Introduction
An antiferromagnetic Ising triangle cannot make all three bonds antiparallel. In the zero-field triangular-lattice model with only a uniform nearest-neighbor coupling J 1 > 0 J_1 > 0 J 1 > 0 , every ground-state triangle has one parallel-spin bond. The many ways of satisfying this local rule give an extensive degeneracy and a nonzero residual entropy per spin [1 , 2 ] . Early studies of further-neighbor interactions identified ordered ground states [3 , 4 , 5 ] . A weak antiferromagnetic second-neighbor interaction J 2 > 0 J_2 > 0 J 2 > 0 favors stripe order [6 ] . Uniform stripes alternate between adjacent rows in at least one lattice direction. On an L × L L \times L L × L triangular-lattice torus, both primitive lattice directions have period L L L . If L L L is odd, that alternation cannot close. States that retain the local ground-state rule must instead contain interfaces between different stripe orientations.
Domain-wall descriptions give an energy cost 2 J 2 2J_2 2 J 2 per unit length and identify networks whose domain areas can change while their total wall length and energy remain fixed [6 ] . Later work linked stripe ordering and its thermal destruction to the topology of domain-wall configurations [7 ] . For the odd-L L L torus, minimizing the second-neighbor energy additionally requires a network that closes periodically while obeying the local triangle rule. The resulting questions are the least possible total wall length, the complete set of states attaining it, and the statistical weights of their orientation compositions.
Periodic boundaries also affect the unperturbed state count. Pairing the centers of adjacent triangles across their shared equal-spin bond gives a dimer representation, but a valid covering need not reconstruct periodic spins. The periodic-spin count provides the reference for quantifying the loss of degeneracy under the second-neighbor interaction. Its logarithm, divided by the number of spins, is the entropy per spin. Exact results already describe its approach to the bulk limit in several regimes. Spectral solutions determine the entropy of infinitely long systems of fixed width [8 , 9 , 10 ] . Dimer methods give finite-size corrections and asymptotic error control on tori whose two periods grow at a bounded ratio [11 ] . These results motivate a complementary question: how accurately can one bound the entropy of any given finite sample, including highly elongated ones? Such a guarantee would both control the comparison with the selected states and specify sample dimensions for a desired entropy accuracy.
In this work, we address the geometry and counting consequences of these periodic constraints. At each fixed odd L ≥ 5 L \ge5 L ≥ 5 , we prove a minimum total interface length of 4 L 4L 4 L within the J 2 = 0 J_2=0 J 2 = 0 ground-state set and classify all L 2 ( L 2 − 1 ) / 2 L^2(L^2-1)/2 L 2 ( L 2 − 1 ) /2 states attaining it. The directional composition is measured by the fractions of triangles with each equal-spin bond direction. The classification gives the same number of selected states in every allowed composition. Under the uniform measure on these states, these fractions therefore approach the uniform distribution over all nonnegative triples summing to one. Figure 1 shows two such states with different proportions and equal interface energies. The extensive entropy is removed, yet the whole-sample proportions do not concentrate at a single composition. We also give a sufficient finite-size coupling interval in which these constrained minimizers are the ground states of the full Hamiltonian.
Our second result concerns the unperturbed model on fully periodic M × N M \times N M × N triangular lattices. Here M M M and N N N are the numbers of lattice spacings along the two primitive periodic directions, so there are M N MN M N spins. The odd tori considered above have M = N = L M=N=L M = N = L . For evevery M , N ≥ 3 M,N \ge3 M , N ≥ 3 , the entropy per spin differs from the bulk residual entropy by at most an explicit constant times min ( M , N ) − 2 \min(M,N)^{-2} min ( M , N ) − 2 . The guarantee does not restrict the aspect ratio and gives sufficient sample dimensions for a prescribed accuracy. We also bound the entropy of all honeycomb dimer coverings, including those that cannot reconstruct periodic spins. Applying the spin bound to an odd L × L L \times L L × L torus gives an explicit finite-size correction to the logarithm of the selected fraction. Its leading exponential decay already follows from the polynomial selected count and the nonzero bulk residual entropy.
Section II defines the model, the two ground-state counts, and the periodic spin reconstruction used to evaluate the unperturbed count. Section III classifies the odd-torus minimizers and their orientation statistics. Section IV gives the finite-size entropy bounds, the strip results used to derive them, and their quantitative consequences. Section V discusses the physical interpretation and unresolved limits. Appendices A–C contain the counting identities, entropy-bound proofs, and geometric classification, respectively.
Model and periodic ground-state counts
Sites are ( x , y ) ∈ Z M × Z N (x,y) \in\mathbb{Z}_M \times\mathbb{Z}_N ( x , y ) ∈ Z M × Z N , with s x , y = ± 1 s_{x,y} = \pm1 s x , y = ± 1 and M , N ≥ 3 M,N \ge3 M , N ≥ 3 . The two primitive lattice directions have periods M M M and N N N . In coordinates with nearest-neighbor directions ( 1 , 0 ) (1,0) ( 1 , 0 ) , ( 0 , 1 ) (0,1) ( 0 , 1 ) , and ( 1 , 1 ) (1,1) ( 1 , 1 ) , the zero-field Hamiltonian is
H 1 , 2 ( s ) = J 1 A ( s ) + J 2 B ( s ) , \mathcal{H}_{1,2}(s) = J_1\mathcal{A}(s) + J_2\mathcal{B}(s), H 1 , 2 ( s ) = J 1 A ( s ) + J 2 B ( s ) ,
A ( s ) = ∑ x , y s x , y ( s x + 1 , y + s x , y + 1 + s x + 1 , y + 1 ) , \mathcal{A}(s) = \sum_{x,y} s_{x,y}(s_{x+1,y} + s_{x,y+1} + s_{x+1,y+1}), A ( s ) = x , y ∑ s x , y ( s x + 1 , y + s x , y + 1 + s x + 1 , y + 1 ) ,
B ( s ) = ∑ x , y s x , y ( s x + 1 , y − 1 + s x + 2 , y + 1 + s x + 1 , y + 2 ) . (1) \mathcal{B}(s) = \sum_{x,y} s_{x,y}(s_{x+1,y-1} + s_{x+2,y+1} + s_{x+1,y+2}). \tag*{(1)} B ( s ) = x , y ∑ s x , y ( s x + 1 , y − 1 + s x + 2 , y + 1 + s x + 1 , y + 2 ) . (1)
The dimensionless bond sums A ( s ) \mathcal{A}(s) A ( s ) and B ( s ) \mathcal{B}(s) B ( s ) collect the nearest- and second-neighbor spin products, respectively; their energy contributions are J 1 A J_1\mathcal{A} J 1 A and J 2 B J_2\mathcal{B} J 2 B . Here J 1 > 0 J_1 > 0 J 1 > 0 and J 2 ≥ 0 J_2 \ge0 J 2 ≥ 0 are the corresponding antiferromagnetic couplings. We set k B = 1 k_B = 1 k B = 1 and use natural logarithms throughout. Coordinates are reduced modulo the two periods. Each sum includes three forward bonds per site; if periodic identification makes distinct bonds share their endpoints, their multiplicities are retained. The second-neighbor selection studied below uses odd M = N = L ≥ 5 M = N = L \ge5 M = N = L ≥ 5 , for which these bonds are distinct.
We distinguish two counts. The number G 0 ( M , N ) G_0(M,N) G 0 ( M , N ) includes all periodic spin ground states at J 2 = 0 J_2 = 0 J 2 = 0 . On an odd L × L L \times L L × L torus, let C L = { s : A ( s ) = − L 2 } C_L = \{s : \mathcal{A}(s) = -L^2\} C L = { s : A ( s ) = − L 2 } be this unperturbed ground-state set. The number G s e l ( L ) G_{\mathrm{sel}}(L) G sel ( L ) counts the states that minimize B \mathcal{B} B within C L C_L C L . Both counts include distinct translations and both global spin signs. Section III determines G s e l G_{\mathrm{sel}} G sel ; Sec. IV bounds G 0 G_0 G 0 and compares the two. The constrained minimum is also the full-model ground-state minimum in a sufficient coupling interval established in Sec. III D.
From the local triangle rule to periodic spins
At J 2 = 0 J_2 = 0 J 2 = 0 , the energy is J 1 A J_1A J 1 A . Each triangle has either one or three aligned bonds. There are 2 M N 2MN 2 M N triangles, and each bond belongs to two of them, so J 1 A ≥ − J 1 M N J_1A \geq-J_1MN J 1 A ≥ − J 1 M N . Equality requires exactly one aligned bond per triangle. Place a dimer between the centers of the two triangles sharing each aligned bond. Each triangle center is then paired with exactly one neighbor. Denote the set of occupied dimer edges by M M M ; each dual vertex belongs to exactly one such edge. A matching compatible with periodic spins exists for every size in our range, as shown below.
To reconstruct spins from a matching, fix one spin and move along the triangular-lattice bonds. A bond crossed by a dimer preserves the spin; any other bond reverses it. The local triangle rule guarantees consistency around each triangle. Along the two independent periodic directions, let ν 1 , ν 2 \nu_1,\nu_2 ν 1 , ν 2 count the reversals on paths that wind once around the torus. An odd count would assign opposite spins to the same site after one circuit. Periodic reconstruction therefore requires both counts to be even. Their parities are the matching holonomies, and matchings with the same parity pair form a holonomy sector.
For the algebraic formulation, identify each dimer with the original bond it crosses. For an original bond e e e , 1 e ∈ M \mathbf{1}_{e\in M} 1 e ∈ M is one if its crossing dual edge carries a dimer and zero otherwise. The binary variable indicating a spin reversal is
c M ( e ) = 1 − 1 e ∈ M ( m o d 2 ) . (2) c_M(e) = 1 - \mathbf{1}_{e\in M} \pmod2. \tag*{(2)} c M ( e ) = 1 − 1 e ∈ M ( mod 2 ) . (2)
This edge variable is a closed Z 2 \mathbb{Z}_2 Z 2 cochain: its sum around each triangle vanishes modulo two, because each triangle has two unmatched edges. Let L 1 L_1 L 1 and L 2 L_2 L 2 be the horizontal cycle at y = 0 y = 0 y = 0 and the vertical cycle at x = 0 x = 0 x = 0 , containing M M M and N N N bonds, respectively, and set
h r ( M ) = ∑ e ∈ L r c M ( e ) ( m o d 2 ) , h ( M ) = ( h 1 , h 2 ) . (3) h_r(M) = \sum_{e\in L_r} c_M(e) \pmod2,\qquad h(M) = (h_1,h_2). \tag*{(3)} h r ( M ) = e ∈ L r ∑ c M ( e ) ( mod 2 ) , h ( M ) = ( h 1 , h 2 ) . (3)
Thus h 1 , h 2 ∈ { 0 , 1 } h_1,h_2 \in\{0,1\} h 1 , h 2 ∈ { 0 , 1 } record the two winding parities. Write D h D_h D h for the matching count with holonomy h h h , D t o t = ∑ h D h D_{\mathrm{tot}} = \sum_h D_h D tot = ∑ h D h , with the size arguments suppressed when unambiguous. Only one of these four matching counts contributes to G 0 G_0 G 0 .
Proposition 1 (Spin reconstruction). A matching lifts to a periodic spin ground state if and only if h = ( 0 , 0 ) h = (0,0) h = ( 0 , 0 ) . There are exactly two lifts, related by a global spin flip. Consequently,
G 0 ( M , N ) = 2 D 00 ( M , N ) . (4) G_0(M,N) = 2D_{00}(M,N). \tag*{(4)} G 0 ( M , N ) = 2 D 00 ( M , N ) . (4)
Proof. For adjacent sites u , v u,v u , v , write t u = ( 1 − s u ) / 2 t_u = (1-s_u)/2 t u = ( 1 − s u ) /2 . Reconstruction is the equation t u + t v = c M ( u v ) t_u+t_v=c_M(uv) t u + t v = c M ( uv ) over Z 2 \mathbb{Z}_2 Z 2 . A closed cochain on the torus is exact precisely when both periods vanish. On this connected graph any two solutions differ by a constant, giving the two spin flips.
Table 1. Exact counts in the conventions of Sec. II. The sectors are labeled by spin holonomy, not by the unshifted slope parity. Direct spin enumeration verifies G 0 G_0 G 0 for all listed sizes.
For example, only 21 of the 417 matchings on the 4 × 4 4\times4 4 × 4 torus have h = ( 0 , 0 ) h=(0,0) h = ( 0 , 0 ) . They give 42 periodic spin ground states, since each matching allows two overall spin signs. Table I gives further small-system counts. Loop descriptions likewise distinguish configurations by their winding around the periodic directions [12 ] .
To calculate these counts, we describe a matching as a pairing of up and down triangles. Up and down triangles have vertices { ( i , j ) , ( i + 1 , j ) , ( i + 1 , j + 1 ) } \{(i,j),(i+1,j),(i+1,j+1)\} {( i , j ) , ( i + 1 , j ) , ( i + 1 , j + 1 )} and { ( i , j ) , ( i , j + 1 ) , ( i + 1 , j + 1 ) } \{(i,j),(i,j+1),(i+1,j+1)\} {( i , j ) , ( i , j + 1 ) , ( i + 1 , j + 1 )} , respectively. Each up triangle is paired across one of its three bonds; label that choice t ( i , j ) ∈ { d , h , v } t(i,j)\in\{d,h,v\} t ( i , j ) ∈ { d , h , v } by the direction of the crossed bond. A down triangle is labeled ( i , j ) (i,j) ( i , j ) by the first vertex in its listed vertex set. Across the d , h , v d,h,v d , h , v bonds of an up triangle ( i , j ) (i,j) ( i , j ) , its neighbors are therefore down triangles ( i , j ) (i,j) ( i , j ) , ( i , j − 1 ) (i,j-1) ( i , j − 1 ) , and ( i + 1 , j ) (i+1,j) ( i + 1 , j ) , respectively. The pairing gives the permutation
π ( i , j ) = ( i + 1 t ( i , j ) = v , j − 1 t ( i , j ) = h ) . (5) \pi(i,j) = (i+1_{t(i,j)=v},j-1_{t(i,j)=h}). \tag*{(5)} π ( i , j ) = ( i + 1 t ( i , j ) = v , j − 1 t ( i , j ) = h ) . (5)
Every down triangle must be paired exactly once. Thus the permutation must be bijective, which is equivalent to
1 t ( k , l ) = d + 1 t ( k − 1 , l ) = v + 1 t ( k , l + 1 ) = h = 1. (6) 1_{t(k,l)=d}+1_{t(k-1,l)=v}+1_{t(k,l+1)=h}=1. \tag*{(6)} 1 t ( k , l ) = d + 1 t ( k − 1 , l ) = v + 1 t ( k , l + 1 ) = h = 1. (6)
Summing this identity over k k k at fixed l l l shows that every fixed-j j j line has the same number H H H of h h h cells. Summing over l l l at fixed k k k similarly gives the same number V V V of v v v cells on every fixed-i i i line. These are also the matched-bond counts on L 1 , L 2 L_1,L_2 L 1 , L 2 . The pair ( H , V ) (H,V) ( H , V ) specifies directional occupations; we refer to it as the matching slope. Thus
h = ( M + H , N + V ) ( m o d 2 ) , N H + M V ≤ M N . (7) h=(M+H,N+V)\pmod2,\qquad NH+MV\leq MN. \tag*{(7)} h = ( M + H , N + V ) ( mod 2 ) , N H + M V ≤ M N . (7)
The inequality counts the h h h and v v v cells, whose totals are N H NH N H and M V MV M V .
A compatible matching always exists. If M M M or N N N is even, the corresponding alternating stripe spin configuration attains − J 1 M N -J_1MN − J 1 M N . If both are odd, choose one permutation cycle with successive steps v M − 1 h N v^{M-1}h^N v M − 1 h N and set every remaining cell to d d d . The cycle visits M + N M+N M + N distinct cells and closes. It has H = V = 1 H=V=1 H = V = 1 , so its holonomy vanishes by (54) ; Proposition 1 supplies a ground state.
Four-sector counts
For the bipartite honeycomb graph, we use the determinant formulation of Kasteleyn’s method and its extensions to graphs on the torus [11 , 13 , 14 , 15 , 16 ] . To extract the periodically reconstructible matchings from the total, we compute four determinants with different boundary signs. Their linear combinations separate the four possible parity pairs h h h . Let B ( σ , τ ) B(\sigma,\tau) B ( σ , τ ) be the M N × M N MN \times MN M N × M N biadjacency matrix corresponding to Eq. (5), with unit d d d edges. A v v v edge crossing i = M − 1 i=M-1 i = M − 1 has weight σ \sigma σ , and an h h h edge crossing j = 0 j=0 j = 0 has weight τ \tau τ . The signs σ , τ = ± 1 \sigma,\tau=\pm1 σ , τ = ± 1 keep track of crossings of these two seams. Define Δ ( σ , τ ) = det B ( σ , τ ) \Delta(\sigma,\tau)=\det B(\sigma,\tau) Δ ( σ , τ ) = det B ( σ , τ ) . Fourier diagonalization gives
Δ ( σ , τ ) = ∏ w M = σ ∏ z N = τ ( 1 + w + z ) = ∏ z N = τ [ ( 1 + z ) M − ( − 1 ) M σ ] . (8) \begin{aligned}
\Delta(\sigma,\tau) &= \prod_{w^M=\sigma}\prod_{z^N=\tau}(1+w+z) \\
&= \prod_{z^N=\tau}\left[(1+z)^M-(-1)^M\sigma\right]. \tag*{(8)}
\end{aligned} Δ ( σ , τ ) = w M = σ ∏ z N = τ ∏ ( 1 + w + z ) = z N = τ ∏ [ ( 1 + z ) M − ( − 1 ) M σ ] . (8)
The product is a polynomial in the twists with integer coefficients. At σ , τ = ± 1 \sigma,\tau=\pm1 σ , τ = ± 1 it is an integer.
For the two counts needed here, the resulting identities are as follows; Appendix A 1 supplies the inversion and the determinant-sign proof.
Proposition 2 (Finite-torus identities). For every M , N ≥ 3 M,N\ge3 M , N ≥ 3 ,
D t o t = 1 2 ∑ σ , τ = ± 1 ∣ Δ ( σ , τ ) ∣ , (9) D_{\mathrm{tot}}=\frac{1}{2}\sum_{\sigma,\tau=\pm1}\left|\Delta(\sigma,\tau)\right|, \tag*{(9)} D tot = 2 1 σ , τ = ± 1 ∑ ∣ Δ ( σ , τ ) ∣ , (9) G 0 = D t o t − ∣ Δ ( 1 , 1 ) ∣ , (10) G_0=D_{\mathrm{tot}}-\left|\Delta(1,1)\right|, \tag*{(10)} G 0 = D tot − ∣ Δ ( 1 , 1 ) ∣ , (10) G 0 = D t o t ⟺ 3 ∣ M and 3 ∣ N . (11) G_0=D_{\mathrm{tot}}\Longleftrightarrow3\mid M\ \text{and}\ 3\mid N. \tag*{(11)} G 0 = D tot ⟺ 3 ∣ M and 3 ∣ N . (11)
When both periods are divisible by three, Eq. (11) gives G 0 = D t o t G_0=D_{\mathrm{tot}} G 0 = D tot . This equality includes the two spin signs per compatible matching: D 00 = D t o t / 2 D_{00}=D_{\mathrm{tot}}/2 D 00 = D tot /2 , while the other three parity groups together contain the remaining half. Thus equal total counts do not mean that every matching reconstructs periodic spins. The spectral zeros enter the counting identity through one determinant rather than through a change in the local ground-state rule.
Slope-resolved counts can also be evaluated using integer polynomial arithmetic. Appendix A 3 gives that formula and the counts at the boundary of the allowed slope range as well as in its interior.
Odd-torus weak-coupling selection
We now ask which periodic nearest-neighbor ground states minimize an antiferromagnetic second-neighbor perturbation. On an odd L × L L\times L L × L torus, a uniform stripe pattern cannot close periodically: the spins reverse after one circuit in an alternating direction. We find the minimum interface length between local stripe orientations and classify every state that attains it.
Set M = N = L M=N=L M = N = L with odd L ≥ 5 L\ge5 L ≥ 5 . We minimize the second-neighbor bond sum B B B from Eq. (1) over C L C_L C L . All states in C L C_L C L have the same nearest-neighbor energy, so the perturbation ranks them by B B B alone. At fixed L L L , the equal-weight measure on these selected states is obtained by taking J 1 / T → ∞ J_1/T\to\infty J 1 / T → ∞ first and then J 2 / T → ∞ J_2/T\to\infty J 2 / T → ∞ . Section III D gives a sufficient coupling interval in which this constrained minimization also gives the ground states of the full Hamiltonian.
Let h ( x , y ) h(x,y) h ( x , y ) , v ( x , y ) v(x,y) v ( x , y ) , and d ( x , y ) d(x,y) d ( x , y ) indicate equal spins on the bonds from ( x , y ) (x,y) ( x , y ) to ( x + 1 , y ) (x+1,y) ( x + 1 , y ) , ( x , y + 1 ) (x,y+1) ( x , y + 1 ) , and ( x + 1 , y + 1 ) (x+1,y+1) ( x + 1 , y + 1 ) , respectively. Each is one for equal endpoint spins and zero otherwise. The triangle constraints are
h ( x , y ) + v ( x + 1 , y ) + d ( x , y ) = 1 , h(x,y)+v(x+1,y)+d(x,y)=1, h ( x , y ) + v ( x + 1 , y ) + d ( x , y ) = 1 ,
h ( x , y + 1 ) + v ( x , y ) + d ( x , y ) = 1. (12) h(x,y+1)+v(x,y)+d(x,y)=1. \tag*{(12)} h ( x , y + 1 ) + v ( x , y ) + d ( x , y ) = 1. (12)
The local constraints conserve the number of equal-spin bonds along each periodic line. Denote the counts on a horizontal, vertical, and diagonal line by H , V , D H,V,D H , V , D , respectively. The first two are those in Eq. (7); the same local argument gives the diagonal count. The total numbers of equal-spin bonds in the three directions are L H , L V , L D LH,LV,LD L H , L V , L D . Thus ( H , V , D ) / L (H,V,D)/L ( H , V , D ) / L gives their fractions and, equivalently, the fractions of triangles with each local stripe orientation. We call a fixed triple a slope sector. It records the directional composition of a state. The holonomy sector of Sec. II instead records the two parities that determine whether a matching reconstructs periodic spins; every state considered here already satisfies that compatibility condition. Periodic spins on an odd line require an odd number of equal bonds, so the possible slope sectors obey
H , V , D ∈ { 1 , 3 , 5 , … } , H + V + D = L . (13) H,V,D\in\{1,3,5,\ldots\},\qquad H+V+D=L. \tag*{(13)} H , V , D ∈ { 1 , 3 , 5 , … } , H + V + D = L . (13)
Figure 1(a),(b) shows two L = 15 L=15 L = 15 minimizers with different directional compositions but the same energy. Panel (c) places their compositions among all the allowed triples. The result below explains both the equal energies and the equal weights of these triples.
An exact minimum interface cost
To express the second-neighbor energy as an interface cost, label each elementary triangle by the direction of its unique equal bond. Adjacent triangles with different labels meet at an interface, as shown in the enlarged portion of Figure 1 . Let W H , W V , W D W_H,W_V,W_D W H , W V , W D count the shared edges of each direction along these interfaces. The opposite corners of two triangles sharing an edge are second neighbors. Their spin product is − 1 + 2 f e + 2 w e -1+2f_e+2w_e − 1 + 2 f e + 2 w e , where f e f_e f e indicates an equal shared bond and w e w_e w e a label change. Since ∑ e f e = L 2 \sum_e f_e=L^2 ∑ e f e = L 2 ,
Figure 1. Different stripe proportions at the same minimum energy. (a),(b) Exact L = 15 L = 15 L = 15 periodic spin states. Filled/open sites show the spins; triangle colors indicate the direction of the unique equal-spin bond. The per-line equal-bond counts are ( H , V , D ) (H,V,D) ( H , V , D ) ; their ratios ( H , V , D ) / L (H,V,D)/L ( H , V , D ) / L are the fractions of triangles with horizontal, vertical, and diagonal equal bonds. The enlarged box is an excerpt of (a), with equal bonds emphasized. Matching boundary letters denote identified edges; pale strips show periodic continuation. Both configurations are ground states of the nearest-neighbor model (J 2 = 0 J_2 = 0 J 2 = 0 ). Among these ground states, they minimize the second-neighbor energy, with total interface length W = 4 L = 60 W = 4L = 60 W = 4 L = 60 . Each unit of interface costs 2 J 2 2J_2 2 J 2 , giving an energy 8 J 2 L 8J_2L 8 J 2 L above the uniform-stripe bulk reference. A uniform stripe cannot itself satisfy these odd periodic boundaries. (c) All positive odd triples H + V + D = L H + V + D = L H + V + D = L , plotted as orientation fractions; rings mark (a),(b). Each point represents 4 L 2 4L^2 4 L 2 selected states and hence has equal weight in the uniform ensemble of all such states. As odd L → ∞ L \to\infty L → ∞ , the fraction distribution approaches the uniform distribution on the simplex.
B = − L 2 + 2 W , W = W H + W V + W D . (14) \mathcal{B}=-L^2+2W,\qquad W=W_H+W_V+W_D. \tag*{(14)} B = − L 2 + 2 W , W = W H + W V + W D . (14)
Each interface edge therefore costs 2 J 2 2J_2 2 J 2 , recovering the wall tension used in Ref. [6 ] . The remaining question is how much interface the odd periodic boundary forces into a state.
Theorem 1 (Odd-torus selection). In every sector in Eq. (13) ,
W H ≥ 2 ( V + D ) , W V ≥ 2 ( H + D ) , W D ≥ 2 ( H + V ) , min W = 4 L , min B = − L 2 + 8 L . (15) \begin{aligned}
W_H &\ge2(V+D), & W_V &\ge2(H+D),\\
W_D &\ge2(H+V), & \min W &= 4L,\\
\min\mathcal{B} &= -L^2+8L. \tag*{(15)}
\end{aligned} W H W D min B ≥ 2 ( V + D ) , ≥ 2 ( H + V ) , = − L 2 + 8 L . W V min W ≥ 2 ( H + D ) , = 4 L , (15) At fixed ( H , V , D ) (H,V,D) ( H , V , D ) , every minimizer is obtained from one of the two configurations in Appendix C 2 a by a translation, possibly followed by a global spin reversal. There are exactly 4 L 2 4L^2 4 L 2 distinct periodic spin minimizers in each sector. Their total number is
G s e l ( L ) = L 2 ( L 2 − 1 ) 2 . (16) G_{\mathrm{sel}}(L)=\frac{L^2(L^2-1)}{2}. \tag*{(16)} G sel ( L ) = 2 L 2 ( L 2 − 1 ) . (16) In particular, L − 2 ln G s e l ( L ) → 0 L^{-2}\ln G_{\mathrm{sel}}(L)\to0 L − 2 ln G sel ( L ) → 0 .
The theorem gives the same minimum second-neighbor energy in every slope sector. Summing its three wall bounds gives W ≥ 4 L W\ge4L W ≥ 4 L , and the classified states attain all three bounds at once. Their degeneracy grows only polynomially, although the unperturbed manifold contains exponentially many states.
Each orientation occupies one region that, in lattice coordinates adapted to that orientation, is a rectangle with one corner rectangle removed. The two families differ in which of two opposite corners is removed from the diagonal-orientation region. At fixed composition they form two distinct orbits under translations and global spin reversal, and spatial inversion exchanges them up to these operations. Figure 2 compares the two families at the same composition. The insets isolate the opposite corner cuts in the green D D D regions. Each family contributes L 2 L^2 L 2 distinct translations and two overall spin signs, giving 4 L 2 4L^2 4 L 2 states per composition. The orbit count is established in Appendix C 2 c; completeness is proved in Appendix C 3.
Figure 2. The two minimizer families on an L = 15 L = 15 L = 15 torus, both with orientation fractions ( H , V , D ) / L = ( 3 , 5 , 7 ) / 15 (H,V,D)/L = (3,5,7)/15 ( H , V , D ) / L = ( 3 , 5 , 7 ) /15 . Triangle colors indicate whether the unique equal-spin bond is horizontal, vertical, or diagonal, as in Figure 1 . Black lines are walls of total length 4 L = 60 4L = 60 4 L = 60 in both panels; circles mark the six three-wall junctions, numbered separately within each panel. Matching p p p and q q q edges are periodically identified. Insets use the ( x , y ) (x,y) ( x , y ) coordinate grid to highlight the opposite missing corners of the green D D D regions. The two families lie in distinct translation/global-spin-reversal orbits and are related by spatial inversion.
The minimum energy exceeds the ideal uniform-stripe value − ( J 1 + J 2 ) L 2 -(J_1+J_2)L^2 − ( J 1 + J 2 ) L 2 by 8 J 2 L 8J_2L 8 J 2 L . A uniform stripe is not an admissible periodic state on this odd torus. This comparison measures the excess imposed by the boundary condition, proportional to the linear size, rather than an excitation gap above an allowed stripe ground state.
Below we outline the lower bound, its saturation, and the classification of the equality case; Appendix C gives the full proof.
For the lower bound, read a fixed spin configuration one diagonal at a time, using t = x − y t=x-y t = x − y and ring coordinate u = x u=x u = x . The following particle motion describes changes between these spatial slices. The local labels h h h , v v v , and d d d represent, respectively, a particle that stays, a particle that moves one site to the right, and a hole; their precise index shifts are given in Appendix C 1. A consecutive run of moving particles displaces one hole to the left by less than L L L . Over a full period the total displacement is L V LV L V . A cyclic shift by k k k hole labels carries total displacement k L kL k L . Periodic closure restores the set of D D D hole positions, but a tracked hole can end at another hole’s initial position. The hole labels therefore shift by V V V . Because the holes cannot pass one another, at least D + V D+V D + V runs are needed for this return (Lemma 1), each contributing two horizontal interfaces. Hence W H ≥ 2 ( D + V ) W_H\ge2(D+V) W H ≥ 2 ( D + V ) ; cyclic permutation gives the other two bounds and W ≥ 4 L W\ge4L W ≥ 4 L . Appendix C 1 proves the lemma and applies it to the spin model.
For saturation, Appendix C 2 a gives explicit supports for both families at every positive odd triple. There are ( L 2 − 1 ) / 8 (L^2-1)/8 ( L 2 − 1 ) /8 such triples; the 4 L 2 4L^2 4 L 2 states per triple give the total count in Eq. (16) .
The completeness step rules out other interface geometries at the same energy. Treat each region of one local stripe orientation as a set of paired triangles. Equality fixes its two directional run counts. A periodic-profile inequality then forces each orientation to occupy one connected component; the three resulting disks have six wall junctions. The conserved diagonal counts then fix the corner-cut sizes, giving precisely the two families described above. Appendix C 3 gives the full argument.
The zero entropy density also follows from the minimum wall count W = 4 L W=4L W = 4 L alone, by the direct counting bound in Appendix C; it does not require the completeness argument.
Related ground-state ordering was studied in extended-range lattice-gas models [5 ] . Finite-cluster enumerations have resolved local motifs and spin-flip connectivity [17 , 18 ] . Here independent integer-transfer traces reproduce the finite counts in Table II. At L = 5 L=5 L = 5 , comparison with the motif census of Ref. [18] gives the same minimum; the traces and motif conversion are included in the verification archive [19 ] .
Equal composition weights and orientation fluctuations
The classification determines both the total degeneracy and the statistical weights of the directional compositions. Each dot in Figure 1 (c) represents 4 L 2 4L^2 4 L 2 selected spin states. Under the equal-weight measure on all constrained minimizers, every dot thus has probability 8 / ( L 2 − 1 ) 8/(L^2-1) 8/ ( L 2 − 1 ) . The normalized slopes H / L H/L H / L , V / L V/L V / L , and D / L D/L D / L are the area fractions of the three local stripe orientations. Writing ( H 1 , H 2 , H 3 ) = ( H , V , D ) (H_1,H_2,H_3)=(H,V,D) ( H 1 , H 2 , H 3 ) = ( H , V , D ) and H i = 1 + 2 n i H_i=1+2n_i H i = 1 + 2 n i
turns the sector measure into the uniform distribution on nonnegative integer triples n 1 + n 2 + n 3 = ( L − 3 ) / 2 n_1+n_2+n_3=(L-3)/2 n 1 + n 2 + n 3 = ( L − 3 ) /2 . In particular,
E H i L = 1 3 , Var H i L = 1 − 9 / L 2 18 . (17) \mathbb{E}\frac{H_i}{L}=\frac{1}{3},\qquad\operatorname{Var}\frac{H_i}{L}=\frac{1-9/L^2}{18}. \tag*{(17)} E L H i = 3 1 , Var L H i = 18 1 − 9/ L 2 . (17)
As L → ∞ L\to\infty L → ∞ through odd sizes, ( H , V , D ) / L (H,V,D)/L ( H , V , D ) / L converges in distribution to the uniform measure on the triangle of nonnegative fractions summing to one (the unit simplex), equivalently Dirichlet( 1 , 1 , 1 ) (1,1,1) ( 1 , 1 , 1 ) . This follows by Riemann sums over the composition lattice. Thus eliminating the extensive entropy does not fix the global stripe proportions. Their nonzero limiting variance quantifies the absence of concentration in this uniform selected-state ensemble.
The same composition fluctuations can be expressed through a height field, an auxiliary scalar whose increments are determined by the matching. The increments sum to zero around a small closed path, but the height can change after going once around a periodic direction. These net changes are called height periods.
Orient dual edges from the up-triangle to the down-triangle sublattice. For the matching M M M associated with the spin state, define the auxiliary integer current
η M ( e ) = 31 e ∈ M − 1 (18) \eta_M(e)=3\mathbf{1}_{e\in M}-1 \tag*{(18)} η M ( e ) = 3 1 e ∈ M − 1 (18)
This current takes the values 2 2 2 on occupied edges and − 1 -1 − 1 on unoccupied edges. Its divergence vanishes because each dual vertex has one occupied and two unoccupied edges, whose signed contributions sum to zero. Rotating this current assigns height increments to the original lattice bonds. Their sum around each elementary triangle is zero. Choose orientations so that its periods around the two length-L L L directions are
Φ 1 = 3 H − L , Φ 2 = 3 V − L . (19) \Phi_1=3H-L,\qquad\Phi_2=3V-L. \tag*{(19)} Φ 1 = 3 H − L , Φ 2 = 3 V − L . (19)
Then ( Φ 1 , Φ 2 ) m o d 2 = h ( M ) (\Phi_1,\Phi_2)\bmod2=h(M) ( Φ 1 , Φ 2 ) mod 2 = h ( M ) . Spin compatibility selects even height periods. Thus the periods record the global winding of a height field whose increments are locally consistent. The height periods express the same fluctuation: E Φ 1 = 0 \mathbb{E}\Phi_1=0 E Φ 1 = 0 and Var ( Φ 1 / L ) = ( 1 − 9 / L 2 ) / 2 \operatorname{Var}(\Phi_1/L)=(1-9/L^2)/2 Var ( Φ 1 / L ) = ( 1 − 9/ L 2 ) /2 . The net height change around a periodic direction therefore fluctuates on the scale of L L L .
Directional couplings select orientations
Directional couplings lift the composition degeneracy left by the isotropic interaction. The same wall bounds determine the selected slopes. Write H H = H H_H=H H H = H , H V = V H_V=V H V = V , H D = D H_D=D H D = D , and let B i B_i B i be the opposite-corner sum for shared nearest-neighbor edges of direction i i i . Thus B H B_H B H uses displacement ( 1 , 2 ) (1,2) ( 1 , 2 ) , B V B_V B V uses ( 2 , 1 ) (2,1) ( 2 , 1 ) , and B D B_D B D uses ( 1 , − 1 ) (1,-1) ( 1 , − 1 ) . Here i ∈ { H , V , D } i\in\{H,V,D\} i ∈ { H , V , D } indexes the three directions, and K i ≥ 0 K_i\geq0 K i ≥ 0 is the second-neighbor coupling multiplying B i B_i B i ; the isotropic model has K H = K V = K D = J 2 K_H=K_V=K_D=J_2 K H = K V = K D = J 2 . The directional identity and lower bound give
B i = − L 2 + 2 L H i + 2 W i ≥ − L 2 + 4 L + 2 ( L − 2 ) H i . (20) B_i=-L^2+2LH_i+2W_i\geq-L^2+4L+2(L-2)H_i. \tag*{(20)} B i = − L 2 + 2 L H i + 2 W i ≥ − L 2 + 4 L + 2 ( L − 2 ) H i . (20)
Since the constructed states saturate all three inequalities, the exact minimum at fixed slope is
min i ∑ K i B i = ( − L 2 + 4 L ) ∑ K i + 2 ( L − 2 ) ∑ K i H i . (21) \min_i\sum K_iB_i=(-L^2+4L)\sum K_i+2(L-2)\sum K_iH_i. \tag*{(21)} i min ∑ K i B i = ( − L 2 + 4 L ) ∑ K i + 2 ( L − 2 ) ∑ K i H i . (21)
It remains to minimize over the allowed slopes. Putting H i = 1 + 2 n i H_i=1+2n_i H i = 1 + 2 n i with ∑ i n i = ( L − 3 ) / 2 \sum_i n_i=(L-3)/2 ∑ i n i = ( L − 3 ) /2 shows that a minimizing sector allocates this total among directions with the smallest K i K_i K i . If there is a unique smallest K i K_i K i , the only minimizing sector has H i = L − 2 H_i=L-2 H i = L − 2 and the other counts equal to one. If exactly two couplings share the minimum, the count in the third direction is one. Equal couplings leave all sectors degenerate. These statements determine the selected sectors for nonnegative couplings; when a coupling vanishes they do not require saturation of its unweighted wall bound. For strictly positive couplings every weighted minimizer saturates all three bounds. The classification then gives degeneracy 4 L 2 4L^2 4 L 2 for a unique least coupling, 2 L 2 ( L − 1 ) 2L^2(L-1) 2 L 2 ( L − 1 ) for exactly two tied least couplings, and L 2 ( L 2 − 1 ) / 2 L^2(L^2-1)/2 L 2 ( L 2 − 1 ) /2 for three equal couplings.
A sufficient window in the full model
So far, minimization has been restricted to C L \mathcal{C}_L C L . To relate it to the full model, compare the selected states with configurations that pay extra nearest-neighbor energy but gain second-neighbor energy. For any spin configuration, the second-neighbor triangle inequality gives B ≥ − L 2 \mathcal{B} \ge-L^2 B ≥ − L 2 . The two second-neighbor triangle types, with vectors ( 1 , − 1 ) (1,-1) ( 1 , − 1 ) , ( 1 , 2 ) (1,2) ( 1 , 2 ) , and their sum ( 2 , 1 ) (2,1) ( 2 , 1 ) , count each such bond twice. Let S S S be the set of down-spin sites and e ( S ) e(S) e ( S ) the number of nearest-neighbor edges with both endpoints in S S S . The nearest-neighbor graph is six-regular. The number c c c of bonds joining S S S to its complement is therefore c = 6 ∣ S ∣ − 2 e ( S ) c=6|S|-2e(S) c = 6∣ S ∣ − 2 e ( S ) , which is even, and A \mathcal{A} A changes in multiples of four. A state outside C L \mathcal{C}_L C L therefore has A ≥ − L 2 + 4 \mathcal{A}\ge-L^2+4 A ≥ − L 2 + 4 . Its energy exceeds the constrained minimum
E ∗ = − ( J 1 + J 2 ) L 2 + 8 J 2 L (22) E_* = -(J_1+J_2)L^2+8J_2L \tag*{(22)} E ∗ = − ( J 1 + J 2 ) L 2 + 8 J 2 L (22)
by at least 4 J 1 − 8 L J 2 4J_1-8LJ_2 4 J 1 − 8 L J 2 . Consequently
0 < J 2 J 1 < 1 2 L (23) 0<\frac{J_2}{J_1}<\frac{1}{2L} \tag*{(23)} 0 < J 1 J 2 < 2 L 1 (23)
is sufficient for the full-model ground states to coincide with the minimizers in Theorem 1. The second-neighbor triangle argument uses bond incidences and remains valid when 3 ∣ L 3\mid L 3 ∣ L , where that graph has disconnected components. Equation (23) is a sufficient finite-size interval. It compares energies at fixed L L L and does not locate a thermodynamic phase boundary. In particular, its decrease with L L L does not determine the physical stability range at fixed coupling or justify exchanging the constrained and thermodynamic limits.
An explicit competitor also provides an upper bound on the first crossing out of C L \mathcal{C}_L C L . Start from the construction in sector ( 1 , 1 , L − 2 ) (1,1,L-2) ( 1 , 1 , L − 2 ) and flip the segment { ( x , 1 ) : 2 ≤ x ≤ L − 1 } \{(x,1):2\le x\le L-1\} {( x , 1 ) : 2 ≤ x ≤ L − 1 } . This raises A \mathcal{A} A by four and lowers B \mathcal{B} B by sixteen, independently of L L L ; Appendix C4 gives a direct bond calculation. If r c ( L ) r_c(L) r c ( L ) is the first ratio J 2 / J 1 J_2/J_1 J 2 / J 1 at which an outside state reaches E ∗ E_* E ∗ , then
r c ( L ) = min s ∉ C L B ( s ) < − L 2 + 8 L A ( s ) + L 2 − L 2 + 8 L − B ( s ) , 1 2 L ≤ r c ( L ) ≤ 1 4 . (24) r_c(L)=\min_{\substack{s\notin\mathcal{C}_L\\ \mathcal{B}(s)<-L^2+8L}}
\frac{\mathcal{A}(s)+L^2}{-L^2+8L-\mathcal{B}(s)},
\qquad
\frac{1}{2L}\le r_c(L)\le\frac{1}{4}. \tag*{(24)} r c ( L ) = s ∈ / C L B ( s ) < − L 2 + 8 L min − L 2 + 8 L − B ( s ) A ( s ) + L 2 , 2 L 1 ≤ r c ( L ) ≤ 4 1 . (24)
For J 2 / J 1 > 1 / 4 J_2/J_1>1/4 J 2 / J 1 > 1/4 this competitor strictly beats every constrained minimizer. Its energy equals E ∗ E_* E ∗ at 1 / 4 1/4 1/4 , but this need not imply full-model ground-state coexistence for general L L L , since another state might cross earlier. Full-spin enumeration at L = 5 L=5 L = 5 gives r c ( 5 ) = 1 / 4 r_c(5)=1/4 r c ( 5 ) = 1/4 , as detailed in Appendix C4; the programs and complete histogram are included in the verification archive [19 ] . These bounds do not determine the general threshold.
Entropy suppression and finite-size control
The second result concerns the nearest-neighbor model before the second-neighbor selection. A finite periodic sample has entropy per spin ln G 0 ( M , N ) / ( M N ) \ln G_0(M,N)/(MN) ln G 0 ( M , N ) / ( M N ) . We bound how far this value can be from the bulk limit, so that the sample dimensions can be chosen for a prescribed accuracy. Figure 3 compares the finite values with the bulk limit and shows the guaranteed errors as the shorter period grows. The same analysis bounds D t o t / ( M N ) D_{\rm tot}/(MN) D tot / ( M N ) for all honeycomb matchings, including those incompatible with periodic spins. Both counts use one triangular-lattice site, or equivalently one honeycomb unit cell, as the normalization. Their common bulk reference is
Figure 3. The shorter period controls the guaranteed accuracy of the periodic-spin entropy on a triangular-lattice torus. Here m = min ( M , N ) m=\min(M,N) m = min ( M , N ) , n = max ( M , N ) n=\max(M,N) n = max ( M , N ) , and G 0 G_0 G 0 counts periodic ground states of the nearest-neighbor model (J 2 = 0 J_2=0 J 2 = 0 ), including both global spin flips. (a) The entropy per spin s m , n = ln G 0 ( m , n ) / ( m n ) s_{m,n}=\ln G_0(m,n)/(mn) s m , n = ln G 0 ( m , n ) / ( mn ) approaches the bulk value I = 0.3230659472 … I=0.3230659472\ldots I = 0.3230659472 … at three aspect ratios; lines join finite-product evaluations as guides to the eye. The inset shows a periodic cell, with opposite edges identified. (b) The absolute deviations (symbols) lie below the rigorous bounds π 3 / ( 3 m 2 ) + ln 25 / ( m n ) \pi\sqrt{3}/(3m^2)+\ln25/(mn) π 3 / ( 3 m 2 ) + ln 25/ ( mn ) (dashed lines) in Eq. (31) ; colors and symbols identify the same aspect ratios in both panels. The bound holds for every integer n ≥ m ≥ 3 n\ge m\ge3 n ≥ m ≥ 3 , including arbitrarily elongated tori.
I = 2 π ∫ 0 π / 3 ln ( 2 cos t ) d t = Cl 2 ( π / 3 ) π = 0.323065947219450514 … (25) \begin{aligned}
I=\frac{2}{\pi}\int_0^{\pi/3}\ln(2\cos t)\,\mathrm{d}t
=\frac{\operatorname{Cl}_2(\pi/3)}{\pi} \\
&=0.323065947219450514\ldots \tag*{(25)}
\end{aligned} I = π 2 ∫ 0 π /3 ln ( 2 cos t ) d t = π Cl 2 ( π /3 ) = 0.323065947219450514 … (25)
The Clausen function is Cl 2 ( x ) = ∑ r ≥ 1 sin ( r x ) / r 2 \operatorname{Cl}_2(x)=\sum_{r\ge1}\sin(rx)/r^2 Cl 2 ( x ) = ∑ r ≥ 1 sin ( r x ) / r 2 . The entropy-density error is less than 5.032676 / min ( M , N ) 2 5.032676/\min(M,N)^2 5.032676/ min ( M , N ) 2 for periodic spins and 5.613369 / min ( M , N ) 2 5.613369/\min(M,N)^2 5.613369/ min ( M , N ) 2 for all matchings. These bounds hold for every M , N ≥ 3 M,N\ge3 M , N ≥ 3 , including arbitrarily large aspect ratios, so they allow dimensions to be chosen before evaluating a finite count. The periodic-spin entropy is the quantity needed for the selected fraction. We first bound the total dimer count using the positive sum of determinant magnitudes in Eq. (9). We then control the subtraction in Eq. (10) to obtain the spin bound and apply it to the selected fraction. Strip growth rates explain how compatibility affects the counts; width corrections and sample-size estimates give further quantitative consequences.
For the full matching count, asymptotics on well-shaped tori resolve the bounded limiting corrections to ln D t o t − M N I \ln D_{\rm tot}-MN I ln D tot − M N I , including arithmetic phases [11 ] . Here we need an error estimate that also covers highly elongated samples and finite sizes. The following bound separates a correction controlled by the shorter period from one controlled by the area.
Theorem 2 (Uniform dimer entropy bound). Let m = min ( M , N ) m=\min(M,N) m = min ( M , N ) and n = max ( M , N ) n=\max(M,N) n = max ( M , N ) , with m ≥ 3 m\ge3 m ≥ 3 .
C ∗ = 2 ∑ j = 0 ∞ [ − ln ( 1 − 2 − 3 j − 1 / 2 ) − ln ( 1 − 2 − 3 j − 5 / 2 ) ] C_* = 2\sum_{j=0}^{\infty}\left[-\ln(1-2^{-3j-1/2})-\ln(1-2^{-3j-5/2})\right] C ∗ = 2 j = 0 ∑ ∞ [ − ln ( 1 − 2 − 3 j − 1/2 ) − ln ( 1 − 2 − 3 j − 5/2 ) ] = 3.10642204076874649 … (26) =3.10642204076874649\ldots \tag*{(26)} = 3.10642204076874649 … (26)
Then
∣ ln D t o t M N − I ∣ ≤ π 3 3 m 2 + C ∗ + ln 2 m n < 5.613369 m 2 . (27) \left|\frac{\ln D_{\mathrm{tot}}}{MN}-I\right|\le\frac{\pi\sqrt{3}}{3m^2}+\frac{C_*+\ln2}{mn}<\frac{5.613369}{m^2}. \tag*{(27)} M N ln D tot − I ≤ 3 m 2 π 3 + mn C ∗ + ln 2 < m 2 5.613369 . (27)
The two terms in Eq. (27) come from comparing the transverse spectral sum with its bulk integral and controlling the finite-length product factors, respectively. The latter are bounded by their distances from the modes with ∣ 1 + z ∣ = 1 |1+z|=1 ∣1 + z ∣ = 1 . Appendix B 3 gives the complete argument, including the enclosure of C ∗ C_* C ∗ .
For example, along M = k 3 M=k^3 M = k 3 , N = k N=k N = k with k ≥ 3 k\ge3 k ≥ 3 , the estimate is π 3 / ( 3 k 2 ) + ( C ∗ + ln 2 ) / k 4 \pi\sqrt{3}/(3k^2)+(C_*+\ln2)/k^4 π 3 / ( 3 k 2 ) + ( C ∗ + ln 2 ) / k 4 , even though the aspect ratio diverges. The unbounded aspect ratio places this sequence outside the well-shaped hypothesis in Eq. (5) of Ref. [11 ] .
Thus the full matching entropy tends to I I I whenever both periods grow. On the subfamily 3 ∣ gcd ( M , N ) 3\mid\gcd(M,N) 3 ∣ g cd( M , N ) , Eq. (11) transfers the same estimate to G 0 G_0 G 0 . For a general torus, G 0 ≤ D t o t G_0\le D_{\mathrm{tot}} G 0 ≤ D tot gives only an upper bound on the spin count. To obtain a two-sided entropy estimate, we must also control the subtraction that selects compatible matchings.
To identify a reference scale for the periodic-spin count, consider the product in Eq. (8). Factoring each mode by max { ∣ 1 + z ∣ M , 1 } \max\{|1+z|^M,1\} max { ∣1 + z ∣ M , 1 } gives a leading logarithmic magnitude M max { ln ∣ 1 + z ∣ , 0 } M\max\{\ln|1+z|,0\} M max { ln ∣1 + z ∣ , 0 } . The remaining factors, including possible zeros on the unit circle, are treated below. Summing the clipped logarithms on the two transverse twist grids defines
Λ + ( N ) = ∑ z N = 1 max { ln ∣ 1 + z ∣ , 0 } , \Lambda_+(N)=\sum_{z^N=1}\max\{\ln|1+z|,0\}, Λ + ( N ) = z N = 1 ∑ max { ln ∣1 + z ∣ , 0 } ,
Λ − ( N ) = ∑ z N = − 1 max { ln ∣ 1 + z ∣ , 0 } . (28) \Lambda_-(N)=\sum_{z^N=-1}\max\{\ln|1+z|,0\}. \tag*{(28)} Λ − ( N ) = z N = − 1 ∑ max { ln ∣1 + z ∣ , 0 } . (28)
A mode with z = − 1 z=-1 z = − 1 contributes zero to the clipped sum. The signs ± \pm ± label the grids z N = ± 1 z^N=\pm1 z N = ± 1 in the counting determinants. Combining these determinants to enforce periodic spin reconstruction leaves the rate Λ − \Lambda_- Λ − ; the next subsection explains the cancellation. At finite size, the corresponding exponential e n Λ − ( m ) e^{n\Lambda_-(m)} e n Λ − ( m ) already approximates G 0 G_0 G 0 within fixed factors. Here m = min ( M , N ) m=\min(M,N) m = min ( M , N ) and n = max ( M , N ) n=\max(M,N) n = max ( M , N ) ; exchanging the two primal coordinates preserves the graph, so G 0 ( M , N ) = G 0 ( n , m ) G_0(M,N)=G_0(n,m) G 0 ( M , N ) = G 0 ( n , m ) .
Theorem 3 (Uniform spin bound). For all integers n ≥ m ≥ 3 n\ge m\ge3 n ≥ m ≥ 3 ,
1 25 < G 0 ( n , m ) e − n Λ − ( m ) < 3. (29) \frac{1}{25}<G_0(n,m)e^{-n\Lambda_-(m)}<3. \tag*{(29)} 25 1 < G 0 ( n , m ) e − n Λ − ( m ) < 3. (29)
Consequently,
− π 3 3 m 2 − ln 25 m n < ln G 0 m n − I < π 3 3 m 2 + ln 3 m n , (30) -\frac{\pi\sqrt{3}}{3m^2}-\frac{\ln25}{mn}<\frac{\ln G_0}{mn}-I<\frac{\pi\sqrt{3}}{3m^2}+\frac{\ln3}{mn}, \tag*{(30)} − 3 m 2 π 3 − mn ln 25 < mn ln G 0 − I < 3 m 2 π 3 + mn ln 3 , (30)
∣ ln G 0 m n − I ∣ ≤ π 3 3 m 2 + ln 25 m n < 5.032676 m 2 . (31) \left|\frac{\ln G_0}{mn}-I\right|\le\frac{\pi\sqrt{3}}{3m^2}+\frac{\ln25}{mn}<\frac{5.032676}{m^2}. \tag*{(31)} mn ln G 0 − I ≤ 3 m 2 π 3 + mn ln 25 < m 2 5.032676 . (31)
The conservative ratio bound controls the entropy loss from spin compatibility. Its proof is computer-assisted: analytic estimates cover m ≥ 1000 m \ge1000 m ≥ 1000 , and integer intervals certify the remaining finite width range. The programs and complete interval records are supplied in the verification archive [19 ] , together with source checksums and reproduction instructions.
The proof isolates a positive contribution from the half-shifted grid and bounds the remaining determinant contributions. The resulting envelope is nonincreasing in the strip length, so a width certificate extends to every larger length. Appendix B 4 gives the full proof and its analytic and finite-width certification steps.
Taking logarithms gives the area correction in Eq. (31) ; the spectral rate supplies the width correction. Together they prove the periodic-spin bulk limit whenever min ( M , N ) → ∞ \min(M,N) \to\infty min ( M , N ) → ∞ , without arithmetic or aspect-ratio restrictions.
We can now compare the selected count from Sec. III with the unperturbed periodic-spin count. For odd L ≥ 5 L \ge5 L ≥ 5 , the uniform measure on C L C_L C L assigns the selected states a fraction P s e l ( L ) = G s e l ( L ) / G 0 ( L , L ) P_{\mathrm{sel}}(L) = G_{\mathrm{sel}}(L)/G_0(L,L) P sel ( L ) = G sel ( L ) / G 0 ( L , L ) . The numerator is known exactly; the spin entropy bound controls the denominator. Equations (16) and (31) give
ln P s e l ( L ) = − I L 2 + 4 ln L − ln 2 + ln ( 1 − L − 2 ) + ξ L , ∣ ξ L ∣ < 5.032676. (32) \ln P_{\mathrm{sel}}(L) = -IL^2 + 4\ln L - \ln2
+ \ln(1-L^{-2}) + \xi_L,
\qquad|\xi_L| < 5.032676. \tag*{(32)} ln P sel ( L ) = − I L 2 + 4 ln L − ln 2 + ln ( 1 − L − 2 ) + ξ L , ∣ ξ L ∣ < 5.032676. (32)
Thus ln P s e l ( L ) = − I L 2 + 4 ln L + O ( 1 ) \ln P_{\mathrm{sel}}(L) = -IL^2 + 4\ln L + O(1) ln P sel ( L ) = − I L 2 + 4 ln L + O ( 1 ) along odd sizes. Isotropic selection retains every allowed odd slope sector, while its minimizers occupy an exponentially small fraction of the full periodic-spin nearest-neighbor ground-state ensemble. Here P s e l P_{\mathrm{sel}} P sel is a static fraction under that uniform measure, not a cooling probability.
Strip rates and sector suppression
How much can the periodic-spin reconstruction condition reduce the number of matchings? To isolate this effect, fix the width N N N and let the length M M M grow. The relevant fraction is p M , N = D 00 / D t o t p_{M,N} = D_{00}/D_{\mathrm{tot}} p M , N = D 00 / D tot ; J 2 = 0 J_2 = 0 J 2 = 0 throughout this comparison. Define the strip entropy densities by
s d ( N ) = lim M → ∞ ln D t o t ( M , N ) M N , s s ( N ) = lim M → ∞ ln G 0 ( M , N ) M N . (33) s_d(N) = \lim_{M\to\infty}\frac{\ln D_{\mathrm{tot}}(M,N)}{MN},
\qquad
s_s(N) = \lim_{M\to\infty}\frac{\ln G_0(M,N)}{MN}. \tag*{(33)} s d ( N ) = M → ∞ lim M N ln D tot ( M , N ) , s s ( N ) = M → ∞ lim M N ln G 0 ( M , N ) . (33)
Theorem 4 (Strip rates). For fixed N ≥ 3 N \ge3 N ≥ 3 ,
s d ( N ) = max ( Λ + , Λ − ) N , s s ( N ) = Λ − N . (34) s_d(N) = \frac{\max(\Lambda_+,\Lambda_-)}{N},
\qquad
s_s(N) = \frac{\Lambda_-}{N}. \tag*{(34)} s d ( N ) = N max ( Λ + , Λ − ) , s s ( N ) = N Λ − . (34)
The ordering is Λ + > Λ − \Lambda_+ > \Lambda_- Λ + > Λ − if 3 ∤ N 3 \nmid N 3 ∤ N and Λ + < Λ − \Lambda_+ < \Lambda_- Λ + < Λ − if 3 ∣ N 3 \mid N 3 ∣ N .
These rates determine the exponential decay of the compatible matching fraction p M , N p_{M,N} p M , N . Whereas P s e l P_{\mathrm{sel}} P sel measures the reduction of the spin ground-state count by
Table 3. Entropy per site in the infinite-length limit at fixed width N N N : s d s_d s d includes all matchings, s s s_s s s counts periodic-spin ground states, and γ N = s d − s s \gamma_N = s_d - s_s γ N = s d − s s is their entropy-density difference. Values are evaluated from the defining spectral sums. All logarithms are natural and k B = 1 k_B = 1 k B = 1 .
the second-neighbor interaction, p M , N p_{M,N} p M , N measures which dimer coverings represent periodic spins. Equations (4) and (34) imply
lim M → ∞ 1 M ln p M , N = − N γ N , γ N = s d ( N ) − s s ( N ) . (35) \lim_{M\to\infty}\frac{1}{M}\ln p_{M,N} = -N\gamma_N,
\qquad
\gamma_N = s_d(N) - s_s(N). \tag*{(35)} M → ∞ lim M 1 ln p M , N = − N γ N , γ N = s d ( N ) − s s ( N ) . (35)
For 3 ∤ N 3 \nmid N 3 ∤ N , this fraction decays exponentially with length. For 3 ∣ N 3 \mid N 3 ∣ N , its exponential rate is zero; the finite-size equality of Eq. (11) still also requires 3 ∣ M 3 \mid M 3 ∣ M . The gap γ N \gamma_N γ N is an entropy-density deficit, not an excitation energy.
Table III illustrates both possibilities: widths divisible by three have equal strip rates, while other widths have a positive entropy-density deficit.
The distinction between the rates comes from determinant cancellation. For the full matching count, Eq. (8) assigns every mode with ∣ 1 + z ∣ ≠ 1 |1+z| \ne1 ∣1 + z ∣ = 1 its clipped logarithm. Modes with ∣ 1 + z ∣ = 1 |1+z| = 1 ∣1 + z ∣ = 1 occur only on the + + + transverse grid, and a suitable choice of longitudinal twist avoids cancellation. For periodic spins, Eq. (10) subtracts a determinant with the dominant rate. To track what survives, write
U M = 1 2 ( ∣ Δ ( 1 , − 1 ) ∣ + ∣ Δ ( − 1 , − 1 ) ∣ ) , R M = 1 2 ( ∣ Δ ( − 1 , 1 ) ∣ − ∣ Δ ( 1 , 1 ) ∣ ) . (36) U_M = \frac{1}{2}\bigl(|\Delta(1,-1)| + |\Delta(-1,-1)|\bigr),
\qquad
R_M = \frac{1}{2}\bigl(|\Delta(-1,1)| - |\Delta(1,1)|\bigr). \tag*{(36)} U M = 2 1 ( ∣Δ ( 1 , − 1 ) ∣ + ∣Δ ( − 1 , − 1 ) ∣ ) , R M = 2 1 ( ∣Δ ( − 1 , 1 ) ∣ − ∣Δ ( 1 , 1 ) ∣ ) . (36)
Then G 0 = U M + R M G_0 = U_M + R_M G 0 = U M + R M and M − 1 ln U M → Λ − M^{-1}\ln U_M \to\Lambda_- M − 1 ln U M → Λ − . For 3 ∤ N 3 \nmid N 3 ∤ N , set δ N = min z ≠ 1 ∣ ln ∣ 1 + z ∣ ∣ \delta_N = \min_{z\ne1}|\ln|1+z|| δ N = min z = 1 ∣ ln ∣1 + z ∣∣ , omitting z = − 1 z=-1 z = − 1 if present. Factoring out the leading magnitude in each determinant shows
R M = O ( e M ( Λ + − δ N ) ) . (37) R_M = O\bigl(e^{M(\Lambda_+ - \delta_N)}\bigr). \tag*{(37)} R M = O ( e M ( Λ + − δ N ) ) . (37)
Appendix B 1 proves δ N > Λ + − Λ − > 0 \delta_N > \Lambda_+ - \Lambda_- > 0 δ N > Λ + − Λ − > 0 . If 3 ∣ N 3 \mid N 3 ∣ N , the simpler R M = O ( e M Λ + ) R_M = O(e^{M\Lambda_+}) R M = O ( e M Λ + ) suffices since Λ + < Λ − \Lambda_+ < \Lambda_- Λ + < Λ − . Thus R M / U M → 0 R_M/U_M \to0 R M / U M → 0 in both cases: the subtraction leaves U M U_M U M as the leading spin contribution, establishing the rate Λ − / N \Lambda_-/N Λ − / N .
This spectral rate can also be read from the fermionic cylinder description of [8 ] , Eqs. (5)–(7): the mode energy is ϵ ( q ) = − ln ( 2 cos ( q / 2 ) ) \epsilon(q) = -\ln(2\cos(q/2)) ϵ ( q ) = − ln ( 2 cos ( q /2 )) , and even fermion number, corresponding to periodic spins, selects half-integer momenta. Filling the negative-energy modes in conjugate pairs gives Λ − \Lambda_- Λ − . Here γ N \gamma_N γ N compares the periodic-spin ensemble with the full matching ensemble; it is distinct from the excitation gaps governing mutual information in Ref. [8 ] .
The same half-shifted rate follows from the zero-temperature cylinder solution of Ref. [10 ] ; Appendix B 1 gives the explicit correspondence. The finite determinant calculation here tracks how periodic-spin compatibility selects this rate.
Finite-width expansions
At fixed width, making the torus longer approaches a strip entropy, which generally differs from I I I . The bulk entropy is recovered by subsequently increasing the width. The width corrections depend on N N N modulo three, which must be retained in a controlled extrapolation. Write χ 3 ( N ) = 0 , 1 , − 1 \chi_3(N)=0,1,-1 χ 3 ( N ) = 0 , 1 , − 1 when N N N has residue 0 , 1 , 2 0,1,2 0 , 1 , 2 modulo three, respectively.
The first terms of the unshifted-grid rate are
Λ + N − I = − 3 π 6 N 2 + 3 π 3 45 N 4 + O ( N − 6 ) , 3 ∣ N , (38) \frac{\Lambda_+}{N}-I=-\frac{\sqrt{3}\pi}{6N^2}+\frac{\sqrt{3}\pi^3}{45N^4}+O(N^{-6}),\qquad3\mid N, \tag*{(38)} N Λ + − I = − 6 N 2 3 π + 45 N 4 3 π 3 + O ( N − 6 ) , 3 ∣ N , (38)
whereas for 3 ∤ N 3\nmid N 3 ∤ N ,
Λ + N − I = − 3 π 18 N 2 + 4 χ 3 ( N ) π 2 81 N 3 − 13 3 π 4 1215 N 4 + O ( N − 5 ) . (39) \begin{aligned}
\frac{\Lambda_+}{N}-I={}&-\frac{\sqrt{3}\pi}{18N^2}+\frac{4\chi_3(N)\pi^2}{81N^3}\\
&-\frac{13\sqrt{3}\pi^4}{1215N^4}+O(N^{-5}). \tag*{(39)}
\end{aligned} N Λ + − I = − 18 N 2 3 π + 81 N 3 4 χ 3 ( N ) π 2 − 1215 N 4 13 3 π 4 + O ( N − 5 ) . (39)
Appendix B 2 gives the coefficients at arbitrary order and an explicit remainder bound in Theorem 5. It uses Fourier sampling and endpoint expansions, complementing existing exact partitions and expansions [20 , 21 , 22 , 23 ] . The low-order terms above show the main effect: finite-width corrections depend on the width modulo three.
The same expansion determines the periodic-spin rate through the exact sampling identity
Λ − ( N ) = Λ + ( 2 N ) − Λ + ( N ) (40) \Lambda_-(N)=\Lambda_+(2N)-\Lambda_+(N) \tag*{(40)} Λ − ( N ) = Λ + ( 2 N ) − Λ + ( N ) (40)
so no second endpoint calculation is needed. On the nondivisible branch this gives
γ N = 3 π 12 N 2 + χ 3 ( N ) π 2 9 N 3 − 13 3 π 3 648 N 4 + O ( N − 5 ) ; (41) \gamma_N=\frac{\sqrt{3}\pi}{12N^2}+\frac{\chi_3(N)\pi^2}{9N^3}-\frac{13\sqrt{3}\pi^3}{648N^4}+O(N^{-5}); \tag*{(41)} γ N = 12 N 2 3 π + 9 N 3 χ 3 ( N ) π 2 − 648 N 4 13 3 π 3 + O ( N − 5 ) ; (41)
on 3 ∣ N 3\mid N 3 ∣ N , γ N = 0 \gamma_N=0 γ N = 0 exactly. For 3 ∤ N 3\nmid N 3 ∤ N , the entropy-density deficit therefore vanishes as the width grows, while its leading suppression exponent per unit length in Eq. (35) is 3 π / ( 12 N ) \sqrt{3}\pi/(12N) 3 π / ( 12 N ) .
The explicit remainder in Appendix B 2 controls these truncations without assuming convergence of the infinite series. Its bound justifies the O ( N − 5 ) O(N^{-5}) O ( N − 5 ) remainder in Eq. (41) .
Quantitative consequences
The compatible matching fraction p M , N = D 00 / D t o t p_{M,N}=D_{00}/D_{\mathrm{tot}} p M , N = D 00 / D tot measures the cost of periodic spin compatibility within the matching ensemble. Figure 4 shows this comparison on strips. Its fraction can be tiny even when the entropy-density difference is small: the logarithm of the fraction accumulates over the strip area, whereas the entropy density is divided by that area.
Figure 4. Periodic-spin compatibility can suppress a matching fraction exponentially with strip length even though its entropy-density cost vanishes with width. (a) Natural logarithm of p M , N = D 00 / D t o t p_{M,N}=D_{00}/D_{\mathrm{tot}} p M , N = D 00 / D tot , the fraction of all honeycomb matchings that lift to periodic spins on a torus of length M M M and width N N N . Points are evaluated from finite products at M = 12 , 24 , … , 240 M=12,24,\ldots,240 M = 12 , 24 , … , 240 . Dashed lines are − ln 2 − M N γ N -\ln2-MN\gamma_N − ln 2 − M N γ N , where γ N = s d ( N ) − s s ( N ) \gamma_N=s_d(N)-s_s(N) γ N = s d ( N ) − s s ( N ) is the strip entropy-density deficit. Compatibility suppresses this fraction exponentially with length for 3 ∤ N 3\nmid N 3 ∤ N ; the displayed N = 6 N=6 N = 6 points have p M , 6 = 1 / 2 p_{M,6}=1/2 p M , 6 = 1/2 exactly. (b) Spectral evaluations of N 2 γ N N^2\gamma_N N 2 γ N for 3 ≤ N ≤ 90 3\leq N\leq90 3 ≤ N ≤ 90 . Dashed curves retain terms through N − 4 N^{-4} N − 4 in Eq. (41) ; the dotted line is 3 π / 12 \sqrt{3}\pi/12 3 π /12 , the common limit of the two nondivisible branches. Thus their entropy-density deficit vanishes as N − 2 N^{-2} N − 2 , while γ N = 0 \gamma_N=0 γ N = 0 exactly for 3 ∣ N 3\mid N 3 ∣ N .
The dashed lines in Figure 4 (a) use the spectral rate and the asymptotic intercept − ln 2 -\ln2 − ln 2 , without a fit. For 3 ∤ N 3\nmid N 3 ∤ N , this intercept follows by removing the leading magnitude from every factor with ∣ 1 + z ∣ ≠ 1 |1+z|\ne1 ∣1 + z ∣ = 1 in Eq. (8). The remaining magnitudes tend to one. The cancellation estimate used in Theorem 4 then gives
p M , N = 1 2 e − M N γ N [ 1 + o ( 1 ) ] , 3 ∤ N , (42) p_{M,N}=\frac{1}{2}e^{-MN\gamma_N}[1+o(1)],\qquad3\nmid N, \tag*{(42)} p M , N = 2 1 e − M N γ N [ 1 + o ( 1 )] , 3 ∤ N , (42)
for fixed N N N and M → ∞ M\to\infty M → ∞ . When 3 ∣ N 3\mid N 3 ∣ N , the half-shifted branch dominates both counts, giving p M , N → 1 / 2 p_{M,N}\to1/2 p M , N → 1/2 . The N = 6 N=6 N = 6 points in the figure have 3 ∣ M 3\mid M 3 ∣ M as well, so they equal 1 / 2 1/2 1/2 exactly by Eq. (11) . For example, p 240 , 4 = 1.5644 × 10 − 17 p_{240,4}=1.5644\times10^{-17} p 240 , 4 = 1.5644 × 1 0 − 17 whereas p 240 , 6 = 1 / 2 p_{240,6}=1/2 p 240 , 6 = 1/2 . Such large differences in compatible fractions are consistent with both strip entropy densities approaching I I I as N → ∞ N\to\infty N → ∞ .
For uniform rejection sampling from all matchings, p M , N p_{M,N} p M , N is the acceptance probability for periodic-spin compatibility; the 240 × 4 240\times4 240 × 4 example requires on average 1 / p 240 , 4 ≃ 6.39 × 10 16 1/p_{240,4}\simeq6.39\times10^{16} 1/ p 240 , 4 ≃ 6.39 × 1 0 16 independent draws per accepted matching.
For width extrapolation, Figure 4 (b) shows the three residue classes separately. The N − 3 N^{-3} N − 3 term in Eq. (41) has opposite signs on the two nondivisible branches. The explicit truncation envelope in Appendix B 2, Eq. (69) , controls the error at finite N N N : at N = 31 N=31 N = 31 , the actual error is 1.76 × 10 − 7 1.76\times10^{-7} 1.76 × 1 0 − 7 , below the guaranteed 4.89 × 10 − 7 4.89\times10^{-7} 4.89 × 1 0 − 7 .
Finally, the uniform bounds give dimensions sufficient for a prescribed accuracy, without a strip extrapolation. For the matching count, define the entropy-density error ϵ M , N = ln D t o t / ( M N ) − I \epsilon_{M,N}=\ln D_{\mathrm{tot}}/(MN)-I ϵ M , N = ln D tot / ( M N ) − I and the aspect ratio ρ = n / m ≥ 1 \rho=n/m\ge1 ρ = n / m ≥ 1 . Theorem 2 gives
C w = π 3 / 3 , C a = C ∗ + ln 2 , C_w=\pi\sqrt{3}/3,\qquad C_a=C_*+\ln2, C w = π 3 /3 , C a = C ∗ + ln 2 ,
m 2 ∣ ϵ M , N ∣ ≤ C w + C a ρ , (43) m^2|\epsilon_{M,N}|\le C_w+\frac{C_a}{\rho}, \tag*{(43)} m 2 ∣ ϵ M , N ∣ ≤ C w + ρ C a , (43)
separating the width contribution C w C_w C w from the area contribution C a / ρ C_a/\rho C a / ρ . This conservative envelope guarantees accuracy before a calculation; it does not estimate the smallest possible correction. For a target entropy-density error ϵ > 0 \epsilon>0 ϵ > 0 , a sufficient condition is
m ≥ max { 3 , ⌈ C w + C a / ρ ϵ ⌉ } . (44) m\ge\max\left\{3,\left\lceil\sqrt{\frac{C_w+C_a/\rho}{\epsilon}}\right\rceil\right\}. \tag*{(44)} m ≥ max { 3 , ⌈ ϵ C w + C a / ρ ⌉ } . (44)
For ϵ = 10 − 3 \epsilon=10^{-3} ϵ = 1 0 − 3 , this condition gives m = 75 m=75 m = 75 at ρ = 1 \rho=1 ρ = 1 , m = 46 m=46 m = 46 at ρ = 16 \rho=16 ρ = 16 , and m = 44 m=44 m = 44 at ρ = 64 \rho=64 ρ = 64 . These sufficient sizes can be chosen before evaluating the finite products. For periodic spins, Eq. (31) replaces C a C_a C a by ln 25 \ln25 ln 25 , with no divisibility condition. At ρ = 16 \rho=16 ρ = 16 , choosing $N=45 and M = 720 M=720 M = 720 guarantees ∣ ln G 0 / ( M N ) − I ∣ < 10 − 3 |\ln G_0/(MN)-I|<10^{-3} ∣ ln G 0 / ( M N ) − I ∣ < 1 0 − 3 . The matching bound would require a minimum dimension of 46 at this aspect ratio. Figure 3 (b) compares this spin-entropy guarantee with the actual errors. Increasing the long period at fixed width reduces the area term but leaves the width term. A guaranteed approach to the bulk entropy therefore requires the shorter period to grow as well.
The evaluations in Figure 4 were compared at 80 and 120 decimal digits, with selected sizes also checked against exact integer determinants. For Figure 3 , evaluations at 100 and 140 decimal digits agreed, and every displayed count was checked by an exact integer determinant calculation. An independent spin-row calculation agrees with 2 D h 2D_h 2 D h for all four periodic or antiperiodic seam choices at the recorded test sizes, and small tori were checked by complete spin enumeration. These checks test the implementation; the proofs establish the bounds. The precision-check summaries, interval records, and input-output files are supplied in the verification archive [19 ] .
Discussion
The odd-period boundary condition frustrates the uniform stripe order favored by the second-neighbor interaction. Within the J 2 = 0 J_2=0 J 2 = 0 ground-state set, interfaces accommodate this mismatch while preserving the local triangle rule. Their minimum cost is proportional to L L L and vanishes per spin. Minimizing that cost at each finite size nevertheless reduces an exponential degeneracy to a polynomial one. The classification also fixes the statistics of the surviving states: each allowed composition has the same weight, giving a continuous limiting distribution of whole-sample orientation fractions. The known freedom to change domain areas at fixed energy does not by itself determine these weights.
The distinction between local order and global composition is essential. Every selected state has 4 L 4L 4 L interface edges, and the uniform measure includes all translations. For a fixed connected patch of triangles, translation averaging bounds the probability of an internal interface by a constant times 4 L / L 2 4L/L^2 4 L / L 2 . Thus a fixed patch has a single stripe orientation with probability tending to one, even though the whole-sample proportions have limiting variance 1 / 18 1/18 1/18 . For comparison, an equal mixture of uniform stripe orientations on a compatible even-period sample has global fractions only at the three simplex vertices, with variance 2 / 9 2/9 2/9 . The Dirichlet law instead has weight throughout the simplex interior. A fixed observation window and an area fraction measured over the whole sample therefore probe different aspects of the limit. The absence of concentration stated here refers to the latter observable in the specified odd-torus ensemble.
The relevant counting comparison is between these selected states and all J 2 = 0 J_2=0 J 2 = 0 ground states with the same periodic spin boundaries. Their ratio has the exponential area dependence in Eq. (32) . A separate comparison is needed when counting spins through dimers: only matchings that reconstruct periodic spins contribute to G 0 G_0 G 0 . On a strip of fixed width, their fraction of all matchings can decay exponentially with length, even though the difference between spin and dimer entropy densities vanishes as the width increases. Keeping the two comparisons distinct is necessary to quantify the entropy removed by the interaction.
The finite-size bounds give conservative accuracy guarantees for residual entropy, including on strongly elongated tori. The strip expansions resolve the additional dependence on width modulo three. These results complement the asymptotics for well-shaped tori in Ref. [11 ] , the sector and cylinder analyses of Refs. [8 , 9 ] , and higher-order toroidal dimer expansions [20 , 21 , 22 ] . The integer-interval certification is essential to the constants in the periodic-spin bound; its analytic reduction is given in Appendix B 4 and its records in the verification archive [19 ] .
The composition limit takes weak coupling at each fixed size before increasing odd L L L . At fixed L L L , the uniform full-model ground-state ensemble agrees with the constrained measure throughout 0 < J 2 / J 1 < 1 / ( 2 L ) 0 < J_2/J_1 < 1/(2L) 0 < J 2 / J 1 < 1/ ( 2 L ) . Its limit as J 2 / J 1 → 0 + J_2/J_1 \to0^+ J 2 / J 1 → 0 + therefore gives the measure studied here. The threshold satisfies 1 / ( 2 L ) ≤ r c ( L ) ≤ 1 / 4 1/(2L) \le r_c(L) \le1/4 1/ ( 2 L ) ≤ r c ( L ) ≤ 1/4 ; these bounds neither establish nor exclude stability at a fixed nonzero coupling ratio as L L L grows. That order of limits is not determined by the present finite-size results. Further directions include the nonvanishing of interior slope coefficients for general rectangular tori beyond the odd-square sectors constructed here, and sharper entropy bounds at unrestricted aspect ratio. Boundary-dependent entropy and vortices in other domains [24 , 25 , 26 ] raise a further comparison. Some of those ground states have no dimer covering, so their vortex counts cannot be directly identified with the height periods used here.
Acknowledgments
We acknowledge Gewu Intelligence Lab for providing the collaborative research environment in which this project was developed. GPT-assisted symbolic, numerical, and editorial work was checked against the derivations and exact counts reported here. The authors take full responsibility for the scientific content and conclusions of this paper.
Data availability
The data and code supporting this article are available in the versioned Zenodo reproducibility archive [19 ] . It includes exact counting records, integer-interval certificates, figure and table inputs, and the programs and instructions for reproducing the computations. Analytic derivations are given in the main text and Appendices A–C.
Matching sectors, signs, and slope census
The determinant formulas in Sec. II count matchings with signs. We give the sector inversion and finite-torus identities, then prove the sign rule and use it to resolve counts at each slope ( H , V ) (H,V) ( H , V ) . The holonomy sectors are recovered by grouping slopes according to Eq. (7). On a general M × N M \times N M × N torus, the height-current construction used in Sec. III gives periods Φ 1 = 3 H − M \Phi_1 = 3H - M Φ 1 = 3 H − M and Φ 2 = 3 V − N \Phi_2 = 3V - N Φ 2 = 3 V − N . Equation (19) is their M = N = L M = N = L M = N = L specialization.
Sector inversion and finite-torus identities
The boundary signs in Eq. (8) distinguish the two crossing parities. To extract an unsigned count, we also need the permutation sign associated with each parity pair. For a = ( a 1 , a 2 ) ∈ Z 2 2 a = (a_1,a_2) \in\mathbb{Z}_2^2 a = ( a 1 , a 2 ) ∈ Z 2 2 set
Q ( a ) = ( − 1 ) a 1 a 2 + a 1 [ 1 + ( N m o d 2 ) ] + a 2 [ 1 + ( M m o d 2 ) ] . (45) Q(a) = (-1)^{a_1a_2+a_1[1+(N \bmod2)]+a_2[1+(M \bmod2)]}. \tag*{(45)} Q ( a ) = ( − 1 ) a 1 a 2 + a 1 [ 1 + ( N mod 2 )] + a 2 [ 1 + ( M mod 2 )] . (45)
The permutation sign of a matching with slope ( H , V ) (H,V) ( H , V ) is Q ( H , V ) Q(H,V) Q ( H , V ) , where arguments of Q Q Q are reduced modulo two. Appendix A 2 gives an elementary derivation in this convention. Thus, with a = h + ( M , N ) ( m o d 2 ) a = h + (M,N) \pmod2 a = h + ( M , N ) ( mod 2 ) , character inversion yields
D h = Q ( a ) 4 ∑ u , v = 0 1 ( − 1 ) v a 1 + u a 2 Δ ( ( − 1 ) u , ( − 1 ) v ) . (46) D_h = \frac{Q(a)}{4}\sum_{u,v=0}^{1}(-1)^{v a_1+u a_2}\Delta((-1)^u,(-1)^v). \tag*{(46)} D h = 4 Q ( a ) u , v = 0 ∑ 1 ( − 1 ) v a 1 + u a 2 Δ (( − 1 ) u , ( − 1 ) v ) . (46)
This is the genus-one homology inversion of surface-dimer theory [14 ] , evaluated with the explicit shift in Eq. (7). The finite-torus identities in Sec. II now follow.
Proof. The only possible zero in Eq. (8) has ∣ z ∣ = ∣ 1 + z ∣ = 1 |z| = |1+z| = 1 ∣ z ∣ = ∣1 + z ∣ = 1 , so z = e ± 2 π i / 3 z = e^{\pm2\pi i/3} z = e ± 2 π i /3 . Substitution shows that only Δ ( 1 , 1 ) \Delta(1,1) Δ ( 1 , 1 ) can vanish, exactly when both dimensions are divisible by three. Nonreal roots pair to give positive factors. The real root z = 1 z = 1 z = 1 , when present, gives 2 M − ( − 1 ) M σ > 0 2^M - (-1)^M \sigma> 0 2 M − ( − 1 ) M σ > 0 ; z = − 1 z = -1 z = − 1 gives − ( − 1 ) M σ -(-1)^M \sigma − ( − 1 ) M σ . Hence a nonzero determinant has sign
sgn Δ ( σ , τ ) = { + 1 , τ = ( − 1 ) N + 1 , − ( − 1 ) M σ , τ = ( − 1 ) N . (47) \operatorname{sgn}\Delta(\sigma,\tau)=
\begin{cases}
+1, & \tau= (-1)^{N+1},\\
-(-1)^M\sigma, & \tau= (-1)^N.
\end{cases} \tag*{(47)} sgn Δ ( σ , τ ) = { + 1 , − ( − 1 ) M σ , τ = ( − 1 ) N + 1 , τ = ( − 1 ) N . (47)
Substituting these signs into Eq. (A2), summing over h h h , and using Eq. (4) gives Eqs. (9) and (10). The zero criterion then gives Eq. (11) . ■
The matching sign in the cell convention
Every nontrivial cycle of the permutation in Eq. (5) consists of v v v steps ( 1 , 0 ) (1,0) ( 1 , 0 ) and h h h steps ( 0 , − 1 ) (0,-1) ( 0 , − 1 ) . It is an embedded essential loop: its vertices are distinct, and a lift has nonzero monotone displacement ( M q , − N p ) (Mq,-Np) ( M q , − N p ) with
@ p a g e 12 / 22 s t a r t s m i d = 1 e n d s m i d = 1 t i n c t i s n e c e s s a r y t o q u a n t i f y t h e e n t r o p y r e m o v e d b y t h e i n t e r a c t i o n . @page 12/22 starts_mid=1 ends_mid=1
tinct is necessary to quantify the entropy removed by the interaction. @ p a g e 12/22 s t a r t s m i d = 1 e n d s m i d = 1 t in c t i s n ecess a r y t o q u an t i f y t h ee n t r o p y r e m o v e d b y t h e in t er a c t i o n .
The finite-size bounds give conservative accuracy guarantees for residual entropy, including on strongly elongated tori. The strip expansions resolve the additional dependence on width modulo three. These results complement the asymptotics for well-shaped tori in Ref. [11 ] , the sector and cylinder analyses of Refs. [8 , 9 ] , and higher-order toroidal dimer expansions [20 , 21 , 22 ] . The integer-interval certification is essential to the constants in the periodic-spin bound; its analytic reduction is given in Appendix B 4 and its records in the verification archive [19 ] .
The composition limit takes weak coupling at each fixed size before increasing odd L L L . At fixed L L L , the uniform full-model ground-state ensemble agrees with the constrained measure throughout 0 < J 2 / J 1 < 1 / ( 2 L ) 0 < J_2/J_1 < 1/(2L) 0 < J 2 / J 1 < 1/ ( 2 L ) . Its limit as J 2 / J 1 → 0 + J_2/J_1 \to0^+ J 2 / J 1 → 0 + therefore gives the measure studied here. The threshold satisfies 1 / ( 2 L ) ≤ r c ( L ) ≤ 1 / 4 1/(2L) \le r_c(L) \le1/4 1/ ( 2 L ) ≤ r c ( L ) ≤ 1/4 ; these bounds neither establish nor exclude stability at a fixed nonzero coupling ratio as L L L grows. That order of limits is not determined by the present finite-size results. Further directions include the nonvanishing of interior slope coefficients for general rectangular tori beyond the odd-square sectors constructed here, and sharper entropy bounds at unrestricted aspect ratio. Boundary-dependent entropy and vortices in other domains [24 , 25 , 26 ] raise a further comparison. Some of those ground states have no dimer covering, so their vortex counts cannot be directly identified with the height periods used here.
Acknowledgments
We acknowledge Gewu Intelligence Lab for providing the collaborative research environment in which this project was developed. GPT-assisted symbolic, numerical, and editorial work was checked against the derivations and exact counts reported here. The authors take full responsibility for the scientific content and conclusions of this paper.
Data availability
The data and code supporting this article are available in the versioned Zenodo reproducibility archive [19 ] . It includes exact counting records, integer-interval certificates, figure and table inputs, and the programs and instructions for reproducing the computations. Analytic derivations are given in the main text and Appendices A–C.
Matching sectors, signs, and slope census
The determinant formulas in Sec. II count matchings with signs. We give the sector inversion and finite-torus identities, then prove the sign rule and use it to resolve counts at each slope ( H , V ) (H,V) ( H , V ) . The holonomy sectors are recovered by grouping slopes according to Eq. (7). On a general M × N M \times N M × N torus, the height-current construction used in Sec. III gives periods Φ 1 = 3 H − M \Phi_1 = 3H - M Φ 1 = 3 H − M and Φ 2 = 3 V − N \Phi_2 = 3V - N Φ 2 = 3 V − N . Equation (19) is their M = N = L M = N = L M = N = L specialization.
Sector inversion and finite-torus identities
The boundary signs in Eq. (8) distinguish the two crossing parities. To extract an unsigned count, we also need the permutation sign associated with each parity pair. For a = ( a 1 , a 2 ) ∈ Z 2 2 a = (a_1,a_2) \in\mathbb{Z}_2^2 a = ( a 1 , a 2 ) ∈ Z 2 2 set
Q ( a ) = ( − 1 ) a 1 a 2 + a 1 [ 1 + ( N m o d 2 ) ] + a 2 [ 1 + ( M m o d 2 ) ] . (48) Q(a) = (-1)^{a_1a_2+a_1[1+(N \bmod2)]+a_2[1+(M \bmod2)]}. \tag*{(48)} Q ( a ) = ( − 1 ) a 1 a 2 + a 1 [ 1 + ( N mod 2 )] + a 2 [ 1 + ( M mod 2 )] . (48)
The permutation sign of a matching with slope ( H , V ) (H,V) ( H , V ) is Q ( H , V ) Q(H,V) Q ( H , V ) , where arguments of Q Q Q are reduced modulo two. Appendix A 2 gives an elementary derivation in this convention. Thus, with a = h + ( M , N ) ( m o d 2 ) a = h + (M,N) \pmod2 a = h + ( M , N ) ( mod 2 ) , character inversion yields
D h = Q ( a ) 4 ∑ u , v = 0 1 ( − 1 ) v a 1 + u a 2 Δ ( ( − 1 ) u , ( − 1 ) v ) . (49) D_h = \frac{Q(a)}{4}\sum_{u,v=0}^{1}(-1)^{v a_1+u a_2}\Delta((-1)^u,(-1)^v). \tag*{(49)} D h = 4 Q ( a ) u , v = 0 ∑ 1 ( − 1 ) v a 1 + u a 2 Δ (( − 1 ) u , ( − 1 ) v ) . (49)
This is the genus-one homology inversion of surface-dimer theory [14 ] , evaluated with the explicit shift in Eq. (7). The finite-torus identities in Sec. II now follow.
Proof. The only possible zero in Eq. (8) has ∣ z ∣ = ∣ 1 + z ∣ = 1 |z| = |1+z| = 1 ∣ z ∣ = ∣1 + z ∣ = 1 , so z = e ± 2 π i / 3 z = e^{\pm2\pi i/3} z = e ± 2 π i /3 . Substitution shows that only Δ ( 1 , 1 ) \Delta(1,1) Δ ( 1 , 1 ) can vanish, exactly when both dimensions are divisible by three. Nonreal roots pair to give positive factors. The real root z = 1 z = 1 z = 1 , when present, gives 2 M − ( − 1 ) M σ > 0 2^M - (-1)^M \sigma> 0 2 M − ( − 1 ) M σ > 0 ; z = − 1 z = -1 z = − 1 gives − ( − 1 ) M σ -(-1)^M \sigma − ( − 1 ) M σ . Hence a nonzero determinant has sign
sgn Δ ( σ , τ ) = { + 1 , τ = ( − 1 ) N + 1 , − ( − 1 ) M σ , τ = ( − 1 ) N . (50) \operatorname{sgn}\Delta(\sigma,\tau)=
\begin{cases}
+1, & \tau= (-1)^{N+1},\\
-(-1)^M\sigma, & \tau= (-1)^N.
\end{cases} \tag*{(50)} sgn Δ ( σ , τ ) = { + 1 , − ( − 1 ) M σ , τ = ( − 1 ) N + 1 , τ = ( − 1 ) N . (50)
Substituting these signs into Eq. (A2), summing over h h h , and using Eq. (4) gives Eqs. (9) and (10). The zero criterion then gives Eq. (11) . ■
The matching sign in the cell convention
Every nontrivial cycle of the permutation in Eq. (5) consists of v v v steps ( 1 , 0 ) (1,0) ( 1 , 0 ) and h h h steps ( 0 , − 1 ) (0,-1) ( 0 , − 1 ) . It is an embedded essential loop: its vertices are distinct, and a lift has nonzero monotone displacement ( M q , − N p ) (Mq,-Np) ( M q , − N p ) with
p , q ≥ 0 p, q \ge0 p , q ≥ 0 . The grid embedding has no edge crossings for M , N ≥ 3 M, N \ge3 M , N ≥ 3 . Disjoint essential simple loops on a torus have parallel primitive homology classes. To recall why primitivity holds, cut along one such loop. The complement is an annulus; a transverse arc joining its two boundary components closes to a curve of algebraic intersection ± 1 \pm1 ± 1 with the original loop. A divisible homology class could not have that intersection number. Parallelism follows either from the annulus or from vanishing pairwise intersection numbers. Monotonicity fixes the same orientation for all the cycles.
If there are s s s nontrivial cycles, they therefore have the same primitive pair ( p , q ) (p,q) ( p , q ) , so H = s p H=sp H = s p , V = s q V=sq V = s q , and s = gcd ( H , V ) s=\gcd(H,V) s = g cd( H , V ) . Their total length is N H + M V NH+MV N H + M V , giving permutation sign ( − 1 ) N H + M V − s (-1)^{NH+MV-s} ( − 1 ) N H + M V − s . For integers H , V H,V H , V , gcd ( H , V ) ≡ H + V + H V ( m o d 2 ) \gcd(H,V)\equiv H+V+HV\pmod2 g cd( H , V ) ≡ H + V + H V ( mod 2 ) : it is even exactly when both entries are even. Taking gcd ( 0 , 0 ) = 0 \gcd(0,0)=0 g cd( 0 , 0 ) = 0 covers the identity permutation. This proves (48) and the constant sign at each slope.
Resolving the count by slope
Let N M , N ( H , V ) \mathcal{N}_{M,N}(H,V) N M , N ( H , V ) count matchings of slope ( H , V ) (H,V) ( H , V ) . These counts resolve the directional occupations more finely than the four holonomy sectors. Their signed generating function is
Δ ( σ , τ ) = ∑ H , V N M , N ( H , V ) Q ( H , V ) τ H σ V . (51) \Delta(\sigma,\tau)=\sum_{H,V}\mathcal{N}_{M,N}(H,V)Q(H,V)\tau^{H}\sigma^{V}. \tag*{(51)} Δ ( σ , τ ) = H , V ∑ N M , N ( H , V ) Q ( H , V ) τ H σ V . (51)
The coefficients can be obtained using integer polynomial arithmetic. Put c = ( − 1 ) M σ c=(-1)^M\sigma c = ( − 1 ) M σ , and let β i \beta_i β i be the M M M roots of ( 1 + x ) M = c (1+x)^M=c ( 1 + x ) M = c . Let e p e_p e p be the degree-p p p elementary symmetric polynomial in β 0 N , … , β M − 1 N \beta_0^N,\ldots,\beta_{M-1}^N β 0 N , … , β M − 1 N , and let [ c V ] [c^V] [ c V ] denote extraction of the coefficient of c V c^V c V . The power sums q k ( c ) q_k(c) q k ( c ) and Newton recursion are
q k ( c ) = M ∑ b = 0 ⌊ N k / M ⌋ ( N k b M ) ( − 1 ) N k − b M c b , (52) q_k(c)=M\sum_{b=0}^{\lfloor Nk/M\rfloor}\binom{Nk}{bM}(-1)^{Nk-bM}c^b, \tag*{(52)} q k ( c ) = M b = 0 ∑ ⌊ N k / M ⌋ ( b M N k ) ( − 1 ) N k − b M c b , (52)
e 0 = 1 , p e p = ∑ k = 1 p ( − 1 ) k − 1 e p − k q k . (53) e_0=1,\qquad pe_p=\sum_{k=1}^{p}(-1)^{k-1}e_{p-k}q_k. \tag*{(53)} e 0 = 1 , p e p = k = 1 ∑ p ( − 1 ) k − 1 e p − k q k . (53)
Then
N M , N ( H , V ) = ∣ [ c V ] e M − H ( c ) ∣ . (54) \mathcal{N}_{M,N}(H,V)=\left|[c^V]e_{M-H}(c)\right|. \tag*{(54)} N M , N ( H , V ) = [ c V ] e M − H ( c ) . (54)
To derive this, write β i = ω i − 1 \beta_i=\omega_i-1 β i = ω i − 1 with ω i M = c \omega_i^M=c ω i M = c . The root-of-unity sum in ∑ i ( ω i − 1 ) N k \sum_i(\omega_i-1)^{Nk} ∑ i ( ω i − 1 ) N k gives eq:A5 , and Newton’s identities give eq:A6 . Alternatively evaluating the resultant in (55) gives
[ τ H σ V ] Δ = ( − 1 ) M N + H + M V [ c V ] e M − H ( c ) . (55) [\tau^H\sigma^V]\Delta=(-1)^{MN+H+MV}[c^V]e_{M-H}(c). \tag*{(55)} [ τ H σ V ] Δ = ( − 1 ) M N + H + M V [ c V ] e M − H ( c ) . (55)
Since the matching sign is constant at each slope, taking the absolute coefficient proves eq:A7 . All divisions by p p p are exact in Z [ c ] \mathbb{Z}[c] Z [ c ] . The arithmetic cost is O ( M 2 N ) O(M^2N) O ( M 2 N ) ring multiplications; this statement does not bound the bit cost of the growing integer coefficients.
The support lies in N H + M V ≤ M N NH+MV\le MN N H + M V ≤ M N , and interchange of the two periods gives
N M , N ( H , V ) = N N , M ( V , H ) . (56) \mathcal{N}_{M,N}(H,V)=\mathcal{N}_{N,M}(V,H). \tag*{(56)} N M , N ( H , V ) = N N , M ( V , H ) . (56)
On the legs, N ( H , 0 ) = ( M H ) \mathcal{N}(H,0)=\binom{M}{H} N ( H , 0 ) = ( H M ) and N ( 0 , V ) = ( N V ) \mathcal{N}(0,V)=\binom{N}{V} N ( 0 , V ) = ( V N ) . For the hypotenuse put g = gcd ( M , N ) g=\gcd(M,N) g = g cd( M , N ) , M = g M ′ M=gM' M = g M ′ , and N = g N ′ N=gN' N = g N ′ . Its lattice points and their counts are
N ( t M ′ , N ′ ( g − t ) ) = ( g t ) , 0 ≤ t ≤ g . (57) \mathcal{N}(tM',N'(g-t))=\binom{g}{t},\qquad0\le t\le g. \tag*{(57)} N ( t M ′ , N ′ ( g − t )) = ( t g ) , 0 ≤ t ≤ g . (57)
Indeed, equality in the slope constraint excludes d d d cells. Every step is an i i i or changes j j j by 1 1 1 . Bijectivity forces cells connected by ( i , j ) ↦ ( i + 1 , j + 1 ) (i,j)\mapsto(i+1,j+1) ( i , j ) ↦ ( i + 1 , j + 1 ) to have the same type. There are g g g such orbits, each of length M N / g MN/g M N / g ; choosing which t t t orbits have type h h h proves eq:A10 . Thus the total hypotenuse count is 2 g 2^g 2 g . The identity specifies boundary occupations exactly, but supplies no proof that every strictly interior lattice point is occupied. Interior nonvanishing remains an additional coefficient problem, not a premise of the sector or entropy results.
This appendix proves the strip ordering, endpoint expansions, and uniform entropy bounds used in Sec. IV. The spin count here includes all periodic nearest-neighbor ground states, before the second-neighbor selection of Sec. III. The determinant estimates control the total matching count; a separate cancellation estimate controls the periodic-spin count.
Strip ordering and cancellation
Put a = π / 3 a=\pi/3 a = π /3 and φ ( t ) = ln ( 2 cos t ) \varphi(t)=\ln(2\cos t) φ ( t ) = ln ( 2 cos t ) on [ 0 , a ] [0,a] [ 0 , a ] . Let f ( t ) = max { ln ( 2 ∣ cos t ∣ ) , 0 } f(t)=\max\{\ln(2|\cos t|),0\} f ( t ) = max { ln ( 2∣ cos t ∣ ) , 0 } , of period π \pi π , and f ^ ( l ) = π − 1 ∫ − π / 2 π / 2 f ( t ) e − 2 i l t d t \widehat f(l)=\pi^{-1}\int_{-\pi/2}^{\pi/2}f(t)e^{-2ilt}\,dt f ( l ) = π − 1 ∫ − π /2 π /2 f ( t ) e − 2 i l t d t . Sampling at the unshifted and half-shifted grids gives
S N : = Λ + − Λ − = 4 N ∑ q ≥ 1 q o d d f ^ ( q N ) . (58) S_N:=\Lambda_+-\Lambda_-=4N\sum_{\substack{q\ge1\\ q\ \mathrm{odd}}}\widehat f(qN). \tag*{(58)} S N := Λ + − Λ − = 4 N q ≥ 1 q odd ∑ f ( q N ) . (58)
Two integrations by parts, and one more for the remainder, give for integer l ≠ 0 l\ne0 l = 0
f ^ ( l ) = − 3 cos ( 2 π l / 3 ) 2 π l 2 + ϵ l , ∣ ϵ l ∣ ≤ 2 3 + 3 4 π ∣ l ∣ 3 . (59) \widehat f(l)=-\frac{\sqrt{3}\cos(2\pi l/3)}{2\pi l^2}+\epsilon_l,\qquad|\epsilon_l|\le\frac{2\sqrt{3}+3}{4\pi|l|^3}. \tag*{(59)} f ( l ) = − 2 π l 2 3 cos ( 2 π l /3 ) + ϵ l , ∣ ϵ l ∣ ≤ 4 π ∣ l ∣ 3 2 3 + 3 . (59)
For the bound, ∣ sin ( 2 π l / 3 ) ∣ ≤ 3 / 2 |\sin(2\pi l/3)|\le\sqrt{3}/2 ∣ sin ( 2 π l /3 ) ∣ ≤ 3 /2 , ∣ φ ′ ′ ( a ) ∣ = 4 |\varphi''(a)|=4 ∣ φ ′′ ( a ) ∣ = 4 , and ∫ 0 a ∣ φ ′ ′ ′ ( t ) ∣ d t = 3 \int_0^a|\varphi'''(t)|\,dt=3 ∫ 0 a ∣ φ ′′′ ( t ) ∣ d t = 3 . The odd cosine sum consequently yields
S N = B N N + E N , ∣ E N ∣ ≤ C E N 2 . (60) S_N=\frac{B_N}{N}+E_N,\qquad|E_N|\le\frac{C_E}{N^2}. \tag*{(60)} S N = N B N + E N , ∣ E N ∣ ≤ N 2 C E . (60)
where C E = ( 2 3 + 3 ) ( 7 / 8 ) ζ ( 3 ) / π < 2.2 C_E=(2\sqrt{3}+3)(7/8)\zeta(3)/\pi<2.2 C E = ( 2 3 + 3 ) ( 7/8 ) ζ ( 3 ) / π < 2.2 , B N = − 3 π / 4 B_N=-\sqrt{3}\pi/4 B N = − 3 π /4 on 3 ∣ N 3\mid N 3 ∣ N and B N = 3 π / 12 B_N=\sqrt{3}\pi/12 B N = 3 π /12 otherwise. This proves S N < 0 S_N<0 S N < 0 on the divisible branch for N ≥ 3 N\ge3 N ≥ 3 , and S N > 0 S_N>0 S N > 0 on the nondivisible branch for N ≥ 5 N\ge5 N ≥ 5 . At N = 4 N=4 N = 4 , S 4 = ln [ 4 / ( 2 + 2 ) ] > 0 S_4=\ln[4/(2+\sqrt{2})]>0 S 4 = ln [ 4/ ( 2 + 2 )] > 0 .
For 3 ∤ N 3\nmid N 3 ∤ N , every point of the unshifted transverse grid is at distance at least π / ( 3 N ) \pi/(3N) π / ( 3 N ) from a kink in the folded t t t variable. Monotonicity of ln ( 2 cos t ) \ln(2\cos t) ln ( 2 cos t ) on each side of the kink implies
δ N ≥ κ 0 N , κ 0 = π 3 tan ( 2 π / 9 ) . (61) \delta_N\ge\frac{\kappa_0}{N},\qquad\kappa_0=\frac{\pi}{3}\tan(2\pi/9). \tag*{(61)} δ N ≥ N κ 0 , κ 0 = 3 π tan ( 2 π /9 ) . (61)
The elementary bounds κ 0 − 3 π / 12 > 0.4 \kappa_0-\sqrt{3}\pi/12>0.4 κ 0 − 3 π /12 > 0.4 and C E < 2.2 C_E<2.2 C E < 2.2 show δ N > S N \delta_N>S_N δ N > S N for N ≥ 6 N\ge6 N ≥ 6 . The remaining cases are exact: δ 4 = ln 2 > S 4 \delta_4=\ln\sqrt{2}>S_4 δ 4 = ln 2 > S 4 and δ 5 = ln ϕ > ln ( 2 ϕ 2 / 5 ) = S 5 > 0 \delta_5=\ln\phi>\ln(2\phi^2/5)=S_5>0 δ 5 = ln ϕ > ln ( 2 ϕ 2 /5 ) = S 5 > 0 , with ϕ = ( 1 + 5 ) / 2 \phi=(1+\sqrt{5})/2 ϕ = ( 1 + 5 ) /2 . This closes the cancellation estimate used in Theorem 4 without assuming numerical separation of the rates.
The half-shifted branch in Eq. (34) agrees with the zero-temperature cylinder solution of Ref. [10 ] . Set all three couplings in its convention to − K -K − K and take K → ∞ K\to\infty K → ∞ . The logarithm’s argument is
F K ( k 1 , k 2 ) = cosh 3 ( 2 K ) − sinh 3 ( 2 K ) + sinh ( 2 K ) [ cos k 1 + cos k 2 + cos ( k 1 + k 2 ) ] = 1 4 e 2 K ∣ 1 + e i k 1 + e − i k 2 ∣ 2 + O ( e − 2 K ) . (62) \begin{aligned}
F_K(k_1,k_2)=\cosh^3(2K)-\sinh^3(2K)
+\sinh(2K)[\cos k_1+\cos k_2+\cos(k_1+k_2)] \\
&=\frac{1}{4}e^{2K}\lvert1+e^{ik_1}+e^{-ik_2}\rvert^2+O(e^{-2K}). \tag*{(62)}
\end{aligned} F K ( k 1 , k 2 ) = cosh 3 ( 2 K ) − sinh 3 ( 2 K ) + sinh ( 2 K ) [ cos k 1 + cos k 2 + cos ( k 1 + k 2 )] = 4 1 e 2 K ∣ 1 + e i k 1 + e − i k 2 ∣ 2 + O ( e − 2 K ) . (62)
Its transverse momenta are 2 π ( k + 1 / 2 ) / N 2\pi(k+1/2)/N 2 π ( k + 1/2 ) / N . This grid has no spectral zero; Jensen’s formula then gives the residual entropy per site Λ − / N \Lambda_-/N Λ − / N , in agreement with s s ( N ) s_s(N) s s ( N ) in Eq. (34) . Thus the strip rate is an explicit specialization of that published formula. The determinant calculation in Sec. IV tracks how the periodic compatibility subtraction produces this rate.
Endpoint expansion and its bound
The bulk value also has the form I = 3 3 L ( 2 , χ 3 ) / ( 4 π ) I=3\sqrt{3}L(2,\chi_3)/(4\pi) I = 3 3 L ( 2 , χ 3 ) / ( 4 π ) , with L ( 2 , χ 3 ) = ∑ r ≥ 1 χ 3 ( r ) / r 2 L(2,\chi_3)=\sum_{r\ge1}\chi_3(r)/r^2 L ( 2 , χ 3 ) = ∑ r ≥ 1 χ 3 ( r ) / r 2 . Here χ 3 ( r ) \chi_3(r) χ 3 ( r ) is zero, 1 1 1 , or − 1 -1 − 1 on residues 0 0 0 , 1 1 1 , or 2 2 2 modulo three, respectively. Put a = π / 3 a=\pi/3 a = π /3 , φ ( t ) = ln ( 2 cos t ) \varphi(t)=\ln(2\cos t) φ ( t ) = ln ( 2 cos t ) , and A j = φ ( j ) ( a ) A_j=\varphi^{(j)}(a) A j = φ ( j ) ( a ) . Define
Z n = ( − 1 ) n + 1 2 2 n − 1 B 2 n ( 2 n ) ! , r k = ( − 1 ) k + 1 2 2 k + 1 B 2 k + 1 ( 1 / 3 ) ( 2 k + 1 ) ! , Z_n=\frac{(-1)^{n+1}2^{2n-1}B_{2n}}{(2n)!},
\qquad
r_k=\frac{(-1)^{k+1}2^{2k+1}B_{2k+1}(1/3)}{(2k+1)!}, Z n = ( 2 n )! ( − 1 ) n + 1 2 2 n − 1 B 2 n , r k = ( 2 k + 1 )! ( − 1 ) k + 1 2 2 k + 1 B 2 k + 1 ( 1/3 ) ,
b n ( N ) = { 1 , 3 ∣ N , ( 3 1 − 2 n − 1 ) / 2 , 3 ∤ N , (63) b_n(N)=
\begin{cases}
1,&3\mid N,\\
(3^{1-2n}-1)/2,&3\nmid N,
\end{cases} \tag*{(63)} b n ( N ) = { 1 , ( 3 1 − 2 n − 1 ) /2 , 3 ∣ N , 3 ∤ N , (63)
where B n ( x ) B_n(x) B n ( x ) are Bernoulli polynomials and B n = B n ( 0 ) B_n=B_n(0) B n = B n ( 0 ) . For k ≥ 0 k\ge0 k ≥ 0 define
c 2 k + 2 ( N ) = ( − 1 ) k A 2 k + 1 b k + 1 ( N ) Z k + 1 2 2 k π 2 k + 1 c_{2k+2}(N)=\frac{(-1)^kA_{2k+1}b_{k+1}(N)Z_{k+1}}{2^{2k}}\pi^{2k+1} c 2 k + 2 ( N ) = 2 2 k ( − 1 ) k A 2 k + 1 b k + 1 ( N ) Z k + 1 π 2 k + 1
c 2 k + 1 ( N ) = χ 3 ( N ) ( − 1 ) k A 2 k r k 2 2 k π 2 k ( k ≥ 1 ) . (64) c_{2k+1}(N)=\frac{\chi_3(N)(-1)^kA_{2k}r_k}{2^{2k}}\pi^{2k}\qquad(k\ge1). \tag*{(64)} c 2 k + 1 ( N ) = 2 2 k χ 3 ( N ) ( − 1 ) k A 2 k r k π 2 k ( k ≥ 1 ) . (64)
Theorem 5 (Expansion with a remainder). For every integer K ≥ 1 K\ge1 K ≥ 1 and N ≥ 3 N\ge3 N ≥ 3 ,
Λ + ( N ) N − I = ∑ j = 2 2 K c j ( N ) N j + R K ( N ) , (65) \frac{\Lambda_+(N)}{N}-I=\sum_{j=2}^{2K}\frac{c_j(N)}{N^j}+R_K(N), \tag*{(65)} N Λ + ( N ) − I = j = 2 ∑ 2 K N j c j ( N ) + R K ( N ) , (65) ∣ R K ( N ) ∣ ≤ 4 3 ∣ A 2 K ∣ ζ ( 2 K ) ( 2 N ) 2 K . (66) \lvert R_K(N)\rvert\le\frac{4}{3}\lvert A_{2K}\rvert\frac{\zeta(2K)}{(2N)^{2K}}. \tag*{(66)} ∣ R K ( N )∣ ≤ 3 4 ∣ A 2 K ∣ ( 2 N ) 2 K ζ ( 2 K ) . (66)
For l ≥ 1 l\ge1 l ≥ 1 repeated integration by parts in φ ^ ( l ) = ( 2 / π ) ∫ 0 a φ ( t ) cos ( 2 l t ) d t \widehat{\varphi}(l)=(2/\pi)\int_0^a\varphi(t)\cos(2lt)\,\mathrm{d}t φ ( l ) = ( 2/ π ) ∫ 0 a φ ( t ) cos ( 2 l t ) d t gives
φ ^ ( l ) = 2 π [ ∑ k = 0 K − 1 ( − 1 ) k A 2 k + 1 cos ( 2 l a ) ( 2 l ) 2 k + 2 + ∑ k = 1 K − 1 ( − 1 ) k A 2 k sin ( 2 l a ) ( 2 l ) 2 k + 1 ] + ρ K , l , (67) \widehat{\varphi}(l)=\frac{2}{\pi}\left[\sum_{k=0}^{K-1}\frac{(-1)^kA_{2k+1}\cos(2la)}{(2l)^{2k+2}}+\sum_{k=1}^{K-1}\frac{(-1)^kA_{2k}\sin(2la)}{(2l)^{2k+1}}\right]+\rho_{K,l}, \tag*{(67)} φ ( l ) = π 2 [ k = 0 ∑ K − 1 ( 2 l ) 2 k + 2 ( − 1 ) k A 2 k + 1 cos ( 2 l a ) + k = 1 ∑ K − 1 ( 2 l ) 2 k + 1 ( − 1 ) k A 2 k sin ( 2 l a ) ] + ρ K , l , (67)
with
∣ ρ K , l ∣ ≤ 2 3 ∣ A 2 K ∣ ( 2 l ) 2 K . (68) \lvert\rho_{K,l}\rvert\le\frac{2}{3}\frac{\lvert A_{2K}\rvert}{(2l)^{2K}}. \tag*{(68)} ∣ ρ K , l ∣ ≤ 3 2 ( 2 l ) 2 K ∣ A 2 K ∣ . (68)
Indeed, φ ( a ) = 0 \varphi(a)=0 φ ( a ) = 0 and all odd derivatives vanish at zero. If P 0 ( u ) = u P_0(u)=u P 0 ( u ) = u and P j + 1 ( u ) = ( 1 + u 2 ) P j ′ ( u ) P_{j+1}(u)=(1+u^2)P'_j(u) P j + 1 ( u ) = ( 1 + u 2 ) P j ′ ( u ) , then φ ( j ) ( t ) = − P j − 1 ( tan t ) \varphi^{(j)}(t)=-P_{j-1}(\tan t) φ ( j ) ( t ) = − P j − 1 ( tan t ) . These polynomials have nonnegative integer coefficients, so the derivative envelope on [ 0 , a ] [0,a] [ 0 , a ] is attained at a a a and equals ∣ A j ∣ \lvert A_j\rvert ∣ A j ∣ . The remaining integral is bounded by its length times this envelope. Also A 2 k ∈ Z A_{2k}\in\mathbb{Z} A 2 k ∈ Z and A 2 k + 1 ∈ 3 Z A_{2k+1}\in\sqrt{3}\mathbb{Z} A 2 k + 1 ∈ 3 Z .
The Fourier coefficients are absolutely summable. Poisson sampling therefore gives Λ + ( N ) / N − I = 2 ∑ q ≥ 1 φ ^ ( q N ) \Lambda_+(N)/N-I=2\sum_{q\ge1}\widehat{\varphi}(qN) Λ + ( N ) / N − I = 2 ∑ q ≥ 1 φ ( q N ) exactly. The cosine channel is ∑ q ≥ 1 cos ( 2 π q N / 3 ) / q 2 n = b n ( N ) ζ ( 2 n ) \sum_{q\ge1}\cos(2\pi qN/3)/q^{2n}=b_n(N)\zeta(2n) ∑ q ≥ 1 cos ( 2 π q N /3 ) / q 2 n = b n ( N ) ζ ( 2 n ) . The sine channel is ∑ q ≥ 1 sin ( 2 π q N / 3 ) / q 2 k + 1 = 3 χ 3 ( N ) L ( 2 k + 1 , χ 3 ) / 2 \sum_{q\ge1}\sin(2\pi qN/3)/q^{2k+1}=\sqrt{3}\chi_3(N)L(2k+1,\chi_3)/2 ∑ q ≥ 1 sin ( 2 π q N /3 ) / q 2 k + 1 = 3 χ 3 ( N ) L ( 2 k + 1 , χ 3 ) /2 . Using ζ ( 2 n ) = Z n π 2 n \zeta(2n)=Z_n\pi^{2n} ζ ( 2 n ) = Z n π 2 n and L ( 2 k + 1 , χ 3 ) = r k π 2 k + 1 / 3 L(2k+1,\chi_3)=r_k\pi^{2k+1}/\sqrt{3} L ( 2 k + 1 , χ 3 ) = r k π 2 k + 1 / 3 proves Eq. (64) . Summing 2 ∣ ρ K , q N ∣ 2\lvert\rho_{K,qN}\rvert 2 ∣ ρ K , q N ∣ proves Eq. (66) . For the periodic-spline rate, Eq. (40) gives coefficient 2 1 − j c j ( 2 N ) − c j ( N ) 2^{1-j}c_j(2N)-c_j(N) 2 1 − j c j ( 2 N ) − c j ( N ) and remainder bounded by 2 ∣ R K ( 2 N ) ∣ + ∣ R K ( N ) ∣ 2\lvert R_K(2N)\rvert+\lvert R_K(N)\rvert 2 ∣ R K ( 2 N )∣ + ∣ R K ( N )∣ . The bound in Theorem 5 has the same power as the last retained even term. To obtain an asymptotic remainder, as prescribed order, use a larger K K K and bound the extra terms; K = 3 K=3 K = 3 suffices for Eq. (41) . On 3 ∣ N 3\mid N 3 ∣ N the sine channel vanishes, so all odd-power coefficients vanish. This is a statement about the expansion, not an assertion that the finite-width remainder is an even function of 1 / N 1/N 1/ N .
The remainder theorem provides an explicit finite-N N N error bar for the truncation in Eq. (41) . Let γ N [ 4 ] \gamma_N^{[4]} γ N [ 4 ] denote its three retained terms and set B 3 ( N ) = 4 3 ∣ A 6 ∣ ζ ( 6 ) / ( 2 N ) 6 B_3(N)=\frac{4}{3}\lvert A_6\rvert\zeta(6)/(2N)^6 B 3 ( N ) = 3 4 ∣ A 6 ∣ ζ ( 6 ) / ( 2 N ) 6 . On 3 ∤ N 3\nmid N 3 ∤ N ,
∣ γ N − γ N [ 4 ] ∣ ≤ 33 ∣ c 5 ( N ) ∣ 16 N 5 + 63 ∣ c 6 ( N ) ∣ 32 N 6 + 65 32 B 3 ( N ) . (69) \lvert\gamma_N-\gamma_N^{[4]}\rvert\le\frac{33\lvert c_5(N)\rvert}{16N^5}+\frac{63\lvert c_6(N)\rvert}{32N^6}+\frac{65}{32}B_3(N). \tag*{(69)} ∣ γ N − γ N [ 4 ] ∣ ≤ 16 N 5 33 ∣ c 5 ( N )∣ + 32 N 6 63 ∣ c 6 ( N )∣ + 32 65 B 3 ( N ) . (69)
Indeed, γ N = 2 [ Λ + ( N ) / N − I ] − 2 [ Λ + ( 2 N ) / ( 2 N ) − I ] \gamma_N=2[\Lambda_+(N)/N-I]-2[\Lambda_+(2N)/(2N)-I] γ N = 2 [ Λ + ( N ) / N − I ] − 2 [ Λ + ( 2 N ) / ( 2 N ) − I ] ; c 5 ( 2 N ) = − c 5 ( N ) c_5(2N)=-c_5(N) c 5 ( 2 N ) = − c 5 ( N ) , c 6 ( 2 N ) = c 6 ( N ) c_6(2N)=c_6(N) c 6 ( 2 N ) = c 6 ( N ) , and B 3 ( 2 N ) = B_3(2N)= B 3 ( 2 N ) = B 3 ( N ) / 64 B_{3}(N)/64 B 3 ( N ) /64 . (69) follows by retaining the next two coefficients in Theorem 5. Here c 5 = − 20 χ 3 ( N ) π 4 / 729 c_{5}=-20\chi_{3}(N)\pi^{4}/729 c 5 = − 20 χ 3 ( N ) π 4 /729 and c 6 = 2662 3 π 5 / 229635 c_{6}=2662\sqrt{3}\pi^{5}/229635 c 6 = 2662 3 π 5 /229635 .
For Theorem 2, we first compare each nonzero determinant with its leading transverse-grid sum, then compare that sum with the bulk entropy I I I . These two errors give the inverse-area and inverse-width-squared terms in the bound. The distance multiplicities in Table IV control the first error uniformly in the aspect ratio.
Table IV. Allowed positive scaled angular distances q k = 3 m d k / ( 2 π ) q_k=3md_k/(2\pi) q k = 3 m d k / ( 2 π ) from the transverse grid to 2 π / 3 2\pi/3 2 π /3 and 4 π / 3 4\pi/3 4 π /3 , where d k d_k d k is the shorter circular distance to these two angles. The grid shift is α = 0 \alpha=0 α = 0 or 1 / 2 1/2 1/2 . Each listed family, with j ≥ 0 j\geq0 j ≥ 0 , has the indicated maximum multiplicity.
Proof. By symmetry, evaluate the length-n n n product on the transverse size-m m m grid. Write ψ ( θ ) = max { ln ∣ 1 + e i θ ∣ , 0 } \psi(\theta)=\max\{\ln|1+e^{i\theta}|,0\} ψ ( θ ) = max { ln ∣1 + e i θ ∣ , 0 } and
J m ( α ) = 1 m ∑ k = 0 m − 1 ψ ( 2 π k + α m ) , α ∈ { 0 , 1 / 2 } . (70) J_m(\alpha)=\frac{1}{m}\sum_{k=0}^{m-1}\psi\left(2\pi\frac{k+\alpha}{m}\right),\qquad\alpha\in\{0,1/2\}. \tag*{(70)} J m ( α ) = m 1 k = 0 ∑ m − 1 ψ ( 2 π m k + α ) , α ∈ { 0 , 1/2 } . (70)
The distributional second derivative of ψ \psi ψ has total variation 2 3 2\sqrt{3} 2 3 : its smooth part has variation 3 \sqrt{3} 3 and its two derivative jumps contribute another 3 \sqrt{3} 3 . With Fourier convention ψ ^ ( l ) = ( 2 π ) − 1 ∫ − π π ψ ( θ ) e − i l θ d θ \widehat{\psi}(l)=(2\pi)^{-1}\int_{-\pi}^{\pi}\psi(\theta)e^{-il\theta}\,d\theta ψ ( l ) = ( 2 π ) − 1 ∫ − π π ψ ( θ ) e − i l θ d θ , its coefficients therefore obey ∣ ψ ^ ( l ) ∣ ≤ 3 / ( π l 2 ) |\widehat{\psi}(l)|\leq\sqrt{3}/(\pi l^{2}) ∣ ψ ( l ) ∣ ≤ 3 / ( π l 2 ) for l ≠ 0 l\ne0 l = 0 . Exact grid sampling gives
∣ J m ( α ) − I ∣ ≤ π 3 3 m 2 . (71) |J_m(\alpha)-I|\leq\frac{\pi\sqrt{3}}{3m^{2}}. \tag*{(71)} ∣ J m ( α ) − I ∣ ≤ 3 m 2 π 3 . (71)
Put z k = e 2 π i ( k + α ) / m z_k=e^{2\pi i(k+\alpha)/m} z k = e 2 π i ( k + α ) / m . A mode z k = − 1 z_k=-1 z k = − 1 has a factor of unit magnitude in Eq. (8) and contributes zero logarithmic error; we omit it from the following estimate. For the other modes of a nonzero determinant let λ k = ln ∣ 1 + z k ∣ \lambda_k=\ln|1+z_k| λ k = ln ∣1 + z k ∣ . After its leading magnitude is removed, a factor with ∣ 1 + z k ∣ ≠ 1 |1+z_k|\ne1 ∣1 + z k ∣ = 1 in Eq. (8) has logarithmic error e k e_k e k bounded by
ln ( 1 − u k ) ≤ e k ≤ ln ( 1 + u k ) , u k = e − n ∣ λ k ∣ . (72) \ln(1-u_k)\leq e_k\leq\ln(1+u_k),\qquad u_k=e^{-n|\lambda_k|}. \tag*{(72)} ln ( 1 − u k ) ≤ e k ≤ ln ( 1 + u k ) , u k = e − n ∣ λ k ∣ . (72)
A mode with ∣ 1 + z k ∣ = 1 |1+z_k|=1 ∣1 + z k ∣ = 1 occurs only on the unshifted grid with 3 ∣ m 3\mid m 3 ∣ m . For a nonzero determinant, the corresponding product factor has magnitude 1 1 1 , 3 \sqrt{3} 3 , or 2 2 2 . Its logarithmic error therefore lies between 0 and ln 2 \ln2 ln 2 ; there are at most two such factors.
Let d k d_k d k be circular distance from the angle of z k z_k z k to { 2 π / 3 , 4 π / 3 } \{2\pi/3,4\pi/3\} { 2 π /3 , 4 π /3 } . Concavity of ln ( 2 cos ( θ / 2 ) ) \ln(2\cos(\theta/2)) ln ( 2 cos ( θ /2 )) on the inner arc, and its slope on the outer arcs, give
∣ λ k ∣ ≥ 3 ln 2 2 π d k . (73) |\lambda_k|\geq\frac{3\ln2}{2\pi}d_k. \tag*{(73)} ∣ λ k ∣ ≥ 2 π 3 ln 2 d k . (73)
If q k = 3 m d k / ( 2 π ) q_k=3md_k/(2\pi) q k = 3 m d k / ( 2 π ) , then u k ≤ 2 − q k u_k\leq2^{-q_k} u k ≤ 2 − q k since n ≥ m n\geq m n ≥ m . The possible distances form the ladders in Table IV . To verify their multiplicities, a point at distance q q q from the marks m , 2 m m,2m m , 2 m modulo 3 m 3m 3 m must be one of m ± q , 2 m ± q m\pm q,2m\pm q m ± q , 2 m ± q ; retaining points in 3 α + 3 Z 3\alpha+3\mathbb{Z} 3 α + 3 Z gives the table, and coincidences decrease the multiplicities.
Sum the upper and lower errors separately. For the off-unit factors define P = ∑ ln ( 1 + 2 − q k ) P=\sum\ln(1+2^{-q_k}) P = ∑ ln ( 1 + 2 − q k ) and Q = ∑ − ln ( 1 − 2 − q k ) Q=\sum-\ln(1-2^{-q_k}) Q = ∑ − ln ( 1 − 2 − q k ) , so P ≤ Q P\leq Q P ≤ Q . The largest ladder bound is Q ≤ C ∗ Q\leq C_* Q ≤ C ∗ on the half-shifted, nondivisible branch. The other branches follow from geometric sums and − ln ( 1 − u ) ≤ u / ( 1 − u ) -\ln(1-u)\leq u/(1-u) − ln ( 1 − u ) ≤ u / ( 1 − u ) : their bounds are respectively 64 / 21 < 3.1 64/21<3.1 64/21 < 3.1 , max { 32 / 49 , 4 / 7 + 2 ln 2 } < 2 \max\{32/49,4/7+2\ln2\}<2 max { 32/49 , 4/7 + 2 ln 2 } < 2 , and 32 u / [ 7 ( 1 − u ) ] < 2.51 32u/[7(1-u)]<2.51 32 u / [ 7 ( 1 − u )] < 2.51 , where u = 1 / ( 2 2 ) u=1/(2\sqrt{2}) u = 1/ ( 2 2 ) in the last branch. The middle bound includes both unit factors. All are smaller than C ∗ C_* C ∗ . Consequently every nonzero determinant satisfies
∣ ln ∣ Δ ∣ m n − J m ( α ) ∣ ≤ C ∗ m n . (74) \left|\frac{\ln|\Delta|}{mn}-J_m(\alpha)\right|\leq\frac{C_*}{mn}. \tag*{(74)} mn ln ∣Δ∣ − J m ( α ) ≤ mn C ∗ . (74)
If R max R_{\max} R m a x is the largest normalized logarithm of the nonzero determinants, Eq. (9) gives e m n R max / 2 ≤ D T ≤ 2 e m n R max e^{mnR_{\max}/2}\leq D_T\leq2e^{mnR_{\max}} e mn R m a x /2 ≤ D T ≤ 2 e mn R m a x . Combining this with Eq. (71) proves Eq. (27) . Finally 3.1 < C ∗ < 3.106422042 3.1<C_*<3.106422042 3.1 < C ∗ < 3.106422042 and π 3 / 3 + C ∗ + ln 2 < 5.613369 \pi\sqrt{3}/3+C_*+\ln2<5.613369 π 3 /3 + C ∗ + ln 2 < 5.613369 . These inequalities can be certified by the positive series − ln ( 1 − u ) = ∑ r ≥ 1 u r / r -\ln(1-u)=\sum_{r\geq1}u^r/r − ln ( 1 − u ) = ∑ r ≥ 1 u r / r , whose tail after L L L terms is at most u L + 1 / [ ( L + 1 ) ( 1 − u ) ] u^{L+1}/[(L+1)(1-u)] u L + 1 / [( L + 1 ) ( 1 − u )] , and geometric bounds on the remaining ladder levels. The rational certificate supplied in the verification archive [19 ] implements these enclosures. ■
To prove Theorem 3, we isolate a positive reference term in the determinant expansion and bound all other terms relative to it. This gives a two-sided comparison for the spin count despite possible cancellations. Analytic estimates cover large widths, and interval arithmetic covers the remaining finite range.
Proof. For τ = ± 1 \tau=\pm1 τ = ± 1 let
E τ ( n ) = ∑ S ⊂ { z : z m = τ } ∣ S ∣ e v e n [ ∏ z ∈ S ( 1 + z ) ] n . (75) E_{\tau}(n)=\sum_{\substack{S\subset\{z:z^{m}=\tau\}\\ |S|\ \mathrm{even}}}\left[\prod_{z\in S}(1+z)\right]^{n}. \tag*{(75)} E τ ( n ) = S ⊂ { z : z m = τ } ∣ S ∣ even ∑ [ z ∈ S ∏ ( 1 + z ) ] n . (75)
Expanding Eq. (8) and applying Eq. (A2) gives
G 0 ( n , m ) = E − ( n ) + ( − 1 ) n E + ( n ) . (76) G_{0}(n,m)=E_{-}(n)+(-1)^{n}E_{+}(n). \tag*{(76)} G 0 ( n , m ) = E − ( n ) + ( − 1 ) n E + ( n ) . (76)
Indeed, ∑ σ P m Δ ( σ , τ ) = 2 ( − 1 ) ( n + 1 ) m E − ( n ) \sum_{\sigma}\mathcal{P}^{m}\Delta(\sigma,\tau)=2(-1)^{(n+1)m}E_{-}(n) ∑ σ P m Δ ( σ , τ ) = 2 ( − 1 ) ( n + 1 ) m E − ( n ) and Q ( n , m ) = ( − 1 ) n m + n + m Q(n,m)=(-1)^{nm+n+m} Q ( n , m ) = ( − 1 ) nm + n + m . These signs include both parities of n n n and m m m .
On each grid, take as reference the subset R τ = { z : z m = τ , ∣ 1 + z ∣ > 1 } R_{\tau}=\{z:z^{m}=\tau,\lvert1+z\rvert>1\} R τ = { z : z m = τ , ∣ 1 + z ∣ > 1 } . The half-grid reference consists of conjugate pairs, has even cardinality, and contributes the positive term L n , m = e n Λ − ( m ) \mathcal{L}_{n,m}=e^{n\Lambda_{-}(m)} L n , m = e n Λ − ( m ) to E − E_{-} E − . The full-grid reference has odd cardinality, because it also contains z = 1 z=1 z = 1 . Define
u z ( n ) = e − n ∣ ln ∣ 1 + z ∣ ∣ , (77) u_z(n)=e^{-n\lvert\ln\lvert1+z\rvert\rvert}, \tag*{(77)} u z ( n ) = e − n ∣ l n ∣ 1 + z ∣∣ , (77)
with u = 0 u=0 u = 0 at z = − 1 z=-1 z = − 1 and u = 1 u=1 u = 1 when ∣ 1 + z ∣ = 1 \lvert1+z\rvert=1 ∣ 1 + z ∣ = 1 . Write u + u_{+} u + and u − u_{-} u − for the lists on the grids z m = 1 z^{m}=1 z m = 1 and z m = − 1 z^{m}=-1 z m = − 1 , respectively. For either list u u u set
P e , o ( u ) = 1 2 [ ∏ z ( 1 + u z ) ± ∏ z ( 1 − u z ) ] , (78) P_{\mathrm{e,o}}(u)=\frac{1}{2}\left[\prod_z(1+u_z)\pm\prod_z(1-u_z)\right], \tag*{(78)} P e , o ( u ) = 2 1 [ z ∏ ( 1 + u z ) ± z ∏ ( 1 − u z ) ] , (78)
where the upper sign selects even subsets. Taking symmetric differences with R τ R_{\tau} R τ gives
∣ E − / L n , m − 1 ∣ ≤ P e ( u − ) − 1 , ∣ E + ∣ / L n , m ≤ e n S m P o ( u + ) . (79) \begin{aligned}
\lvert E_{-}/\mathcal{L}_{n,m}-1\rvert&\le P_{\mathrm{e}}(u_{-})-1,\\
\lvert E_{+}\rvert/\mathcal{L}_{n,m}&\le e^{nS_m}P_{\mathrm{o}}(u_{+}). \tag*{(79)}
\end{aligned} ∣ E − / L n , m − 1 ∣ ∣ E + ∣ / L n , m ≤ P e ( u − ) − 1 , ≤ e n S m P o ( u + ) . (79)
Here S m = Λ + ( m ) − Λ − ( m ) S_m=\Lambda_{+}(m)-\Lambda_{-}(m) S m = Λ + ( m ) − Λ − ( m ) . Thus, for F m ( n ) = P e ( u − ) + e n S m P o ( u + ) F_m(n)=P_{\mathrm{e}}(u_{-})+e^{nS_m}P_{\mathrm{o}}(u_{+}) F m ( n ) = P e ( u − ) + e n S m P o ( u + ) ,
2 − F m ( n ) ≤ G 0 ( n , m ) / L n , m ≤ F m ( n ) . (80) 2-F_m(n)\le G_0(n,m)/\mathcal{L}_{n,m}\le F_m(n). \tag*{(80)} 2 − F m ( n ) ≤ G 0 ( n , m ) / L n , m ≤ F m ( n ) . (80)
Every nonconstant half-grid term decreases with n n n . Each weighted full-grid odd term has exponent n [ S m − ∑ z ∈ X ∣ ln ∣ 1 + z ∣ ∣ ] n[S_m-\sum_{z\in X}\lvert\ln\lvert1+z\rvert\rvert] n [ S m − ∑ z ∈ X ∣ ln ∣ 1 + z ∣∣] . If 3 ∤ m 3\nmid m 3 ∤ m , then S m < 0 S_m<0 S m < 0 ; otherwise any nonzero odd term has excitation energy at least δ m > S m \delta_m>S_m δ m > S m , as proved in Appendix B1. Hence F m ( n ) F_m(n) F m ( n ) is nonincreasing in n n n , including the unit-mode cases. Appendix B4a proves F m ( m ) < 39 / 20 F_m(m)<39/20 F m ( m ) < 39/20 for m ≥ 1000 m\ge1000 m ≥ 1000 . For 3 ≤ m ≤ 999 3\le m\le999 3 ≤ m ≤ 999 , the integer-interval certificate proves 2 − F m ( m ) > 1 / 25 2-F_m(m)>1/25 2 − F m ( m ) > 1/25 , except at widths 4 and 7. It proves instead 2 − F 4 ( 6 ) > 1 / 25 2-F_4(6)>1/25 2 − F 4 ( 6 ) > 1/25 , 2 − F 7 ( 8 ) > 1 / 25 2-F_7(8)>1/25 2 − F 7 ( 8 ) > 1/25 , and F 4 ( 4 ) , F 7 ( 7 ) < 3 F_4(4),F_7(7)<3 F 4 ( 4 ) , F 7 ( 7 ) < 3 . Monotonicity then leaves only the three lower-bound cases ( n , m ) = ( 4 , 4 ) , ( 5 , 4 ) , ( 7 , 7 ) (n,m)=(4,4),(5,4),(7,7) ( n , m ) = ( 4 , 4 ) , ( 5 , 4 ) , ( 7 , 7 ) . Their exact counts are 42, 712, and 2620800. Since e Λ − ( 4 ) = 2 + 2 < 4 e^{\Lambda_{-}(4)}=2+\sqrt{2}<4 e Λ − ( 4 ) = 2 + 2 < 4 and e Λ − ( 7 ) = 16 cos 2 ( π / 14 ) cos 2 ( 3 π / 14 ) < 12 e^{\Lambda_{-}(7)}=16\cos^2(\pi/14)\cos^2(3\pi/14)<12 e Λ − ( 7 ) = 16 cos 2 ( π /14 ) cos 2 ( 3 π /14 ) < 12 , the integer inequalities 25 G 0 ( n , m ) > 4 n 25G_0(n,m)>4^n 25 G 0 ( n , m ) > 4 n or 25 G 0 ( n , m ) > 12 n 25G_0(n,m)>12^n 25 G 0 ( n , m ) > 1 2 n settle these cases. (29) follows. Finally ∣ Λ − ( m ) / m − I ∣ ≤ π 3 / ( 3 m 2 ) \lvert\Lambda_{-}(m)/m-I\rvert\le\pi\sqrt{3}/(3m^2) ∣ Λ − ( m ) / m − I ∣ ≤ π 3 / ( 3 m 2 ) by Eq. (71) . Taking logarithms proves Eqs. (30) and (31) . The rounded constant is enclosed by a positive series for ln 25 \ln25 ln 25 , Machin’s formula for π \pi π , and a rational bound on 3 \sqrt{3} 3 . ■
Since F m ( n ) F_m(n) F m ( n ) is nonincreasing, it suffices to bound it at n = m n=m n = m , apart from the explicitly handled short lengths at widths 4 and 7.
Large widths
Fold transverse coordinates to x = ( k + α ) / m ∈ [ − 1 / 2 , 1 / 2 ) x=(k+\alpha)/m\in[-1/2,1/2) x = ( k + α ) / m ∈ [ − 1/2 , 1/2 ) and put ϕ ( x ) = ln ( 2 cos π x ) \phi(x)=\ln(2\cos\pi x) ϕ ( x ) = ln ( 2 cos π x ) . Its zeros are at ± 1 / 3 \pm1/3 ± 1/3 . The mode z = − 1 z=-1 z = − 1 has ∣ 1 + z ∣ = 0 \lvert1+z\rvert=0 ∣ 1 + z ∣ = 0 and activity zero, so it is omitted here. If the index distance d d d to a zero is at most m / 100 m/100 m /100 , then
∣ ϕ ′ ( x ) ∣ ≥ π 3 − 4 π 2 100 > 5 , m ∣ ϕ ( x ) ∣ ≥ 5 d . (81) \lvert\phi'(x)\rvert\ge\pi\sqrt{3}-\frac{4\pi^2}{100}>5,\qquad m\lvert\phi(x)\rvert\ge5d. \tag*{(81)} ∣ ϕ ′ ( x )∣ ≥ π 3 − 100 4 π 2 > 5 , m ∣ ϕ ( x )∣ ≥ 5 d . (81)
On the inner interval this follows from sec 2 θ ≤ 4 \sec^2\theta\le4 sec 2 θ ≤ 4 up to θ = π / 3 \theta=\pi/3 θ = π /3 ; on the outer interval the tangent is larger. Outside these neighborhoods monotonicity gives ∣ ϕ ∣ ≥ 1 / 20 \lvert\phi\rvert\ge1/20 ∣ ϕ ∣ ≥ 1/20 . For m ≥ 1000 m\ge1000 m ≥ 1000 , the sum of all far-mode activities on either grid is therefore bounded by
m e − m / 20 ≤ 1000 e − 50 < 10 − 6 . (82) me^{-m/20}\le1000e^{-50}<10^{-6}. \tag*{(82)} m e − m /20 ≤ 1000 e − 50 < 1 0 − 6 . (82)
The estimate for S m S_m S m in Appendix B1 gives m S m < 0.456 mS_m<0.456 m S m < 0.456 if 3 ∤ m 3\nmid m 3 ∤ m , and m S m < − 1.35 mS_m<-1.35 m S m < − 1.35 if 3 ∣ m 3\mid m 3 ∣ m .
On the nondivisible branch, half-grid distances from the two zeros lie in j + 1 / 6 , j + 5 / 6 j+1/6,j+5/6 j + 1/6 , j + 5/6 , each with multiplicity at most two. Full-grid distances lie in j + 1 / 3 , j + 2 / 3 j+1/3,j+2/3 j + 1/3 , j + 2/3 , also with multiplicity at most two. Isolate two nearest modes on each grid. Their activities and the sums of the remaining activities obey
a < 0.435 , V − ≤ 2 ( e − 25 / 6 + e − 35 / 6 ) 1 − e − 5 + 10 − 6 < 0.038 , a<0.435,\qquad V_{-}\le\frac{2(e^{-25/6}+e^{-35/6})}{1-e^{-5}}+10^{-6}<0.038, a < 0.435 , V − ≤ 1 − e − 5 2 ( e − 25/6 + e − 35/6 ) + 1 0 − 6 < 0.038 ,
b < 0.189 , V + ≤ 2 ( e − 10 / 3 + e − 20 / 3 ) 1 − e − 5 + 10 − 6 < 0.075. (83) b<0.189,\qquad V_{+}\le\frac{2(e^{-10/3}+e^{-20/3})}{1-e^{-5}}+10^{-6}<0.075. \tag*{(83)} b < 0.189 , V + ≤ 1 − e − 5 2 ( e − 10/3 + e − 20/3 ) + 1 0 − 6 < 0.075. (83)
Extending near-mode ladders to infinity and adding the far-mode sum overcounts rather than omits modes. For a list with sum V V V , its elementary symmetric coefficient of order j j j is at most V j / j ! V^j/j! V j / j ! . The remaining even and odd partitions are consequently bounded by cosh V \cosh V cosh V and sinh V \sinh V sinh V . Hence
P e ( u − ) − 1 ≤ ( 1 + a 2 ) cosh V − + 2 a sinh V − − 1 < 0.224 , P_{\mathrm{e}}(u_{-})-1\le(1+a^2)\cosh V_{-}+2a\sinh V_{-}-1<0.224, P e ( u − ) − 1 ≤ ( 1 + a 2 ) cosh V − + 2 a sinh V − − 1 < 0.224 ,
P o ( u + ) ≤ ( 1 + b 2 ) sinh V + + 2 b cosh V + < 0.457. (84) P_{\mathrm{o}}(u_{+})\le(1+b^2)\sinh V_{+}+2b\cosh V_{+}<0.457. \tag*{(84)} P o ( u + ) ≤ ( 1 + b 2 ) sinh V + + 2 b cosh V + < 0.457. (84)
Since e m S m < 1.58 e^{mS_m}<1.58 e m S m < 1.58 ,
F m ( m ) − 1 < 0.224 + 1.58 ( 0.457 ) = 0.94606 < 19 20 . (85) F_m(m)-1<0.224+1.58(0.457)=0.94606<\frac{19}{20}. \tag*{(85)} F m ( m ) − 1 < 0.224 + 1.58 ( 0.457 ) = 0.94606 < 20 19 . (85)
If 3 ∣ m 3\mid m 3 ∣ m , the half-grid distances are j + 1 / 2 j+1/2 j + 1/2 , with multiplicity at most four. The full-grid nonunit distances are positive integers, also with multiplicity at most four. Their activity sums satisfy
U − ≤ 4 e − 5 / 2 1 − e − 5 + 10 − 6 < 0.335 , U_{-}\le\frac{4e^{-5/2}}{1-e^{-5}}+10^{-6}<0.335, U − ≤ 1 − e − 5 4 e − 5/2 + 1 0 − 6 < 0.335 ,
U + ≤ 4 e − 5 1 − e − 5 + 10 − 6 < 0.028. (86) U_{+}\le\frac{4e^{-5}}{1-e^{-5}}+10^{-6}<0.028. \tag*{(86)} U + ≤ 1 − e − 5 4 e − 5 + 1 0 − 6 < 0.028. (86)
The two full-grid unit modes have activity one, so P o ( u + ) = 2 ∏ n o n u n i t ( 1 + u ) P_{\mathrm{o}}(u_{+})=2\prod_{\mathrm{nonunit}}(1+u) P o ( u + ) = 2 ∏ nonunit ( 1 + u ) . It follows that
P e ( u − ) − 1 < 0.060 , P_{\mathrm{e}}(u_{-})-1<0.060, P e ( u − ) − 1 < 0.060 ,
e m S m P o ( u + ) < 0.26 2 1 − 0.028 < 0.535. (87) e^{mS_m}P_{\mathrm{o}}(u_{+})<0.26\frac{2}{1-0.028}<0.535. \tag*{(87)} e m S m P o ( u + ) < 0.26 1 − 0.028 2 < 0.535. (87)
Thus F m ( m ) − 1 < 0.595 F_m(m)-1<0.595 F m ( m ) − 1 < 0.595 on this branch. Both branches give 1 / 20 < G 0 / L n , m < 2 1/20<G_{0}/L_{n,m}<2 1/20 < G 0 / L n , m < 2 for every n ≥ m ≥ 1000 n\ge m\ge1000 n ≥ m ≥ 1000 .
All decimal comparisons in this subsection are rationally certified. For x ≥ 0 x\ge0 x ≥ 0 , a positive Taylor partial sum of e x e^x e x gives a lower bound, hence an upper bound for e − x e^{-x} e − x by reciprocation. The degree-20 sums suffice for the near-mode enclosures, and degree 10 at x = 50 x=50 x = 50 suffices for the far tail. The elementary rational bounds
cosh v ≤ ( 1 − v 2 / 2 ) − 1 , 0 ≤ v < 2 , \cosh v\le(1-v^2/2)^{-1},\qquad0\le v<\sqrt{2}, cosh v ≤ ( 1 − v 2 /2 ) − 1 , 0 ≤ v < 2 ,
sinh v ≤ v ( 1 − v 2 / 6 ) − 1 , 0 ≤ v < 6 , \sinh v\le v(1-v^2/6)^{-1},\qquad0\le v<\sqrt{6}, sinh v ≤ v ( 1 − v 2 /6 ) − 1 , 0 ≤ v < 6 ,
e x ≤ 1 + x + x 2 2 ( 1 − x / 3 ) , 0 ≤ x < 3 (88) e^x\le1+x+\frac{x^2}{2(1-x/3)},\qquad0\le x<3 \tag*{(88)} e x ≤ 1 + x + 2 ( 1 − x /3 ) x 2 , 0 ≤ x < 3 (88)
follow by bounding factorials termwise on these ranges. The exact rational program supplied in the verification archive [19 ] checks these comparisons, the activity sums, and the derivative and strip-gap constants.
Finite widths
For 3 ≤ m ≤ 999 3\le m\le999 3 ≤ m ≤ 999 , the certificate represents interval endpoints as integers in units of 10 − 50 10^{-50} 1 0 − 50 . Every arithmetic operation rounds outward. Machin’s formula π = 16 arctan ( 1 / 5 ) − 4 arctan ( 1 / 239 ) \pi=16\arctan(1/5)-4\arctan(1/239) π = 16 arctan ( 1/5 ) − 4 arctan ( 1/239 ) is enclosed by alternating partial sums through indices 60 and 61. At the folded spectral arguments 0 ≤ x < π / 2 0\le x<\pi/2 0 ≤ x < π /2 , alternating cosine sums through indices 60 and 61 enclose 2 cos x 2\cos x 2 cos x . The tail terms decrease at these truncations. Exact grid indices determine whether a radius is above, equal to, or below one; no uncertain numerical classification enters the proof. Positive interval products, reciprocals, and integer powers then enclose 2 − F m ( n ) 2-F_m(n) 2 − F m ( n ) .
The program certifies 2 − F m ( m ) > 1 / 25 2-F_m(m)>1/25 2 − F m ( m ) > 1/25 at every width in this range except 4 and 7. At those widths it certifies 2 − F 4 ( 6 ) > 1 / 25 2-F_4(6)>1/25 2 − F 4 ( 6 ) > 1/25 , 2 − F 7 ( 8 ) > 1 / 25 2-F_7(8)>1/25 2 − F 7 ( 8 ) > 1/25 , and F 4 ( 4 ) , F 7 ( 7 ) < 3 F_4(4),F_7(7)<3 F 4 ( 4 ) , F 7 ( 7 ) < 3 . The smallest ordinary lower endpoint is at m = n = 10 m=n=10 m = n = 10 :
2 − F 10 ( 10 ) > 0.04600277656683344116 > 1 / 25. (89) 2-F_{10}(10)>0.04600277656683344116>1/25. \tag*{(89)} 2 − F 10 ( 10 ) > 0.04600277656683344116 > 1/25. (89)
The JSON supplied in the verification archive [19 ] lists all 997 width records and the additional upper checks. The three remaining short cases are the exact counts used in the proof above; the program verifies their integer resultants with Bareiss elimination and exact divisibility checks. Thus the finite step is an exhaustive interval certificate over a specified range, followed by a proved extension in length, rather than an extrapolation from sampled sizes.
Odd-torus walls, minimizers, and full-model threshold
Throughout this appendix, L ≥ 5 L\ge5 L ≥ 5 is odd, and a constrained minimizer minimizes B \mathcal{B} B on the nearest-neighbor ground-state manifold C L = { s : A ( s ) = − L 2 } \mathcal{C}_L=\{s:\mathcal{A}(s)=-L^2\} C L = { s : A ( s ) = − L 2 } . The wall lower bound and explicit constructions establish the minimum energy. The equality-case analysis then classifies all constrained minimizers and gives their exact count. The last two subsections compare this constrained minimum with an outside state in the full model and compute moments of the equal-weight measure on the slope triples ( H , V , D ) (H,V,D) ( H , V , D ) .
Wall lower bound and an entropy consequence
The triangle constraints can be read as an exclusion dynamics on successive diagonals. In this description, a run of moving particles displaces one hole. The following lemma bounds the number of such runs needed for periodic closure; it supplies the directional wall bound in Theorem 1.
Lemma 1. Consider d ≥ 1 d\ge1 d ≥ 1 ordered holes on a ring of integer length L L L . Each jump moves a single hole left by a positive distance less than L L L , without passing another hole. Suppose C C C jumps return the unlabeled hole positions to their initial set, with final ordered lifts z i ′ = z i − v z'_i=z_{i-v} z i ′ = z i − v for v ≥ 1 v\ge1 v ≥ 1 , where z i + d = z i + L z_{i+d}=z_i+L z i + d = z i + L . Then C ≥ d + v C\ge d+v C ≥ d + v .
Proof. Every hole moves, since z i − v < z i z_{i-v}<z_i z i − v < z i . Choose the hole whose last jump occurs first among all last jumps, and call it i i i . Write v = q d + m v=qd+m v = q d + m , 0 ≤ m < d 0\le m<d 0 ≤ m < d . At that moment its position is z i − m − q L z_{i-m}-qL z i − m − q L . The m m m preceding holes i − r i-r i − r , 1 ≤ r ≤ m 1\le r\le m 1 ≤ r ≤ m , have initial positions at least z i − m z_{i-m} z i − m and now lie strictly to its left. Each has moved more than q L qL q L and thus made at least q + 1 q+1 q + 1 jumps. For m < r < d m<r<d m < r < d , its initial position is greater than z i − m − L z_{i-m}-L z i − m − L , so it has moved more than ( q − 1 ) L (q-1)L ( q − 1 ) L and made at least q q q jumps. When q = 0 q=0 q = 0 this last bound is just zero. Hole i i i has moved at least q L qL q L , with positive displacement also when q = 0 q=0 q = 0 , and therefore has made at least q + 1 q+1 q + 1 jumps. Through this moment the total number of jumps is at least
( q + 1 ) + m ( q + 1 ) + ( d − 1 − m ) q = v + 1. (90) (q+1)+m(q+1)+(d-1-m)q=v+1. \tag*{(90)} ( q + 1 ) + m ( q + 1 ) + ( d − 1 − m ) q = v + 1. (90)
Each of the other d − 1 d-1 d − 1 holes still has its last jump to make. Hence C ≥ v + d C\ge v+d C ≥ v + d , including d = 1 d=1 d = 1 and v ≥ d v\ge d v ≥ d .
For the spin problem set t = x − y t=x-y t = x − y and u = x u=x u = x , both modulo L L L , and define
a t ( u ) = h ( u , u − t ) , a_t(u)=h(u,u-t), a t ( u ) = h ( u , u − t ) ,
b t ( u ) = v ( u + 1 , u − t ) , b_t(u)=v(u+1,u-t), b t ( u ) = v ( u + 1 , u − t ) ,
c t ( u ) = d ( u , u − t ) . (91) c_t(u)=d(u,u-t). \tag*{(91)} c t ( u ) = d ( u , u − t ) . (91)
(12) implies
a t ( u ) + b t ( u ) + c t ( u ) = 1 , a_t(u)+b_t(u)+c_t(u)=1, a t ( u ) + b t ( u ) + c t ( u ) = 1 ,
a t + 1 ( u ) + b t + 1 ( u ) = a t ( u ) + b t ( u − 1 ) . (92) a_{t+1}(u)+b_{t+1}(u)=a_t(u)+b_t(u-1). \tag*{(92)} a t + 1 ( u ) + b t + 1 ( u ) = a t ( u ) + b t ( u − 1 ) . (92)
Thus a a a denotes particles that stay and b b b particles that move one site to the right. The number of occupied sites is conserved. Averaging over t t t gives H + V H+V H + V occupied sites and D D D holes in each slice. A maximal run of b = 1 b=1 b = 1 has a hole immediately to its right: a stationary particle there would cause a collision in Eq. (92) . A run of length r r r moves that hole left by r r r . Since D > 0 D>0 D > 0 , r < L r<L r < L . Distinct runs use disjoint sites, so these jumps can be performed sequentially without passing holes.
The horizontal wall count is
W H = ∑ x , y ∣ v ( x + 1 , y ) − v ( x , y − 1 ) ∣ = ∑ t , u ∣ b t ( u ) − b t ( u − 1 ) ∣ = 2 C , (93) W_H=\sum_{x,y}\lvert v(x+1,y)-v(x,y-1)\rvert
=\sum_{t,u}\lvert b_t(u)-b_t(u-1)\rvert=2C, \tag*{(93)} W H = x , y ∑ ∣ v ( x + 1 , y ) − v ( x , y − 1 )∣ = t , u ∑ ∣ b t ( u ) − b t ( u − 1 )∣ = 2 C , (93)
where C C C is the number of runs over L L L time steps. The hole positions return to their initial set after those steps. Their total leftward distance is ∑ t , u b t ( u ) = L V \sum_{t,u}b_t(u)=LV ∑ t , u b t ( u ) = L V . If the ordered hole indices shift by k k k , their total displacement is ∑ i = 1 D ( z i − z i − k ) = k L \sum_{i=1}^{D}(z_i-z_{i-k})=kL ∑ i = 1 D ( z i − z i − k ) = k L , so k = V k=V k = V . Lemma 1 with d = D d=D d = D , v = V v=V v = V now gives W H ≥ 2 ( D + V ) W_H\ge2(D+V) W H ≥ 2 ( D + V ) . The invertible lattice map ( x , y ) ↦ ( − y , x − y ) (x,y)\mapsto(-y,x-y) ( x , y ) ↦ ( − y , x − y ) cyclically permutes the three undirected nearest-neighbor directions on the L × L L\times L L × L torus. The same argument therefore gives the other two bounds in Eq. (15) .
An entropy bound without classification
For odd L ≥ 5 L\ge5 L ≥ 5 , the wall lower bound above and the saturating configurations in Appendix C 2a imply that every constrained minimizer has exactly 4 L 4L 4 L wall edges. This does not require the complete classification. There are 3 L 2 3L^2 3 L 2 possible wall edges. Fixing their set and one triangle label determines every other label: across a nonwall edge it stays the same, and across a wall of direction i i i the labels must be the other two directions. There are at most three choices for the initial label and two spin lifts. The connected triangle adjacency graph then gives G s e l ( L ) ≤ 6 3 L 2 / ( 4 L ) G_{\mathrm{sel}}(L)\le6^{3L^2/(4L)} G sel ( L ) ≤ 6 3 L 2 / ( 4 L ) . In particular,
ln G s e l ( L ) L 2 ≤ ln 6 L 2 + 4 L ln 3 e L 4 . (94) \frac{\ln G_{\mathrm{sel}}(L)}{L^2}\le\frac{\ln6}{L^2}+\frac{4}{L}\ln\frac{3eL}{4}. \tag*{(94)} L 2 ln G sel ( L ) ≤ L 2 ln 6 + L 4 ln 4 3 e L . (94)
Explicit constructions and finite counts
For every allowed directional composition ( H , V , D ) (H,V,D) ( H , V , D ) , the two families below attain all three wall bounds. Their same-composition geometries are shown in Figure 2 . We now give the supports and spin recursion, verify periodic reconstruction, and count the disjoint translation orbits. For the displayed representatives, the square-grid D D D -region boxes are aligned by translations. With all coordinates taken modulo 15, the spin relation in those displayed coordinates is s B ( x , y ) = s A ( 12 − x , 10 − y ) s_B(x,y)=s_A(12-x,10-y) s B ( x , y ) = s A ( 12 − x , 10 − y ) . The finite transfer calculation at the end checks the counts at the sizes in Table II; completeness for all odd L ≥ 5 L\ge5 L ≥ 5 is proved separately in Appendix C 3.
Explicit minimizers in every odd slope sector
For any positive odd H , V , D H,V,D H , V , D with H + V + D = L H+V+D=L H + V + D = L , let I ( a , m ) = { a , a + 1 , … , a + m − 1 } ( m o d L ) I(a,m)=\{a,a+1,\ldots,a+m-1\}\pmod L I ( a , m ) = { a , a + 1 , … , a + m − 1 } ( mod L ) and I ( a , 0 ) = ∅ I(a,0)=\varnothing I ( a , 0 ) = ∅ . Let R R R shift a set by one site in the positive x x x direction, R ( S ) = { x + 1 : x ∈ S } ( m o d L ) R(S)=\{x+1:x\in S\}\pmod L R ( S ) = { x + 1 : x ∈ S } ( mod L ) . A subscript y y y denotes the set of x x x positions where the corresponding bond indicator equals one in row y y y . Two choices of diagonal support are
d y A = { I ( H , D ) , 0 ≤ y < H , I ( H − V , V + D ) , H ≤ y < H + D , ∅ , H + D ≤ y < L . (95) d_y^A=
\begin{cases}
I(H,D), & 0\le y<H,\\
I(H-V,V+D), & H\le y<H+D,\\
\varnothing, & H+D\le y<L.
\end{cases} \tag*{(95)} d y A = ⎩ ⎨ ⎧ I ( H , D ) , I ( H − V , V + D ) , ∅ , 0 ≤ y < H , H ≤ y < H + D , H + D ≤ y < L . (95)
and
d y B = { I ( H , D ) , 0 ≤ y < H , ∅ , H ≤ y < H + V , I ( H , V + D ) , H + V ≤ y < L . (96) d_y^B=
\begin{cases}
I(H,D), & 0\le y<H,\\
\varnothing, & H\le y<H+V,\\
I(H,V+D), & H+V\le y<L.
\end{cases} \tag*{(96)} d y B = ⎩ ⎨ ⎧ I ( H , D ) , ∅ , I ( H , V + D ) , 0 ≤ y < H , H ≤ y < H + V , H + V ≤ y < L . (96)
The spin configuration associated with either bond assignment is obtained from
s ( x , 0 ) = ( − 1 ) max ( 0 , x − H ) , 0 < x < L , s(x,0)=(-1)^{\max(0,x-H)},\qquad0<x<L, s ( x , 0 ) = ( − 1 ) m a x ( 0 , x − H ) , 0 < x < L ,
s ( x , y + 1 ) = [ 2 d ( x − 1 , y ) − 1 ] s ( x − 1 , y ) . (97) s(x,y+1)=[2d(x-1,y)-1]s(x-1,y). \tag*{(97)} s ( x , y + 1 ) = [ 2 d ( x − 1 , y ) − 1 ] s ( x − 1 , y ) . (97)
Its initial horizontal support is I ( 0 , H ) I(0,H) I ( 0 , H ) . The horizontal supports for both choices through the first block are
h y = I ( y , H − y ) ∪ I ( H + D , y ) , 0 ≤ y ≤ H . (98) h_y=I(y,H-y)\cup I(H+D,y),\qquad0\le y\le H. \tag*{(98)} h y = I ( y , H − y ) ∪ I ( H + D , y ) , 0 ≤ y ≤ H . (98)
The intervals are disjoint, also when the second wraps around, because its wrapped endpoint is y − V − 1 < y y-V-1<y y − V − 1 < y . For A the remaining supports are
h y A = { I ( H + D , H ) , H ≤ y ≤ H + D , I ( y , H ) , H + D ≤ y ≤ L . (99) h_y^A=
\begin{cases}
I(H+D,H), & H\le y\le H+D,\\
I(y,H), & H+D\le y\le L.
\end{cases} \tag*{(99)} h y A = { I ( H + D , H ) , I ( y , H ) , H ≤ y ≤ H + D , H + D ≤ y ≤ L . (99)
and for B they are
h y B = { I ( y + D , H ) , H ≤ y ≤ H + V , I ( 0 , H ) , H + V ≤ y ≤ L . (100) h_y^B=
\begin{cases}
I(y+D,H), & H\le y\le H+V,\\
I(0,H), & H+V\le y\le L.
\end{cases} \tag*{(100)} h y B = { I ( y + D , H ) , I ( 0 , H ) , H ≤ y ≤ H + V , H + V ≤ y ≤ L . (100)
The formulas agree at their shared endpoints. Directly in each row block they satisfy
h y ∩ d y = ∅ , d y ⊂ R ( h y ∪ d y ) , h y + 1 = R ( h y ∪ d y ) ∖ d y . (101) h_y\cap d_y=\varnothing,\qquad d_y\subset R(h_y\cup d_y),\qquad h_{y+1}=R(h_y\cup d_y)\setminus d_y. \tag*{(101)} h y ∩ d y = ∅ , d y ⊂ R ( h y ∪ d y ) , h y + 1 = R ( h y ∪ d y ) ∖ d y . (101)
With v y = R [ Z L ∖ ( h y ∪ d y ) ] v_y=R[\mathbb{Z}_L\setminus(h_y\cup d_y)] v y = R [ Z L ∖ ( h y ∪ d y )] , these identities give both triangle constraints in Eq. (12) . They also identify the actual equality indicators in the spin recursion: the vertical ratio is 2 d ( x − 1 , y ) − 1 2d(x-1,y)-1 2 d ( x − 1 , y ) − 1 times the horizontal ratio, and taking the horizontal ratio at the next row yields Eq. (101) .
Closure and sector
The initial row closes because it has L − H = V + D L-H=V+D L − H = V + D unequal bonds, an even number. Every diagonal x − y = q x-y=q x − y = q ... contains exactly D D D equal diagonal bonds. To see this, the column x = H + j x=H+j x = H + j , j = 0 , … , D − 1 j=0,\ldots,D-1 j = 0 , … , D − 1 , in the first H H H rows supplies diagonal indices I ( j + 1 , H ) I(j+1,H) I ( j + 1 , H ) . The remaining D D D -row block, after indexing its rows by j = D − 1 − k j=D-1-k j = D − 1 − k , supplies I ( j + H + 1 , L − H ) I(j+H+1,L-H) I ( j + H + 1 , L − H ) in both A and B. These two intervals partition Z L \mathbb{Z}_L Z L for each j j j . Hence every diagonal contains D D D contributions. After L L L steps of Eq. (97) , the multiplier is ( − 1 ) L − D = 1 (-1)^{L-D}=1 ( − 1 ) L − D = 1 . Both spin configurations are periodic.
Each horizontal row has H H H equal bonds. The diagonal supports have total size H D + D ( V + D ) = L D HD+D(V+D)=LD H D + D ( V + D ) = L D . Vertical row sizes are V V V in the first block, zero in the D D D -row block, and V + D V+D V + D in the V V V -row block, giving total H V + V ( V + D ) = L V HV+V(V+D)=LV H V + V ( V + D ) = L V . The sector is therefore exactly ( H , V , D ) (H,V,D) ( H , V , D ) .
Wall counts and disjoint orbits
Let △ \triangle △ denote symmetric difference of sets. The wall counts are
W D = ∑ y ∣ h y △ h y + 1 ∣ , W_D=\sum_y |h_y\triangle h_{y+1}|, W D = y ∑ ∣ h y △ h y + 1 ∣ ,
W V = ∑ y ∣ R h y △ h y + 1 ∣ , W_V=\sum_y |Rh_y\triangle h_{y+1}|, W V = y ∑ ∣ R h y △ h y + 1 ∣ ,
W H = ∑ y ∣ R − 1 v y △ v y − 1 ∣ . (102) W_H=\sum_y |R^{-1}v_y\triangle v_{y-1}|. \tag*{(102)} W H = y ∑ ∣ R − 1 v y △ v y − 1 ∣. (102)
During each of the first H H H steps, h y h_y h y loses one site and gains one distinct site. During each of the V V V empty-diagonal steps it shifts a proper interval by one, and during the D D D steps it stays fixed. Thus W D = 2 ( H + V ) W_D=2(H+V) W D = 2 ( H + V ) . In comparison with R h y Rh_y R h y , the next support differs at two endpoints in the H H H and D D D blocks and agrees in the V V V block. This gives W V = 2 ( H + D ) W_V=2(H+D) W V = 2 ( H + D ) . Inside each nonzero vertical block v y v_y v y shifts by one; only the three block interfaces contribute to W H W_H W H . Their sizes are V , V + D , D V,V+D,D V , V + D , D for A and D , V + D , V D,V+D,V D , V + D , V for B. Hence W H = 2 ( V + D ) W_H=2(V+D) W H = 2 ( V + D ) . Both choices attain W = 4 L W=4L W = 4 L and the minimum in Theorem 1.
The cyclic diagonal row-width words are ( D H , ( V + D ) D , 0 V ) (D^H,(V+D)^D,0^V) ( D H , ( V + D ) D , 0 V ) for A and ( D H , 0 V , ( V + D ) D ) (D^H,0^V,(V+D)^D) ( D H , 0 V , ( V + D ) D ) for B, where superscripts give repetition counts. The three widths are distinct and each block is nonempty, so the cyclic orders are different. Within either choice, each directed width transition occurs once; a stabilizing translation must therefore have zero vertical shift. A row of width D D D is a single proper cyclic interval and has no nonzero horizontal translation stabilizer. Each choice consequently has L 2 L^2 L 2 distinct translations, and their two orbits are disjoint. A translation sending a spin state to its negative would stabilize its bond array and hence be the identity, which cannot flip a spin state. Including both spin signs yields exactly 4 L 2 4L^2 4 L 2 constructed minimizers per sector.
Exact finite counts
The independent calculation in Table 2 uses a transfer state consisting of two labeled spin rows ( a , b ) (a,b) ( a , b ) . An allowed pair has neither type of monochromatic elementary triangle between its rows. The transition ( a , b ) → ( b , c ) (a,b)\to(b,c) ( a , b ) → ( b , c ) has integer cost
Table 2. Minimum second-neighbor bond sum and number of selected states within the periodic nearest-neighbor ground-state manifold C L \mathcal{C}_L C L . The last column is obtained from an exact integer transfer trace, independent of the classification proof, and agrees with Eq. (16) .
∑ x [ b x + 1 c x + b x c x + 2 + a x c x + 1 ] . (103) \sum_x [b_x+{}_{1}c_x+b_xc_{x+2}+a_xc_{x+1}]. \tag*{(103)} x ∑ [ b x + 1 c x + b x c x + 2 + a x c x + 1 ] . (103)
Over a closed path of L L L transitions the three terms count every second-neighbor bond exactly once. Dynamic programming retains the least cost and the number of paths attaining it. Summing the diagonal trace over all source row pairs counts labeled torus spin arrays; no vertical translation quotient is taken. Horizontal translations and spin inversion reduce only that outer source sum, with each source orbit’s multiplicity restored. The calculation splits into conserved H H H blocks and uses overflow-checked integer additions and orbit weights. It contains no assumed general formula for G s e l ( L ) G_{\mathrm{sel}}(L) G sel ( L ) . The source, completed output, and independent construction checks are supplied in the verification archive [19 ] .
Completeness of the two minimizer families
We prove that simultaneous equality in the three wall bounds forces the two supports in Appendix C 2 a. The four steps establish finite phase components, use a periodic-profile inequality to prove their uniqueness, determine the six-junction geometry, and fix the remaining corner sizes by diagonal counts. The argument does not use finite counts.
Step 1: Finite monotone phase components
Assume simultaneous equality in the three wall bounds. We first show that each connected region of one stripe orientation has a finite lift, with interval rows and columns in coordinates adapted to that orientation. On the infinite lift the triangle equations give an integer potential
p ( x + 1 , y ) − p ( x , y ) = h ( x , y ) , p(x+1,y)-p(x,y)=h(x,y), p ( x + 1 , y ) − p ( x , y ) = h ( x , y ) ,
p ( x , y + 1 ) − p ( x , y ) = v ( x , y ) , p(x,y+1)-p(x,y)=v(x,y), p ( x , y + 1 ) − p ( x , y ) = v ( x , y ) ,
p ( x + 1 , y + 1 ) − p ( x , y ) = 1 − d ( x , y ) . (104) p(x+1,y+1)-p(x,y)=1-d(x,y). \tag*{(104)} p ( x + 1 , y + 1 ) − p ( x , y ) = 1 − d ( x , y ) . (104)
Its periods are H , V H,V H , V . The height normalization used in Sec. III is 3 p ( x , y ) − x − y 3p(x,y)-x-y 3 p ( x , y ) − x − y , up to an additive constant; its periods are 3 H − L 3H-L 3 H − L and 3 V − L 3V-L 3 V − L . The affine extension of p p p over each elementary triangle has gradient ( 1 , 0 ) , ( 0 , 1 ) , ( 0 , 0 ) (1,0),(0,1),(0,0) ( 1 , 0 ) , ( 0 , 1 ) , ( 0 , 0 ) in phases H , V , D H,V,D H , V , D , respectively. Pair the two triangles sharing an equal bond. In each phase these pairs become unit squares in the following unimodular coordinates:
For H , V H,V H , V the support index is the upper-right corner of its square; for D D D it is the lower-left corner. The resulting common unit shift does not affect line counts. In all three coordinate systems Q Q Q is nondecreasing in a , b a,b a , b , with directional derivatives zero or one; both vanish precisely for the designated phase gradient. Each candidate square is split into the two triangles sharing an edge of that direction. Constant Q Q Q therefore labels both triangles by that edge, making it their equal bond. Conversely a paired phase cell has constant Q Q Q . Continuity makes Q Q Q constant on each edge-connected phase component. For two coordinatewise comparable cells in that component, every point of their bounding rectangle lies coordinatewise between corners with the same Q Q Q value. Monotonicity therefore makes Q Q Q constant throughout this rectangle; the preceding cell criterion puts the rectangle in the same component.
The local wall rule also gives
W H = ∑ ∣ d ( x , y ) − d ( x , y − 1 ) ∣ = ∑ ∣ v ( x + 1 , y ) − v ( x , y − 1 ) ∣ , W V = ∑ ∣ d ( x , y ) − d ( x − 1 , y ) ∣ = ∑ ∣ h ( x − 1 , y ) − h ( x , y + 1 ) ∣ , W D = ∑ ∣ h ( x , y ) − h ( x , y + 1 ) ∣ = ∑ ∣ v ( x + 1 , y ) − v ( x , y ) ∣ . (105) \begin{aligned}
W_H &= \sum|d(x,y)-d(x,y-1)| \\
&= \sum|v(x+1,y)-v(x,y-1)|,\\
W_V &= \sum|d(x,y)-d(x-1,y)| \\
&= \sum|h(x-1,y)-h(x,y+1)|,\\
W_D &= \sum|h(x,y)-h(x,y+1)| \\
&= \sum|v(x+1,y)-v(x,y)|. \tag*{(105)}
\end{aligned} W H W V W D = ∑ ∣ d ( x , y ) − d ( x , y − 1 ) ∣ = ∑ ∣ v ( x + 1 , y ) − v ( x , y − 1 ) ∣ , = ∑ ∣ d ( x , y ) − d ( x − 1 , y ) ∣ = ∑ ∣ h ( x − 1 , y ) − h ( x , y + 1 ) ∣ , = ∑ ∣ h ( x , y ) − h ( x , y + 1 ) ∣ = ∑ ∣ v ( x + 1 , y ) − v ( x , y ) ∣. (105)
with sums over all cells. No phase has a full coordinate line. For D D D , a full horizontal or vertical line would force H = 0 H=0 H = 0 or V = 0 V=0 V = 0 . For H H H , full vertical or diagonal support would force V = 0 V=0 V = 0 or D = 0 D=0 D = 0 ; for V V V , full horizontal or diagonal support would force H = 0 H=0 H = 0 or D = 0 D=0 D = 0 . All contradict Eq. (13) . Thus each cyclic variation is twice the number of occupied runs. At equality their totals are
Each total is less than L L L . Since every nonempty line contributes a run, there is an empty line in each coordinate direction. Cutting there places every lifted phase component in a rectangle with side lengths less than L L L . Its projection to the torus is injective. This finiteness uses equality of the wall bounds, not height monotonicity alone.
Rectangle filling makes every row and column of a component an interval. Connectivity makes its occupied row and column indices consecutive. If its bounding box has cell widths w , h w,h w , h , it contributes w + h w+h w + h runs. Its northwest and southeast corner cells are present, since a minimum-a a a cell and a maximum-b b b cell are comparable, and likewise a minimum-b b b cell and a maximum-a a a cell. The slices t = a − b t=a-b t = a − b consequently have interval support of length s = w + h − 1 s=w+h-1 s = w + h − 1 . Their cell counts n ( t ) n(t) n ( t ) start and end at one and are 1-Lipschitz. Indeed, if their occupied a a a intervals on slices t , t + 1 t,t+1 t , t + 1 are [ l , r ] , [ l ′ , r ′ ] [l,r],[l',r'] [ l , r ] , [ l ′ , r ′ ] , rectangle filling gives l ′ − l , r ′ − r ∈ { 0 , 1 } l'-l,r'-r\in\{0,1\} l ′ − l , r ′ − r ∈ { 0 , 1 } : any larger jump or negative jump would fill a cell beyond one of the stated endpoints. For example, l ′ > l + 1 l'>l+1 l ′ > l + 1 would fill the cell ( l + 1 , l − t ) (l+1,l-t) ( l + 1 , l − t ) on slice t + 1 t+1 t + 1 . The slice counts therefore differ by at most one.
Step 2: One component per phase
The component profiles from Step 1 must add to the conserved directional count on every periodic line. The next lemma bounds their total support length; together with the equality run counts, it forces one component per phase.
Lemma 2. Let L , K L,K L , K be integers with 1 ≤ K < L 1\le K<L 1 ≤ K < L . Let nonzero nonnegative integer profiles n i n_i n i have finite interval supports of lengths s i s_i s i , endpoint values one, and adjacent differences of absolute value at most one. If, for every integer t t t ,
∑ i ∑ j ∈ Z n i ( t + j L ) = K , 1 ≤ K < L , (106) \sum_i\sum_{j\in\mathbb{Z}} n_i(t+jL)=K,\qquad1\le K<L, \tag*{(106)} i ∑ j ∈ Z ∑ n i ( t + j L ) = K , 1 ≤ K < L , (106) then ∑ i s i ≥ L + K − 1 \sum_i s_i\ge L+K-1 ∑ i s i ≥ L + K − 1 .
Proof. Let m ( t ) m(t) m ( t ) count all positive translated profiles at time t t t , counting overlapping translates of the same profile separately. It is L L L -periodic, m ( t ) ≥ 1 m(t)\ge1 m ( t ) ≥ 1 , and ∑ t m o d L m ( t ) = S : = ∑ i s i \sum_{t\ \mathrm{mod}\ L}m(t)=S:=\sum_i s_i ∑ t mod L m ( t ) = S := ∑ i s i . Choose a longest profile A and a lift starting at t 0 t_0 t 0 , with value one. For each other lift B positive there write k B = n B ( t 0 ) k_B=n_B(t_0) k B = n B ( t 0 ) , so ∑ B k B = K − 1 \sum_B k_B=K-1 ∑ B k B = K − 1 . Endpoint values and the Lipschitz bound imply k B ≤ ⌊ ( s B + 1 ) / 2 ⌋ ≤ s A k_B\le\lfloor(s_B+1)/2\rfloor\le s_A k B ≤ ⌊( s B + 1 ) /2 ⌋ ≤ s A . Thus both A and B remain positive for k B k_B k B times starting at t 0 t_0 t 0 . Charge one overlap for B at each of those times. At any charged time the charges are at most m ( t ) − 1 m(t)-1 m ( t ) − 1 , with A left uncharged. All charged times lie in fewer than L L L consecutive positions, since k B ≤ K − 1 < L k_B\le K-1<L k B ≤ K − 1 < L ; their residues are distinct. Hence S − L ≥ ∑ B k B ≥ K − 1 S-L\ge\sum_B k_B\ge K-1 S − L ≥ ∑ B k B ≥ K − 1 . For K = 1 K=1 K = 1 there are no charges, and m ( t ) ≥ 1 m(t)\ge1 m ( t ) ≥ 1 gives the same conclusion. ◼
For a phase with sector count K ∈ { H , V , D } K\in\{H,V,D\} K ∈ { H , V , D } , choose one lift of each of its C C C torus components. Its sector lines are a − b a-b a − b modulo L L L , so conservation of their cell counts gives Eq. (106) . The run table yields ∑ i ( w i + h i ) = L + K \sum_i(w_i+h_i)=L+K ∑ i ( w i + h i ) = L + K , whereas S = ∑ i ( w i + h i − 1 ) = L + K − C S=\sum_i(w_i+h_i-1)=L+K-C S = ∑ i ( w i + h i − 1 ) = L + K − C . Lemma 2 therefore gives L + K − C ≥ L + K − 1 L+K-C\ge L+K-1 L + K − C ≥ L + K − 1 , hence C ≤ 1 C\le1 C ≤ 1 . Since K > 0 K>0 K > 0 , each phase has exactly one component.
Three disks and six junctions
The finite component has no hole, since a row through a hole would have two separated occupied intervals. It has no diagonal pinch either. For example, cells ( − 1 , 0 ) , ( 0 , − 1 ) (-1,0),(0,-1) ( − 1 , 0 ) , ( 0 , − 1 ) with ( 0 , 0 ) , ( − 1 , − 1 ) (0,0),(-1,-1) ( 0 , 0 ) , ( − 1 , − 1 ) absent put the row-0 interval wholly to the left of the row-− 1 -1 − 1 interval. These intervals share no a a a coordinate, so no component edge crosses between the two rows. Any edge-connected path between the stated cells would have to cross there, a contradiction. Reflection covers the other diagonal pinch. At every boundary vertex the remaining cell boundary then has degree two, so it consists of simple cycles. Connectivity gives one outer cycle, and any additional cycle would enclose a hole. Thus each phase is an injectively embedded disk.
The local possibilities follow from the six neighbors of a spin: its same-spin neighbors form an independent set of their six-cycle. If their number is k k k , the wall degree is six at k = 0 k=0 k = 0 , three at k = 1 , 3 k=1,3 k = 1 , 3 , zero for opposite pairs at k = 2 k=2 k = 2 , and two collinear edges for the other pairs at k = 2 k=2 k = 2 . At k = 0 k=0 k = 0 each phase has two diagonally touching cells with both intervening cells absent, contradicting its single component without pinches. This excludes degree six. At k = 1 k=1 k = 1 one phase has a reflex corner and the other two convex corners; at k = 3 k=3 k = 3 all three are convex. Angles are measured in each phase’s square coordinates. These cases exhaust the 18 independent sets of a six-cycle.
The wall graph bounds three disk faces and is connected: a disk’s attaching circle cannot join two graph components to form a connected surface. With U U U degree-two vertices and T T T junctions, its edge count is U + 3 T / 2 U+3T/2 U + 3 T /2 . The torus Euler relation gives U + T − ( U + 3 T / 2 ) + 3 = 0 U+T-(U+3T/2)+3=0 U + T − ( U + 3 T /2 ) + 3 = 0 , hence T = 6 T=6 T = 6 . If N k N_k N k counts sites with k k k equal incident bonds, double counting the L 2 L^2 L 2 equal bonds gives
∑ k k N k = 2 L 2 = 2 ∑ k N k , \sum_k kN_k=2L^2=2\sum_k N_k, k ∑ k N k = 2 L 2 = 2 k ∑ N k ,
or N 3 = N 1 + 2 N 0 N_3=N_1+2N_0 N 3 = N 1 + 2 N 0 . Since N 0 = 0 N_0=0 N 0 = 0 and N 1 + N 3 = T = 6 N_1+N_3=T=6 N 1 + N 3 = T = 6 , we have N 1 = N 3 = 3 N_1=N_3=3 N 1 = N 3 = 3 . The local corner rule therefore leaves exactly three reflex corners among all phase boundaries.
The phase bounding boxes have widths and heights ( V + D , H + D ) (V+D,H+D) ( V + D , H + D ) , ( H + V , H + D ) (H+V,H+D) ( H + V , H + D ) , and ( V + D , H + V ) (V+D,H+V) ( V + D , H + V ) for D , H , V D,H,V D , H , V . The phase areas are L D , L H , L V LD,LH,LV L D , L H , L V , so their respective deficits from the box areas are H V , V D , H D HV,VD,HD H V , V D , H D , all positive. For each simple orthogonal boundary, the total turn gives c − r = 4 c-r=4 c − r = 4 , where c , r c,r c , r count convex and reflex corners. Thus a nonrectangular phase has r ≥ 1 r\ge1 r ≥ 1 . The three available reflex corners give exactly one per phase, hence five convex corners and six sides per boundary. Tracing these six axis-aligned sides gives a bounding rectangle with a single corner rectangle removed. For phase D D D , rectangle filling keeps the northwest and southeast corners, so only the southwest or northeast corner can be removed.
Diagonal counts fix the shape
For the southwest case, let w , h w,h w , h denote the box width and height in square-grid cells, and put the D D D box at 0 ≤ x < w = V + D , 0 ≤ y < h = H + D 0\le x<w=V+D,0\le y<h=H+D 0 ≤ x < w = V + D , 0 ≤ y < h = H + D , and remove 0 ≤ x ≤ a , 0 ≤ y ≤ b 0\le x\le a,0\le y\le b 0 ≤ x ≤ a , 0 ≤ y ≤ b , with 0 ≤ a ≤ V , 0 ≤ b ≤ H 0\le a\le V,0\le b\le H 0 ≤ a ≤ V , 0 ≤ b ≤ H . The polynomial F ( z ) = ∑ ( x , y ) ∈ D r e g i o n z x − y F(z)=\sum_{(x,y)\in D\ \mathrm{region}}z^{x-y} F ( z ) = ∑ ( x , y ) ∈ D region z x − y counts cells on each diagonal modulo L L L . Since every diagonal has D D D cells, F ( z ) = D ( 1 + z + ⋯ + z L − 1 ) F(z)=D(1+z+\cdots+z^{L-1}) F ( z ) = D ( 1 + z + ⋯ + z L − 1 ) in Z [ z , z − 1 ] / ( z L − 1 ) \mathbb{Z}[z,z^{-1}]/(z^L-1) Z [ z , z − 1 ] / ( z L − 1 ) . Multiplying by ( 1 − z ) ( 1 − z − 1 ) (1-z)(1-z^{-1}) ( 1 − z ) ( 1 − z − 1 ) and subtracting the corner rectangle yields
z a + z − b + z V − H = z − H + z V + z a − b . (107) z^a+z^{-b}+z^{V-H}=z^{-H}+z^V+z^{a-b}. \tag*{(107)} z a + z − b + z V − H = z − H + z V + z a − b . (107)
The residue monomials are a basis, so these exponent multisets agree modulo L L L , including multiplicities. The exponent a a a cannot match − H -H − H , since 0 < a < L − H 0<a<L-H 0 < a < L − H , or a − b a-b a − b , since 0 < b < L 0<b<L 0 < b < L . Thus a = V a=V a = V . After canceling z V z^V z V , − b -b − b cannot match V − b V-b V − b , so b = H b=H b = H . The northeast case follows by inversion. These are precisely the two diagonal supports of Appendix C2a, up to translation.
Finally, the triangle equations determine h , v h,v h , v from this d d d . Two choices of h h h with the same d d d differ by δ h ( x , y + 1 ) = δ h ( x − 1 , y ) \delta h(x,y+1)=\delta h(x-1,y) δ h ( x , y + 1 ) = δ h ( x − 1 , y ) , constant on each x − y x-y x − y diagonal. Every such diagonal has a d = 1 d=1 d = 1 cell, where both choices of h h h must agree. Hence δ h = 0 \delta h=0 δ h = 0 , and v v v is forced as well. Connected spin reconstruction leaves exactly the two global signs. The disjoint-orbit count in Appendix C2c therefore exhausts all minimizers.
A full-model competing configuration
We now compare an explicit state outside C L C_L C L with the energy E ∗ E_* E ∗ in (22) , the energy of an actual constrained minimizer. The construction A in sector ( H , V , D ) = ( 1 , 1 , L − 2 ) (H,V,D)=(1,1,L-2) ( H , V , D ) = ( 1 , 1 , L − 2 ) has the simple spin formula u ( 0 , 0 ) = 1 u(0,0)=1 u ( 0 , 0 ) = 1 and u ( x , y ) = ( − 1 ) x + y − 1 u(x,y)=(-1)^{x+y-1} u ( x , y ) = ( − 1 ) x + y − 1 at all other sites, using representatives 0 ≤ x , y < L 0\le x,y<L 0 ≤ x , y < L . It has A ( u ) = − L 2 \mathcal{A}(u)=-L^2 A ( u ) = − L 2 and B ( u ) = − L 2 + 8 L \mathcal{B}(u)=-L^2+8L B ( u ) = − L 2 + 8 L by Appendix C2. Flip its spins on F = { ( x , 1 ) : 2 ≤ x ≤ L − 1 } F=\{(x,1):2\le x\le L-1\} F = {( x , 1 ) : 2 ≤ x ≤ L − 1 } to obtain u F u^F u F . For any bond direction, its bond-sum change is minus twice the sum of u i u j u_i u_j u i u j over bonds with exactly one endpoint in F F F . Counting forward and backward neighbors of each site in F F F counts each crossing bond once. Their sums and changes are
Table 1 (PDF p. 21).
These include all periodic wraps. For example, the ( 2 , 1 ) (2,1) ( 2 , 1 ) forward neighbors of x = L − 2 , L − 1 x=L-2,L-1 x = L − 2 , L − 1 and the backward neighbor of x = 2 x=2 x = 2 have product + 1 +1 + 1 ; all other products in that direction are − 1 -1 − 1 . The latter backward neighbor is the exceptional spin at ( 0 , 0 ) (0,0) ( 0 , 0 ) . In the ( 1 , 2 ) (1,2) ( 1 , 2 ) direction the forward wrap contributes + 1 +1 + 1 , the other L − 3 L-3 L − 3 forward products are − 1 -1 − 1 , and all L − 2 L-2 L − 2 backward products to row L − 1 L-1 L − 1 are + 1 +1 + 1 . The first three changes sum to four and the last three to minus sixteen. Thus
H 1 , 2 ( u F ) − E ∗ = 4 J 1 − 16 J 2 , (108) \mathcal{H}_{1,2}(u^{F})-E_* = 4J_1-16J_2, \tag*{(108)} H 1 , 2 ( u F ) − E ∗ = 4 J 1 − 16 J 2 , (108)
which proves the upper bound in Eq. (24) for every odd L ≥ 5 L\ge5 L ≥ 5 .
For L = 5 L=5 L = 5 an exhaustive full-spin calculation fixes one spin and visits all 2 24 2^{24} 2 24 remaining arrays. A Gray-code calculation and an independent five-row enumeration give the same complete joint histogram of ( A , B ) (\mathcal{A},\mathcal{B}) ( A , B ) ; the programs and complete histogram are supplied in the verification archive [19 ] .
Its minimum of 4 A + B 4\mathcal{A}+\mathcal{B} 4 A + B is − 85 = − 5 L 2 + 8 L -85=-5L^2+8L − 85 = − 5 L 2 + 8 L . For any outside state with B < 15 \mathcal{B}<15 B < 15 , this implies A + 25 ≥ ( 15 − B ) / 4 \mathcal{A}+25\ge(15-\mathcal{B})/4 A + 25 ≥ ( 15 − B ) /4 . Together with u F u^{F} u F it proves r c ( 5 ) = 1 / 4 r_c(5)=1/4 r c ( 5 ) = 1/4 . No corresponding all-size inequality is assumed here.
Moments of the sector measure
The exact count 4 L 2 4L^2 4 L 2 per slope triple makes the constrained-minimizer measure uniform over the positive odd compositions of L L L . For m = ( L − 3 ) / 2 m=(L-3)/2 m = ( L − 3 ) /2 , the uniform compositions n 1 + n 2 + n 3 = m n_1+n_2+n_3=m n 1 + n 2 + n 3 = m give E n i = m / 3 \mathbb{E}n_i=m/3 E n i = m /3 and E [ n i ( n i − 1 ) ] = m ( m − 1 ) / 6 \mathbb{E}[n_i(n_i-1)]=m(m-1)/6 E [ n i ( n i − 1 )] = m ( m − 1 ) /6 by summing the composition counts. Thus Var n i = m ( m + 3 ) / 18 \operatorname{Var}n_i=m(m+3)/18 Var n i = m ( m + 3 ) /18 . With H i = 1 + 2 n i H_i=1+2n_i H i = 1 + 2 n i this is Eq. (17) . The covariances of different normalized slopes are − ( 1 − 9 / L 2 ) / 36 -(1-9/L^2)/36 − ( 1 − 9/ L 2 ) /36 , since their sum is fixed at one.