Abstract

We define a 2-dimensional Ising model on a triangulated sphere, $\mathbb{S}^{2}$, designed to approach the exact conformal field theory (CFT) in the continuum limit. Surprisingly, the derivation leads to a set of geometric constraints that the lattice field theory must satisfy. Monte Carlo simulations are in agreement with the exact Ising CFT on $\mathbb{S}^{2}$. We discuss the inherent benefits of using non-uniform simplicial lattices and how these methods may be generalized for use with other quantum theories on curved manifolds.

1 Introduction

Lattice Monte Carlo has proven to be a powerful method to numerically solve non-perturbative quantum field theories for condensed matter and relativistic high energy physics – most celebrated in the QCD sector of the standard model. But for the most part the high precision results  [1] are restricted to flat Euclidean space discretized on hypercubic lattices. Extensions to curved manifolds would open up new frontiers. For example the simulation of conformal  [2] or near conformal theories [3] on discretization of a Riemann [4] sphere $\mathbb{S}^{d}$ or on a cylinder, $\mathbb{R}\times\mathbb{S}^{d}$ for radial quantization  [5]. For conformal field theories, unlike toroidal lattices, the finite volume errors for spatial images are removed, giving direct access to conformal parameters.

However the problem with curved manifolds, even for spherical manifolds in 2 or more dimensions  [6], is that there are only a finite number of uniform discretizations. For the 2 sphere there are 5 Platonic solids. The largest, the icosahedron, consists of 20 equilateral triangles invariant under the finite 120 element subgroup, $A_{5}\times Z_{2}$, of $O(3)$. For the 3 sphere, the largest uniform sub-manifold consists of 600 equilateral tetrahedrons in the 14400 element Coxeter group $H_{4}$ subgroup of O$(4)$. Similar restrictions apply to Anti-de Sitter space [7, 8]. This means that for a uniform discretization on constant curvature manifolds, there is a fundamental minimum lattice spacing relative to the curvature. To avoid this limitation we seek a method for simulating a theory on a non-uniform lattices, which exactly approaches the symmetries of a smooth manifold as the effective lattice spacing goes to zero.

For classical physics, expressed as a system of partial differential equation, the problem of discretization on curved manifolds is solved by the Finite Element Method (FEM) [9]. Solutions to the discrete equations of motion on piecewise flat manifolds, compose of d dimensional simplicies (e.g. 2d triangles, 3d tetrahedrons, etc.), converge to the exact continuum on a suitable sequence of refined lattices. The same simplicial lattices were also introduced in 1961 for the metric field in Regge Calculus (RC) [10] to discretize the Einstein-Hilbert action resulting in powerful tool to construct numerical solutions to Einstein gravity. However quantum fields on these discrete manifolds are not so forgiving. Finite element methods fail.

One way to see the difficulty for quantum fields theories is to note that the ultraviolet divergences are sensitive to lattice cut-off, which is no longer independent of position. As a consequence in general there is no global second order fixed point. Removing the UV cut-off fails to converge to the quantum field theory or even to give the continuum renormalized perturbation theory. Our first remedy implemented in the Quantum Finite Elements (QFE) project [11] was to modify the FEM discretization by additional quantum counter terms to cancel the local cut-off dependence of UV divergent perturbation loops. Implemented for the super renormalizable $\phi^{4}$ theory enabled accurate lattice simulations for CFT data for 2d $\phi^{4}$ on $\mathbb{S}^{2}$ [4] and for 3d on $\mathbb{R}\times\mathbb{S}^{2}$ [2, 12]. However subsequent studies suggested that exact results for lattice $\lambda_{0}\phi^{4}$ theory in the continuum limit also required tuning the bare dimensionless coupling ($\lambda_{0}$) to zero by fixing UV renormalized parameters.

Here we seek a more general approach, applicable to strong coupling infrared fixed points. To this end, we studied the universally equivalent lattice Ising model. Surprisingly we found a solution for the 2d Ising model on a triangulated sphere with only nearest neighbor couplings. By smoothing the simplicial geometry, combined with matching the lattice coupling, this enabled simulations to reach the CFT in the continuum. A first step to this matching condition was identified in Ref. [13] in the analysis of the Ising model in flat space as a function of a global affine transformation. To restore Poincaré invariance of quantum correlators fixes the map from the geometry (edge lengths of triangles) to the couplings on the triangular Ising lattice as function (3). This map was found analytically using the star-triangle relation [14, 15] and the free fermion sector of the trivalent dual Ising model [16].

In Sec. 2 , we show how to generalize flat affine analysis to any triangulation of the Riemann sphere. The Kramers Wannier transformation maps the triangulated Ising model to a dual Ising model on a trivalent lattice, which is subsequently mapped to free Wilson-Majorana fermions [16]. All three representations give equivalent correlators. With suitable smoothness assumptions, the fermion representation is used to determine the geometry at short distances on each tangent plane of the curved manifold. The reader may prefer first to skip the technical details in subsections. 2.1 - 2.3, returning when needed. In Sec. 3, this map is applied to our target application on $\mathbb{S}^{2}$ as a sequence of refined lattices approaching the continuum. Extensive Monte Carlo simulations agree with the exact Ising minimal $c=1/2$ CFT two point function on the sphere. In Sec. 3.4 we discuss further tests and generalizations. We conjecture that a local affine transformation to the tangent planes is sufficient for any lattice field on a smooth manifold, when implemented with a suitable ancillary numerical algorithm.

2 Geometry of 2d Ising Graphs

The heart of the problem for lattice field theory on curved manifolds is how to reconcile two conflicting representations of geometry. In Regge calculus, the geometry consists of a collection of simplicies (e.g. triangles in 2d) with flat interiors. The edge lengths, $\ell_{ij}$, on the graph are a replacement for metric, $g_{\mu\nu}(x)$. Hence the title of Regge’s seminal 1961 paper [10] on General relativity without coordinates! The continuum requires taking edge lengths to zero and re-introducing a differentiable metric $g_{\mu\nu}(x)$. G. Feinberg, R. Friedberg, T. D. Lee, and H. C. Ren [17] give a detail procedure applied to spherical manifolds. In contrast a lattice quantum field theory on this same simplicial graph is defined by local dimensionless couplings on the edges, $K_{ij}$, with no metric. The quantum field geometry emerges dynamically determined by computing correlators as one approaches a second order phase boundary. To reconcile Regge geometry and lattice fields on a curved manifold requires finding a map coupling constants, $K_{ij}$, and edge lengths, $\ell_{ij}$,

$$\mbox{Map:}\\ \quad K_{ij}\rightarrow\ell_{ij}\;,$$ (1)

consistent with the target continuum geometry.

Interestingly this problem already arises in flat space for the 2d Ising model studied in [13]. An affine transformation on an equilateral triangle takes it to a general simplex with 3 distinct edge lengths, $\ell_{l},\ell_{2},\ell_{3}$. For a scalar field in this uniform affine lattice, it is reasonable to introduce 3 corresponding couplings, $K_{1},K_{2},K_{3}$ as illustrated on the right in Fig. 1. This preserves the discrete translational invariance of on the lattice.

Figure 1: On the left is the triangular graph (with black dots, solid lines) and its hexagonal dual graph (with open circles, dashed lines). In Regge Calculus the geometry is determine by assigning edge length, $\{\ell_{ij}\}$, to the each simplex. On the right for affine flat space example, the lattice action has 3 independent couplings, $\{K_{1},K_{2},K_{3}\}$ identical triangles 3 edge lengths, $\{\ell_{l},\ell_{2},\ell_{3}\}$.

For free scalar field theory with lattice action: $S[\phi_{j}]=\sum_{\hat{i}=1,2,3}K_{i}(\phi_{j+\hat{i}}-\phi_{j})^{2}$, linear finite elements provides the correct map,

$$2K_{i}=\frac{\ell^{*}_{i}}{\ell_{i}}\;.$$ (2)

The couplings are expressed as the ratio of Voronoi dual lengths, $\ell^{*}_{i}$, between circumcenter of two adjacent triangles with common edge length $\ell_{i}$. This restores rotational symmetry in the continuum for the resultant FEM Laplace-Beltrami operator. Incidentally this is an example of the elegant Discrete Exterior Calculus [9] (DEC) which applies to general simplicial manifold in any dimension.

In contrast for the Affine Ising lattice [13] the correct map is

$$\sinh 2K_{i}=\dfrac{\ell_{i}^{*}}{\ell_{i}}\;,$$ (3)

to restore rotational symmetry for critical Ising CFT on $\mathbb{R}^{2}$. This is a non-perturbative quantum effect at the Wilson Fisher fixed point, not part of classical FEM methods. The map in Eq. (3) also fixes the critical surface at

$$p_{1}p_{2}+p_{2}p_{3}+p_{3}p_{1}=1\quad,\quad p_{i}=e^{-2K_{i}}\;.$$ (4)

Both are consequence of the star-triangle identity [14, 15] between the triangular lattice with couplings $K_{1},K_{2},K_{3}$ and its Kramers Wannier dual with with couplings $L_{1},L_{2},L_{3}$ fixed by $e^{-2K_{i}}=\tanh(L_{i})$. The geometry is then fixed by the loop expansion for free Wilson-Majorana fermions on the trivalent dual lattice.

The take away is that even on $\mathbb{R}^{2}$, restoring Poincaré invariance require solving the the quantum theory. The problem we now face is how to generalize this to curved manifolds, forgoing the dependance on analytical tools of the 2d flat space Ising model. The hope is that the map in Eq. (3) can still be apply locally on each tangent plane as we take the continuum limit.

2.1 Kramers Wannier Duality

A general triangulated surface with no boundaries is defined by a set of $N$ vertices, $E$ edges, and $F$ triangular faces satisfying Euler’s theorem, $F-E+N=2-2g$. The topology is fixed by the genus , $g$. Constant curvature metric exist for the positive curvature sphere at $g=0$, zero curvature torus at $g=1$ and negative curvature hyperbolic Riemann surfaces for $g>1$.

The graph dual to any triangulated Riemann surface is trivalent graph as illustrated in Fig. 2 with a one to one match of edges and dual edges.

Figure 2: An illustration of a loop configuration for a non-uniform simplicial lattice and its trivalent dual lattice on portion of Riemann manifold. The bold solid lines show two dual loops ($\Gamma_{i}$) in the low temperature.

The Ising model triangular graph can be define the sum over the edges of the triangulated graphs with action,

$$S_{\triangle}=-\sum_{\langle ij\rangle}K_{ij}s_{i}s_{j}\;.$$ (5)

Similarly on the dual graph we can define a dual Ising spin system

$$S_{\text{dual}}=-\sum_{\langle ij\rangle}L_{ij}s_{i}s_{j}\;.$$ (6)

Properly the dual Ising spins are now disorder variables, $\mu_{i^{*}}=\pm 1$, at dual sites $i^{*}$ but for notational simplicity we label them by $\mu_{i^{*}}\rightarrow s_{i}$ without the star on dual sites $i^{*}\rightarrow i$. The coupling constants for the triangle and trivalent dual Ising models are paired as illustrated in Fig. 3. In passing we note the special simplicity the trivalent dual of 2d triangle simplicial complex for any Riemann triangulated surface is at the heart of the random graph solution to 2d string theory as a large N matrix model in the double scaling limit [18].

For any fixed triangulation, there is a general Kramers Wannier map [19] found by comparing strong and weak loop expansions.

Figure 3: Couplings for a single edge on both the simplicial lattice (closed circles, solid lines) and its trivalent dual (open circles, dashed lines).

On the triangular graph, the low temperature expansion is a power series in the number of broken bonds,

$$Z_{\triangle}[K]=Tr[e^{K_{ij}s_{i}s_{j}}]=\sum_{\{s_{i}\}=\pm 1}\prod_{\langle ij\rangle}e^{K_{ij}}[e^{-2K_{ij}}+(1-e^{-2K_{ij}})\delta_{s_{i},s_{j}}]=2\prod_{\langle ij\rangle}e^{K_{ij}}\sum_{\{\Gamma\}}\prod_{\langle ij\rangle\in\Gamma}e^{-2K_{ij}}$$ (7)

Since the dual lattice is trivalent the path of broken bounds form non-intersecting loops and $\Gamma$ enumerates the loops ensemble.

These same loop enumeration also give the high temperature for aligned spins $s_{i}=s_{j}$ on the dual lattice for sphere ($g=0$) or the topologically equivalent $CP(1)$ complex plane,

$$Z_{\text{dual}}[L]=Tr[e^{L_{ij}s_{i}s_{j}}]=\sum_{\{s_{i}\}=\pm 1}\prod_{\langle ij\rangle}\cosh L_{ij}(1+s_{i}s_{j}\tanh L_{ij})=2^{F}\prod_{\langle ij\rangle}\cosh L_{ij}\sum_{\{\Gamma\}}\prod_{\langle ij\rangle\in\Gamma}\tanh L_{ij}$$ (8)

where $F$ is the number of faces on the triangular sphere. For $g=0$, partition functions in Eqs, (7-8) are equivalent if we assert the identity,

$$e^{-2K_{ij}}=\tanh L_{ij}$$ (9)

or its equivalent to the symmetric form

$$\sinh 2K_{ij}\sinh 2L_{ij}=1\;.$$ (10)

Although we use the term ”expansion”, both Eq. (7) and Eq. (8) are exact. We note that for higher genus Riemann surfaces ($g\geq 1$) such as the modular torus, the non-contractible loops require addition care with boundary conditions.

This is a powerful map central to our construction. It fixes the ratio of the partition functions or equivalently the difference of the free energies ,

$$F_{\triangle}[K]=F_{\text{dual}}[L]+\frac{1}{2}\sum_{\langle i,j\rangle}\log(\sinh(2L_{ij}))+(F-1)\log(2)\;.$$ (11)

Taking derivatives with respect to the link coupling, $K_{ij}$, all correlation function for $Z_{2}$ even Ising energy operators on the edges of triangular lattice are mapped to dual correlation functions. For example if we consider the connected two point energy-energy correlators between any two edges $\langle 1,2\rangle$ and $\langle 3,4\rangle$, we have

$$\partial_{K_{12}}\partial_{K_{34}}F_{\triangle}[K]=\langle(s_{1}s_{2})(s_{3}s_{4})\rangle_{\triangle}$$ (12)

which is mapped to the dual by the identity,

$$\langle(s_{1}s_{2})(s_{3}s_{4})\rangle_{\text{dual}}=\sinh(2L_{12})\sinh(2L_{34})\;\langle(s_{1}s_{2})(s_{3}s_{4})\rangle_{\triangle}$$ (13)

using the chain rule, it $dL/dK=-\sinh(2L)$ on Eq. (10).

At present we don’t have enough information to determine the geometry of the quantum system. The positions of the vertices, the lengths of the edges, and the locations of the dual sites within the triangular faces are all unknown quantities. Only if the lattice model has a well-defined continuum limit at a second order phase boundary, can we find the emergent quantum geometry through its correlation functions. To implement this, we now map the trivalent Ising partition function to an equivalent Wilson-Majorana free fermion partition function.

2.2 Wilson-Majorana fermion loop expansion

We introduce Wilson-Majorana fermion on the trivalent dual lattice, with fields $\psi_{i}^{\alpha}$ ($\alpha=1,2$). Each field is a two-component spinor that obeys the charge conjugation constraint $\bar{\psi}_{i}^{\beta}=\psi^{\alpha}_{i}\epsilon^{\alpha\beta}$, i.e. $\bar{\psi}^{1}_{i}=-\psi^{2}_{i}$ and $\bar{\psi}^{2}_{i}=\psi^{1}_{i}$. The action is

$$S_{\psi}=\dfrac{1}{2}\sum_{i}\bar{\psi}_{i}\psi_{i}-\sum_{\langle ij\rangle}\kappa_{ij}\bar{\psi}_{i}P_{ij}\psi_{j}\;,$$ (14)

with local hop** parameters, $\kappa_{ij}$ and Wilson spinor propagator,

$$P_{ij}=\dfrac{1}{2}(1+\hat{e}_{ij}\cdot\vec{\sigma})\;.$$ (15)

The vectors, $\hat{e}_{ij}$, are defined as a unit vector on the dual links from $i$ to $j$, so that Wilson factor are projection operator, $P^{2}_{ij}=P_{ij}$, with two eigenvalues $0,1$. Wilson-Majorana fermion with real Grassmann variables has a symmetry property identifying directed links $i\rightarrow j\equiv\{i,j\}$ with the reflected direction $j\rightarrow i\equiv\{j,i\}$ :

$$\bar{\psi}_{i}(1+\hat{e}_{ij}\cdot\vec{\sigma})\psi_{j}=\bar{\psi}_{j}(1+\hat{e}_{ji}\cdot\vec{\sigma})\psi_{i}$$ (16)

where $\hat{e}_{ji}=-\hat{e}_{ij}$. This is an additional reflection symmetry, independent of so called $\gamma_{5}$ hermiticity for Wilson Fermions. So there is no need to add links in both directions. This is similar to the convention introduced in Ref. [11] for Wilson fermions, except now the fermions are moved from the triangulated simplex to sites on the dual lattice. The geodesics between dual sites are straight lines illustrated in Fig.4.

Figure 4: The angles at each of the trilinear sites provide a discrete rotation of the Wilson-Majorana propagator. This forms a Voronoi polyhedron dual to the central vertex ($\bullet$). with area $A^{*}_{\bullet}$.

Now we show that free fermion partition function,

$$Z_{\psi}=\int\mathcal{D}\psi e^{-S_{\psi}}\;,$$ (17)

with Grassmann path integral measure

$$\mathcal{D}\psi\equiv\prod_{i}d\psi_{i}^{1}d\psi_{i}^{2}$$ (18)

with appropriate hop** parameters: $\kappa_{ij}$ is equivalent to the Ising partition function on the dual lattice (8). We can expand the exponential in Eq. (17),

$$Z_{\psi}=\int\mathcal{D}\psi_{i}\prod_{i}\left(1-\dfrac{1}{2}\bar{\psi}_{i}\psi_{i}\right)\prod_{\langle ij\rangle}\left(1+\dfrac{1}{2}\bar{\psi}_{i}(1+\hat{e}_{ij}\cdot\vec{\sigma})\psi_{j}\right)\;,$$ (19)

discarding all terms above first order which vanish in a Grassmann integral. Now, following the standard rules for integration of Grassmann variables, the only nonzero terms are those which include every pair $\bar{\psi}_{i}\psi_{i}=\epsilon^{\alpha\beta}\psi^{\alpha}_{i}\psi^{\beta}_{i}$ exactly once on each site $i$. The set of nonzero terms is exactly described by the same set of unique loop configurations $\{\Gamma\}$ that we used in Sec. 2.1. Due to Eq. (16), it does not matter which direction we choose to traverse each loop, the result will be the same either way. Discarding irrelevant constant factors, the fermion partition function becomes

$$Z_{\psi}=\sum_{\{\Gamma\}}(-1)^{n}\operatorname{tr}\left[\prod_{\langle ij\rangle\in\Gamma}\dfrac{1}{2}\kappa_{ij}(1+\hat{e}_{ij}\cdot\vec{\sigma})\right]$$ (20)

where $n$ is the number of closed loops in $\Gamma$.

The evaluation depends on the product of adjacent projection operators. We may choose an arbitrary coordinate system with $\hat{e}=[e_{x},e_{y}]$ across each edge, given as a rotation from the dyadic along the x axis

$$P(\vec{n})=\frac{1}{2}(1+\hat{n}\cdot\vec{\sigma})=e^{\textstyle i\theta\sigma^{z}/2}\big{[}\frac{1+\sigma^{x}}{2}\big{]}e^{\textstyle-i\theta\sigma^{z}/2}$$ (21)

or introducing complex variable $z=n_{x}+i\;n_{y}=e^{i\theta}$ as dyadic,

$$P(\theta)=\frac{1}{2}\begin{bmatrix}e^{i\theta/2}&\\ 0&e^{-i\theta/2}\end{bmatrix}\begin{bmatrix}1&1\\ 1&1\end{bmatrix}\begin{bmatrix}e^{-i\theta/2}&\\ 0&e^{i\theta/2}\end{bmatrix}=\frac{1}{2}\begin{bmatrix}\sqrt{z}\\ \sqrt{z^{*}}\end{bmatrix}\begin{bmatrix}\sqrt{z^{*}}&\sqrt{z}\end{bmatrix}\equiv|z\rangle\langle z|\;,$$ (22)

where have used the constraint $zz^{*}=1$. In the loop expansion each corner has the spinor matrix, $\langle z_{i}|z_{j}\rangle=e^{\theta_{ij}\sigma^{z}/2}$ which is evaluated between the non-zero Grassmann integral $\epsilon^{\alpha\beta}\psi^{\alpha}\psi^{\beta}$. This takes the trace, $Tr[e^{\theta_{ij}\sigma^{z}}/2]=2\cos(\theta_{ij}/2)$, resulting in the loop expansion,

$$Z_{\psi}=\sum_{\{\Gamma\}}\prod_{\langle ij\rangle\in\Gamma}\kappa_{ij}\cos\dfrac{\theta_{ij}}{2}\;.$$ (23)

Here our notation is a bit compressed. Fermions live on dual sites $i$ and propagate on directed links, $i\rightarrow j\equiv\{i,j\}$, along unit vectors $\hat{e}_{ij}$ with couplings $\kappa_{ij}$. For a fixed loop as illustrated in Fig. 4 we may relabel the sites in sequence a $1\rightarrow 2\rightarrow 3\rightarrow\cdots$ with angles $\theta_{i+1,i}$ at each vertex $i$. Of course the vertex angles depend on the path. So a proper notation without relabeling sites for the rotation at vertex $k$ for a general path $i\rightarrow k\rightarrow j$ is determined by the scalar product $\hat{e}_{ik}\cdot\hat{e}_{kj}=\cos(\theta_{ik}-\theta_{kj})$. This requires 3 site labels for the rotation angle: $\theta^{k}_{ij}\equiv\theta_{ik}-\theta_{kj}$. The sign is undetermined because the loop is unchanged if we reverse the direction due the reflection identity in Eq. (16).

In the affine case there are 3 distinct links enumerated by 3 out-going (or in-going) vectors, $\hat{n}_{1},\hat{n}_{2},\hat{n}_{3}$ in a bipartite graph. In Ref. [13], we proved the identity

$$t_{2}/\kappa_{2}=\frac{\cos(\theta_{12}/2)\cos(\theta_{23}/2)}{\cos(\theta_{31}/2)}=\frac{\cos((\theta_{1}-\theta_{2})/2)\cos((\theta_{2}-\theta_{3})/2)}{\cos((\theta_{3}-\theta_{1})/2)}=\frac{\langle z_{2}|z_{1}\rangle\langle z_{3}|z_{1}\rangle}{\langle z_{2}|z_{3}\rangle}$$ (24)

and cyclic permutations. Remarkably this can be generalized for any trivalent lattice to the identity

$$\tanh^{2}L_{ij}=\kappa_{ij}^{2}\dfrac{\cos(\alpha_{2}/2)\cos(\alpha_{3}/2)\cos(\beta_{2}/2)\cos(\beta_{3}/2)}{\cos(\alpha_{1}/2)\cos(\beta_{1}/2)}$$ (25)

where the six angles are defined in Fig. 5. It follows by matching terms in loop expansion (8) and (23) or proven algebraically by extending the analysis above for each of the 4 paths through the link $\langle i,j\rangle$ with $\alpha_{2}$ or $\alpha_{3}$ on the left and $\beta_{2}$ or $\beta_{3}$ on the right. In flat space this simplifies to square of the affine identity in Eq. (24), with the condition that $\alpha_{i}=\beta_{i}$. We can use Eq. (10) to relate the triangular lattice coupling $K_{ij}$ to the corresponding dual lattice coupling $L_{ij}$ and by this association we now have three different models with equivalent partition functions.

Figure 5: Definition of the 6 angles used in Eq. (25). The open circles are the circumcenters of the two triangles.

2.3 Lattice continuum limit

We have now shown that the Ising model on the simplicial triangular lattice, its trivalent dual and the free Wilson-Majorana fermion all have equivalent partition functions, and therefore they all describe systems with equivalent dynamics. But we haven’t yet identified the critical point of the system, nor have we determined the appropriate geometry of the continuum theory (assuming such a theory exists). This requires a smooth local lattice theory, most likely close to the uniform flat space triangulated lattice studied in Ref. [13].

The first step is we need to have a smooth definition of the scalar curvature that at present are delta function singularities at the triangular manifold. This can be done to $O(a^{2})$ using the theorem in the remarkable paper on Lattice Gravity Near the Continuum Limit  [17] that proves that in 2d the scalar curvature, $R=2\epsilon_{i}/A^{*}_{i}+O(a^{2})$, is given by the ratio of the deficit angle $\epsilon_{i^{*}}$ at the vertices of the triangulated manifold and the dual area $A^{*}_{i}$ on the trivalent lattice as illustrated in Fig. 4. Thus we can we can introduce a smooth approximation to the manifold by spreading the singular curvature uniformly over the dual area and introducing a modified fermion action analogous to the Dirac fermions in Ref. [11] for the simplicial lattices,

$$\widetilde{S}_{\psi}=\dfrac{1}{2}\sum_{i}\bar{\psi}_{i}\psi_{i}-\dfrac{1}{2}\sum_{\langle ij\rangle}\kappa_{ij}\bar{\psi}_{i}(1+\hat{e}_{ij}\cdot\vec{\sigma})\Omega_{ij}\psi_{j}\;.$$ (26)

Now $\hat{e}_{ij}$ is a unit tangent vector at site $i$ on the geodesic from $i$ to $j$ and $\Omega_{ij}=\Omega^{\dagger}_{ji}$ is the rotation (or discrete spin connection) to the vector $\hat{e}_{ji}$ at site $j$ on the reverse geodesic to $i$. Due to local curvature there are two distinct tangent vectors at each end of each link that are no longer reflected: $\hat{e}_{ji}\neq-\hat{e}_{ij}$. Instead they obey the discrete tetra hypothesis introduced in Ref. [11],

$$\hat{e}_{ji}=-\Omega^{\dagger}_{ij}\hat{e}_{ij}\Omega_{ij}\;.$$ (27)

Here the local curvature between dual lattice points requires replacing the in Eq 16 by

$$\bar{\psi}_{i}(1+\hat{e}_{ij}\cdot\vec{\sigma})\Omega_{ij}\psi_{j}=\bar{\psi}_{j}\Omega_{ji}(1+\hat{e}_{ji}\cdot\vec{\sigma})\psi_{j}$$ (28)

Now in a smooth gauge with $\Omega_{ij}=1+O(a)$, we can expand (26) in the effective lattice spacing $a$ using

$$\psi_{j}=\psi_{i}+\vec{l}_{ij}^{*}\cdot\vec{\nabla}\psi_{i}+\mathcal{O}(a^{2})$$ (29)

and

$$\Omega_{ij}=e^{-i\vec{l}_{ij}^{*}\cdot\vec{\omega}}=1-i\vec{l}_{ij}^{*}\cdot\vec{\omega}+\mathcal{O}(a^{2})$$ (30)

where $\vec{l}_{ij}^{*}$ is a vector from $i$ to $j$ with length $l_{ij}^{*}=\mathcal{O}(a)$ and $\vec{\omega}$ is the continuum spin connection. The action becomes

$$\begin{split}\widetilde{S}_{\psi}&=\dfrac{1}{2}\sum_{i}\bar{\psi}_{i}\psi_{i}-\dfrac{1}{2}\sum_{\langle ij\rangle}\kappa_{ij}\bar{\psi}_{i}(1+\hat{e}_{ij}\cdot\vec{\sigma})(1-i\vec{l}_{ij}^{*}\cdot\vec{\omega})(1+\vec{l}_{ij}^{*}\cdot\vec{\nabla})\psi_{i}+\mathcal{O}(a^{2})\\ &=\dfrac{1}{2}\sum_{i}\left(1-\dfrac{1}{2}\sum_{j\in\langle ij\rangle}\kappa_{ij}\right)\bar{\psi}_{i}\psi_{i}+\dfrac{1}{2}\sum_{\langle ij\rangle}\kappa_{ij}\bar{\psi}_{i}(\hat{e}_{ij}\cdot\vec{\sigma})(\vec{l}_{ij}^{*}\cdot\vec{D})\psi_{i}\end{split}$$ (31)

where $\vec{D}=\vec{\nabla}-i\vec{\omega}$ is the covariant derivative for the spinor field and in the second line we have used $\bar{\psi}_{i}\vec{\sigma}\psi_{i}=0$ and discarded all terms of $\mathcal{O}(a^{2})$.

We would like to show that this is equivalent to the continuum action for a free fermion on a Riemannian manifold,

$$S_{\text{cont}}=\dfrac{1}{2}\int d^{2}x~{}\sqrt{g}\bar{\psi}(\vec{x})(m+\vec{\sigma}\cdot\vec{D})\psi(\vec{x})\;.$$ (32)

In order to get Eq. (31) into this form, we need the $\vec{\sigma}$ and $\vec{D}$ in the lattice action to be contracted, which is not possible for an arbitrary choice of lattice vectors in a non-uniform lattice. However, provided that all adjacent triangles have the same circumradius, the following identity is true for a lattice which is the circumcenter dual of a non-uniform simplicial lattice in 2 dimensions:

$$\sum_{j\in\langle ij\rangle}l_{ij}^{\mu}l_{ij}^{*\nu}=2A_{i}\epsilon^{\mu\nu}$$ (33)

where $\vec{l}_{ij}$ is the lattice vector in the simplicial lattice dual to $\vec{l}^{*}_{ij}$ and $A_{i}$ is the area of the triangular face dual to site $i$.

In order to apply this identity, it is necessary for $\kappa_{ij}$ to be proportional to the corresponding triangular lattice length $l_{ij}$. Substituting this all into the action we obtain

$$\widetilde{S}_{\psi}=\dfrac{1}{2}\sum_{i}A_{i}\bar{\psi}_{i}\left(m_{i}+\vec{\sigma}\cdot\vec{D}\right)\psi_{i}+\mathcal{O}(a^{2})$$ (34)

with a dimensionless mass parameter

$$m_{i}=\dfrac{1}{A_{i}}\left(1-\dfrac{1}{2}\sum_{j\in\langle ij\rangle}\kappa_{ij}\right)\;.$$ (35)

In the limit $a\to 0$ we identify the limit of the discrete integration measure $A_{i}\to d^{2}x\sqrt{g}$ and find that the lattice action converges to the continuum action in Eq.  (32), as required. To restore local rotational symmetry for the continuum fermion implies that the critical hop** parameter values for each link are defined by

$$\kappa_{ij}=\dfrac{2l_{ij}}{\sum_{k\in\langle ik\rangle}l_{ik}}\;,$$ (36)

Now, in order to have a well-defined critical point, we must satisfy this relation for every link on the lattice simultaneously, but upon inspection we find that this is only possible if all of the triangular faces have equal perimeter. Combining Eq.  (25) and (36) we obtain an expression which relates the dual lattice Ising coupling $L_{ij}$ to the geometry of the lattice

$$\tanh^{2}L_{ij}=\dfrac{4l_{ij}^{2}\cos(\alpha_{2}/2)\cos(\alpha_{3}/2)\cos(\beta_{2}/2)\cos(\beta_{3}/2)}{P_{\triangle}^{2}\cos(\alpha_{1}/2)\cos(\beta_{1}/2)}$$ (37)

where $P_{\triangle}$ is the triangle perimeter. We can also recover the critical couplings $K_{ij}$ for the triangular lattice through Eq. ( 10).

The two geometrical constraints of equal triangle circumradius and equal triangle perimeter place strict limitations on the set of simplicial lattices for which it is possible to define a critical theory with a well-defined continuum limit. They are natural generalizations of the flat space affine solution  [13], which gave exact Ising correlator on flat space and on the genus $g=1$ modular torus. The zero mass condition $m_{i}=0$ in Eq. (35) reduces to Eq.  (4).

3 Ising model on a 2-sphere

We now proceed to test whether these local constraints are sufficient in the continuum to give the critical Ising model on the Riemann sphere.

We define a base discretization of a 2-sphere illustrated in Fig. 6 by introducing equilateral triangles on each of the 20 faces of one of the simplicial Platonic solids. These are then projecting radially onto a unit sphere define by 3 vectors $\vec{r}_{i}$ in $\mathbb{R}^{3}$. The result is smooth but non-uniform triangulation of the sphere that in the continuum limit approaches an affine map to each tangent plane. At finite refinement the triangulated Regge manifold [10] consist of piecewise flat triangle, with all curvature is confined to singularities at the vertices.

Figure 6: Steps in the basic discretization of $\mathbb{S}^{2}$ using an icosahedral base, shown here for a refinement of L = 3 into $L^{2}$ triangle on each of the 20 icosahedral faces subsequently projected onto the sphere.

First we note that the projected triangulated sphere illustrate on right in Fig.6 does give critical Ising model in the continuum, but it fails to recover spherical isometries for the 2d Ising CFT on $\mathbb{S}^{2}$. To fix this we proceed to move the position of the vertices on the sphere to to minimize non-uniformities in the triangle circumradius and perimeter as suggested in Sec.  2.3. We show that Monte Carlo simulations extrapolated to the continuum on these modified smooth spherical lattices are in good agreement with the exact solution of the 2d Ising CFT on $S^{2}$. The geometric constraints of uniform circumradius and perimeter are indeed required to reach the desired continuum limit quantum field theory.

We use the simplicial Ising action (5) with coupling constants,

$$\sinh(2K_{ij})=\frac{\ell^{*}_{ij}}{\ell_{ij}}$$ (38)

as a function of the edge lengths, $\ell_{ij}=|\vec{r}_{i}-\vec{r}_{j}|$, and its circumcenter dual lengths, $\ell^{*}_{ij}$, perpendicular to the edge $\langle i,j\rangle$. It is important to note that the dual lattice edge lengths $l_{ij}^{*}$ are not calculated by using the geodesic distance between triangle circumcenters in the 3-dimensional embedding space. Under the conventions of Regge calculus, the manifold is defined on the flat triangular faces of the simplicial complex so the dual lattice lengths are computed by following intrinsic Regge geodesics as straight lines between flat triangular faces. The only curvature consists of delta function singularities at the vertex proportional to the deficit angle from adjacent triangles. For piecewise Regge manifolds in high dimensions singular curvature are on 2d co-dimension hinges (2d points , 3d edges, 4d triangles etc.). Because the perimeters of pairs of triangles which share an edge are not always equal we instead define $P_{\triangle}$ in Eq.  (37) as the geometric mean of the perimeters of the two triangles

3.1 Geometric uniformity

The solution requires two steps. One is to establish a more uniform Regge geometric manifold to match the continuum spherical manifold. This is necessary to related the discrete simplicial geometry to a differential co-ordinate on the sphere as described Ref. [17] and utilized to go local continuum on tangent planes described in Sec. 2.3 above.

The only simplicial discretizations of a 2-sphere which can be constructed from uniform triangles are the platonic solids: tetrahedron, octahedron, and icosahedron. We are therefore unable to construct arbitrarily refined discretizations of a sphere which satisfy the constraints of uniform circumradius and perimeter exactly. Instead, we will generate a sequence of lattices such that the non-uniformity in these two quantities goes to zero in the continuum limit.

We define the non-uniformity in the circumradius $R$ and perimeter $P$ as

$$E_{R}=\dfrac{\langle R^{2}\rangle}{\langle R\rangle^{2}}-1\qquad\text{and}\qquad E_{P}=\dfrac{\langle P^{2}\rangle}{\langle P\rangle^{2}}-1$$ (39)

where here the angle brackets denote an average over all triangles in the simplicial complex. These quantities can be understood as the normalized variance in the circumradius and perimeter over the entire lattice. Our goal is to move the vertices of the basic discretization (without adding or removing any edges) so that both of these quantities are minimized while retaining the point group symmetry of the original lattice. We do this by minimizing the sum $E=E_{R}+E_{P}$. In general, there are not enough degrees of freedom to find a solution such that $E=0$, however we will show that the minimum value of $E$ goes to zero as the square of the effective lattice spacing. As shown in Sec. 2.3, this is sufficient to restore local rotational symmetry in the continuum close to each tangent plane of the critical Ising model.

3.2 Projection for Spherical Symmetry

The spin-spin correlation function on the sphere is calculated by projecting onto spherical harmonics to test spherical symmetry. This is analogous to the discrete Fourier expansion on a hypercubic lattice to recover Poincaré invariance at long distances. However the transform on a simplicial sphere is a bit more involved as explained in detail in Appendix A . The result is the expansion,

$$\langle s(\hat{n}_{i})s(\hat{n}_{j})\rangle=\sum_{\ell^{\prime},m^{\prime};\ell,m}C_{\ell^{\prime}m^{\prime};\ell,m}Y^{*}_{\ell^{\prime}m^{\prime}}(\hat{n}_{i})Y_{\ell m}(\hat{n}_{j})\;,$$ (40)

with $i,j=1,\cdots,N$. The coefficients are well approximate by

$$C_{\ell^{\prime}m^{\prime};\ell,m}=\dfrac{1}{N^{2}}\sum_{ij}\sqrt{g_{i}g_{j}}Y_{\ell^{\prime}m^{\prime}}(\hat{n}_{i})Y^{*}_{\ell,m}(\hat{n}_{j})\langle s_{i}s_{j}\rangle+O(a^{2})\;,$$ (41)

for $\ell,\ell^{\prime}\ll\sqrt{N}$. In the continuum limit spherical symmetry demands that only the diagonal terms survive,

$$C_{\ell^{\prime}m^{\prime};\ell,m}=\delta_{\ell^{\prime},\ell}\delta_{m^{\prime},m}C_{\ell,m}$$ (42)

and with $C_{\ell,m}$ independent of m, the scalar two point function only depends on the geodesic through $\hat{n}_{i}\cdot\hat{n}_{j}=\cos(\theta_{ij})$. In practice we have only performed tests on the m dependent diagonal term $\ell^{\prime}=\ell,m^{\prime}=m$ for harmonics $\ell\leq 12$.

Here, $N$ is the number of lattice sites and the quantity $\sqrt{g_{i}}$ is a discrete integration measure for each site. Following the circumcenter-based conventions of discrete exterior calculus [9], it is proportional to the area of the dual simplex associated with the site (i.e. the Voronoi area) as shown in Figure 7.

Figure 7: The shaded region is the Voronoi area, $\sqrt{g_{i}}=A_{i}$, at the lattice site $i$. The open circles are the circumcenters of the adjacent triangles.

Because we are using a non-uniform lattices, we adopt a lattice spacing definition as a global average over all of the sites. For a lattice discretization of a sphere with $N=2+10L^{2}$ vertices, we simply define the lattice spacing to be equal to the square root of the average Voronoi area,

$$\dfrac{a^{2}}{r^{2}}=\dfrac{1}{N}\sum_{i}\sqrt{g_{i}}$$ (43)

where $r$ is the radius of the sphere.

3.3 Restoring Rotational Symmetry

The 2-sphere is invariant under the orthogonal group O(3), which has an infinite set of irreducible representations labeled by the familiar $\ell$ index used to describe the spherical harmonics, $Y_{\ell m}(\theta,\phi)$. For the spin-spin correlation function in our lattice model, restoration of rotational symmetry in the continuum limit requires that for each $\ell$, all of the measured $C_{\ell m}$ coefficients for $m\in\{-\ell,...,\ell\}$ must become degenerate as $a\to 0$. To check this, we define a measurement of rotational symmetry breaking

$$\delta C_{\ell}=1-\dfrac{C_{\ell m}^{\textrm{(min)}}}{C_{\ell m^{\prime}}^{\textrm{(max)}}}$$ (44)

which captures the maximum deviation between the $C_{\ell m}$ coefficients for a given value of $\ell$.

Because our lattices have been constructed to be symmetric under a discrete subgroup of O(3), some of these coefficients will automatically be degenerate. The full octahedral symmetry group $O_{h}$ contains a 3-dimensional irreducible representation $T_{1u}$ which maps exactly onto the $\ell=1$ irreducible representation of O(3). Similarly, the full icosahedral symmetry group $I_{h}$ contains the 3-dimensional and 5-dimensional irreducible representation $T_{1u}$ and $H_{g}$ which map exactly onto the $\ell=1$ and $\ell=2$ irreducible representation of O(3), respectively. The corresponding coefficients of the 2-point function on lattices with octahedral or icosahedral symmetry do not break rotational symmetry (up to statistical errors), and therefore they provide a good indication for what the symmetry-breaking measurement should look like for an unbroken irreducible representation.

In Fig. 8 we show the rotational symmetry breaking measurement as a function of the lattice spacing using the basic discretization of the icosahedron. The measurement is clearly approaching a nonzero value in the continuum limit for $\ell\geq 2$, which indicates that this construction does not restore rotational symmetry.

Figure 8: On left rotational symmetry breaking of the spin-spin correlation function using the basic icosahedral discretization of $S^{2}$ up to a refinement of 96. On the right severe breaking of rotational symmetry for Tegmark’s octahedral based equal area pixelization of the celestial sphere [20].

Because the triangular faces of the basic discretization of the 2-sphere fail to satisfy the constraints of equal circumradius and equal perimeter, it is perhaps unsurprising that the higher $\ell$ coefficients of the spin-spin correlation function do not converge to zero in the continuum limit. One might have hoped that because the triangular faces in the basic discretization have locally uniform circumradius and perimeter (i.e. deviations in the circumradius and perimeter of neighboring triangles go to zero in the continuum limit), this might have been sufficient to satisfy the geometric constraints required by our derivation in Sec. 3.1. However, it is clear from our measurements here that variations in triangle geometry over long distances result in a lattice theory which fails to fully restore continuum symmetries. We therefore conclude that the critical couplings derived in Sec. 2.3 do not result in a lattice theory with a well-defined continuum limit if these geometrical constraints are only satisfied locally.

Figure 9: Rotational symmetry breaking of the spin-spin two-point function as a function of lattice spacing using the modified discretization up to a refinement of 64 for the sphere using an icosahedral (left) compared to octahedral (right) base lattice .

To remedy this issue, we repeat the Monte Carlo simulations of the previous section, but this time we use the modified lattice construction described in Appendix B.2, which explicitly minimizes non-uniformities in the circumradius and perimeter of triangular faces so that global variations in these quantities go to zero in the continuum limit. We again plot the symmetry breaking measurement $\delta C_{\ell}$ as a function of lattice spacing in Fig. 9. For both the octahedral and icosahedral discretizations, the rotational symmetry-breaking measurement goes to zero for all measured irreducible representation within statistical uncertainty. This result indicates that global uniformity in the triangle geometry is necessary to generate a sequence of lattices for which this model restores rotational symmetry in the continuum limit. It is especially promising that this procedure works even for the octahedral lattice, which requires much larger variations in triangle shape in order to fully tesselate the sphere, which can be seen explicitly in Fig. 10 comparing the octahedral and icosahedral examples.

Figure 10: Simplicial discretizations of the sphere after applying the optimization procedure described in this appendix. On the left is the icosahedral lattice with a refinement of L= 12, compared to on the right the octahedral lattice with a refinement of L =16.

For a quantitative comparison of the unmodified basic lattice and the modified lattice, in Fig. 11 we show the non-uniformity measure $E$ as a function of the lattice spacing as defined in Eq. (43) for both the octahedral and icosahedral lattices. We can clearly see that the non-uniformity in the unmodified lattice approaches a nonzero value in the continuum limit, whereas in the modified lattice the non-uniformity goes to zero roughly quadratically in the lattice spacing.

Figure 11: Non-uniformity measure for the octahedral and icosahedral discretizations of $S^{2}$ as a function of lattice spacing before and after being modified using the optimization procedure described in this appendix.

It is worth noting that although the triangle circumradius and perimeter are the quantities that appeared in our derivation of the Ising critical couplings, in general there are many other ways to define the measure of uniformity for a simplicial lattice. As an example, we tested our simulation of the Ising model on a discretized sphere using the optimization procedure described in [20], which adjusts the positions of the lattice sites so that the Voronoi areas of all sites become approximately equal. However, just as we saw when using the basic discretization, this construction fails to restore rotational symmetry in the continuum limit. The rotational symmetry breaking is especially strong using a octahedral base lattice, as shown earlier in Fig. 8. We also tested several other methods for adjusting the vertices with similar results [21, 22, 23, 24]. We therefore conclude that the definition of uniformity based on the circumradius and perimeter is indeed necessary to ensure a valid continuum limit for the critical Ising model.

Figure 12: Breaking of rotational symmetry in critical $\phi^{4}$ theory with local perturbative counter-terms using the basic icosahedral discretization of $S^{2}$.

It is worth noting that the restoration of rotational symmetries in a lattice model of a sphere is a challenging problem which has been encountered in previous works. In [25] attempts were made to simulate the 3d Ising CFT on the manifold $\mathbb{R}\times\mathbb{S}^{2}$ of radial quantization using an icosahedral discretization of the sphere and uniform couplings. Although this construction was shown to have a well-defined critical point, rotational symmetry was not restored in the continuum limit. In [4, 2], simulations of critical $\phi^{4}$ theory were performed using a generalization of the finite element method. Again, it was shown that a well-defined critical point exists after adding a perturbative mass counter-term to account for the non-uniform UV divergence inherent in the lattice. However, as shown in Fig. 12, higher precision measurements of the scalar 2-point function reveal a slight breaking of rotational symmetry in the continuum limit, very similar to the same results presented in Fig. 8 for the Ising model on the basic discretization of the sphere. It therefore seems likely that some additional geometric constraints may be necessary in order to restore rotational symmetry in these theories as well. It’s possible that the same constraints used here (uniform triangle circumradius and perimeter) may also work for other theories, but although we have begun to study this possibility it is unclear if this is the case.

3.4 Agreement with the Ising CFT

Now that we have confirmed that the spin-spin correlation function in our lattice theory becomes rotationally symmetric in the continuum limit, we would like to check that it also agrees with the exactly known analytical result for the 2d Ising CFT on a 2-sphere:

$$\langle s_{i}s_{j}\rangle\propto\dfrac{1}{(1-\hat{n}_{i}\cdot\hat{n}_{j})^{\Delta_{s}}}=\sum_{\ell}\dfrac{2\ell+1}{2}F_{\ell}P_{\ell}(\hat{n}_{i}\cdot\hat{n}_{j})$$ (45)

where $\Delta_{s}=1/8$ and we have expanded in a series of Legendre polynomials which have coefficients

$$\dfrac{F_{\ell}}{F_{0}}=\dfrac{\Gamma(\Delta_{s}+\ell)\Gamma(2-\Delta_{s})}{\Gamma(\Delta_{s})\Gamma(2-\Delta_{s}+\ell)}$$ (46)

which we can measure directly from lattice configurations generated by our Monte Carlo simulations. Eliminating $\Delta_{s}$ and solving for $\ell$ we find

$$\ell=\dfrac{(F_{0}-F_{1})(F_{\ell-1}+F_{\ell})}{(F_{0}+F_{1})(F_{\ell-1}-F_{\ell})}\;.$$ (47)

To check that the lattice simulation agrees with the analytic result, we first calculate a conformal symmetry breaking measurement

$$\delta\ell=1-\dfrac{(F_{0}-F_{1})(F_{\ell-1}+F_{\ell})}{\ell(F_{0}+F_{1})(F_{\ell-1}-F_{\ell})}$$ (48)

which should go to zero in the continuum limit. This quantity is plotted as a function of lattice spacing in Fig. 13, showing the expected behavior for both the octahedral and icosahedral lattice for all irrupts of O(3) up to $\ell=12$.

Figure 13: Conformal symmetry breaking measurement for the spin-spin two-point function as a function of lattice spacing using the modified discretization of the sphere comparing our icosahedral (left) lattice to the octahedral (right) lattice up to a refinement of $L=64$.

Finally, we measure the scaling exponent of the $s$ operator on the lattice via

$$\Delta_{s}=\dfrac{2F_{1}}{F_{1}+F_{0}}$$ (49)

which is plotted as a function of lattice spacing in Fig. 14. Extrapolating to the continuum limit by fitting to a quadratic polynomial in $a$, we obtain the results $\Delta_{s}=0.125048(44)$ with $\chi^{2}/\text{dof}=1.8$ for the octahedral lattice and $\Delta_{s}=0.124985(47)$ with $\chi^{2}/\text{dof}=1.8$ for the icosahedral lattice, both in excellent agreement with the exact value of $1/8$ and with a relative uncertainty of about $0.03\%$.

Figure 14: Continuum extrapolation of the scaling exponent $\Delta_{s}$ for the modified icosahedral (left) lattice compared to modified octahedral (right).

Our results indicate that by using the modified discretization of $S^{2}$, our lattice simulation accurately captures the properties of the lowest $\mathbb{Z}_{2}$-odd operator $s$. The 2d Ising CFT of course contains an infinite set of operators related by the Virasoro algebra. Measurement of additional operators using our lattice model is a subject for future study outside the present scope, however based on the high level of accuracy of our results and the delicate nature of a properly-tuned conformal field theory, we believe it is likely that our lattice formulation is sufficient to capture the properties of all of the operators in the 2d Ising CFT in the continuum limit, given sufficient computational resources.

Another approach taken recently is to use the so-called “fuzzy sphere” which projects the states of the theory in the basis of spherical harmonics instead of using a traditional spatial lattice [26, 27]. In this case, the lattice cutoff is replaced by a maximum spherical harmonic quantum number, which is a more manifest way to ensure that rotational symmetry is restored in the continuum limit. This method shows promising results for the lowest operators of the 3d Ising CFT, though it is unclear how difficult it would be to generalize to other theories.

4 Conclusion

We have demonstrated that there is a nearest neighbor Ising model properly implemented on triangulation of the sphere which we conjecture yields the exact $c=1/2$ minimal Ising CFT in the continuum limit. This is based on both theoretical support and Monte Carlo simulations. There are more stringent numerical tests and theoretical arguments being pursued to support this conjecture. We are increasing the statistical accuracy for the $\langle s_{x}s_{y}\rangle$ correlation function as well as computing the scalar 4 point function and the correlator for the energy momentum tensor. The energy momentum tensor plays a crucial theoretical role as the exact marginal operator conjugate to the metric tensor $g_{\mu\nu}$. Our working hypothesis is that smoothing the geometry is the first step to extract the spherical differential from a Regge piecewise flat manifold.

The benefit of the 2d Ising soluble theory is both its theoretical foundation in our affine analysis in flat space in Ref.[13] and comparison with the exact solution on $\mathbb{S}^{2}$. The 3 state Potts model and SUSY tricritical Ising model (TIM) provide interesting further examples to try. To understand this in a larger context beyond solvable lattice models, we are develo** similar analysis for $\lambda_{0}\phi^{4}$ theory where the map from the geometry of edge lengths $\ell_{ij}$ to coupling is no longer known analytically. We expect a smooth interpolation of the map from FEM map (2) at zero bare coupling $\lambda_{o}$ to the Ising map (3) at infinite coupling. But the general dependence on the bare coupling, $\lambda_{0}$, must be found numerically. This is a first step in develo** efficient numerical algorithms of the affine map for general renormalizable lattice field theories. To go to higher dimensions, we are starting with the 3d Ising theory in radial quantized form on $\mathbb{R}\times\mathbb{S}^{2}$ and on the Riemann sphere, $\mathbb{S}^{3}$, as simplicial refinement of the 600 cell.

In conclusion it is useful to summarize our intuitive picture for the central role of the affine parameter in constructing more general solutions to lattice field theories on smooth curved manifolds. In d-dimension a global affine transformation, $x^{\mu}=A_{\mu i}\xi^{i}+b^{\mu}$, extends the $d(d+1/2)/2$ parameters of the Euclidean Poincaré group by an additional $d(d+1)/2$ parameters. These are equivalent to a general metric tensor, $ds^{2}=g_{ij}d\xi^{i}d\xi^{j}=[A^{T}A]_{ij}d\xi^{i}d\xi^{j}$ on each tangent plane. The discrete simplicial geometry has been extensively studied [28] in the context of Regge Calculus on higher dimensional triangular lattices. Each piecewise flat simplex is an affine subspace with the parameters represented by $d(d+1)/2$ edge lengths of the d-simplex. The gauge potential (or Christoffel symbol) define the parallel transport between adjacent simplicies across the share boundary.

In the quantum field theory the affine transformation is generated by the $d(d+1)/2$ components of the conserved energy momentum operator. In a conformal theory, it is a marginal traceless operator. The trace providing the relevant operator, breaking conformal invariance. We believe the fundamental geometric problem is constructing the map to the tangent plane, re-introduce the differential metric in the continuum limit of the Regge manifold[17] that matches the coordinate induced by the energy momentum operator in the quantum field theory. While extensions to general lattice field theories are both non-trivial theoretically and algorithmically challenging, we are optimistic that systematic sequence of solutions for higher dimensional examples, including gauge and fermion fields, will systematically developed the requisite tools with the ultimate goal of high performance 4d gauge theories leveraging methods for lattice QCD at the Exascale.

Appendix A Simplicial Spherical Harmonics

On a hypercubic lattice the analysis of translational and rotational symmetries using discrete Fourier transform is obligatory. Here the analogous procedure on a triangulated sphere is the transform to discrete spherical harmonics. This requires a careful discussion of integration over a piecewise simplicial manifold.

For example on a 2d simplicial manifold for the sphere, the integral of a smooth function $\phi(x)$ with values $\phi_{i}$ on each vertex is defined by

$$\int d\Omega\phi(x)=\sum^{N}_{i=1}\sqrt{g_{i}}\phi_{i}+O(a^{2})\ ,$$ (50)

where the measure $\sqrt{g_{i}}$ is the scaled circumcenter Delaunay dual area: $\sum_{i}\sqrt{g_{i}}=4\pi$. For any reasonable refinement the error is $O(a^{2})$ with with $a\sim 1/\sqrt{N}$ given in Eq. (43). Indeed this is a natural generalization of the trapezoidal rule on a simplicial complex for $d>1$. Many higher order integration schemes are possible but are not needed for our current analysis.

We now apply this to spherical harmonic, $Y_{lm}(\hat{n}_{i})$, evaluated on each site on the sphere $\hat{n}_{i}=\vec{r}_{i}/R$. The latice addition theorem,

$$P_{\ell}(\hat{n}_{i}\cdot\hat{n}_{j})=\frac{4\pi}{2\ell+1}\sum^{m=\ell}_{m=-\ell}Y^{*}_{\ell m}(\hat{n}_{i})Y_{\ell,m}(\hat{n}_{j})\;,$$ (51)

is still exact and the integration rule gives orthogonality,

$$\sum_{i}\sqrt{g_{i}}Y^{*}_{\ell^{\prime}m^{\prime}}(\hat{n}_{i})Y_{\ell m}(\hat{n}_{i})=\delta_{\ell^{\prime},\ell}\delta_{m^{\prime},m}+O(a^{2})\;,$$ (52)

for smooth modes with $\ell\ll L_{max}=O(1/a)$. This is applied to scalar two point function,

$$\langle\widetilde{\phi}^{*}_{\ell^{\prime}m^{\prime}}\widetilde{\phi}_{\ell m}\rangle=\sum_{\hat{n}_{i}}\sum_{\hat{n}_{j}}\sqrt{g_{i}}Y^{*}_{\ell^{\prime}m^{\prime}}(\hat{n}_{i})\langle\phi(\hat{n}_{i})\phi(\hat{n}_{j})\rangle\sqrt{g_{j}}Y_{\ell m}(\hat{n}_{j})\;.$$ (53)

where $\widetilde{\phi}_{lm}=\sum^{N}_{i=1}\phi_{i}Y_{lm}(\hat{n}_{i})$. The integration procedure allows us to invert the expansion,

$$\langle\phi(\hat{n}_{j})\phi(\hat{n}_{i})\rangle=\sum_{\ell^{\prime},m^{\prime};\ell,m}C_{\ell^{\prime}m^{\prime};\ell,m}Y^{*}_{\ell^{\prime}m^{\prime}}(\hat{n}_{i})Y_{\ell m}(\hat{n}_{j})\;,$$ (54)

with expansion coefficient for low modes modes $\ell,\ell^{\prime}\ll L_{max}$ given by

$$C_{\ell^{\prime}m^{\prime};\ell,m}=\dfrac{1}{N^{2}}\sum_{ij}\sqrt{g_{i}g_{j}}Y^{*}_{\ell^{\prime}m^{\prime}}(\hat{n}_{i})Y_{\ell,m}(\hat{n}_{j})\langle\phi(\hat{n}_{i})\phi(\hat{n}_{j})\rangle+O(a^{2})\;.$$ (55)

Using the addition theorem, spherical symmetry implies

$$C_{\ell^{\prime}m^{\prime};\ell,m}=\delta_{\ell^{\prime},\ell}\delta_{m^{\prime},m}C_{\ell}$$ (56)

so that in the continuum $\langle\phi(\hat{n}_{j})\phi(\hat{n}_{i})\rangle$ is function the spherical geodesic through $\hat{n}_{i}\cdot\hat{n}_{j}=\cos(\theta_{ij})$. In Sec.3.2, substituting $\phi(\hat{n}_{i})\rightarrow s_{i}$, this procedure applies equally well to smooth quantum Ising correlators as one approaches a second order critical surface.

Appendix B Smoothing the Simplicial Manifold.

B.1 Vertex degrees of freedom

On our graph illustrated in Fig. 6 with refinement level $L$ has $F=20L^{2}$ faces, $E=30L^{2}$ edges and $N=2+10L^{2}$ vertices. Embedded on the sphere the 3d unit vectors represent $2N=4+20L^{2}$ dof. After removing the 3 rotations and fixing the scale by the sphere’s radius, the independent degrees of freedom are equal to the number of coupling constants $K_{ij}$ or the corresponding edge lengths $\ell_{ij}=|r_{i}-r_{j}|$.

In our implementation, we also seek to retain as much symmetry as possible, namely the discrete subgroup of $S(3)$ of the original platonic solid. This can be accomplished by a judicious choice of the degrees of freedom of the problem, which we will describe here for an icosahedral discretization. It is important to note that the simplicial graph remains fixed during this procedure. This method works equally well for discretizations with octahedral or tetrahedral symmetry, and can also be generalized for discretizations of $S^{3}$. It is also possible to generalize to non-spherical manifolds, though we have not yet had a reason to do so.

We first identify “orbits” which are sets of vertices which transform into one another under the icosahedral group transformations. We use the full icosahedral group which includes rotations and reflections (120 group elements), but this method will also work for the chiral icosahedral group which includes only rotations (60 group elements). The number of distinct vertices in an orbit (the orbit’s degeneracy) depends on whether its vertices lie on any symmetry axes of the icosahedron. For example, an orbit with vertices at the midpoints of the icosahedral edges has a degeneracy of 30 (one per icosahedral edge), while an orbit with vertices on one of the reflection axes of an icosahedral face has a degeneracy of 60 (three per icosahedral face). An orbit which does not lie on any symmetry axes has a degeneracy of 120 (six per icosahedral face).

We parameterize an orbit’s position by its barycentric coordinates within an icosahedral face. The barycentric coordinates describe a position on an icosahedral face, which is then projected onto the unit sphere. Using barycentric coordinates allows us to define the coordinates of all vertices in an orbit simultaneously on all 20 icosahedral faces. In addition, permuting the order of the 3 barycentric coordinate values generates all of the vertices within a single icosahedral face.

In order to preserve icosahedral symmetry, we require that orbits with vertices on a symmetry axis remains on that symmetry axis during the minimization procedure. Thus, though a point on a sphere has two degrees of freedom in general, this constraint reduces the number of degrees of freedom for some orbits (e.g. an orbit with vertices at the midpoints of the icosahedral edges has no degrees of freedom). The reduced set of orbit degrees of freedom, denoted $\xi_{i}$ where $i$ runs over all of the remaining degrees of freedom, can be freely adjusted without breaking icosahedral symmetry. Once we determine the values of $\xi_{i}$ which minimize $E$, we simply use the action of the icosahedral group elements to compute the vertex coordinates of all of the lattice sites in each orbit as described above.

This reduction by imposing symmetries reduces the dof continues to match the constraint of a single triangle term. However as described in the text we consider the over constrained system with both area (A) and perimeter (P) forms. The use of over constraint optimization appears to be important for these highly non-linear optimization appears to be important.

B.2 Iterative method for generating uniform spherical meshes

In Sec. 3.3, we show that rotational symmetry of the spin-spin correlation function was broken in the continuum limit for simulations of the critical Ising model on the basic discretization of a sphere. The failure to restore rotational symmetry in the continuum limit can be traced back to the fact that our derivation of the critical couplings on a simplicial lattice required all of the triangles to have equal circumradius and perimeter, which is not the case for the basic discretization of the sphere. In this appendix I will describe an iterative method for modifying the basic discretization of $S^{2}$ to make the triangles more uniform.

After identifying the reduced set of degrees of freedom $\xi_{i}$, we use Newton’s method to solve the non-linear system of equations

$$\dfrac{\partial E}{\partial\xi_{i}}=0\;.$$ (57)

We use the barycentric coordinates of the vertices from the basic discretization as the initial guess $\xi_{i}^{(0)}$, then the $(k+1)$-th iteration of Newton’s method [29] sets

$$\xi_{i}^{(k+1)}=\xi_{i}^{(k)}-\left[\dfrac{\partial^{2}E^{(k)}}{\partial\xi_{i}^{(k)}\partial\xi_{j}^{(k)}}+\mu^{(k)}\delta_{ij}\right]^{-1}\dfrac{\partial E^{(k)}}{\partial\xi_{j}^{(k)}}$$ (58)

where $E^{(k)}$ is the non-uniformity measure computed on the lattice with the vertex coordinates determined by the orbit coordinates $\xi_{i}^{(k)}$. We compute the partial derivatives via first-order finite differences with a step size of $10^{-5}$, which is efficient for double-precision floating point numbers. The parameter $\mu^{(k)}$ is a preconditioning factor and is chosen to ensure that $E^{(k)}$ is strictly decreasing for successive iterations. We continue the iterative process until the quantity $1-E^{(k+1)}/E^{(k)}$ becomes less than $10^{-10}$.