The eight-vertex model (Baxter 1971)

Intuition Beginner

Think back to the grid of streets with an arrow on every segment. The six-vertex model of 08.13.02 kept one strict law at each crossing: exactly two arrows in and two out. That law is tidy, but it is also restrictive. The eight-vertex model loosens it. Now a crossing is allowed whenever an even number of arrows point inward, which means two-in-two-out as before, but also all-four-in and all-four-out. Those two extra patterns are the new arrangements that give the model its name: six old vertices plus two new ones makes eight.

Why bother adding them? Because the two new vertices carry a fresh weight, written $d$, that the six-vertex model simply did not have. Tuning $d$ links the lattice of arrows directly to two interleaved Ising magnets sitting on the same grid. So the eight-vertex model is a single dial that turns one famous solved model into another, sweeping through a whole landscape in between.

The reason this matters is that Baxter solved it exactly in 1971. The trick he found, making slightly different stamping machines agree, is the seed of a large part of modern mathematical physics.

Visual Beginner

The figure shows all eight allowed crossings. The first six are the familiar two-in-two-out patterns, grouped by weight into the pairs $a$, $b$, and $c$. The last two are the newcomers: one crossing with all four arrows pointing in, one with all four pointing out, sharing the weight $d$.

Reading the picture: when $d=0$ the two newcomers are forbidden and you fall back to the six-vertex model. Switching on $d$ lets arrow lines start and stop, so the neat count of down-arrows per row no longer stays fixed as you climb the lattice. That broken count is exactly what makes the older counting method stall, and what forces Baxter's deeper idea.

Worked example Beginner

Take one crossing with periodic edges, so the arrow leaving the right re-enters from the left and the arrow leaving the top re-enters from the bottom. Count the patterns with an even number of inward arrows.

First the two-in-two-out patterns: there are six of them, the same six from the ice model. Now the all-in pattern: every one of the four arrows points toward the centre, and that is a single arrangement. Likewise the all-out pattern is one more. So the even rule allows $6+1+1=8$ patterns.

Give weights by class. The two horizontal-line patterns each carry weight $a$, the two vertical-line patterns weight $b$, the two mixed patterns weight $c$, and the all-in and all-out patterns weight $d$. The total weight of this one crossing, summed over allowed patterns, is $2a+2b+2c+2d$. Set every weight to $1$ and the total is $8$, matching the eight allowed vertices.

What this tells us: the eight-vertex model is again a weighted count of arrow patterns, but with two more patterns than ice. The single new weight $d$ is the whole difference, and at $d=0$ the count drops back to the six of the ice model.

Check your understanding Beginner

Formal definition Intermediate+

Place arrows on the edges of an $N\times N$ square lattice with periodic boundaries. At each vertex admit the eight configurations with an even number of inward arrows; the eight with an odd number are excluded. Group the eight into four classes related by reversing every arrow, and assign Boltzmann weights $a$, $b$, $c$, $d$ to the four classes: $a$ to the two horizontal-flow vertices, $b$ to the two vertical-flow vertices, $c$ to the two mixed vertices, and $d$ to the two sources/sinks (all-in and all-out). The partition function is

$$Z=\sum _{\text{even configs}}{a}^{n_a}{b}^{n_b}{c}^{n_c}{d}^{n_d},$$

with $n_a,n_b,n_c,n_d$ counting vertices of each class. The case $d=0$ recovers the six-vertex model of 08.13.02; the case $a=b=c=d$ counts all even configurations ^{[Baxter Ch. 10]}.

The row-to-row transfer matrix $T$ acts on the $2^N$-dimensional space spanned by the horizontal-arrow states of one row, following the construction of 08.03.02; its matrix element between an incoming and an outgoing row is the product of vertex weights summed over the vertical-arrow states consistent with the even rule. As usual $Z=\mathrm{tr}{T}^N$, and the free energy per site is fixed by the largest eigenvalue ${\Lambda }_{\max }$:

$$f=-k_{B}T\underset{N\to \infty }{\lim }\frac{1}{N^2}\log Z=-k_{B}T\underset{N\to \infty }{\lim }\frac{1}{N}\log {\Lambda }_{\max },$$

exactly as for the Ising transfer matrix 08.03.01. Unlike the six-vertex case, the $d$-vertices break the conservation of reversed-arrow number, so $T$ is no longer block-diagonal in $m$ and the coordinate Bethe ansatz of 08.13.02 does not apply.

The solvable surface is the locus on which the weights solve the star–triangle relation of 08.13.01. Following Baxter, parametrise the four weights by Jacobi theta functions of a spectral parameter $u$ and a fixed crossing parameter $\eta$ and elliptic modulus $k$:

$$a:b:c:d=\Theta (\eta )\Theta (u-\eta )H(u):\Theta (\eta )H(u-\eta )\Theta (u):H(\eta )\Theta (u-\eta )\Theta (u):H(\eta )H(u-\eta )H(u),$$

where $H$ and $\Theta$ are Jacobi theta functions of nome $q$, matched to the conventions of 06.06.05. Two weight-independent invariants single out the line within this family: the anisotropy

$$\Gamma =\frac{ab-cd}{ab+cd},\Delta =\frac{a^2+b^2-c^2-d^2}{2(ab+cd)}.$$

Both $\Gamma$ and $\Delta$ are constant along the commuting family; varying $u$ at fixed $(\Gamma ,\Delta )$ sweeps the family, while $\eta$ and $k$ fix the line. At $d=0$ the parameter $\Delta$ reduces to the six-vertex anisotropy $(a^2+b^2-c^2)/(2ab)$.

Key derivation Intermediate+

Theorem (commuting eight-vertex transfer matrices and the $TQ$ relation). Let the weights $a,b,c,d$ be parametrised as above on an elliptic curve with fixed modulus $k$ and crossing parameter $\eta$, and let $T(u)$ be the row transfer matrix at spectral parameter $u$. Then $[T(u),T(u^{\prime })]=0$ for all $u,u^{\prime }$, and there is a one-parameter family $Q(u)$ of matrices commuting with each other and with every $T(u^{\prime })$ such that

$$T(u)Q(u)=\varphi (u-\eta )Q(u+2\eta )+\varphi (u+\eta )Q(u-2\eta ),$$

where $\varphi (u)=[\Theta (0)H(u)\Theta (u){]}^N$ is a scalar built from the same theta functions.

Derivation. Commutation is the content of 08.13.01: the star–triangle relation, in component form, equates two ways of routing three lines, and the elliptic parametrisation is chosen precisely so that the relation becomes the theta-function addition theorem

$$H(u+v)H(u-v)\Theta (0{)}^2=H(u{)}^2\Theta (v{)}^2-\Theta (u{)}^{2}H(v{)}^2,$$

an identity among Jacobi theta functions. Verifying that the three star–triangle equations close reduces to this single addition formula evaluated at the three rapidities; the spectral parameter $u$ is the device that turns the algebraic identity into a continuous commuting family, as proved in 08.13.01.

For the functional equation, Baxter constructs an auxiliary matrix $Q(u)$ by a second transfer matrix built from an auxiliary diagonal-to-diagonal lattice with a free internal parameter. The key structural facts are: (i) $Q(u)$ commutes with $T(u^{\prime })$ for all $u^{\prime }$, because both are built from $R$-matrices on the same elliptic curve and intertwine through the same star–triangle relation; and (ii) $Q(u)$ is nonsingular for generic $u$, so it can be inverted. Acting with $T(u)$ on an eigenvector shared with the whole commuting family, and using the local decomposition of the product $T(u)Q(u)$ at each site into two terms — one shifting the auxiliary rapidity up by $2\eta$, one down — produces the scalar two-term recursion above. The decomposition at a single site is the statement that the elliptic weights satisfy a fusion identity: the product of a physical and an auxiliary weight collapses onto two shifted auxiliary weights, with scalar coefficients $\varphi (u\mp \eta )$. Multiplying along the row of $N$ sites raises $\varphi$ to the $N$th power and assembles the matrix relation.

Because $T$ and $Q$ commute, restrict to a joint eigenvector: $T(u)$ acts by $\Lambda (u)$ and $Q(u)$ by $q(u)$, scalars. The matrix equation becomes the scalar $TQ$ relation

$$\Lambda (u)q(u)=\varphi (u-\eta )q(u+2\eta )+\varphi (u+\eta )q(u-2\eta ).$$

The eigenvalue $\Lambda (u)$ is an entire (theta-)function with $N$ quasi-periods; analyticity forces $q(u)$ to have the form $q(u)={\prod }_{j=1}^{N/2}H(u-u_j)\Theta (u-u_j)$ for zeros $u_j$, and requiring $\Lambda (u)$ to stay entire — that the apparent poles from $1/q$ cancel — yields the Bethe-type equations for the $u_j$:

$${\left[\frac{\Theta (u_j-\eta )H(u_j-\eta )}{\Theta (u_j+\eta )H(u_j+\eta )}\right]}^N=-\prod _{\ell =1}^{N/2}\frac{H(u_j-u_{\ell }-2\eta )\Theta (u_j-u_{\ell }-2\eta )}{H(u_j-u_{\ell }+2\eta )\Theta (u_j-u_{\ell }+2\eta )}.$$

These select the eigenvalue $\Lambda (u)$ for each state; the largest controls the free energy. $\square$

Bridge. This derivation builds toward the free-energy computation of the Master tier, where the inversion relation collapses $\Lambda (u)$ to a closed elliptic product. The $Q$-operator construction is exactly the structural replacement for the Bethe ansatz that failed once the $d$-vertices broke arrow conservation: where 08.13.02 diagonalised the transfer matrix by an explicit eigenvector, here the foundational reason for solvability is again that the family $\{T(u)\}$ commutes, and the $TQ$ relation is the device that extracts the spectrum without ever writing the eigenvector down. This is exactly the same star–triangle content of 08.13.01, now carried by an elliptic rather than a trigonometric $R$-matrix. Putting these together, the eight-vertex model generalises the six-vertex solution by trading the rational/trigonometric kernel for the elliptic one, and the bridge is the theta addition formula that closes the star–triangle relation. The same $TQ$ structure appears again in the corner-transfer-matrix order-parameter computation and identifies the lattice spectrum with the analytic data of a single elliptic curve.

Exercises Intermediate+

Advanced results Master

Baxter's free energy follows from the commuting family together with an inversion relation, a functional equation independent of the $TQ$ analysis. The partition-function-per-site $\kappa (u)={\Lambda }_{\max }(u{)}^{1/N}$ in the thermodynamic limit satisfies two relations dictated by the elliptic weights: a crossing (reflection) symmetry $\kappa (u)=\kappa (\eta -u)$ inherited from the star–triangle structure, and an inversion relation $\kappa (u)\kappa (-u)=\rho (u)$ with $\rho$ a known scalar built from the theta weights, valid because $T(u)T(-u)$ is scalar at the inversion point. Together with the analyticity and quasi-periodicity of $\kappa$ in $u$, these fix $\kappa (u)$ uniquely as an infinite product of theta functions. Evaluating at the physical value of $u$ gives the free energy of the eight-vertex model in closed form ^{[Baxter 1972]}; the six-vertex free energy of 08.13.02 is the $d\to 0$ degeneration of this elliptic product, and the Onsager Ising free energy ^{[Onsager 1944]} is the free-fermion specialisation $\Delta =0$.

The principal physical surprise is non-universal criticality. The model becomes critical when the elliptic modulus degenerates, $k\to 0$ or equivalently when one weight equals the sum of the others, e.g. $a=b+c+d$. Near criticality the specific heat diverges with an exponent that depends continuously on the crossing parameter $\mu$, defined through $\Delta =-\cos 2\mu$ (equivalently $\eta =\mu$ in suitable units). Baxter found

$$\alpha =2-\frac{\pi }{\mu },\beta =\frac{\pi }{16\mu },\nu =\frac{\pi }{2\mu },$$

so the exponents vary along the critical surface rather than taking fixed universality-class values. At the Ising point $\mu =\pi /4$ one recovers $\alpha =0$ (the logarithmic Onsager divergence, the $\alpha \to 0$ limit of $2-\pi /\mu$), $\beta =1/8$, $\nu =1$. The exponents satisfy the scaling and hyperscaling relations $\alpha +2\beta +\gamma =2$ and $2-\alpha =2\nu$ identically in $\mu$, so the weak universality of fixed ratios survives even though the individual exponents drift. This is the precise integrable counterexample that refines the universality picture of 08.05.01: universality of exponents is a property of generic fixed points, and the eight-vertex line is a marginal direction along which a continuous family of fixed points sits, with a marginal operator (the four-spin coupling) carrying the variation.

The lattice maps to a quantum spin chain by the standard transfer-matrix-to-Hamiltonian limit: the logarithmic derivative of $T(u)$ at the shift point is the XYZ Heisenberg Hamiltonian

$$H_{XYZ}=-\frac{1}{2}\sum _n(J_x{\sigma }_n^x{\sigma }_{n+1}^x+J_y{\sigma }_n^y{\sigma }_{n+1}^y+J_z{\sigma }_n^z{\sigma }_{n+1}^z),$$

with the three anisotropic couplings $J_x:J_y:J_z=(1+\Gamma ):(1-\Gamma ):\Delta$ fixed by the eight-vertex invariants. The six-vertex / XXZ chain of 08.13.02 is the $J_x=J_y$ (i.e. $\Gamma =0$, $d=0$) degeneration. The algebra controlling the elliptic $R$-matrix is the Sklyanin algebra, the elliptic deformation of $U({\mathfrak{sl}}_2)$ whose representation theory replaces the quantum-affine $U_q({\widehat{\mathfrak{sl}}}_2)$ of the trigonometric six-vertex case; this is the apex of the rational-trigonometric-elliptic ladder catalogued in 08.13.01.

Synthesis. Putting these together, the eight-vertex model identifies three previously separate facts as one. It is the apex solvable lattice model, whose free energy the inversion relation and the $TQ$ relation jointly compute; it is the integrable counterexample to exponent universality, the central insight being that continuously varying exponents arise from a marginal four-spin operator along an integrable line, with weak universality (fixed exponent ratios) surviving the drift; and it is the XYZ spin chain, dual to the elliptic Sklyanin algebra exactly as the six-vertex model is dual to the trigonometric quantum group. The foundational reason all three coincide is the single elliptic solution of the star–triangle relation of 08.13.01: solvability, non-universality, and the elliptic algebra are three readings of the same theta-parametrised $R$-matrix. This is exactly the structural principle that the rational and trigonometric cases of 08.13.02 anticipate, and the eight-vertex result generalises the Onsager solution of 08.03.01 by embedding it as the free-fermion slice of a continuous integrable family.

Full proof set Master

Proposition (the Ising decoupling at the free-fermion point). On the free-fermion surface $a^2+b^2=c^2+d^2$ (equivalently $\Delta =0$), the eight-vertex partition function factorises as a product of two Ising-type Pfaffians, and at the further symmetric point $a=b$, $c=d$ it equals the square of a single Onsager Ising partition function up to a known scalar.

Proof. Encode each arrow configuration by an Ising spin on every face of the lattice, with the rule that an arrow points so as to keep the face spins to its two sides in a fixed relative orientation; the even rule on arrows becomes the requirement that the product of the four face spins around each vertex is constrained, which is automatically satisfied for face-spin assignments. The Boltzmann weight of a vertex becomes a function of the four surrounding face spins, expandable as $w=w_0+w_1(\text{two-spin terms})+w_2{\sigma }_1{\sigma }_2{\sigma }_3{\sigma }_4$, where the four-spin coefficient $w_2$ is a fixed combination of $c$ and $d$. The free-fermion condition $a^2+b^2=c^2+d^2$ is exactly the vanishing of the obstruction to writing the vertex weight as a Gaussian (quadratic) form in Grassmann variables placed on the edges. With a Gaussian weight, the partition sum is a Grassmann integral whose value is a Pfaffian, as in the dimer/Pfaffian solution of the Ising model used by Onsager's successors ^{[Onsager 1944]}. The lattice's two interpenetrating face sublattices give two decoupled Gaussian sectors, hence a product of two Pfaffians. At $a=b$, $c=d$ the two sectors carry identical couplings, so the two Pfaffians are equal and the partition function is the square of a single Ising partition function, with the proportionality scalar absorbing the vertex normalisation. $\square$

Proposition (weak universality of the eight-vertex exponents). With $\Delta =-\cos 2\mu$ and Baxter's exponents $\alpha =2-\pi /\mu$, $\beta =\pi /16\mu$, $\nu =\pi /2\mu$, the Rushbrooke and hyperscaling relations $\alpha +2\beta +\gamma =2$ and $2-\alpha =2\nu$ hold for every $\mu$, with $\gamma =7\pi /8\mu$.

Proof. Hyperscaling: $2\nu =2\cdot \pi /2\mu =\pi /\mu$, and $2-\alpha =2-(2-\pi /\mu )=\pi /\mu$, so the two agree identically in $\mu$. Rushbrooke: substitute $\gamma =7\pi /8\mu$, giving $\alpha +2\beta +\gamma =(2-\pi /\mu )+2\cdot \pi /16\mu +7\pi /8\mu =2-\pi /\mu +\pi /8\mu +7\pi /8\mu =2-\pi /\mu +8\pi /8\mu =2-\pi /\mu +\pi /\mu =2$. The cancellation is exact for all $\mu$, so although each exponent slides with the crossing parameter, the scaling combinations are pinned. The fixed-ratio statement — that $\beta /\nu =1/8$ and $\gamma /\nu =7/4$ independent of $\mu$ — is read directly from the formulae and is the weak-universality content: the correlation-function exponents (which depend only on ratios such as $\beta /\nu$) are universal, while the thermal exponents (tied to the diverging length scale) are not. $\square$

The eigenvalue functions entering the inversion relation are constrained by the same analyticity used in the $TQ$ analysis; the inversion relation, the crossing symmetry, and quasi-periodicity overdetermine $\kappa (u)$, and Baxter's verification that the resulting infinite product is consistent with the $TQ$ spectrum is the cross-check that the two solution routes agree ^{[Baxter 1972]}.

Connections Master

The transfer-matrix formalism, the trace $Z=\mathrm{tr}{T}^N$, and the largest-eigenvalue free energy are imported from 08.03.02; this unit pushes that machinery to its hardest classical instance, where no explicit eigenvector exists and the spectrum is reached only through the commuting auxiliary $Q$-operator.

The commuting family and the elliptic star–triangle relation are the apparatus of 08.13.01: the eight-vertex $R$-matrix is the elliptic apex solution of the Yang–Baxter equation, and the theta addition formula that closes the star–triangle relation is what makes the whole construction work. The six-vertex model and Bethe ansatz of 08.13.02 are the trigonometric degeneration recovered at $d=0$, and the failure of that Bethe ansatz here — broken arrow conservation — is precisely what the $Q$-operator repairs.

The free-fermion slice $\Delta =0$ reproduces the Onsager solution 08.03.01 and the Ising model, while the continuously varying exponents are the integrable counterexample that sharpens the universality discussion of 08.05.01: the eight-vertex line is a marginal direction carrying a continuous family of fixed points, with weak universality surviving as fixed exponent ratios.

The elliptic weights are parametrised by the Jacobi theta functions of 06.06.05, a lateral link to the Riemann-surfaces chapter: the spectral parameter lives on an elliptic curve, and the modulus degeneration that drives criticality is the same nome limit that flattens that curve to a cylinder.

Historical & philosophical context Master

Rodney Baxter announced the exact solution of the eight-vertex model in a 1971 letter, reporting the free energy and the striking continuously varying critical exponents, and gave the full commuting-transfer-matrix derivation the following year ^{[Baxter 1971; Baxter 1972]}. The model had been introduced by Sutherland, Fan, and Wu as the natural generalisation of the six-vertex models that Lieb had solved in 1967 ^{[Lieb 1967]}, and its mapping to two coupled Ising lattices was understood before the solution; what was missing was a method, because the broken arrow-number conservation defeated the Bethe ansatz. Baxter's resolution — that the transfer matrices commute when the weights solve a star–triangle relation, and that a second commuting matrix $Q$ then satisfies a functional equation with $T$ — supplied the method and, in retrospect, the organising principle of the whole field.

The non-universal exponents were the conceptual shock. Onsager's 1944 Ising solution had established a fixed logarithmic specific-heat singularity ^{[Onsager 1944]}, and the renormalisation-group programme of the 1970s explained such fixed exponents as universal fixed-point data. Baxter's exponents instead slid continuously with the crossing parameter, which Kadanoff and Wegner reconciled with the renormalisation group by identifying a marginal operator — the inter-sublattice four-spin coupling — along whose direction a line of fixed points runs. The elliptic algebra behind the solution was abstracted by Sklyanin after 1979 into the algebra that now bears his name, completing the rational-trigonometric-elliptic hierarchy of integrable structures of which the eight-vertex model is the elliptic summit.

Bibliography Master

@article{Baxter1971eightvertex,
  author  = {Baxter, Rodney J.},
  title   = {Eight-Vertex Model in Lattice Statistics},
  journal = {Physical Review Letters},
  volume  = {26},
  year    = {1971},
  pages   = {832--833}
}

@article{Baxter1972,
  author  = {Baxter, Rodney J.},
  title   = {Partition Function of the Eight-Vertex Lattice Model},
  journal = {Annals of Physics},
  volume  = {70},
  year    = {1972},
  pages   = {193--228}
}

@article{Onsager1944,
  author  = {Onsager, Lars},
  title   = {Crystal Statistics. I. A Two-Dimensional Model with an Order-Disorder Transition},
  journal = {Physical Review},
  volume  = {65},
  year    = {1944},
  pages   = {117--149}
}

@article{Lieb1967f,
  author  = {Lieb, Elliott H.},
  title   = {Exact Solution of the F Model of an Antiferroelectric},
  journal = {Physical Review Letters},
  volume  = {18},
  year    = {1967},
  pages   = {1046--1048}
}

@article{SutherlandFanWu1973,
  author  = {Sutherland, Bill and Fan, Chungpeng and Wu, Fa-Yueh},
  title   = {Exact Solution of a Model of Two-Dimensional Ferroelectrics in an Arbitrary External Electric Field},
  journal = {Physical Review B},
  volume  = {7},
  year    = {1973},
  pages   = {4006--4009}
}

@article{KadanoffWegner1971,
  author  = {Kadanoff, Leo P. and Wegner, Franz J.},
  title   = {Some Critical Properties of the Eight-Vertex Model},
  journal = {Physical Review B},
  volume  = {4},
  year    = {1971},
  pages   = {3989--3993}
}

@article{Sklyanin1982,
  author  = {Sklyanin, E. K.},
  title   = {Some Algebraic Structures Connected with the Yang--Baxter Equation},
  journal = {Functional Analysis and Its Applications},
  volume  = {16},
  year    = {1982},
  pages   = {263--270}
}

@book{Baxter1982,
  author    = {Baxter, Rodney J.},
  title     = {Exactly Solved Models in Statistical Mechanics},
  publisher = {Academic Press},
  address   = {London},
  year      = {1982}
}