fermion models [22], their ground states are generically not exact matrix-product states, as their entanglement spectra have infinite rank [19, 23]. However, taking the BDI class for definiteness, we can precisely characterise the subset of this family with exact MPS ground states. Each model in the BDI class corresponds to a Laurent polynomial f(z), and the model has an exact MPS ground state if and only if f(z)=zpg(z)2h(z) for some integer p, polynomial g(z), and Laurent polynomial h(z), where h(z)=h(1/z). The sub-family satisfying this condition constitutes the MPS skeleton: a structure underlying the entire phase diagram. In contrast to finding an MPS approximation to the ground state of a general gapped Hamiltonian, the ground states along this skeleton can be represented exactly by an MPS with finite bond dimension in the thermodynamic limit. While lying exactly on the skeleton is fine-tuned, it is dense in the following sense: we can find a sequence of models lying on the skeleton with ground states that give an arbitrarily good approximation to the energy density of a generic ground state in the class. Hence, we have an analytic approach for approximating those ground states in the phase diagram that have an infinite-rank entanglement spectrum [19, 20]. It is natural to ask if there are other families of models where the underlying MPS skeleton can be identified. It is of particular interest to go beyond free-fermionic cases. In this work, we construct MPS skeletons underlying interacting families of Onsager-integrable chiral clock models. These are families of integrable models on chains of N-state sites, where the Onsager algebra yields local conserved charges [24–26]. Taking N=2, we recover the BDI class (in its spin chain form). Onsager’s 1944 computation of the free energy in the 2D classical Ising model utilised this algebra, and it is thus naturally connected to the 1D quantum transverse-field Ising model and its free-fermion solvability. Nevertheless, other representations of this algebra appear as N-state generalisations of the quantum Ising model. For N>2, these models are interacting chiral clock models [27–30]. Commutators of the Onsager paramagnet, denoted A0, and the Onsager ferromagnet, denoted A1, generate this algebra. Beyond the superintegrable chiral Potts chain A0+λA1, the analogue of the transverse-field Ising model, such models remain relatively unexplored. Ref. [26] examines the Hamiltonian family constructed as a linear combination of the generators, denoted Ak, while Refs. [31, 32] analyse their SPT physics. Note that in some cases the Onsager generators are considered as a symmetry algebra of some other Hamiltonian, rather than as Hamiltonians themselves [33–36]. As demonstrated in [30], we can express generators of (general) representations of the Onsager algebra as a direct sum of representations of sl(2). In principle, we can use this to diagonalise these models, establishing the form of the spectrum [26, 30, 37]. In the chiral clock models, these representations are all two-dimensional [38, 39]. Further analysis is required to determine the spectrum from functional equations in the chiral clock case [29, 38–42]—see also Refs. [37, 43, 44] for more recent work on eigenvectors. These eigenvectors are, roughly speaking, found by filling an interacting Fermi sea [38, 40, 45]. While the ground-state phase diagram for the N=2 case is well understood [46, 47], we have a far more limited understanding of the phase boundaries for N>2. However, we expect gapless regions comprising a sequence of first-order transitions between different Onsager sectors, as occurs when tuning A0+λA1 [45, 48, 49]. Numerical results from [31] demonstrate this extended gapless region in a two-parameter phase diagram of A0+λA1+μA2 (see also Fig. 1). Schematic phase diagrams for the Hamiltonian H=αA0+βA1+γA2 with normalisation α+β+γ=1. The winding number of f(z) is denoted by ω. The trivial phase has ω=0, while ω=1 corresponds to the ferromagnetic (FM) phase. For even N, ω=2 gives an SPT; for odd N, we instead have a representation-SPT (RSPT). The coloured regions are gapped and connected to fixed-point Ak, therefore corresponding to the intersection of the set S with the phase diagram. The white regions are gapless and outside S. For N=2, the phase diagram is well-understood [17, 18]. Diagrams for N=3,4 were found numerically in Ref. [31]. For N=4, it is unclear whether the trivial and SPT phases have a direct transition at the black square. The dashed line represents a leg of the skeleton parameterised by H=(1-λ)2A0+2λ(1-λ)A1+λ2A2. For N=2, the ground state is an MPS along the entire path. We show that the analogous state is an eigenstate for general N, and is the ground state in the gapped regions. The crosses represent points where the MPS construction given in this paper has a singularity; this coincides with a gap closing in the sector S. For each N, we consider the family of Hamiltonians For N=2, these spin chains are generalised cluster models [46, 50–55] where σj are the usual Pauli operators on site j. For the chiral clock models, we show below that each Am is a local Hamiltonian with a range of |m| sites (i.e. is supported on |m|+1 sites), and we impose that the overall model is finite-range. This means we allow only finitely many non-zero tm. Within the family (1), we let i.e., S is the set of Hamiltonians that are in the gapped regions surrounding each of the Onsager generators Ak. The Ak are fixed-point Hamiltonians, with zero-correlation-length ground states [31]. Then, writing f(z)=∑mtmzm, we show that if f(z)=±zpg(z)2, and we are in S, then we can construct an exact MPS ground state except at a measure zero set of points. Thus, we have an MPS skeleton for N>2, giving a significant generalisation of the N=2 results to the interacting case. However, importantly, we note the implication is in one direction, and there may be other Hamiltonians in S with exact MPS ground states (this is certainly the case for N=2, as we have excluded cases with h(z)≠const.). We also show that this skeleton is dense within S, giving us an analytic approach to approximating the ground state by MPS in this region. A further result of our analysis is the first excited state at zero momentum for models on the skeleton, along with the lowest-energy eigenstate in each non-zero momentum sector, up to the points where these states become degenerate with others in the same sector. This is a direct analogue to the appearance of an exact MPS ground state in the zero momentum sector: usually, we can approximate gapped ground states in finite-momentum sectors to be of this form up to small corrections [56, 57]. As with the spectrum, computing correlations in the ground state of HA is not straightforward [58]. Remarkable closed formulae exist for the long-distance behaviour of the order and disorder parameters along the line A0+λA1 [42, 48, 59]. While the class HA can be analysed using the Bethe ansatz, MPS-solvability gives an alternative approach to computing order parameters in the more general family HA, and we demonstrate this for HA=A0+2aA1+a2A2. Many of our results rely only on the algebraic relations, and hence apply in principle to other Onsager-integrable Hamiltonians (subject to some additional assumptions given in Section 2.4). Of course, any unitary transformation applied to the generators Ak will preserve the algebra and the interpretation of the Ak as Hamiltonians. Our results can be simply applied to such cases. For example, in Ref. [26], the authors generate other families of Hamiltonians by combining unitary pivots [31, 60] with Kramers-Wannier duality [61, 62]. It would be interesting to find a model where the sl(2) representations have dimension greater than two [38], and to connect to the constructions in Refs. [63–65]. For the remainder of the paper we focus our attention on the chiral clock representations. We structure the paper as follows. In Section 2, we summarise key properties of our model and state our main results. Section 3 illustrates these results with some explicit examples. Section 4 gives more details of the diagonalisation of Onsager-integrable Hamiltonians. In Section 5, we prove the results of Section 2. We conclude in Section 6 with a discussion of possible directions for future work. Key Definitions and Statement of Results Onsager-Integrable Hamiltonians In Ref. [24], Onsager solved the two-dimensional classical Ising model by computing the spectrum of the transfer matrix, giving the exact partition function. Of interest to us are the corresponding quantum Hamiltonians on a chain, which can be determined by taking a strongly anisotropic limit of the two-dimensional classical model. His solution utilised what is now called the Onsager algebra: the infinite-dimensional Lie algebra obeyed by the generators {Al,Gm|l,m∈Z} given by We observe that the mappings Am↦Am+1,Gl↦Gl and Am↦-Am,Gl↦Gl both leave Eq. (4) invariant. From this structure, it follows that A0 and A1 obey the Dolan-Grady conditions [25, 66–68] In fact, Eqs. (4) and (5) are equivalent [66, 67], and Eq. (5) remains true replacing A0↦Al, A1↦Am [31]. Ref. [69] demonstrates that these conditions give rise to operators that we rewrite as obeying the commutation relations We can therefore interpret Rm,l as ladder operators for Al; this structure ensures that the spectrum of Al consists of sectors with unit level spacing. We can construct A0 and A1 that are periodic, L-site, ZN-symmetric chiral clock chains with local Hilbert space dimension N, and that obey the Onsager algebra. We will take the chain length L to be finite. Moreover, for convenience, we will choose L=0modN, which we discuss further in Appendix D. Let us define the single-site “shift” and “clock” operators Xj and Zj obeying (Xj)N=(Zj)N=1 and XjZk=ωδjkZkXj, where ω=e2πi/N. In the Z-diagonal basis, with states |a⟩ defined modulo N, we have The first two generators of the Onsager algebra, related by Kramers-Wannier duality, are given by for complex couplings αm=(1-ωm)-1. In the Z-diagonal basis, the eigenstates of A1 are {|a1a2⋯aL⟩} and the eigenstates of A0 are {|v1(n1)v2(n2)⋯vL(nL)⟩} for The ground state of A0 is the unique state |v1(0)v2(0)⋯vL(0)⟩, while A1 has the N ferromagnetic ground states {|τa⟩=|aa⋯a⟩} for a∈{0,1,⋯,N-1}. These states all have the same energy E0=-L(N-1)/2N. The remaining generators can be found using either the commutation relations Eq. (4) or the pivot procedure described in Ref. [31], utilising the relation Consequently, all Al for even (odd) l are related by a unitary transformation to A0 (A1), since for Um=e-iπAm. Hence, all of these Hamiltonians have the same ground state energy E0, and the ground state |ψl⟩ of Al is unique (N-fold degenerate) for even (odd) l. Additionally, by considering the light cone of (U1U0)k, it follows that each Am has interaction range at most |m|; our generators are therefore local operators in this representation. We emphasise that while any Al or Gm could be calculated using this procedure, for N>2, a general closed-form expression is not known; however, for A-1 and A2, closed formulae have been determined [26, 31, 33]. Recall that we work with the family of Hamiltonians of the form It is convenient to define the corresponding Laurent polynomial As shown in Results 1 and 2 of Ref. [19], for N=2, if we can write and some integer p, then we can write the ground state as an exact MPS.1 Hamiltonians parameterised by such coefficients s→=(s0,s1,s2,⋯,sd) are said to “lie on the skeleton”. In fact, for N=2, the ground state can be expressed as an MPS if and only if f(z)=zpg(z)2h(z), where h(z) is any Laurent polynomial obeying h(z)=h(1/z). This is loosely motivated by noting that, as discussed in Section 4, the spectrum of HA[{tm},N] involves terms with coefficients ∼f(z)f(1/z). For Hamiltonians corresponding to Eq. (15), these terms simplify to ∼g(z)g(1/z); the lack of a square-root is suggestive of simplifications in the analysis. We aim to generalise the results of Ref. [19] by demonstrating that, for N>2, Hamiltonians satisfying Eq. (15) will have exact MPS ground states in certain gapped regions. However, we will not claim that such a condition is necessary for the ground state to be an exact MPS. As illustrated in Fig. 1, the N>2 phase diagrams have extended gapless regions; this contrasts with the N=2 case. We will find that, on the N>2 skeletons, we can construct MPS eigenstates analogous to the N=2 ground state except at a measure-zero set of points. However, we will find that these eigenstates are only ground states in gapped regions containing fixed-point Hamiltonians Al. We represent these by the coloured regions in Fig. 1. It is therefore convenient to define the set S as the set of Hamiltonians in the class HA that lie within such gapped regions, as in Eq. (3). Finally, we note that the structure of the Onsager algebra imposes that where |ψ0⟩ is the ground state of A0 and We will use these results in Section 5.1, and they are proved in Appendix A using the ladder operators Eq. (7). Remarks on the Phase Diagram The family of models, HA, has a D2N=ZN⋊Z2CPT symmetry. The ZN generator is ∏jXj, while the CPT symmetry combines charge conjugation, lattice inversion, and complex conjugation [31]. For all N, the A2k+1 are symmetry-breaking phases, with N-fold ground-state degeneracy. In the symmetric phases we have a unique ground state, and it is of interest to consider its SPT class [9]. For odd N, D2N has no SPT phases, and each A2k is in the trivial symmetric phase. For even N, A4k is trivial while A4k+2 is a non-trivial D2N SPT; this was shown using charge pumps in Ref. [32]. For odd N, A2 has some SPT-like features, and, while trivial, the nearby region is labelled a representation-SPT (RSPT) [31, 70, 71], due to a non-trivial linear symmetry-fractionalisation in the leading entanglement levels. The above analysis relates to the SPT class when allowing arbitrary symmetric perturbations to our Hamiltonian. In the N=2 case, the BDI class has an integer topological invariant, coinciding with the winding number of f(z). This cannot change without a phase transition, provided we stay within the BDI class of free-fermion models [46]. In fact, the set S inside the family HA also has the same integer invariant. As we discuss in Section 4, zeros of f(z) on the unit circle correspond to a bulk gap closing within S. Unlike the N=2 case, we can leave S without a zero crossing the unit circle due to a first order transition to a different sector. However, if we remain in S, we cannot have a zero of f(z) cross the unit circle, and so the winding number is invariant. This is one way to understand the distinct symmetric phases appearing for N=3 in Fig. 1: they have the same SPT invariant, but different winding numbers. Thus, S consists of disjoint gapped regions around each of the Ak, each dual under Kramers-Wannier and unitary pivots. Main Results Here, we state our main results; these will be proved in Section 5. We characterise exact eigenstates of Hamiltonians on connected networks f(z)=±zpg(z)2 within the family Eq. (13) and show that they are ground states in certain regions of the phase diagram. Moreover, these states can be written in exact MPS form. To write down the closed form of the wavefunction, we must first compute a collection of coefficients {bk} from the coefficients {sk} of g(z)=∑kskzk. This can be done algorithmically [19], which will be discussed in Section 5.1. We note that this algorithm is identical to the Schur-Cohn algorithm [72, 73], which is used to determine the distribution of the roots of a polynomial with respect to the unit circle. Furthermore, we require the fixed point ground states |ψp±⟩ of ±Ap. For the clock models, these can be written in exact MPS form using the procedure given in Ref. [31], and in the symmetry-breaking case the result holds for any of the N symmetric ground states |ψp(i)⟩. Result 1 (Exact eigenstate) If we can express the Laurent polynomial in the form f(z)=±zpg(z)2 for p∈Z and then the corresponding Hamiltonian has the eigenstate Here, M(k)=exp(∓βkAp+k) for βk=2arctanh(bk), and |ψp±⟩ is the ground state of the Hamiltonian ±Ap. Note this breaks down if any |bk|=1. The states |φ⟩ can be expressed in MPS form,2 and these Hamiltonians therefore constitute the MPS skeleton; however, a measure-zero set of points, where any |bk|=1, must be excluded. We prove this result in Section 5.1. Result 2 (Exact ground state) For all vectors s→=(s0,⋯,sd) that do not give any |bk|=1, the eigenstate |φ⟩ is the ground state of HA(s→) when HA(s→)∈S. This follows straightforwardly from the perturbative arguments we provide in Section 5.2. Note that |bk|=1 occurs in gapless models (see Appendix E), but can also arise for Hamiltonians in S. In such cases, the method used to construct the MPS breaks down, but we can argue that we have an exact MPS ground state by considering a limit of models, lying in S and on the skeleton, where we can explicitly construct the ground state; Refs. [19, 20] give related discussions. In the N=2 case, the distinct methods of Ref. [20] prove the existence of MPS ground states for f(z)=±zpg(z)2 with no restrictions on the {bk}. Although we cannot generally express f(z)=±zpg(z)2 for a finite polynomial g(z), we can always define and truncate an appropriate “square-root” series at some finite power D of z [19, 20]. This gives an approximate Hamiltonian HA(D) corresponding to the Laurent polynomial fD(z)=±zkgD(z)2 (note k≠p in general). For example, the chiral Potts model A0+λA1, has Laurent polynomial f(z)=1+λz. This can be approximated by for g3(z)=1+λz/2-λ2z2/8+λ3z3/16. Note that HA(3) is then the chiral Potts model, plus a “small” perturbation by longer-range models. The ground state of the Hamiltonian HA(D) can be used to approximate the ground state of HA. Indeed, the skeleton constitutes a dense subset of our family of Hamiltonians in the following sense. Result 3 (The skeleton is dense) Consider a general chiral clock Hamiltonian HA of the form Eq. (13), with ground state |φ⟩, and let EA be the corresponding ground state energy density of . We can define a Hamiltonian HA(D) on the skeleton with ground state |φD⟩, such that for any fixed ε>0 and sufficiently large D and L, we find that the energy density EA(D) of |φD⟩ with respect to satisfies . Analogous statements hold for the ground state in any fixed sector. The proof is given in Section 5.3. Heuristically, this tells us that the ground state of any Hamiltonian in S can be approximated by a sequence of states of the form Eq. (19). Ref. [19] gives a derivation of the scaling dimension of the disorder operator in the Ising conformal field theory using such a path of skeleton states. We expect that this can be made precise in the thermodynamic limit using the results of [74]. Result 4 (More exact eigenstates) For even p, Hamiltonians with Laurent polynomials f(z)=zpg(z)2 have the exact eigenstates where For f(z)=-zpg(z)2 and even p, an equivalent set of eigenstates exist where we replace Zj†↦Zj and |ψ0+⟩↦|ψ0-⟩. This result is proved in Section 5.4, where we also discuss the necessity of even p in f(z)=±zpg(z)2 for this to hold. These states can also be represented in exact MPS form. Further Remarks We note that Result 1 can be proved if our model obeys the Onsager algebra and when |ψp±⟩ is unique.3 For |φ⟩ to have an exact MPS representation, we additionally require that A0 and A1 have exact MPS ground states and are finite-range. Moreover, Result 2 requires that our model obeys a finite-dimensional Onsager algebra and that each Ap is gapped. Result 3 holds generally for any finite-dimensional representation of the Onsager algebra, provided that the energy density remains finite in the thermodynamic limit. Result 4, however, is specific to the clock models. Finally, for brevity, we will henceforth write |ψp+⟩ as |ψp⟩. We remark that, just as the N=2 case can be viewed as a family of free-fermion models, for N>2, our family HA has a dual parafermionic form [75, 76]. Many of our results will carry over simply to the parafermion picture, although care must be taken with defining parafermionic MPS [77]. Following our discussion about the winding number in Section 2, it would be interesting to understand if there is a connection between this winding and the number of edge modes [76]. Examples In this section, we motivate our results by deriving Result 1, Result 2, and Result 4 explicitly for the simplest non-trivial case. We then demonstrate how the ground state can be expressed as an exact MPS. Finally, we show how this result can be used to evaluate the disorder operator on the skeleton, generalising known results for N=2 to any even integer N. Eigenstates for d=1 Here, we illustrate our results in the case of d=1 using a variation of the method used in the proof for general d. The initial steps of the argument are illustrative of the general method, but we are able to make certain computations explicit in this particular case. We also utilise a known expression for the ground state energy in S that bypasses the perturbative argument of Result 2. Let us consider the d=1 Hamiltonian which has p=0 and g(z)=1+az. The non-interacting N=2 case of HA was studied in detail in Refs. [17, 18]. Defining H~A=eβA1HAe-βA1 for β=2arctanha, the pivot relation Eq. (11) gives Consider the ground state |ψ0⟩ of A0, which—in the chiral clock representation—can be written as the product state [31] and the translationally invariant basis of single-particle excitations {|ϕE0+1/N,p=0(P)⟩} labelled by momentum P. We have Moreover, from Eq. (16), we have The states |ψ0⟩ and |ϕE0+1/N,p=0(P)⟩ are therefore eigenstates of H~A. In fact, we can show (A1+G1)|ψ0⟩=0 by noting that there exists an operator W satisfying both W|ψ0⟩=|ψ0⟩ and {W,A2k+1}=0 [40].4 Thus, reversing the transformation, we have for Therefore, for all a, |φ⟩ and {|χp=0(P)⟩|P∈{0,2π/L,⋯,2(L-1)π/L}} are eigenstates of HA. The state |φ⟩ has energy εφ=(a2+1)E0. From Refs. [26, 30], we know that this is the ground state energy provided HA∈S. Thus, for d=1, Result 1 and Result 2 hold. Ground State as an MPS for d=1 Let us write where are commuting operators. The ground state |φ⟩ can therefore be represented by the circuit in Fig. 2. This can be interpreted as a matrix-product operator (MPO) acting on a product state, giving an MPS. We obtain the (unnormalised) MPS form with the MPS tensors of bond dimension χ=N. (a) The circuit diagram for the d=1 ground state. This can be interpreted as an MPO acting on a product state, giving an MPS. The blue boxes represent each unitary gate Uj,j+1, and the |vj⟩ are the single-site states. (b) The MPS tensor from Eq. (33). Disorder Operator on the d=1 Skeleton For even N, we define the unitary disorder operator For N=2, and for g(z) such that all |bk|<1, the expectation of the disorder operator in the ground state is [19] For g(z)=1+az, with |a|<1, this becomes ⟨μk⟩N=2=1-a2. Here, we outline how this d=1 result generalises to any even N. Full details of the calculation are given in Appendix B. We aim to evaluate5 If we define then we can use the commutation relations XjZk=ωδjkZkXj and the fact that μk|ψ0⟩=|ψ0⟩ to express the numerator as If we write |ψ0⟩ in terms of the states |a⟩ that are diagonal in the Z-basis then ⟨μk⟩ can be written as the ratio of two sums over configurations of states in the Z-basis. We can evaluate these expressions with a transfer matrix method, writing for two matrices T (the transfer matrix) and B. We find that these matrices are circulant, and so their spectra can be determined from standard results [78]. In the thermodynamic limit L→∞, the largest eigenvalue of T dominates, and we obtain for all even N, as desired. The Spectrum of Onsager-Integrable Chiral Clock Chains In this section, we summarise how the Onsager algebra can be used to examine the structure of the spectrum. As stated in Result 2, the eigenstate |φ⟩ is the ground state of Hamiltonians in the region S, defined by Eq. (3). As this region is determined by the spectral gap of the Hamiltonians, it is important to understand some properties of their spectra. The reader may wish to skip this section on a first read, noting only Eq. (46) which is used in Section 5.5. Firstly, we note that chiral clock Hamiltonians commute with the translation operator and the D2N=ZN⋊Z2CPT generators described in Section 2. The Hamiltonian can therefore be block-diagonalised into sectors labelled by the quantum numbers of momentum and ZN-charge, denoted P and Q respectively. Additionally, as discussed in Refs. [30, 37–39, 79], the Hamiltonians can be further block-diagonalised into Onsager sectors, with multiple Onsager sectors per (P, Q) sector. We define the “ground-state Onsager sector”, lying within the (0, 0) sector, to be the Onsager sector containing the ground state of A0. As discussed in Refs. [26, 30], given a finite-dimensional representation of the Onsager algebra, it is possible to write ∑m=-nnαmAm=0 for some integer n and coefficients satisfying αm=±α-m; determining these coefficients is, however, non-trivial. Nevertheless, one can define z∓k=e±imθk as the roots of the polynomial h(z)=∑m=-nnαmzm+n. The generators of any finite-dimensional representation of the Onsager algebra can then be rewritten, up to sector-dependent additive constants, with the invertible transformation6 where the Onsager algebra imposes that The operators {Ej±,Hk} can therefore be interpreted as sl(2) generators. The roots of h(z) lie on the unit circle to ensure the self-adjointness of our Hamiltonian, and they are also sector-dependent. In general, determining the n values {θk} is difficult and must be done by solving functional equations [28–30, 40, 41]. (These are referred to as BAMP polynomials in Ref. [42] and chiral Potts polynomials in Ref. [80].) However, without calculating the values θk, we can easily establish the structure of the spectra of Onsager-integrable Hamiltonians by writing where we have defined Ek±=Jxk±iJyk. This can be diagonalised by rotating in the (Jx,Jy) plane to give where the final line follows only if we are on the skeleton, where f(z)=±zpg(z)2. Given that we know the eigenvalues of the spin-operators Jx,newk, the form of the spectrum in each sector, up to additive constants, is Here, each of the mk can take values from {-S,⋯,S-1,S} for the spin-S representation of sl(2); in the clock models we have S=1/2 [30]. Again, we emphasise that this form is up to sector-dependent additive constants. For chiral-clock Hamiltonians HA∈S, the ground state lies in the “ground-state Onsager sector”. This follows from perturbative arguments and pivoting and duality relations. In this sector, the {θk} values can be calculated [28, 29, 38] and have multiplicity one.7 The ground state energy density for chiral clock Hamiltonians HA∈S in the thermodynamic limit is then given by [28] where and 0≤θ≤π. Since we have f(e±iθ) in Eq. (45), this range of θ covers the unit circle. Thus Eq. (45) connects gapless modes in this sector to zeros of f(eiθ) on the unit circle. This means that, even when we do not have a first-order transition to a different sector, if we tune a zero to the unit circle, HA cannot remain in S. This is consistent with previous analysis of the superintegrable chiral Potts line [81]. As mentioned above, some results for the eigenvectors are also known [37, 38, 40, 43, 44]. These are typically given for the model A0+λA1, but we expect they could be generalised. However, this would correspond in the N=2 case to knowledge of the diagonal free-fermionic modes. For N=2, going from the mode representation to the MPS representation is non-trivial [19], and our simple construction of the MPS eigenstates in these interacting models gives a new perspective and a direct approach that is mostly decoupled from the exact solutions above. Analysis In this section, we prove the results given in Section 2. Given that Hamiltonians with f(z)=±zpg(z)2 are all related to the Hamiltonian with f(z)=±g(z)2 via some combination of Kramers-Wannier duality and unitary transformations, we first prove our results for p=0. Proof of Result 1 We begin by noting that the pivot relation Eq. (11), following directly from the Onsager algebra, means that abstractly has coefficients {aα,gα} independent of our choice of representation. Therefore, if we find the explicit form of H~A in one representation, the form must hold for all representations provided we do not have any representation-specific cancellations of the generators.8 Thus, we will study a particular (free-fermion) representation below. Define the Majorana fermions where Xj,Yj,Zj reduce to the usual 2×2 Pauli matrices. A representation of the Onsager algebra has generators This is a slightly different N=2 representation of the Onsager algebra, related to our definition of the N=2 chiral clock model by a unitary transformation (see Section 2.1). Using results from Ref. [19], Hamiltonians in this representation with f(z)=zpg(z)2 can be expressed as where In contrast with the approach taken in Ref. [19], we treat Ωn and Γn as separate objects rather than writing Ωn=Γn† and considering only transformations on Γn. Begin by defining M(k) as in Result 1. We wish to compute which can be done by taking To evaluate these objects, we observe that which, provided all |bk|≠1, gives Define Algorithm 1 [19] for the coefficients {bk}. This is designed such that each layer of the algorithm acting on Γn eliminates the term γn+α-iγ~n-α with the largest α. As mentioned, this coincides with the Schur-Cohn algorithm [72, 73]. Compute {bk} from vector s→=(s0,⋯,sd) To demonstrate this, note that where sα(d)=sα-bdsd-α. This leads to the recursion where for initial condition sα(d+1)=sα. After all d layers have been applied, this collapses onto Γ~n∝γn-iγ~n. No terms, however, are necessarily eliminated in Ωn, and each layer instead generates a new term γn+α+iγ~n-α with α=k-d. Therefore, for N=2, the transformation leads to for some constants {rα(s→)}. These are real functions of s→. While we do not need the explicit form of these coefficients, the computation can be done algorithmically. Algorithm 2 in Appendix C gives these coefficients up to an overall rescaling. We may wonder if there has been any cancellation of coefficients in this representation that would give a non-vanishing contribution in general. However, for any linear dependence (aside from Gk=G-k, which is true for all representations) to arise, we would need the transformation Eq. (54) to produce generators Al,Gm with l, m of order L. However, Eq. (60) is true for all L. If we consider L≫d,p, then Eq. (54) cannot produce sufficiently large l, m. Therefore, rα=aα=gα is the only possible set of L-independent coefficients generated by Eq. (54) for N=2. Moreover, Eq. (49) contains no constant term, and no linear combination of Al,Gm can give rise to this. Consequently, Algorithm 2 must give r0=|s→|2, and holds for all N. Note that the {rα(s→)} are independent of N and L. As noted in Section 2.1, |ψ0⟩ is an eigenstate of (Aα+Gα). Thus, it is an eigenstate of H~A. Reversing the transformation Eq. (54), we see that |φ⟩=M(d)⋯M(1)|ψ0⟩ must therefore be an eigenstate of HA, provided that |bk|≠1 for all k. Following a similar procedure to that presented in Ref. [19], this can be expressed in MPS form: each operator M(k) is a local operator that can be expressed as an MPO with finite bond dimension, and |ψ0⟩ is a product state. Acting with a series of finite bond dimension MPOs on a product state gives an exact MPS. Thus far, we have restricted ourselves to Hamiltonians with f(z)=g(z)2. However, the Hamiltonians f(z)=-g(z)2 take each Ak↦-Ak. If {Al,Gm|m,l∈Z} obey the Onsager algebra, then the generators {-Al,Gm|m,l∈Z} obey the same structure. Therefore, to obtain the ground state of f(z)=-g(z)2, we merely replace each Ak↦-Ak in every M(k), and replace the ground state |ψp⟩ of Ap with the ground state |ψp-⟩ of -Ap. Generalising this proof to any even p=2k is trivial: the mapping Ak↦Ak+1,Gk↦Gk preserves the structure of the Onsager algebra and Eq. (12) ensures that the ground state |ψ2k⟩ is unique. However, extending to any odd p=2k+1 requires more care, as Kramers-Wannier duality tells us that the ground state of A1 is N-fold degenerate. Nevertheless, Result 1 holds for all p; Aα+Gα commutes with r, and is therefore diagonalised for all α by the N eigenstates of r with distinct charges 0≤Q≤N-1. We give a more detailed discussion in Appendix D. Proof of Result 2 We have established that, provided |bk|≠1 for all k, the state |φ⟩ given in Eq. (19) is an eigenstate of the Hamiltonian HA with f(z)=±zpg(z)2. Results of perturbation theory in finite-dimensional Hilbert spaces from Ref. [82] demonstrate that, provided the gap does not close, the ground state is an analytic function of s→. We apply this below to show that |φ(s→)⟩ is the ground state provided HA∈S. The Ground State in the Gapped Region Containing A0 We begin by noting that |φ(S→)⟩=|ψ0⟩ and HA(S→)=A0 for the vector S→=(1,0,0,⋯). Thus, |φ(s→)⟩ is the ground state of HA(s→) at s→=S→. Provided that the gap does not close, the ground state must be an analytic function of s→. Therefore, since |φ(s→)⟩ is an analytic function of s→ containing the ground state at s→=S→, it is the ground state everywhere in the gapped region surrounding A0. Note that when the gap closes, we find degeneracies, and this argument breaks down. We have therefore proved only that |φ(s→)⟩ is the ground state of HA(s→) in the gapped region containing A0. The Ground State in the Gapped Region Containing Al The gapped region containing A0 has every |bk|<1. If we instead have some |bk|>1, then we can use the identity to write where we have defined We can use the relation Uke-βqAq=e-βqA2k-qUk from Eq. (11) to move all the unitary operators Uk so that they act on the initial state |ψ0⟩. As shown in Ref. [31], this gives the ground state |ψl⟩ of another fixed-point Hamiltonian Al. Therefore, we are always able to write where l∈Z and for |b~k|<1 and some integer n. Although the state |φ(s→)⟩ is not smoothly connected to |ψ0⟩ along a path of gapped Hamiltonians HA if any of the |bk|>1, it is always smoothly connected to a ground state |ψl⟩ of a Hamiltonian Al for some l∈Z. At this point, HA(s→)=Al. Therefore, the eigenstate |φ(s→)⟩ is the ground state provided HA∈S. Proof of Result 3 We can write the Laurent polynomial Eq. (14) describing a general Hamiltonian HA in the family Eq. (13) as where σ∈R, P0∈Z, {Pζ,Pz} are non-negative integers, |zj|<1, and |ζk|>1. Note that this polynomial has no zeros on the unit circle. Such zeros correspond to a gap closing in S and, as discussed in Appendix E, the existence of zeros on the unit circle will give a |bk|=1, which we exclude. If we now define then, for q=P0+Pz, we have f(z)=±zqg(z)2. As overall scale is unimportant, we normalise the constant coefficient to ±1. This polynomial g(z) can be expanded as a Laurent series which is well-defined in the annulus A given by Now, truncating this series, we obtain and which defines the Hamiltonian HA(D). From this, we can define the new Hamiltonian with Laurent polynomial In any sector, consider the ground states |φD⟩ and |φ⟩ of HA(D) and HA, respectively. Define the energy densities From variational principles, we have By definition of the spectral norm, for any normalised state |ψ⟩, we have |⟨ψ|δHA(D)|ψ⟩|≤||δHA(D)||. Therefore, using the triangle inequality, the energy densities differ by where Emax is a bound on the energy density of the Onsager generators Al (assumed in general to be O(1)). In the chiral clock model, Emax=(N-1)/2N [31]. The coefficients in Eq. (73) can be evaluated using residue calculus. This allows us to bound the magnitude of each term as follows: where maxj{|zj|}<ρ0 we can fix a sufficiently large D (and in the background an L≫D) such that Mρ±(D)<ϵ, giving and therefore that Hence, the MPS skeleton is dense: for any Hamiltonian HA∈S in the Onsager-integrable chiral clock class, we can construct a Hamiltonian HA(D) on the skeleton with a ground state of the form with an energy density with respect to HA approaching the ground-state energy density of HA. Thus, we have a method for approximating the ground state of any Hamiltonian in S. Moreover, by the same reasoning, the ground state of HA(D) in any particular sector approximates the ground state of HA in that sector for sufficiently large D. Proof of Result 4 As shown in Section 5.1, we can write As noted in Section 2.1, the states |ϕE0+1/N,p=0(P)⟩ are eigenstates of (Aα+Gα). Thus, these states are eigenstates of H~A. Reversing the transformation Eq. (54), we see that |χp=0(P)⟩=M(d)⋯M(1)|ϕE0+1/N,p=0(P)⟩ are eigenstates of HA, provided that |bk|≠1 for all k. Similarly to the ground state, |ϕE0+1/N,p=0(P)⟩ has an exact MPS representation. In analogy with the W-state [6], the most efficient MPS representation is where we have defined the vectors and the site-dependent matrices As discussed in Section 5.1, each M(k) can be expressed as an MPO with finite bond dimension. Applying a series of such MPOs to an exact MPS yields another MPS of finite bond dimension. Thus, the eigenstate |χp=0(P)⟩ has a representation as an exact MPS. Moreover, for p≠0, keeping p even, we can use Eq. (12) to obtain eigenstates as defined in Eq. (21). Note that the pivot procedure defined in Ref. [31] ensures that |ϕE0+1/N,p(P)⟩ also has an exact MPS representation. Unlike in the ground state case, it is unclear whether this could be generalised for odd p, as Kramers-Wannier dual Hamiltonians with periodic boundary conditions cannot have a single domain wall. Constraints on Coefficients {rα} In this section, we identify some constraints on the coefficients rα appearing above. We have Hence, the energy of the eigenstate |φ⟩ of the Hamiltonian HA with f(z)=g(z)2 is given by as the ladder operators Eq. (7) impose that ⟨ψ|Aα|ψ⟩=⟨ψ|A-α|ψ⟩ for any eigenstate |ψ⟩ of A0. Defining Ek±=Jxk±iJyk, it follows from Eq. (42) that for spin S. For these Hamiltonians HA, we can use this and Eq. (46) to write the ground state energy as We know that |φ⟩ is the ground state of the “ground-state Onsager sector” for Hamiltonians HA. Thus, in this region, εφ=ε0. From this, we can conclude that r0=|s→|2 (as we had already argued) and that Algorithm 2 will therefore give coefficients satisfying this relationship. Outlook In this work, we uncovered dense MPS skeletons in general classes of Onsager-integrable Hamiltonians, each analogous to that in Ref. [19]. While certain results hold for any representation, we deal primarily with the Onsager-integrable chiral clock chains—these are the key physical representations of this class. Due to the interacting nature of these models, any analysis is more involved than in the free-fermion classes considered previously. Nevertheless, we constructed the ground state explicitly for chiral clock Hamiltonians lying on the skeleton in gapped regions of the phase diagram containing any fixed-point Hamiltonian Al. Moreover, we showed that the constructed state remains an eigenstate everywhere on the skeleton, excluding a measure-zero set of cases. Given that the skeleton is dense, we also have a method for approximating the ground state of any model in these gapped regions as the limit of a sequence of states of the form Eq. (19). To prove the existence of the MPS skeleton, we worked in a specific representation of the Onsager algebra, considering Hamiltonians of the form HA=±∑k,qsksqAk+q+p. We emphasise that, even fixing N=2, the method used differs from the proof presented in Ref. [19]. However, it is not fully independent, since we utilise features of the frustration-free form of the Hamiltonian derived in this work. We note that the methods used in earlier work on the BDI skeleton rely heavily on the free-fermion representation. It would be interesting to explore whether our models are frustration-free for N>2, potentially using the parafermionic representation, and to understand whether the Witten conjugation method can be used in this context [83]. Note that any such application will be complicated by the level crossing transitions at the boundary of S. We calculated the ground state expectation of the disorder operator for the d=1 leg of the skeleton, showing how the well-understood N=2 formula generalises to all even N. Whether the same generalisation occurs for larger d remains an interesting question that we plan to explore in future work. Our derivation of the disorder operator on the skeleton gives a new route to understanding correlations in this family of models. Away from the superintegrable chiral Potts line, a formula for the disorder parameter is not known, despite the analogous result being established for the N=2 case using Toeplitz determinant theory [47]. Baxter gives a determinantal formula for this operator on the superintegrable chiral Potts line [84]. We expect that it will be possible to construct similar determinantal results away from this line using the MPS skeleton: if we establish a closed formula on the skeleton, we could use the density of the skeleton to obtain a general formula. Importantly, we contrast our results with existing literature on MPS in Bethe ansatz solvable models, as Bethe ansatz eigenstates can be placed in an MPS framework [85–89]. While this gives an MPS form for all eigenstates, the corresponding bond dimension can diverge in the thermodynamic limit (dependent on the magnetisation of the state), and naturally relies on a U(1) symmetry labelling this magnetisation. In HA, the charge is conserved modulo N. Thus, we cannot apply these results to find exact MPS ground states of our family of models. It would be interesting to see if we could use our exact ground states to feed into a further Bethe ansatz analysis, or to connect to the eigenvector constructions of Refs. [43, 44]. Our MPS skeleton states could also be useful initial states for numerics. Our MPS construction gives an upper bound on the number of non-zero entanglement eigenvalues of the ground state. In Refs. [19, 20], this construction was conjectured to be optimal for N=2, which could be proved in certain cases. It is unclear whether this optimality also applies to the clock models. In Ref. [90], quantum scar states are constructed via imaginary time evolutions with certain Onsager algebra elements. These are MPS eigenstates in a non-integrable variant of the model in Ref. [33]. It would be interesting to understand if our construction would be useful in this context. Finally, we constructed a further set of L exact MPS eigenstates. These are excited states in the aforementioned gapped regions of the phase diagram. We could use these states to place partial constraints on the location of the phase boundaries: the gap will certainly close when any one of these states becomes degenerate with |φ⟩. However, this would not determine the phase boundaries entirely. Therefore, we leave the characterisation of the phase boundaries for further work. Acknowledgements We are grateful to Paul Fendley and Riccardo Senese for illuminating discussions. We thank Abhishodh Prakash, Ruben Verresen and Harvey Weinberger for insightful discussions and for comments on the manuscript. Properties of (An+Gn) To prove Eq. (16), we note that the Onsager algebra admits the ladder operators and Rm,l† given by [69] Given an eigenstate |ψ⟩ of Al with energy E, these obey It follows that However, since |ψ0⟩ is the ground state of A0, there exists no eigenstate with energy E0-1. Therefore, Rn,0|ψ0⟩=0, and we have Thus, acting with (An+Gn) on |ψ0⟩ keeps us within the ground state sector of A0. Given that this ground state is unique, we must have Additionally, any eigenstate |ψ⟩ of A0 with energy E that is annihilated by Rn,0 obeys Thus, (An+Gn)|ψ⟩ acting on |ψ⟩ keeps us within the eigenspace of A0 with energy E. Hence, if all states within this eigenspace obey Rn,0|ψ⟩=0, it is possible to find a basis {|ϕE(i)⟩} such that Given that the lowest charge m excitations have energy E0+m/N [31], and that Rn,0|ϕE⟩∝|ϕE-1⟩, states with m2, leading us to conclude the wrong general form of {aα,gα}. We will argue that such cancellations do not occur in our calculations. Note that we have defined g(z) from f(z) such that the zeroth order term in g(z) does not vanish. Therefore, Tg(z) is always defined. Publisher's Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Data Availability Statement No data were created or analysed in this study.