Introduction
Metasurfaces (MTSs) are planar arrays of subwavelength scatterers that control the amplitude, phase, and polarization of electromagnetic waves through the engineered response of their unit cells [1 , 2 , 3 , 4 , 5 ] . Beyond shaping free-space wavefronts, suitably designed metasurfaces can also support and guide bound surface waves [6 ] . In a homogenised description, the MTS is represented by an effective sheet boundary condition rather than treated element by element [7 , 8 ] , in which the entire unit-cell geometry is reduced to a small set of effective parameters. For capacitive metallic patch arrays, the isotropic response is described by a scalar sheet capacitance, for which simple and accurate closed forms are available for canonical square grids [9 , 10 ] . Such a scalar description underlies many analysis and synthesis approaches for surface-wave MTSs, from high-impedance surfaces [11 ] to modulated-MTS antennas [12 ] . The more general homogenisation challenge, however, is not merely to reproduce a particular electromagnetic observable, such as a dispersion curve, but to establish a direct mapping from the unit-cell geometry to its effective sheet response. Such a constitutive description can then be employed independently of the particular electromagnetic problem in which the sheet is embedded.
A scalar sheet capacitance, however, presumes that the induced current is parallel to the applied tangential field for every in-plane direction. This holds when the unit cell is sufficiently symmetric [13 , 14 ] . A patch with three-fold or higher-rotational symmetry responds isotropically, whereas any lower-symmetry cell couples the two tangential field components and requires a full 2 × 2 2 \times2 2 × 2 surface-capacitance dyadic. Anisotropic sheet responses are an established design resource across the spectrum. At microwave frequencies, tensor holographic impedance surfaces shape aperture fields through modulated anisotropy [15 , 16 ] , and printed-circuit tensor impedance surfaces have been characterised by retrieving the effective sheet tensor from full-wave simulations [17 ] , and spatially varying sheet response has been proposed as a platform for surface-wave transformation optics [18 ] . At optical frequencies, extreme sheet anisotropy produces hyperbolic surface-wave regimes with open isofrequency contours, topological transitions [19 , 20 , 21 ] , and strongly enhanced light–matter interaction [22 , 23 , 24 , 25 ] , a regime realized not only in structured sheets but natively in anisotropy van der Waals materials [26 , 27 , 28 ] . Self-complementary metasurfaces can also exhibit dual-mode hyperbolicity and supercanalization [29 ] . Closed-form dispersion models exist for canonical anisotropic elements such as elliptical patches [30 , 31 ] .
Despite their different formulations, these approaches generally rely on canonical geometries, per-geometry full-wave characterization, or reduced models whose parameters must be determined for the specific structure under consideration. For general patch geometries, including irregular, multiply connected, or random shapes, a direct geometry-to-sheet relation is generally unavailable. A second limitation, common to both scalar and dyadic local sheet models, arises as the Bloch wavevector approaches the Brillouin-zone boundary. In this regime, the higher Floquet orders of the lattice are no longer deeply evanescent and the sheet response becomes non-local, i.e., spatially dispersive [32 , 33 , 34 ] , with related non-local homogenization effects also arising in structured metallic media such as space–time modulated wire media [35 ] . Such non-locality is increasingly exploited as a design degree of freedom in its own right [36 ] . The local dispersion branch then departs from the actual Bloch band [37 ] . Multimodal equivalent-network models explicitly account for higher-order Floquet contributions to the response of periodic patch and aperture arrays [38 , 39 ] . Accurate semianalytical Bloch approaches for square and hexagonal unit cells [40 , 41 ] rely on full-wave unit-cell characterization, which the present geometry-driven framework avoids. Such approaches provide an accurate description of the Bloch spectrum, but they do not directly establish a reduced geometry-to-sheet relation from which both the local anisotropic response and its leading spatial-dispersion correction can be obtained.
The practical consequence is that obtaining the surface-wave band diagram for a new irregular patch still typically requires a full-wave eigenmode computation, or an equivalent multi-modal Floquet analysis re-built for that geometry. The purpose of this work is to replace that step. Once the patch polygon and its point-group class are given, the local capacitance dyadic and the leading Floquet correction are computed from geometry alone, and the dispersion relation is solved without full-wave simulation. Full-wave eigenmode data are used only for validation.
In this paper, we present a geometry-driven framework that constructs the local anisotropic surface-capacitance response and its leading nonlocal correction directly from a general gap-connected patch geometry on a hexagonal lattice. The framework distinguishes three aspects of the homogenization problem. First, the two-dimensional Bravais periodicity defines two independent macroscopic tangential loading directions, associated with the two primitive translation directions of the lattice. In the periodically identified unit cell, these correspond to the two independent winding directions of the flat-torus representation [42 , 43 ] . For the patch geometries considered here, the gap remains connected across both periodic directions, so that both macroscopic loadings produce non-degenerate electrostatic responses and the resulting surface-capacitance dyadic is full rank. Second, the admissible form of the dyadic is dictated by the point-group symmetry of the complete unit-cell geometry [13 , 44 ] , spanning the symmetry classes from the fully symmetric hexagonal cell down to a patch with no point symmetry. Third, the numerical values of the allowed entries follow from a quasi-static boundary-value problem posed on a single periodic cell, based on homogenization theory for composite media [45 , 46 ] . These steps establish the local effective sheet response from geometry before any particular propagation problem is introduced.
Surface-wave propagation is then used to develop and validate the electromagnetic consequences of this effective description. The resulting dyadic drives the anisotropic surface-wave dispersion through its projection onto the transverse-electric (TE) polarization of the guided mode. Near the Brillouin-zone boundary, the six Floquet orders of the first reciprocal shell are retained explicitly [38 , 39 ] . Their coupling to the fundamental guided mode is computed from the reconstructed patch boundary. A boundary Fourier overlap, motivated by coupled-mode theory [47 , 48 , 49 ] , determines the coupling strength of each retained order, while the discrete turning of the boundary, where the induced edge charge concentrates [50 ] , provides its directional weight. The detuning of each order is determined by its own sheet-loaded decay constant. Mirror symmetries of the patch enter through the coherent pairing of the sidebands, as in the symmetry analysis of photonic crystal bands [51 , 52 ] . Thus, both the local anisotropic response and its leading spatial-dispersion correction are derived from geometric quantities rather than fitted to the Bloch spectrum. The resulting chain from geometry to dispersion is validated against full-wave eigenmode simulations for representative structures spanning the symmetry classes considered in this work.
The paper is organized as follows. Section II develops the geometry-driven surface-capacitance dyadic. The two-dimensional periodicity defines the macroscopic loading space. The connectivity of the gap ensures a non-degenerate response. The tensor form is constrained by symmetry. A quasi-static cell problem determines the independent entries and their calibration to an equivalent sheet capacitance. Section III uses the effective sheet-capacitance dyadic to derive the local anisotropic surface-wave dispersion relation and its quasi-TE reduction. Section IV develops the geometry-derived first-shell Floquet correction, incorporating the coupling to higher-order harmonics through boundary Fourier overlaps and directional turning measures. Section V validates the resulting model against full-wave eigenmode simulations across the symmetry classes considered, and Section VI concludes.
We represent the patch-type metasurface (MTS) in the quasi-static regime by an infinitesimally thin, purely electric sheet at z = 0 z = 0 z = 0 . The macroscopic fields satisfy the jump conditions [7 , 8 ]
z ^ × [ H ] = J s = Y ‾ ‾ s ⋅ E t , \hat{\mathbf{z}} \times[\mathbf{H}] = \mathbf{J}_{s} = \overline{\overline{Y}}_{s} \cdot\mathbf{E}_{t}, z ^ × [ H ] = J s = Y s ⋅ E t ,
z ^ × [ E ] = 0. \hat{\mathbf{z}} \times[\mathbf{E}] = 0. z ^ × [ E ] = 0.
Here [ F ] = F + − F − [\mathbf{F}] = \mathbf{F}^{+} - \mathbf{F}^{-} [ F ] = F + − F − denotes the field jump, with F + \mathbf{F}^{+} F + and F − \mathbf{F}^{-} F − evaluated on the upper and lower sides of the sheet, respectively. The quantities J s \mathbf{J}_{s} J s and Y ‾ ‾ s \overline{\overline{Y}}_{s} Y s are the equivalent electric surface-current density and the sheet-admittance dyadic. Equation (1b) implies continuity of the tangential electric field, E t + = E t − ≡ E t \mathbf{E}_{t}^{+} = \mathbf{E}_{t}^{-} \equiv\mathbf{E}_{t} E t + = E t − ≡ E t .
For a lossless reciprocal patch array in the quasi-static capacitive regime, we write the induced tangential surface polarization as
P s , t = C ‾ ‾ e q v ⋅ E t , (1) \mathbf{P}_{s,t} = \overline{\overline{C}}_{\mathrm{eqv}} \cdot\mathbf{E}_t, \tag*{(1)} P s , t = C eqv ⋅ E t , (1)
where C ‾ ‾ e q v \overline{\overline{C}}_{\mathrm{eqv}} C eqv is the equivalent surface-capacitance dyadic, in units of F. With the e j ω t e^{j\omega t} e j ω t time convention, the associated polarization current is J s = j ω P s , t \mathbf{J}_s = j\omega\mathbf{P}_{s,t} J s = j ω P s , t . Substitution into Eq. (1a) gives
Y ‾ ‾ s = j ω C ‾ ‾ e q v . (2) \overline{\overline{Y}}_s = j\omega\overline{\overline{C}}_{\mathrm{eqv}}. \tag*{(2)} Y s = j ω C eqv . (2)
Thus, C ‾ ‾ e q v \overline{\overline{C}}_{\mathrm{eqv}} C eqv is the equivalent sheet quantity entering the macroscopic electromagnetic boundary conditions. In the following, we determine this dyadic directly from the unit-cell geometry for gap-connected patch arrays on a hexagonal lattice.
An example of the unit-cell geometry considered in this work is shown in Figure 1 . The metallic (i.e., patch) region Σ \Sigma Σ repeats on a hexagonal lattice generated by the primitive vectors a 1 \mathbf{a}_1 a 1 and a 2 \mathbf{a}_2 a 2 , and the gap is Ω = U ∖ Σ \Omega= U \setminus\Sigma Ω = U ∖ Σ , with U U U being the entire unit cell. For a patch having three-fold rotation symmetry (C 3 C_3 C 3 ) or higher, the sheet response is a scalar surface capacitance [9 , 10 ] . For a patch with lower symmetry, the induced charge depends on the direction of the applied tangential field, and the response is generally described by a symmetric 2 × 2 2 \times2 2 × 2 surface-capacitance dyadic [15 , 16 ] .
Figure 1. An example of the geometry of the periodic patch layer. The unit cell U U U (black outline) repeats on a hexagonal lattice generated by the primitive vectors a 1 \mathbf{a}_1 a 1 and a 2 \mathbf{a}_2 a 2 . Inside the unit cell, the metallic patch Σ \Sigma Σ (dark blue) is a random geometry, and Ω = U ∖ Σ \Omega= U \setminus\Sigma Ω = U ∖ Σ is the gap.
To determine C ‾ ‾ e q v \overline{\overline{C}}_{\mathrm{eqv}} C eqv from this geometry, the quasi-static cell problem first produces the auxiliary two-dimensional dyadic
C ‾ ‾ = [ C x x C x y C x y C y y ] . (3) \overline{\overline{C}} =
\begin{bmatrix}
C_{xx} & C_{xy} \\
C_{xy} & C_{yy}
\end{bmatrix}. \tag*{(3)} C = [ C xx C x y C x y C y y ] . (3)
in units of (F/m). Its entries characterize the electric energy stored in Ω \Omega Ω under tangential excitation along the two in-plane directions (Sec. II C). Because the quasi-static cell problem is two-dimensional, C ‾ ‾ \overline{\overline{C}} C is a per-unit-length quantity and is not yet the equivalent sheet capacitance appearing in Eq. (2) . The calibration introduced in Sec. II D converts it into equivalent sheet-capacitance dyadic C ‾ ‾ e q v \overline{\overline{C}}_{\mathrm{eqv}} C eqv , in units of (F), which describes the electromagnetic response of the patch layer. The auxiliary dyadic C ‾ ‾ \overline{\overline{C}} C and its calibrated counterpart C ‾ ‾ e q v \overline{\overline{C}}_{\mathrm{eqv}} C eqv therefore characterize the effective quasi-static response of the unit-cell geometry, with C ‾ ‾ e q v \overline{\overline{C}}_{\mathrm{eqv}} C eqv providing the geometry-dependent input to the local surface-wave dispersion model of Sec. III.
The derivation relies on three complementary elements. The two-dimensional Bravais periodicity defines two independent macroscopic tangential loading directions, while connectivity of the gap across both periodic directions ensures that the corresponding quasi-static response is non-degenerate (Sec. II A). The point-group symmetry of the patch constrains the admissible form of the dyadic (Sec. II B), and a quasi-static boundary-value problem posed on a single periodic unit cell determines its independent entries (Sec. II C). Once C ‾ ‾ \overline{\overline{C}} C has been obtained and calibrated, C ‾ ‾ e q v \overline{\overline{C}}_{\mathrm{eqv}} C eqv completely determines the local homogenised response. The nonlocal Floquet correction of Sec. IV additionally requires the reconstructed boundary ∂ Ω \partial\Omega ∂ Ω , which supplies the first-shell overlaps and directional turning measures.
Periodic loading space and non-degenerate capacitance response
We consider a periodic patch-type MTS invariant under the two-dimensional Bravais lattice generated by the primitive translation vectors a 1 \mathbf{a}_1 a 1 and a 2 \mathbf{a}_2 a 2 ,
R n 1 n 2 = n 1 a 1 + n 2 a 2 , n 1 , n 2 ∈ Z , (4) \mathbf{R}_{n_1 n_2} = n_1 \mathbf{a}_1 + n_2 \mathbf{a}_2,\qquad n_1,n_2 \in\mathbb{Z}, \tag*{(4)} R n 1 n 2 = n 1 a 1 + n 2 a 2 , n 1 , n 2 ∈ Z , (4)
as depicted in Figure 1 . The infinite periodic structure can be represented exactly by a single primitive unit cell equipped with periodic boundary conditions. Identifying opposite edges of the cell maps the periodically repeated plane onto a flat torus [42 ] , providing a compact representation of the infinite periodic problem [Figure 2 (a)].
Figure 2. Periodic loading directions and connectivity of the surface-capacitance problem. (a) Identifying opposite edges of the unit cell maps the periodic plane onto a flat torus. The two highlighted winding cycles, γ 1 \gamma_1 γ 1 (red) and γ 2 \gamma_2 γ 2 (blue), correspond to the two primitive translation directions of the Bravais lattice and therefore to the two independent macroscopic tangential loadings. (b) Representative hexagonal patch geometry for which the gap Ω = U ∖ Σ \Omega= U \setminus\Sigma Ω = U ∖ Σ remains connected across both periodic directions. The existence of winding gap paths in both directions allows independent macroscopic potential variations to be sustained and gives a non-degenerate 2 × 2 2 \times2 2 × 2 capacitance response for the class of geometries considered here.
The significance of this construction for the present homogenisation problem is that a two-dimensional periodic sheet admits two independent macroscopic tangential loading directions. These may be represented by the two primitive lattice directions, or equivalently by any two independent Cartesian directions in the tangential plane. A prescribed macroscopic electrostatic loading therefore has two independent components, and the corresponding linear effective response is naturally represented by a 2 × 2 2 \times2 2 × 2 surface-capacitance dyadic C ‾ ‾ \overline{\overline{C}} C .
In the torus representation shown in Figure 2 (a), the primitive periodic directions are associated with two independent non-contractible winding cycles, γ 1 \gamma_1 γ 1 and γ 2 \gamma_2 γ 2 , corresponding respectively to translations along a 1 \mathbf{a}_1 a 1 and a 2 \mathbf{a}_2 a 2 [42 ] . These cycles identify the two independent macroscopic directions in which a potential variation may be imposed across the periodically identified cell.
The topology of the gap region enters separately through its connectivity. Let
Ω = U ∖ Σ (5) \Omega= U \setminus\Sigma \tag*{(5)} Ω = U ∖ Σ (5)
denote the unmetallized area of the unit-cell gap, where U U U is the complete unit cell and Σ \Sigma Σ is the metallic patch. A capacitive response to a macroscopic loading in a given periodic direction requires the gap region to remain connected across the cell in that direction, so that a macroscopic potential variation can be sustained. For every geometry considered in this work, Ω \Omega Ω contains connected winding paths across both primitive directions, as illustrated schematically in Figure 2 (b).
Consequently, neither of the two independent macroscopic loadings is suppressed by the geometry, and the corresponding two column responses of the capacitance dyadic remain linearly independent. Therefore for this class of patch geometries considered here,
rank ( C ‾ ‾ ) = 2. (6) \operatorname{rank}\left(\overline{\overline{C}}\right) = 2. \tag*{(6)} rank ( C ) = 2. (6)
Thus, the two-dimensional Bravais periodicity fixes the dimension of the macroscopic tangential loading space, whereas connectivity of the gap across both periodic directions ensures that the resulting 2 × 2 2 \times2 2 × 2 capacitive response is non-degenerate.
It is useful to distinguish these macroscopic winding directions from the complete topology of the gap domain. Additional holes, indentations, or disconnected metallic components may introduce further closed paths within Ω \Omega Ω , but they do not increase the dimension of the imposed macroscopic tangential loading space, which remains two-dimensional. Such geometric features instead modify the solution of the quasi-static cell problem and therefore the numerical entries of C ‾ ‾ \overline{\overline{C}} C .
Equation (7) fixes only the number of independent macroscopic responses; it does not prescribe the individual entries of C ‾ ‾ \overline{\overline{C}} C . The admissible tensor form is instead constrained by the point-group symmetry of the complete unit-cell geometry, as discussed in Sec. II B and summarised in Table I and Figure 3 . The numerical values of the independent entries are subsequently determined by the quasi-static cell problem of Sec. II C.
Figure 3. Directional-capacitance indicator C ( u ^ ) = u ^ ⋅ C ‾ ⋅ u ^ / C ‾ C(\hat{\mathbf{u}})=\hat{\mathbf{u}}\cdot\overline{C}\cdot\hat{\mathbf{u}}/\overline{C} C ( u ^ ) = u ^ ⋅ C ⋅ u ^ / C , normalised by the mean eigenvalue C ‾ = 1 2 tr ( C ‾ ) \overline{C}=\frac{1}{2}\operatorname{tr}(\overline{C}) C = 2 1 tr ( C ) , for the four symmetry classes of Table I. The light-blue filled region bounded by the solid blue curve is the directional indicator; purple and orange arrows mark the two principal eigenvectors scaled by their normalised eigenvalues; the dashed grey circle is the isotropic reference ( λ 1 / λ 2 = 1 ) (\lambda_1/\lambda_2=1) ( λ 1 / λ 2 = 1 ) . The inset in each panel shows a representative patch geometry Σ \Sigma Σ belonging to that symmetry class. (a) C 6 v / C 3 v C_{6v}/C_{3v} C 6 v / C 3 v : full rotational symmetry forces C ‾ = C I \overline{C}=C I C = C I , giving a perfect circle. (b) C 2 v C_{2v} C 2 v : two distinct eigenvalues with eigenvectors aligned to the lattice axes, giving an axis-aligned directional-capacitance indicator. (c) C s C_s C s : eigenvectors rotated to the mirror-axis orientation φ \varphi φ , giving a rotated directional-capacitance indicator. (d) C 1 C_1 C 1 : a nonzero off-diagonal entry C x y = C y x C_{xy}=C_{yx} C x y = C y x rotates the principal axes freely; at the strong anisotropy of this example the indicator develops a pronounced waist, reflecting the fully general three-parameter dyadic.
Let G G G denote the in-plane point group of the complete unit-cell geometry, including both the patch and the underlying lattice. The shape of the patch Σ \Sigma Σ determines the symmetry of the structure. Since the hexagonal lattice can have at most six-fold symmetry, G G G is always a subgroup of C 6 v C_{6v} C 6 v .
The point groups relevant here are the subgroups of the regular hexagonal cell. The symmetry classes relevant to the present analysis are C 1 C_1 C 1 , C s C_s C s , C 2 C_2 C 2 , C 2 v C_{2v} C 2 v , C 3 C_3 C 3 , C 3 v C_{3v} C 3 v , C 6 C_6 C 6 , and C 6 v C_{6v} C 6 v . In Schönflies notation [44 ] , C n C_n C n denotes a group generated by an n n n -fold rotation about the axis normal to the MTS, whereas C n v C_{nv} C n v additionally contains vertical mirror symmetries. The groups C s C_s C s and C 1 C_1 C 1 denote, respectively, a single-mirror group and the trivial group containing only the identity. Since C n C_n C n denotes the cyclic group generated by a rotation of 2 π / n 2\pi/n 2 π / n about the axis normal to the sheet, C n = { E , R 2 π / n , … , R 2 π ( n − 1 ) / n } C_n = \{E, R_{2\pi/n}, \ldots, R_{2\pi(n-1)/n}\} C n = { E , R 2 π / n , … , R 2 π ( n − 1 ) / n } has order n n n , where E E E is the identity. Adding a mirror plane containing that axis (a vertical mirror, σ v \sigma_v σ v ) generates a group C n v C_{nv} C n v , of the order 2 n 2n 2 n , which contains the n n n rotations of C n C_n C n together with n n n vertical mirror reflections. The group C s = { E , σ } C_s = \{E, \sigma\} C s = { E , σ } contains the identity and a single mirror, and C 1 = { E } C_1 = \{E\} C 1 = { E } is the trivial group with no symmetry beyond the identity. The subgroup chain relevant to a hexagonal lattice is therefore
C 1 ⊂ C 2 , C s ⊂ C 2 v , C 3 ⊂ C 3 v , C 6 ⊂ C 6 v (7) C_1 \subset C_2,\ C_s \subset C_{2v},\ C_3 \subset C_{3v},\ C_6 \subset C_{6v} \tag*{(7)} C 1 ⊂ C 2 , C s ⊂ C 2 v , C 3 ⊂ C 3 v , C 6 ⊂ C 6 v (7)
where C 3 ⊂ C 6 ⊂ C 6 v C_3 \subset C_6 \subset C_{6v} C 3 ⊂ C 6 ⊂ C 6 v and C 3 ⊂ C 3 v ⊂ C 6 v C_3 \subset C_{3v} \subset C_{6v} C 3 ⊂ C 3 v ⊂ C 6 v , so that every group containing a three-fold rotation inherits the isotropy constraint C ‾ ‾ = C I ‾ ‾ \overline{\overline{C}} = C\overline{\overline{I}} C = C I .
Because C ‾ ‾ \overline{\overline{C}} C is an equivalent observable of the sheet, it must be invariant under every symmetry operation g ∈ G g \in G g ∈ G , which in the tangential plane acts as a 2 × 2 2 \times2 2 × 2 orthogonal tensor representation D ‾ ‾ ( g ) \overline{\overline{D}}(g) D ( g ) [13 , 14 , 53 ]
D ‾ ‾ ( g ) ⋅ C ‾ ‾ ⋅ D ‾ ‾ T ( g ) = C ‾ ‾ , ∀ g ∈ G . (8) \overline{\overline{D}}(g) \cdot\overline{\overline{C}} \cdot\overline{\overline{D}}^{T}(g) = \overline{\overline{C}}, \qquad\forall g \in G. \tag*{(8)} D ( g ) ⋅ C ⋅ D T ( g ) = C , ∀ g ∈ G . (8)
where the superscript T T T denotes transpose. The explicit application of Eq. (8) to the symmetry classes considered here, together with the resulting tensor constraints, is given in Sec. S1 of the Supplemental Material [54 ] .
When G G G contains at least a three-fold rotation
C ‾ = C I , (9) \overline{C}=C I, \tag*{(9)} C = C I , (9)
so that the sheet is isotropic and described by a single scalar C C C , with I I I being the identity matrix. Reducing G G G below C 3 C_3 C 3 allows the two eigenvalues of C ‾ \overline{C} C to differ, and the principal axes to rotate, introducing at most two additional degrees of freedom. Table I lists the dyadic form prescribed by group theory [44 ] for each subgroup of C 6 v C_{6v} C 6 v , together with the resulting shape of the directional-capacitance indicator C ( u ^ ) = u ^ ⋅ C ‾ ⋅ u ^ / C ‾ C(\hat{\mathbf{u}})=\hat{\mathbf{u}}\cdot\overline{C}\cdot\hat{\mathbf{u}}/\overline{C} C ( u ^ ) = u ^ ⋅ C ⋅ u ^ / C , where u ^ = ( cos ( θ ) , sin ( θ ) ) \hat{\mathbf{u}}=(\cos(\theta),\sin(\theta)) u ^ = ( cos ( θ ) , sin ( θ )) in the in-plane tangential vector along the applied tangential field Figure 3 .
Note that these are prescribed forms, not values. The explicit integral that evaluates each independent entry from the unit-cell geometry is derived in Sec. II C.
Quasi-static computation of the dyadic entries
For canonical geometries, such as circular and elliptical patches, and for arbitrary-shaped elements, the anisotropic MTS impedance can be obtained through reduced analytical or aggregate spectral-domain models [31 , 55 ] . Tensor impedance surfaces have also been modeled at the printed-circuit level by retrieving the effective tensor sheet impedance from full-wave simulations [17 ] . In contrast, the cell problem developed here computes the full quasi-static capacitive dyadic directly from the metal topology Σ \Sigma Σ , without assuming a canonical patch shape.
To compute the entries of C ‾ \overline{C} C , we impress over the whole unit cell a uniform static electric field of amplitude E 0 E_0 E 0 , directed along e ^ x \hat{\mathbf{e}}_x e ^ x or e ^ y \hat{\mathbf{e}}_y e ^ y . The impressed field induces surface charge on the metal patch, which redistributes until the tangential field on the metal is canceled; each connected metal component thereby becomes an equipotential surface. In the gap Ω = U ∖ Σ \Omega=U\setminus\Sigma Ω = U ∖ Σ the total field is the impressed field distorted by the induced charge,
E ξ c e l l ( r ) = E 0 [ e ^ ξ + ∇ w ξ ( r ) ] , r ∈ Ω , ξ ∈ { x , y } , (10) \mathbf{E}^{\mathrm{cell}}_{\xi}(\mathbf{r})=E_0\left[\hat{\mathbf{e}}_{\xi}+\nabla w_{\xi}(\mathbf{r})\right],\qquad\mathbf{r}\in\Omega,\quad\xi\in\{x,y\}, \tag*{(10)} E ξ cell ( r ) = E 0 [ e ^ ξ + ∇ w ξ ( r ) ] , r ∈ Ω , ξ ∈ { x , y } , (10)
in units of (V/m). Here e ^ ξ \hat{\mathbf{e}}_{\xi} e ^ ξ is the dimensionless unit vector of the imposed direction, and w ξ ( r ) w_{\xi}(\mathbf{r}) w ξ ( r ) is the periodic corrector potential [46 ] , in units of (m), whose gradient is dimensionless; the bracket is therefore the dimensionless normalised cell field E c e l l / E 0 \mathbf{E}^{\mathrm{cell}}/E_0 E cell / E 0 , and the amplitude E 0 E_0 E 0 carries the physical units. Equivalently, for a general dimensionless imposed direction v = v x e ^ x + v y e ^ y \mathbf{v}=v_x\hat{\mathbf{e}}_x+v_y\hat{\mathbf{e}}_y v = v x e ^ x + v y e ^ y , the associated electrostatic potential is
Φ v ( r ) = − E 0 [ v ⋅ r + w v ( r ) ] , (11) \Phi_{\mathbf{v}}(\mathbf{r})=-E_0\left[\mathbf{v}\cdot\mathbf{r}+w_{\mathbf{v}}(\mathbf{r})\right], \tag*{(11)} Φ v ( r ) = − E 0 [ v ⋅ r + w v ( r ) ] , (11)
so that E v c e l l = − ∇ Φ v = E 0 ( v + ∇ w v ) \mathbf{E}^{\mathrm{cell}}_{\mathbf{v}}=-\nabla\Phi_{\mathbf{v}}=E_0(\mathbf{v}+\nabla w_{\mathbf{v}}) E v cell = − ∇ Φ v = E 0 ( v + ∇ w v ) , and by linearity of the electrostatic problem w v = v x w x + v y w y w_{\mathbf{v}}=v_xw_x+v_yw_y w v = v x w x + v y w y , with w x = w e ^ x w_x=w_{\hat{\mathbf{e}}_x} w x = w e ^ x and w y = w e ^ y w_y=w_{\hat{\mathbf{e}}_y} w y = w e ^ y the two basis correctors of Eq. (10) . The corrector is determined by a constrained electrostatic cell problem. Let Σ = ⋃ m = 1 N c Σ m \Sigma=\bigcup_{m=1}^{N_c}\Sigma_m Σ = ⋃ m = 1 N c Σ m denote the decomposition of the metallic region into N c N_c N c electrically disconnected perfect-electric-conductor (PEC) components, where connectivity is understood on the periodically identified cell. The corrector associated with the imposed direction v \mathbf{v} v is defined by the minimisation
w v = arg min w ∈ A v 1 2 ∣ U ∣ ∫ Ω ∣ v + ∇ w ∣ 2 d A , (12) w_{\mathbf{v}}=\underset{w\in A_{\mathbf{v}}}{\operatorname{arg\,min}}\ \frac{1}{2|U|}\int_{\Omega}\left|\mathbf{v}+\nabla w\right|^2\,\mathrm{d}A, \tag*{(12)} w v = w ∈ A v arg min 2∣ U ∣ 1 ∫ Ω ∣ v + ∇ w ∣ 2 d A , (12)
where A v A_{\mathbf{v}} A v is the admissible space of periodic finite-energy potentials
A v = { w ∈ H p e r 1 ( Ω ) : v ⋅ r + w ( r ) = c m on ∂ Σ m , ∫ Ω w d A = 0 } . (13) A_{\mathbf{v}}=\left\{w\in H^1_{\mathrm{per}}(\Omega):\mathbf{v}\cdot\mathbf{r}+w(\mathbf{r})=c_m\ \text{on }\partial\Sigma_m,\quad\int_{\Omega}w\,\mathrm{d}A=0\right\}. \tag*{(13)} A v = { w ∈ H per 1 ( Ω ) : v ⋅ r + w ( r ) = c m on ∂ Σ m , ∫ Ω w d A = 0 } . (13)
c m c_m c m is the normalised equipotential level of the m m m th floating PEC component, m = 1 , … , N c m=1,\ldots,N_c m = 1 , … , N c . The corresponding equivalent potential is − E 0 c m -E_0c_m − E 0 c m . Since the normalised cell field is v + ∇ w v \mathbf{v}+\nabla w_{\mathbf{v}} v + ∇ w v , it is the total affine potential v ⋅ r + w v \mathbf{v}\cdot\mathbf{r}+w_{\mathbf{v}} v ⋅ r + w v , rather than the corrector alone, that is constant on each PEC component. Free variation of the c m c_m c m imposes the componentwise charge-neutrality conditions
∫ ∂ Σ m n ^ ⋅ ( v + ∇ w v ) d s = 0 , m = 1 , … , N c , (14) \int_{\partial\Sigma_m}\hat{\mathbf{n}}\cdot\left(\mathbf{v}+\nabla w_{\mathbf{v}}\right)\,\mathrm{d}s=0,\qquad m=1,\ldots,N_c, \tag*{(14)} ∫ ∂ Σ m n ^ ⋅ ( v + ∇ w v ) d s = 0 , m = 1 , … , N c , (14)
so that each floating component polarises without acquiring net charge. The Euler–Lagrange equations, affine-periodic boundary conditions, discrete potential basis, gauge fixing, and triangular-graph discretisation used in the numerical solution are detailed in § 3 of the Supplemental Material [54 ] .
For the free-standing patch arrays considered here, the gap region Ω \Omega Ω is filled with air, of permittivity ε 0 \varepsilon_0 ε 0 . Multiplying the minimum value of the functional in Eq. (12) by ε 0 E 0 2 \varepsilon_0E_0^2 ε 0 E 0 2 gives the cell-averaged electrostatic energy density associated with the auxiliary two-dimensional cell problem,
W e ( v ) = ε 0 E 0 2 2 ∣ U ∣ ∫ Ω ∣ v + ∇ w v ∣ 2 d A = 1 2 E 0 2 v ⋅ C ‾ ⋅ v , (15) W_{\mathrm{e}}(\mathbf{v})=\frac{\varepsilon_0E_0^2}{2|U|}\int_{\Omega}\left|\mathbf{v}+\nabla w_{\mathbf{v}}\right|^2\mathrm{d}A=\frac{1}{2}E_0^2\,\mathbf{v}\cdot\overline{\mathbf{C}}\cdot\mathbf{v}, \tag*{(15)} W e ( v ) = 2∣ U ∣ ε 0 E 0 2 ∫ Ω ∣ v + ∇ w v ∣ 2 d A = 2 1 E 0 2 v ⋅ C ⋅ v , (15)
in units of (J/m3 ^3 3 ). Identifying the bilinear form, and writing the entries in terms of the dimensionless normalised cell fields E ξ c e l l / E 0 = e ^ ξ + ∇ w ξ \mathbf{E}^{\mathrm{cell}}_{\xi}/E_0=\hat{\mathbf{e}}_{\xi}+\nabla w_{\xi} E ξ cell / E 0 = e ^ ξ + ∇ w ξ of Eq. (10) yields the capacitance integrals
C ξ η [ Σ ] = ε 0 ∣ U ∣ ∫ Ω ( e ^ ξ + ∇ w ξ ) ⋅ ( e ^ η + ∇ w η ) d A , ξ , η ∈ { x , y } , (16) C_{\xi\eta}[\Sigma]=\frac{\varepsilon_0}{|U|}\int_{\Omega}\left(\hat{\mathbf{e}}_{\xi}+\nabla w_{\xi}\right)\cdot\left(\hat{\mathbf{e}}_{\eta}+\nabla w_{\eta}\right)\,\mathrm{d}A,\qquad\xi,\eta\in\{x,y\}, \tag*{(16)} C ξ η [ Σ ] = ∣ U ∣ ε 0 ∫ Ω ( e ^ ξ + ∇ w ξ ) ⋅ ( e ^ η + ∇ w η ) d A , ξ , η ∈ { x , y } , (16)
in which the impressed amplitude E 0 E_0 E 0 has canceled. For a patch array printed between air and a semi-infinite dielectric substrate, ε 0 \varepsilon_0 ε 0 in Eqs. (15) and (17) may be replaced by the quasi-static effective permittivity
ε e f f = ε + + ε − 2 = ε 0 + ε s 2 = ε 0 1 + ε r , s 2 . (17) \varepsilon_{\mathrm{eff}}=\frac{\varepsilon^{+}+\varepsilon^{-}}{2}=\frac{\varepsilon_0+\varepsilon_s}{2}=\varepsilon_0\frac{1+\varepsilon_{r,s}}{2}. \tag*{(17)} ε eff = 2 ε + + ε − = 2 ε 0 + ε s = ε 0 2 1 + ε r , s . (17)
where ε s = ε 0 ε r , s \varepsilon_s=\varepsilon_0\varepsilon_{r,s} ε s = ε 0 ε r , s is the substrate permittivity. The dyadic defined by Eq. (16) is symmetric and positive definite, with units of (F/m). It is positive definite for the gap-connected geometries of Sec. II A, and it reproduces the tensor form of Table 1 from the symmetry of Σ \Sigma Σ . Figure 4 (a) and (Figure 4 b) show the resulting energy-density distributions ∣ e ^ x + ∇ w x ∣ 2 |\hat{\mathbf{e}}_x+\nabla w_x|^2 ∣ e ^ x + ∇ w x ∣ 2 and ∣ e ^ y + ∇ w y ∣ 2 |\hat{\mathbf{e}}_y+\nabla w_y|^2 ∣ e ^ y + ∇ w y ∣ 2 . The capacitance integrals over the gap region Ω \Omega Ω in Eq. (16) , give the entries of the capacitance dyad C ‾ \overline{\mathbf{C}} C . The diagonal entries C x x C_{xx} C xx and C y y C_{yy} C y y are the self-energy integrals of the x x x - and y y y -directed cell fields with themselves, while the off-diagonal entries C x y = C y x C_{xy}=C_{yx} C x y = C y x is determined from the cross-field overlap. Equation (16) is evaluated numerically for arbitrary patch geometries in § 2.3 .
Table 1. Group-theoretically prescribed form of C ‾ \overline{C} C for the principal subgroups of C 6 v C_{6v} C 6 v . Mirror symmetries pin the eigenvectors of C ‾ \overline{C} C : along the mirror axes for C 2 v C_{2v} C 2 v , and along and perpendicular to the mirror (orientation φ \varphi φ ) for C s C_s C s , with the two eigenvalues free in both cases. For C 1 C_1 C 1 and C 2 C_2 C 2 both eigenvalues and eigenvector orientation are free, since a two-fold rotation acts in the plane as − I -\mathit{I} − I and places no constraint on a symmetric dyadic. Degrees of freedom are independent real scalars for a lossless reciprocal sheet.
Figure 4. Normalised energy-density distributions for a representative irregular patch geometry Σ \Sigma Σ , computed from Eqs. (10) –(15) . The gap Ω \Omega Ω is shown; the deep-blue tessellated region is the metal patch Σ \Sigma Σ . (a) Distribution of ∣ e ^ x + ∇ w x ∣ 2 |\hat{\mathbf{e}}_x+\nabla w_x|^2 ∣ e ^ x + ∇ w x ∣ 2 for the x x x -directed loading. (b) Distribution of ∣ e ^ y + ∇ w y ∣ 2 |\hat{\mathbf{e}}_y+\nabla w_y|^2 ∣ e ^ y + ∇ w y ∣ 2 for the y y y -directed loading. Both distributions are normalised and displayed on logarithmic colour scales. Field concentration near narrow opening and irregular patch edges produces enhanced energy density, increasing the local contributions to the corresponding diagonal capacitance response. The different spatial distributions under the two loading directions illustrate the direction-dependent response of the irregular geometry, while C x y = C y x C_{xy}=C_{yx} C x y = C y x is determined by the cross-field overlap in Eq. (16)
From geometric integral to equivalent sheet capacitance
Equation (17) defines the symmetric cell-response dyadic C ‾ [ Σ ] \overline{\mathbf{C}}[\Sigma] C [ Σ ] , in units of (F/m), for the auxiliary two-dimensional electrostatic problem. Its construction uses the periodic-corrector framework of homogenization [56 ] and the variational-energy description of effective tensors [46 ] , with the PEC constraints specified in the preceding subsection. The cell response includes the isotropic background of the empty cell as well as the metal-induced contribution. It is distinct from the equivalent sheet capacitance dyadic C ‾ e q v \overline{\mathbf{C}}_{\mathrm{eqv}} C eqv , in units of (F), which enters the transmission-line model: equivalent sheet coefficients belong to an interface homogenization description [57 ] . We first subtract the background and normalise the remaining cell response, then introduce an explicit reference-energy matching assumption to set its equivalent sheet scale.
Figure 5 shows the empty cell, the reference patch Σ r e f \Sigma_{\mathrm{ref}} Σ ref , and the target geometry Σ \Sigma Σ . In the homogeneous free-space background, the empty-cell correctors vanish, giving
Figure 5. Cell and reference geometries used in the conversion from the geometric electrostatic response to the equivalent sheet capacitance. (a) Empty cell: the cell integral gives the isotropic background contribution C ‾ [ ∅ ] = ε 0 I ‾ \overline{\mathbf{C}}[\varnothing] = \varepsilon_0\overline{\mathbf{I}} C [ ∅ ] = ε 0 I . (b) Regular-hexagonal reference patch Σ r e f \Sigma_{\mathrm{ref}} Σ ref , used to define the scalar reference excess p r e f p_{\mathrm{ref}} p ref and the equivalent reference scale. (c) Arbitrary gap-connected patch geometry Σ \Sigma Σ , here comprising two electrically disconnected PEC components Σ 1 ∪ Σ 2 \Sigma_1 \cup\Sigma_2 Σ 1 ∪ Σ 2 . Its auxiliary cell response is converted into a model sheet capacitance using the same reference scale.
C ‾ [ ∅ ] = ε 0 I ‾ . (18) \overline{\mathbf{C}}[\varnothing] = \varepsilon_0\overline{\mathbf{I}}. \tag*{(18)} C [ ∅ ] = ε 0 I . (18)
We define the metal-induced excess by
C ‾ e x c [ Σ ] = C ‾ [ Σ ] − C ‾ [ ∅ ] . (19) \overline{\mathbf{C}}_{\mathrm{exc}}[\Sigma] = \overline{\mathbf{C}}[\Sigma] - \overline{\mathbf{C}}[\varnothing]. \tag*{(19)} C exc [ Σ ] = C [ Σ ] − C [ ∅ ] . (19)
Thus C ‾ e x c [ ∅ ] = 0 \overline{\mathbf{C}}_{\mathrm{exc}}[\varnothing] = \mathbf{0} C exc [ ∅ ] = 0 . Subtracting an isotropic dyadic preserves the principal directions of the raw cell response, but generally changes its eigenvalue ratio. For the isolated PEC components considered here, the equipotential boundary condition permits each corrector to be extended continuously into the metal, with
∇ w ξ = − e ^ ξ in Σ . \nabla w_{\xi} = -\hat{\mathbf{e}}_{\xi}\quad\text{in }\Sigma. ∇ w ξ = − e ^ ξ in Σ.
The normalised total field e ^ ξ + ∇ w ξ \hat{\mathbf{e}}_{\xi} + \nabla w_{\xi} e ^ ξ + ∇ w ξ therefore vanishes there, so the integration domain in (16) can be extended from Ω \Omega Ω to the complete cell U U U . Periodicity of the extended corrector gives
∫ U ∇ w ξ d A = 0. \int_U \nabla w_{\xi}\,\mathrm{d}A = 0. ∫ U ∇ w ξ d A = 0.
On expanding the integrand, both cross terms consequently vanish, while the constant-field term gives the empty-cell contribution. Subtracting it according to (19) yields
( C ‾ e x c [ Σ ] ) ξ η = ε 0 ∣ U ∣ ∫ U ∇ w ξ ⋅ ∇ w η d A , ξ , η ∈ { x , y } . (20) \left(\overline{\mathbf{C}}_{\mathrm{exc}}[\Sigma]\right)_{\xi\eta}
=
\frac{\varepsilon_0}{|U|}
\int_U \nabla w_{\xi}\cdot\nabla w_{\eta}\,\mathrm{d}A,
\qquad\xi,\eta\in\{x,y\}. \tag*{(20)} ( C exc [ Σ ] ) ξ η = ∣ U ∣ ε 0 ∫ U ∇ w ξ ⋅ ∇ w η d A , ξ , η ∈ { x , y } . (20)
This integral includes the corrector extension inside the metal; its domain is U U U , not only the gap Ω \Omega Ω . For every real in-plane vector q = q x x ^ + q y y ^ \mathbf{q}=q_x\hat{\mathbf{x}}+q_y\hat{\mathbf{y}} q = q x x ^ + q y y ^ ,
q ⋅ C ‾ e x c ⋅ q = ε 0 ∣ U ∣ ∫ U ∣ q x ∇ w x + q y ∇ w y ∣ 2 d A ≥ 0. (21) \mathbf{q}\cdot\overline{\mathbf{C}}_{\mathrm{exc}}\cdot\mathbf{q}
=
\frac{\varepsilon_0}{|U|}
\int_U \left|q_x\nabla w_x+q_y\nabla w_y\right|^2\,\mathrm{d}A
\ge0. \tag*{(21)} q ⋅ C exc ⋅ q = ∣ U ∣ ε 0 ∫ U ∣ q x ∇ w x + q y ∇ w y ∣ 2 d A ≥ 0. (21)
The excess is therefore symmetric and positive semidefinite within the stated two-dimensional problem. A charge-free gap cavity enclosed entirely by one connected PEC component has constant boundary potential and hence zero field by electrostatic uniqueness [58 ] . Filling only such cavities H k H_k H k therefore leaves the auxiliary cell response unchanged:
Σ e f f = Σ ∪ ⋃ k H k , C ‾ [ Σ e f f ] = C ‾ [ Σ ] . (22) \Sigma_{\mathrm{eff}}=\Sigma\cup\bigcup_k H_k,
\qquad
\overline{\mathbf{C}}[\Sigma_{\mathrm{eff}}]=\overline{\mathbf{C}}[\Sigma]. \tag*{(22)} Σ eff = Σ ∪ k ⋃ H k , C [ Σ eff ] = C [ Σ ] . (22)
We retain Σ \Sigma Σ below. This equivalence is restricted to the auxiliary two-dimensional problem: open notches and boundary-connected slots are not filled, and apertures in a physically thin sheet are not thereby equivalent to solid metal.
To display the anisotropy, each nonzero dyadic M ‾ ∈ { C ‾ [ Σ ] , C ‾ [ ∅ ] , C ‾ e x c [ Σ ] } \overline{\mathbf{M}}\in\{\overline{\mathbf{C}}[\Sigma],\overline{\mathbf{C}}[\varnothing],\overline{\mathbf{C}}_{\mathrm{exc}}[\Sigma]\} M ∈ { C [ Σ ] , C [ ∅ ] , C exc [ Σ ]} is represented by the normalized directional-capacitance indicator
M ~ ( u ^ ) = u ^ ⋅ M ‾ ⋅ u ^ , M ~ = M ‾ 1 2 tr ( M ‾ ) , (23) \widetilde{M}(\hat{\mathbf{u}})=\hat{\mathbf{u}}\cdot\overline{\mathbf{M}}\cdot\hat{\mathbf{u}},
\qquad
\widetilde{\mathbf{M}}=\frac{\overline{\mathbf{M}}}{\frac{1}{2}\operatorname{tr}(\overline{\mathbf{M}})}, \tag*{(23)} M ( u ^ ) = u ^ ⋅ M ⋅ u ^ , M = 2 1 tr ( M ) M , (23)
plotted as the radial coordinate for
u ^ ( θ ) = cos θ x ^ + sin θ y ^ . (24) \hat{\mathbf{u}}(\theta)=\cos\theta\,\hat{\mathbf{x}}+\sin\theta\,\hat{\mathbf{y}}. \tag*{(24)} u ^ ( θ ) = cos θ x ^ + sin θ y ^ . (24)
Each nonzero dyadic above is positive semidefinite and therefore has positive trace. This trace normalisation removes the absolute scale while preserving the eigenvalue ratio and principal axes. Figure 6 compares the raw response, the analytically isotropic empty-cell background, and the excess.
Figure 6. Empty-cell subtraction at the dyadic level Eq. (19) . Each dyadic is represented by its normalised directional-capacitance indicator and compared with the dashed isotropic reference circle. Normalisation by one half of the trace removes the absolute scale and highlights the anisotropy and principal axes. (a) Raw cell-response dyadic C ‾ ~ [ Σ ] \widetilde{\overline{C}}[\Sigma] C [ Σ ] . (b) Empty-cell response C ‾ ~ [ ∅ ] = I ‾ \widetilde{\overline{C}}[\varnothing] = \overline{I} C [ ∅ ] = I , which coincides with the reference circle. (c) Excess dyadic C ‾ ~ e x c [ Σ ] \widetilde{\overline{C}}_{\mathrm{exc}}[\Sigma] C exc [ Σ ] , obtained by subtracting the isotropic background before normalisation. Its principal directions coincide with those of the raw response, whereas its eigenvalue ratio generally differs.
The target and reference cell problems use the same lattice, background permittivity, and boundary conditions. The regular-hexagonal reference has an isotropic excess by the symmetry of the complete cell, consistent with the tensor constraints in Table I. Define
S ‾ [ Σ ] = C ‾ e x c [ Σ ] p r e f , p r e f = 1 2 tr ( C ‾ e x c [ Σ r e f ] ) > 0. (25) \overline{S}[\Sigma] = \frac{\overline{C}_{\mathrm{exc}}[\Sigma]}{p_{\mathrm{ref}}},
\qquad
p_{\mathrm{ref}} = \frac{1}{2}\operatorname{tr}\left(\overline{C}_{\mathrm{exc}}[\Sigma_{\mathrm{ref}}]\right) > 0. \tag*{(25)} S [ Σ ] = p ref C exc [ Σ ] , p ref = 2 1 tr ( C exc [ Σ ref ] ) > 0. (25)
Here p r e f p_{\mathrm{ref}} p ref has units of (F/m), whereas S ‾ \overline{S} S is dimensionless. Consequently,
S ‾ [ Σ r e f ] = I ‾ , S ‾ [ ∅ ] = 0 ‾ . (26) \overline{S}[\Sigma_{\mathrm{ref}}] = \overline{I},
\qquad
\overline{S}[\varnothing] = \overline{0}. \tag*{(26)} S [ Σ ref ] = I , S [ ∅ ] = 0 . (26)
These are exact anchors of the continuum normalisation; their numerical recovery requires consistent target and reference discretisations. The conversion to equivalent sheet units requires an additional modeling assumption. For a real tangential loading E 0 v E_0\mathbf{v} E 0 v , where E 0 E_0 E 0 has units of (V/m) and v \mathbf{v} v is a dimensionless in-plane vector, the auxiliary excess energy density is
Δ W e ( v ) = E 0 2 2 v ⋅ C ‾ e x c ⋅ v , (27) \Delta W_{\mathrm{e}}(\mathbf{v}) = \frac{E_0^2}{2}\,\mathbf{v}\cdot\overline{C}_{\mathrm{exc}}\cdot\mathbf{v}, \tag*{(27)} Δ W e ( v ) = 2 E 0 2 v ⋅ C exc ⋅ v , (27)
in units of (J/m3 ^3 3 ). We approximate the sheet energy per unit area by a common conversion length ℓ r e f > 0 \ell_{\mathrm{ref}} > 0 ℓ ref > 0 multiplying this excess energy density:
U s m o d e l ( v ) = E 0 2 2 v ⋅ C ‾ e q v ⋅ v = ℓ r e f Δ W e ( v ) . (28) \mathcal{U}_{\mathrm{s}}^{\mathrm{model}}(\mathbf{v}) = \frac{E_0^2}{2}\,\mathbf{v}\cdot\overline{C}_{\mathrm{eqv}}\cdot\mathbf{v}
= \ell_{\mathrm{ref}}\,\Delta W_{\mathrm{e}}(\mathbf{v}). \tag*{(28)} U s model ( v ) = 2 E 0 2 v ⋅ C eqv ⋅ v = ℓ ref Δ W e ( v ) . (28)
The same macroscopic tangential-field convention is used on both sides. This reference-energy ansatz assumes one scalar conversion length for both in plane directions and all target geometries in the selected family. It does not follow from the two-dimensional cell problem or the normalisation anchors alone, and is not a substitute for an equivalent interface-cell derivation [57 ] . Since Eq. (28) holds for every v \mathbf{v} v , equality of the quadratic forms gives C ‾ e q v = ℓ r e f C ‾ e x c \overline{C}_{\mathrm{eqv}} = \ell_{\mathrm{ref}}\overline{C}_{\mathrm{exc}} C eqv = ℓ ref C exc . Matching the isotropic reference to its prescribed equivalent sheet capacitance C r e f e q v C_{\mathrm{ref}}^{\mathrm{eqv}} C ref eqv fixes
ℓ r e f p r e f = C r e f e q v , ℓ r e f = C r e f e q v p r e f . (29) \ell_{\mathrm{ref}}p_{\mathrm{ref}} = C_{\mathrm{ref}}^{\mathrm{eqv}},
\qquad
\ell_{\mathrm{ref}} = \frac{C_{\mathrm{ref}}^{\mathrm{eqv}}}{p_{\mathrm{ref}}}. \tag*{(29)} ℓ ref p ref = C ref eqv , ℓ ref = p ref C ref eqv . (29)
The resulting equivalent sheet-capacitance model is
C ‾ e q v [ Σ ] = C r e f e q v S ‾ [ Σ ] = C r e f e q v p r e f ( C ‾ [ Σ ] − ε 0 I ‾ ) . (30) \begin{aligned}
\overline{C}_{\mathrm{eqv}}[\Sigma] = C_{\mathrm{ref}}^{\mathrm{eqv}}\overline{S}[\Sigma] \\
&= \frac{C_{\mathrm{ref}}^{\mathrm{eqv}}}{p_{\mathrm{ref}}}\left(\overline{C}[\Sigma] - \varepsilon_0\overline{I}\right). \tag*{(30)}
\end{aligned} C eqv [ Σ ] = C ref eqv S [ Σ ] = p ref C ref eqv ( C [ Σ ] − ε 0 I ) . (30)
The ratio in Eq. (29) has units of length (m), so Eq. (30) has units of (F). We provide the detailed derivation of the excess dyadic and reference normalization in Sec. S3 of the Supplemental Material [54 ] , where we also establish the uniqueness of the sheet calibration within the assumed scalar-affine class and clarify the scope of the common conversion-length assumption. No nonlinear eigenvalue transformation is introduced. In the orthonormal principal basis of the symmetric excess dyadic [59 ] ,
S ‾ [ Σ ] = ∑ i = 1 2 s i v ^ i ⊗ v ^ i , s i ≥ 0 , (31) \overline{S}[\Sigma] = \sum_{i=1}^{2} s_i\,\hat{\mathbf{v}}_i \otimes\hat{\mathbf{v}}_i,
\qquad
s_i \ge0, \tag*{(31)} S [ Σ ] = i = 1 ∑ 2 s i v ^ i ⊗ v ^ i , s i ≥ 0 , (31)
the equivalent principal capacitances are C i e q v = C r e f e q v s i C_i^{\mathrm{eqv}} = C_{\mathrm{ref}}^{\mathrm{eqv}}s_i C i eqv = C ref eqv s i . Positive reference scaling therefore preserves the excess dyadic’s principal directions, its eigenvalue ratio when both eigenvalues are positive, its nonnegative energy, and the symmetry constraints of Table I. The reference value is an independent input. For the free-standing regular-hexagonal array, we retain the reference
C r e f e q v = 2 ε 0 A i s o , A i s o = 3 D h 2 π ln [ csc ( π g 2 D h ) ] . (32) C_{\mathrm{ref}}^{\mathrm{eqv}} = 2\varepsilon_0 A_{\mathrm{iso}},
\qquad
A_{\mathrm{iso}} = \frac{\sqrt{3}D_h}{2\pi}\ln\left[\csc\left(\frac{\pi g}{2D_h}\right)\right]. \tag*{(32)} C ref eqv = 2 ε 0 A iso , A iso = 2 π 3 D h ln [ csc ( 2 D h π g ) ] . (32)
where D h D_h D h and g g g denote the period parameter and uniform interpatch gap of the reference geometry, respectively. The logarithmic gap dependence is motivated by the canonical electrically dense grid models of Luukkonen et al. [10 ] , which treat strips and square patches. The hexagonal prefactor in Eq. (32) is part of the specified reference approximation, not an exact consequence of that square-patch result. Its accuracy must be established independently for the reference array. The relation C ‾ r e f = 2 ε 0 A i s o \overline{C}_{\mathrm{ref}}=2\varepsilon_0 A_{\mathrm{iso}} C ref = 2 ε 0 A iso sets the free-space grid-length convention used in the dispersion model. A separately computed equivalent reference value can replace Eq. (32) without changing the normalisation in Eq. (25) .
For any fixed positive reference value, the complete conversion preserves the two anchors:
C ‾ e q v [ ∅ ] = 0 ‾ , C ‾ e q v [ Σ r e f ] = C r e f I ‾ . (33) \overline{C}_{\mathrm{eqv}}[\varnothing]=\overline{0},\qquad\overline{C}_{\mathrm{eqv}}[\Sigma_{\mathrm{ref}}]=C_{\mathrm{ref}}\overline{I}. \tag*{(33)} C eqv [ ∅ ] = 0 , C eqv [ Σ ref ] = C ref I . (33)
The construction combines the energy-derived excess dyadic with one common reference scale, without eigenvalue compression or geometry-dependent conversion factors. Symmetry, nonnegative excess, and closed-cavity equivalence are properties of the auxiliary cell problem. The accuracy of the reference scale and of the common conversion length remains separate from those exact two-dimensional results.
Quasi-static surface-wave dispersion
The equivalent sheet-capacitance dyadic C ‾ e q v [ Σ ] \overline{C}_{\mathrm{eqv}}[\Sigma] C eqv [ Σ ] from Eq. (30) enters the electromagnetic model through the averaged boundary condition at the MTS plane at z = 0 z=0 z = 0 . In the quasi-static limit, the grid admittance for a free-standing patch array is purely capacitive [10 ] . Hence, the homogenised sheet-admittance dyadic is
Y ‾ s ( ω ) = j ω C ‾ e q v [ Σ ] . (34) \overline{Y}_{s}(\omega)=j\omega\overline{C}_{\mathrm{eqv}}[\Sigma]. \tag*{(34)} Y s ( ω ) = j ω C eqv [ Σ ] . (34)
For a transverse-electric (TE) surface wave propagating along k t = k x x ^ + k y y ^ \mathbf{k}_{t}=k_x\hat{\mathbf{x}}+k_y\hat{\mathbf{y}} k t = k x x ^ + k y y ^ , with k ^ t = k t / ∣ k t ∣ \hat{\mathbf{k}}_{t}=\mathbf{k}_{t}/|\mathbf{k}_{t}| k ^ t = k t /∣ k t ∣ , the in-plane electric field is directed along e ^ T E = ( z ^ × k t ) / k t \hat{\mathbf{e}}_{\mathrm{TE}}=(\hat{\mathbf{z}}\times\mathbf{k}_{t})/k_t e ^ TE = ( z ^ × k t ) / k t . The sheet-capacitance response sampled by the modal electric field is therefore the quadratic projection e ^ T E ⋅ C ‾ e q v ⋅ e ^ T E \hat{\mathbf{e}}_{\mathrm{TE}}\cdot\overline{C}_{\mathrm{eqv}}\cdot\hat{\mathbf{e}}_{\mathrm{TE}} e ^ TE ⋅ C eqv ⋅ e ^ TE of the C ‾ e q v [ Σ ] \overline{C}_{\mathrm{eqv}}[\Sigma] C eqv [ Σ ] dyadic.
Transverse resonance conditions
For an isotropic sheet, or for propagation along a principal axis of an anisotropic sheet, the TE and TM polarisations are uncoupled. In this limit, a bound surface wave with transverse decay constant γ = k t 2 − k 0 2 \gamma=\sqrt{k_t^2-k_0^2} γ = k t 2 − k 0 2 (k t = ∣ k t ∣ k_t=|\mathbf{k}_t| k t = ∣ k t ∣ ; k t > k 0 k_t>k_0 k t > k 0 ) satisfies the two-sided scalar transverse-resonance condition on a free-standing sheet [9 , 10 ]
2 Y 0 p + Y g p = 0 , p ∈ { T E , T M } , (35) 2Y_0^p+Y_g^p=0,\qquad p\in\{\mathrm{TE},\mathrm{TM}\}, \tag*{(35)} 2 Y 0 p + Y g p = 0 , p ∈ { TE , TM } , (35)
where Y 0 p Y_0^p Y 0 p is the one-sided free-space line admittance and Y g p Y_g^p Y g p is the grid admittance for the transverse-electric (TE) or transverse-magnetic (TM) polarisation. Equation (35) is the factorised scalar form of the more general anisotropic matrix condition introduced in Sec. III.B.
For an uncoupled TM mode supported by a scalar capacitive patch array, both admittance of a capacitive sheet in Eq. (35) are positive imaginary
Y 0 T M = j ω ε 0 γ , Y g T M = j 2 α g η 0 . Y_0^{\mathrm{TM}}=j\frac{\omega\varepsilon_0}{\gamma},\qquad Y_g^{\mathrm{TM}}=j\frac{2\alpha_g}{\eta_0}. Y 0 TM = j γ ω ε 0 , Y g TM = j η 0 2 α g .
Consequently the TM condition cannot be equated to 0, and hence no bound TM solution exists. The uncoupled isotropic or principal-axis capacitive sheet model therefore supports a bound TE mode but not a bound TM mode [11 ] .
For the TE polarisation, whose electric field is perpendicular to k t \mathbf{k}_t k t , the free-space line admittance of a bound wave is
Y 0 T E = − j γ ω μ 0 . Y_0^{\mathrm{TE}}=-j\frac{\gamma}{\omega\mu_0}. Y 0 TE = − j ω μ 0 γ .
The Luukkonen grid admittance for a capacitive patch array in free space is [10 ]
Y g T E = j 2 α g η 0 F L ( k t ) , (36) Y_g^{\mathrm{TE}}=j\frac{2\alpha_g}{\eta_0}F_L(k_t), \tag*{(36)} Y g TE = j η 0 2 α g F L ( k t ) , (36)
where F L ( k t ) = [ 1 − k t 2 / ( 2 k 0 2 ) ] F_L(k_t)=[1-k_t^2/(2k_0^2)] F L ( k t ) = [ 1 − k t 2 / ( 2 k 0 2 )] is the k t k_t k t -dependent correction factor, and α g = k 0 A g \alpha_g=k_0A_g α g = k 0 A g is the dimensionless grid parameter, with A g A_g A g (m) the Luukkonen grid length of the patch array. Substituting Eq. (36) into Eq. (35) and using η 0 = ω μ 0 / k 0 \eta_0=\omega\mu_0/k_0 η 0 = ω μ 0 / k 0 gives the scalar TE dispersion condition
γ = k 0 α g F L ( k t ) . (37) \gamma=k_0\alpha_g F_L(k_t). \tag*{(37)} γ = k 0 α g F L ( k t ) . (37)
Equation (37) is valid in the window k 0 < k t < 2 k 0 k_0<k_t<\sqrt{2}k_0 k 0 < k t < 2 k 0 where F L ( k t ) > 0 F_L(k_t)>0 F L ( k t ) > 0 . Substituting k t 2 = k 0 2 + γ 2 k_t^2=k_0^2+\gamma^2 k t 2 = k 0 2 + γ 2 and solving the resulting quadratic in γ \gamma γ gives the closed-form root
γ = k 0 α g ( 1 + α g 2 − 1 ) , k t = k 0 1 + γ 2 k 0 2 . (38) \gamma=\frac{k_0}{\alpha_g}\left(\sqrt{1+\alpha_g^2}-1\right),\qquad k_t=k_0\sqrt{1+\frac{\gamma^2}{k_0^2}}. \tag*{(38)} γ = α g k 0 ( 1 + α g 2 − 1 ) , k t = k 0 1 + k 0 2 γ 2 . (38)
The solution of Eq. (38) automatically satisfies γ < k 0 \gamma<k_0 γ < k 0 and hence remains inside the validity window. Written in terms of the geometry-fixed grid length,
γ = 1 + ( k 0 A g ) 2 − 1 A g . \gamma=\frac{\sqrt{1+(k_0A_g)^2}-1}{A_g}. γ = A g 1 + ( k 0 A g ) 2 − 1 .
The branch remains confined to the light line at low frequency and breaks-off quadratically as k 0 A g k_0A_g k 0 A g grows, with the geometry entering through the single length A g A_g A g .
General local anisotropic condition and quasi-TE reduction
The transverse-resonance conditions of Sec. III.A are parameterised by a single geometry-fixed scalar A g A_g A g . For a patch geometry whose point group does not enforce in-plane isotropy, however, the sheet capacitance is dyadic and the effective response depends on the direction of propagation relative to its principal axes. Accordingly, the scalar grid length must be replaced by a direction-dependent projection of the capacitance dyadic. The equivalent sheet-capacitance dyadic defines the symmetric grid-length dyadic
A ‾ ‾ = C ‾ ‾ e q v [ Σ ] 2 ε 0 , (39) \overline{\overline{A}}=\frac{\overline{\overline{C}}_{\mathrm{eqv}}[\Sigma]}{2\varepsilon_0}, \tag*{(39)} A = 2 ε 0 C eqv [ Σ ] , (39)
in units of (m). For the isotropic reference patch,
A ‾ ‾ [ Σ r e f ] = A i s o I ‾ ‾ , \overline{\overline{A}}[\Sigma_{\mathrm{ref}}]=A_{\mathrm{iso}}\overline{\overline{I}}, A [ Σ ref ] = A iso I ,
recovering the scalar grid length introduced in Sec. III A. The symmetry constraints on the entries and principal axes of A ‾ ‾ \overline{\overline{A}} A for the different point groups are summarised in Table 1 .
For a nonzero in-plane wavevector
k t = k x x ^ + k y y ^ , k t = ∣ k t ∣ , \mathbf{k}_{t}=k_x\hat{\mathbf{x}}+k_y\hat{\mathbf{y}},\qquad k_t=\lvert\mathbf{k}_t\rvert, k t = k x x ^ + k y y ^ , k t = ∣ k t ∣ ,
we introduce the propagation-fixed directions
e ^ T M = k t k t , e ^ T E = z ^ × k t k t . (40) \hat{\mathbf{e}}_{\mathrm{TM}}=\frac{\mathbf{k}_t}{k_t},\qquad\hat{\mathbf{e}}_{\mathrm{TE}}=\frac{\hat{\mathbf{z}}\times\mathbf{k}_t}{k_t}. \tag*{(40)} e ^ TM = k t k t , e ^ TE = k t z ^ × k t . (40)
Here e ^ T M \hat{\mathbf{e}}_{\mathrm{TM}} e ^ TM and e ^ T E \hat{\mathbf{e}}_{\mathrm{TE}} e ^ TE denote the tangential electric-field directions of the TM and TE waves, respectively. At Γ \Gamma Γ , where k t = 0 \mathbf{k}_t=0 k t = 0 , these directions are understood in the one-sided limit along the selected propagation path.
For a general anisotropic sheet, rotation into this propagation-fixed basis produces longitudinal, transverse, and cross projections of A ‾ ‾ \overline{\overline{A}} A . Consequently, the local eigenmode is in general hybrid TE–TM rather than strictly TE. The corresponding local transverse-resonance condition may be written compactly as
det ( T ‾ ‾ 0 ) = 0 , (41) \det\left(\overline{\overline{T}}_0\right)=0, \tag*{(41)} det ( T 0 ) = 0 , (41)
where T ‾ ‾ 0 \overline{\overline{T}}_0 T 0 is the 2 × 2 2\times2 2 × 2 transverse-resonance dyadic in the { e ^ T M , e ^ T E } \{\hat{\mathbf{e}}_{\mathrm{TM}},\hat{\mathbf{e}}_{\mathrm{TE}}\} { e ^ TM , e ^ TE } basis. Its explicit construction, the resulting TE–TM cross-coupling term, and the corresponding Schur-complement reduction are given in Sec. S4 of the Supplemental Material [54 ] .
The reduced model used here follows the dominant transverse component of the TE-dominant branch. The effective grid length sampled by this component is therefore the quadratic projection
A 0 ( k t ) ≡ e ^ T E ⋅ A ‾ ‾ ⋅ e ^ T E = A x x k y 2 − 2 A x y k x k y + A y y k x 2 k t 2 . (42) A_0(\mathbf{k}_t)\equiv\hat{\mathbf{e}}_{\mathrm{TE}}\cdot\overline{\overline{A}}\cdot\hat{\mathbf{e}}_{\mathrm{TE}}=\frac{A_{xx}k_y^2-2A_{xy}k_xk_y+A_{yy}k_x^2}{k_t^2}. \tag*{(42)} A 0 ( k t ) ≡ e ^ TE ⋅ A ⋅ e ^ TE = k t 2 A xx k y 2 − 2 A x y k x k y + A y y k x 2 . (42)
For an isotropic sheet, A 0 = A g A_0=A_g A 0 = A g for every propagation direction. For an anisotropic sheet, A 0 A_0 A 0 varies continuously as k t \mathbf{k}_t k t rotates relative to the principal axes of A ‾ ‾ \overline{\overline{A}} A .
Using (39) , the local sheet-admittance dyadic is
Y ‾ ‾ s = j ω C ‾ ‾ e q v = j 2 k 0 η 0 A ‾ ‾ . (43) \overline{\overline{Y}}_s=j\omega\overline{\overline{C}}_{\mathrm{eqv}}=j\frac{2k_0}{\eta_0}\overline{\overline{A}}. \tag*{(43)} Y s = j ω C eqv = j η 0 2 k 0 A . (43)
Neglecting the explicit TE–TM hybridisation correction while retaining the dominant transverse response reduces the local eigencondition to the quasi-TE closure
2 Y 0 T E + j 2 k 0 η 0 A 0 ( k t ) F L ( k t ) = 0 , (44) 2Y_0^{\mathrm{TE}}+j\frac{2k_0}{\eta_0}A_0(\mathbf{k}_t)F_L(\mathbf{k}_t)=0, \tag*{(44)} 2 Y 0 TE + j η 0 2 k 0 A 0 ( k t ) F L ( k t ) = 0 , (44)
where F L ( k t ) F_L(\mathbf{k}_t) F L ( k t ) is the k t \mathbf{k}_t k t -dependent correction factor introduced in Sec. III A. (44) is the anisotropic extension of the scalar patch-array condition: the complete geometric anisotropy enters through A 0 ( k t ) A_0(\mathbf{k}_t) A 0 ( k t ) , while the scalar k t \mathbf{k}_t k t -dependent correction is retained through F L ( k t ) F_L(\mathbf{k}_t) F L ( k t ) . No tensorial extension of F L F_L F L to the TE–TM cross-coupling term is assumed.
The reduction is exact when the propagation direction is aligned with a principal axis of A ‾ ‾ \overline{\overline{A}} A , for which the TE–TM cross projection vanishes. Away from a principal axis, Eq. (45) approximates the hybrid condition Eq. (42). By Schur reduction in Sec. S4 of the Supplemental Material [54 ] , the hybrid eigencondition is equivalent to
T T E − T x 2 T T M = 0 , (45) T_{\mathrm{TE}}-\frac{T_x^2}{T_{\mathrm{TM}}}=0, \tag*{(45)} T TE − T TM T x 2 = 0 , (45)
whereas (44) retains only T T E = 0 T_{\mathrm{TE}}=0 T TE = 0 . We quantify the neglected hybridisation by the relative frequency shift between the quasi-TE root and the hybrid root at fixed k t \mathbf{k}_t k t ,
δ h y b ( k t ) = ∣ f T E ( k t ) − f h y b r i d ( k t ) ∣ f h y b r i d ( k t ) , (46) \delta_{\mathrm{hyb}}(\mathbf{k}_t)=\frac{\left|f_{\mathrm{TE}}(\mathbf{k}_t)-f_{\mathrm{hybrid}}(\mathbf{k}_t)\right|}{f_{\mathrm{hybrid}}(\mathbf{k}_t)}, \tag*{(46)} δ hyb ( k t ) = f hybrid ( k t ) ∣ f TE ( k t ) − f hybrid ( k t ) ∣ , (46)
which vanishes on principal axes (T x = 0 T_x=0 T x = 0 ) and is largest for low-symmetry cells. The full hybrid construction is given in Sec. S4 of the Supplemental Material [54 ] .
Substituting α g → k 0 A 0 ( k t ) \alpha_g\to k_0A_0(\mathbf{k}_t) α g → k 0 A 0 ( k t ) into (37) and retaining F L ( k t ) F_L(\mathbf{k}_t) F L ( k t ) gives the quasi-TE evanescent decay constant
Γ T L ( k 0 , k t ) = 1 + [ k 0 A 0 ( k t ) ] 2 − 1 A 0 ( k t ) , (47) \Gamma_{\mathrm{TL}}(k_0,\mathbf{k}_t)=\frac{\sqrt{1+[k_0A_0(\mathbf{k}_t)]^2}-1}{A_0(\mathbf{k}_t)}, \tag*{(47)} Γ TL ( k 0 , k t ) = A 0 ( k t ) 1 + [ k 0 A 0 ( k t ) ] 2 − 1 , (47)
in units of m − 1 \mathrm{m}^{-1} m − 1 . The k t \mathbf{k}_t k t -dependent correction factor is already contained in this root and must therefore not be applied a second time.
The corresponding anisotropic quasi-TE residual is
R 0 ( k 0 , k t ) = k t 2 − k 0 2 − Γ T L ( k 0 , k t ) , (48) R_0(k_0,\mathbf{k}_t)=\sqrt{k_t^2-k_0^2}-\Gamma_{\mathrm{TL}}(k_0,\mathbf{k}_t), \tag*{(48)} R 0 ( k 0 , k t ) = k t 2 − k 0 2 − Γ TL ( k 0 , k t ) , (48)
and the dispersion condition is
R 0 ( k 0 , k t ) = 0. R_0(k_0,\mathbf{k}_t)=0. R 0 ( k 0 , k t ) = 0.
Setting A 0 = A g A_0=A_g A 0 = A g recovers the scalar result of Sec. III A.
Since A 0 A_0 A 0 depends on the propagation direction, the resulting quasi-TE isofrequency contours are generally anisotropic, with principal directions inherited from A ‾ ‾ \overline{\overline{A}} A . In the weak-anisotropy limit they are approximately elliptical. For the positive-definite capacitive responses considered here the contours remain closed, in contrast to the hyperbolic contours produced by sign-indefinite anisotropic sheet responses [23 , 24 , 60 ] .
Geometry-derived first-shell Floquet correction
The condition R 0 = 0 R_0=0 R 0 = 0 provides the local dispersion law while the higher Floquet harmonics remain strongly evanescent. As k t \mathbf{k}_t k t approaches the Brillouin-zone boundary, however, the nearest nonzero reciprocal-lattice harmonics can become progressively less confined and provide the leading spatial-dispersion correction [32 , 33 , 61 , 62 ] . We therefore retain only the complete first reciprocal-shell as we will see in § 2.3 ,
G p ∈ { ± b 1 , ± b 2 , ± ( b 1 + b 2 ) } , p = 1 , … , 6 , (49) \mathbf{G}_p \in\{\pm\mathbf{b}_1,\ \pm\mathbf{b}_2,\ \pm(\mathbf{b}_1+\mathbf{b}_2)\}, \qquad p=1,\ldots,6, \tag*{(49)} G p ∈ { ± b 1 , ± b 2 , ± ( b 1 + b 2 )} , p = 1 , … , 6 , (49)
where the reciprocal basis is defined by a i ⋅ b j = 2 π δ i j \mathbf{a}_i\cdot\mathbf{b}_j=2\pi\delta_{ij} a i ⋅ b j = 2 π δ ij . The six vectors in Eq. (49) have the same reciprocal magnitude and constitute the complete nearest nonzero shell. The particular member, or pair of members, that becomes least evanescent depends on the propagation direction; retaining all six orders explicitly therefore avoids introducing a path-dependent spectral truncation [38 , 39 ] . The construction of the first three reciprocal shells, together with their Floquet propagation thresholds and evanescent decay lengths at Γ \Gamma Γ , M M M , and K K K , is given in Sec. S5 of the Supplemental Material [54 ] .
For the p p p th retained sideband, the shifted transverse wavevector and free-space decay constant are
q p = k t + G p , γ h p = ∣ q p ∣ 2 − k 0 2 , (50) \mathbf{q}_p=\mathbf{k}_t+\mathbf{G}_p,\qquad\gamma_{h_p}=\sqrt{\lvert\mathbf{q}_p\rvert^2-k_0^2}, \tag*{(50)} q p = k t + G p , γ h p = ∣ q p ∣ 2 − k 0 2 , (50)
for an evanescent order. Each sideband samples the same local sheet dyadic through its own TE polarization,
e ^ p = z ^ × q p ∣ q p ∣ , A p = e ^ p ⋅ A ‾ ‾ ⋅ e ^ p , (51) \hat{\mathbf{e}}_p=\frac{\hat{\mathbf{z}}\times\mathbf{q}_p}{\lvert\mathbf{q}_p\rvert},\qquad A_p=\hat{\mathbf{e}}_p\cdot\overline{\overline{A}}\cdot\hat{\mathbf{e}}_p, \tag*{(51)} e ^ p = ∣ q p ∣ z ^ × q p , A p = e ^ p ⋅ A ⋅ e ^ p , (51)
with the corresponding sheet-loaded decay Γ p \Gamma_p Γ p obtained from Eq. (47) evaluated at A p A_p A p . Each retained order also carries its own modal admittance [63 ] . Two quantities, both computed directly from the reconstructed boundary ∂ Σ \partial\Sigma ∂ Σ , then determine its coupling to the fundamental channel. Figure 7 shows the reconstructed polygonal boundary and the one-sided tangent–normal frame used to define these geometric coupling measures. The complete six-order construction, including the reciprocal-shell geometry, polygon-edge evaluation of the complex overlap, point-wise normalization of the directional-turning multipliers, is detailed in Sec. S6 of the Supplemental Material [54 ] . The reciprocal-shell geometry and its mirror pairing are illustrated in Supplemental Figure S2 [54 ] .
Figure 7. Random metallic patch inside the hexagonal unit cell and local boundary geometry at a representative vertex v \mathbf{v} v . The one-sided tangents t ^ ν ± \hat{\mathbf{t}}_{\nu}^{\pm} t ^ ν ± , outward normals n ^ ν ± \hat{\mathbf{n}}_{\nu}^{\pm} n ^ ν ± , and exterior turning angle Δ θ ν \Delta\theta_{\nu} Δ θ ν define the discrete-curvature contribution used in Eqs. (61) –(62) . The same reconstructed boundary ∂ Σ \partial\Sigma ∂ Σ is used in the Fourier overlap of Eq. (53) .
Figure S2. Complete first reciprocal shell and its mirror action on the Γ \Gamma Γ –K K K path. (a) The six vectors G p ∈ { ± b 1 , ± b 2 , ± ( b 1 + b 2 ) } \mathbf{G}_p \in\{\pm\mathbf{b}_1,\pm\mathbf{b}_2,\pm(\mathbf{b}_1+\mathbf{b}_2)\} G p ∈ { ± b 1 , ± b 2 , ± ( b 1 + b 2 )} and the shifted Floquet wavevectors q p = k t + G p \mathbf{q}_p=\mathbf{k}_t+\mathbf{G}_p q p = k t + G p at the representative point k t = 0.82 K \mathbf{k}_t=0.82\mathbf{K} k t = 0.82 K . (b) Reflection about the Γ \Gamma Γ –K K K propagation axis pairs the retained orders according to π = ( 1 3 ) ( 2 5 ) ( 4 6 ) \pi=(1\,3)(2\,5)(4\,6) π = ( 1 3 ) ( 2 5 ) ( 4 6 ) . As derived in Sec. S6E, the associated TE-compatible projector has rank three.
Boundary Fourier overlap of edge-charge profiles
We describe coupling between the fundamental Floquet order and the six retained first-shell orders through the normalized overlap of model edge-charge profiles. Charge-based modal descriptions motivate representing interactions through spatial charge distributions [64 , 65 ] , while Floquet expansions of surface-current profiles underlie related analytical models of metasurfaces [66 ] .
We parameterize the reconstructed boundary by arc-length ℓ ∈ [ 0 , L ] \ell\in[0,L] ℓ ∈ [ 0 , L ] , where L L L is its total length, and write r ( ℓ ) ∈ ∂ Σ \mathbf{r}(\ell)\in\partial\Sigma r ( ℓ ) ∈ ∂ Σ . For several boundary loops, their arclength parameterizations are concatenated in a fixed order. The in-plane outward normal n ^ ( ℓ ) \hat{\mathbf{n}}(\ell) n ^ ( ℓ ) is piecewise constant on each open polygon edge.
The field behaviour and charge concentration near conducting edges motivate a boundary-based description of the capacitive response [50 , 58 ] . Within the present reduced model, we adopt the dimensionless edge-charge profiles
ρ ~ a ( ℓ ; k t ) = [ e ^ a ⋅ n ^ ( ℓ ) ] e − i ( k t + G a ) ⋅ r ( ℓ ) , a = 0 , … , 6. (52) \widetilde{\rho}_a(\ell;\mathbf{k}_t)=\left[\hat{\mathbf{e}}_a\cdot\hat{\mathbf{n}}(\ell)\right]e^{-i(\mathbf{k}_t+\mathbf{G}_a)\cdot\mathbf{r}(\ell)},\qquad a=0,\ldots,6. \tag*{(52)} ρ a ( ℓ ; k t ) = [ e ^ a ⋅ n ^ ( ℓ ) ] e − i ( k t + G a ) ⋅ r ( ℓ ) , a = 0 , … , 6. (52)
Here G 0 = 0 \mathbf{G}_0=\mathbf{0} G 0 = 0 , while G p \mathbf{G}_p G p , p = 1 , … , 6 p=1,\ldots,6 p = 1 , … , 6 , are the retained first reciprocal-shell vectors. We take e ^ 0 ≡ e ^ T E \hat{\mathbf{e}}_0\equiv\hat{\mathbf{e}}_{\mathrm{TE}} e ^ 0 ≡ e ^ TE , and define each sideband TE polarization e ^ p \hat{\mathbf{e}}_p e ^ p relative to its own tangential wavevector k t + G p \mathbf{k}_t+\mathbf{G}_p k t + G p . The polarization vectors are real. The modal amplitudes are factored out, and the normal projections provide geometric weights for the assumed charge-profile shapes.
At fixed real k t \mathbf{k}_t k t , the cumulative cross moment is
B 0 p ( ℓ ) = ∫ 0 ℓ ρ ~ 0 ∗ ( ℓ ′ ; k t ) ρ ~ p ( ℓ ′ ; k t ) d ℓ ′ = ∫ 0 ℓ [ e ^ 0 ⋅ n ^ ( ℓ ′ ) ] [ e ^ p ⋅ n ^ ( ℓ ′ ) ] e − i G p ⋅ r ( ℓ ′ ) d ℓ ′ . (53) \begin{aligned}
B_{0p}(\ell)&=\int_0^\ell\widetilde{\rho}_0^{*}(\ell';\mathbf{k}_t)\widetilde{\rho}_p(\ell';\mathbf{k}_t)\,\mathrm{d}\ell'\\
&=\int_0^\ell\left[\hat{\mathbf{e}}_0\cdot\hat{\mathbf{n}}(\ell')\right]\left[\hat{\mathbf{e}}_p\cdot\hat{\mathbf{n}}(\ell')\right]e^{-i\mathbf{G}_p\cdot\mathbf{r}(\ell')}\,\mathrm{d}\ell'. \tag*{(53)}
\end{aligned} B 0 p ( ℓ ) = ∫ 0 ℓ ρ 0 ∗ ( ℓ ′ ; k t ) ρ p ( ℓ ′ ; k t ) d ℓ ′ = ∫ 0 ℓ [ e ^ 0 ⋅ n ^ ( ℓ ′ ) ] [ e ^ p ⋅ n ^ ( ℓ ′ ) ] e − i G p ⋅ r ( ℓ ′ ) d ℓ ′ . (53)
and the squared profile norms are
N a ( ℓ ) = ∫ 0 ℓ ∣ ρ ~ a ( ℓ ′ ; k t ) ∣ 2 d ℓ ′ = ∫ 0 ℓ [ e ^ a ⋅ n ^ ( ℓ ′ ) ] 2 d ℓ ′ , a ∈ { 0 , p } . (54) N_a(\ell)=\int_0^\ell\left|\tilde{\rho}_a(\ell';\mathbf{k}_t)\right|^2\,\mathrm{d}\ell'
=\int_0^\ell\left[\hat{\mathbf{e}}_a\cdot\hat{\mathbf{n}}(\ell')\right]^2\,\mathrm{d}\ell',\qquad a\in\{0,p\}. \tag*{(54)} N a ( ℓ ) = ∫ 0 ℓ ∣ ρ ~ a ( ℓ ′ ; k t ) ∣ 2 d ℓ ′ = ∫ 0 ℓ [ e ^ a ⋅ n ^ ( ℓ ′ ) ] 2 d ℓ ′ , a ∈ { 0 , p } . (54)
The common Bloch phase cancels in the cross moment, leaving the reciprocal phase e − i G p ⋅ r e^{-i\mathbf{G}_p\cdot\mathbf{r}} e − i G p ⋅ r .
The full-boundary quantities are B 0 p = B 0 p ( L ) B_{0p}=B_{0p}(L) B 0 p = B 0 p ( L ) and N a = N a ( L ) N_a=N_a(L) N a = N a ( L ) . For nonzero profile norms, their normalized overlap is
η 0 p ( k t ) = B 0 p ( L ) N 0 ( L ) N p ( L ) , ∣ η 0 p ( k t ) ∣ ≤ 1. (55) \eta_{0p}(\mathbf{k}_t)=\frac{B_{0p}(L)}{\sqrt{N_0(L)N_p(L)}},\qquad\left|\eta_{0p}(\mathbf{k}_t)\right|\leq1. \tag*{(55)} η 0 p ( k t ) = N 0 ( L ) N p ( L ) B 0 p ( L ) , ∣ η 0 p ( k t ) ∣ ≤ 1. (55)
Since ∣ e − i G p ⋅ r ∣ = 1 \left|e^{-i\mathbf{G}_p\cdot\mathbf{r}}\right|=1 e − i G p ⋅ r = 1 , the Cauchy–Schwarz inequality applied to the two boundary profiles gives ∣ B 0 p ( L ) ∣ 2 ≤ N 0 ( L ) N p ( L ) \left|B_{0p}(L)\right|^2\leq N_0(L)N_p(L) ∣ B 0 p ( L ) ∣ 2 ≤ N 0 ( L ) N p ( L ) , and hence ∣ η 0 p ( k t ) ∣ ≤ 1 \left|\eta_{0p}(\mathbf{k}_t)\right|\leq1 ∣ η 0 p ( k t ) ∣ ≤ 1 . The Floquet phase enters the cross moment but not the norms.
The overlap measures the complex correlation between the fundamental and sideband model charge profiles: contributions from different edges reinforce or cancel according to their normal projections and reciprocal phases. It provides a geometric analogue of a coupled-mode overlap [47 , 48 , 49 ] . Its use as an electromagnetic coupling measure is part of the adopted edge-projection approximation. The full-boundary quantities are independent of the starting points and ordering of the boundary loops, although the cumulative functions depend on that choice.
The geometry dependence can also be expressed through the boundary Fourier tensor
F ‾ ∂ Σ ( G p ) = ∮ ∂ Σ n ^ n ^ T e − i G p ⋅ r d ℓ . (56) \overline{\mathbf{F}}_{\partial\Sigma}(\mathbf{G}_p)=\oint_{\partial\Sigma}\hat{\mathbf{n}}\hat{\mathbf{n}}^{\mathsf{T}}e^{-i\mathbf{G}_p\cdot\mathbf{r}}\,\mathrm{d}\ell. \tag*{(56)} F ∂ Σ ( G p ) = ∮ ∂ Σ n ^ n ^ T e − i G p ⋅ r d ℓ . (56)
The full-boundary cross moment is then
B 0 p ( L ) = e ^ 0 T F ‾ ∂ Σ ( G p ) e ^ p . (57) B_{0p}(L)=\hat{\mathbf{e}}_0^{\mathsf{T}}\overline{\mathbf{F}}_{\partial\Sigma}(\mathbf{G}_p)\hat{\mathbf{e}}_p. \tag*{(57)} B 0 p ( L ) = e ^ 0 T F ∂ Σ ( G p ) e ^ p . (57)
This tensor contains the spatial distribution and orientation of the boundary edges. Its contraction need not vanish for G p ≠ 0 \mathbf{G}_p\ne0 G p = 0 : complete-cell Floquet orthogonality concerns an unweighted area inner product, whereas the present overlap uses normal projections integrated over the boundary. The detailed derivation is given in Sec. S6 of the Supplemental Material [54 ] .
For the polygonal boundaries of Sec. II, Eq. (53) is evaluated exactly edge by edge. The complex phase of η 0 p \eta_{0p} η 0 p is retained throughout.
For the visualizations in Figure 8 and Figure 9 , we select the representative interior point k t = 0.82 K \mathbf{k}_t=0.82\,\mathbf{K} k t = 0.82 K . The dispersion calculation evaluates the overlaps and turning multipliers pointwise over the complete sampled Γ \Gamma Γ –K \mathbf{K} K segment.
Figure 8. Complex boundary Fourier overlap for the six retained first-shell orders at the representative point k t = 0.82 K \mathbf{k}_t=0.82\,\mathbf{K} k t = 0.82 K on the Γ \Gamma Γ –K \mathbf{K} K path. Boundary colour denotes the phase of the edgewise contribution to B 0 p B_{0p} B 0 p . Each inset shows the cumulative complex-plane trajectory of the boundary contributions, whose endpoint determines the normalized overlap η 0 p \eta_{0p} η 0 p . The complex phase is retained in the coherent sideband combination of Eq. (70) .
Figure 9. Directional turning participation for the six retained first-shell orders at k t = 0.82 K \mathbf{k}_{\mathrm{t}}=0.82\mathbf{K} k t = 0.82 K . Marker colour and size indicate the vertex summands of Eq. (61) ; all boundary vertices are shown. The values below each unit cell give T p r a w T_p^{\mathrm{raw}} T p raw and the pointwise amplitude multiplier m ~ p ( k t ) \widetilde{m}_p(\mathbf{k}_{\mathrm{t}}) m p ( k t ) .
The edge-resolved contributions for the complete first reciprocal-shell are illustrated in Figure 8 . The boundary colour records the phase of the local complex contribution, while each inset traces its cumulative sum in the complex plane.
Directional multipliers from boundary turning
The directional multiplier m ~ p ( k t ) \tilde{m}_p(\mathbf{k}_t) m ~ p ( k t ) weights fundamental-to-sideband coupling according to the participation of the model edge-charge profiles at changes in the boundary orientation. The field behaviour near conducting corners motivates including boundary turning in this description [67 ] . We adopt the following dimensionless geometric construction within the reduced coupling model.
At each vertex v v v , let t ^ v ± \hat{\mathbf{t}}_v^{\pm} t ^ v ± and n ^ v ± \hat{\mathbf{n}}_v^{\pm} n ^ v ± denote the one-sided unit tangents and normals. The signed exterior turning angle is
Δ θ v = atan2 ( z ^ ⋅ [ t ^ v − × t ^ v + ] , t ^ v − ⋅ t ^ v + ) . (58) \Delta\theta_v=\operatorname{atan2}\left(\hat{\mathbf{z}}\cdot[\hat{\mathbf{t}}_v^{-}\times\hat{\mathbf{t}}_v^{+}],\hat{\mathbf{t}}_v^{-}\cdot\hat{\mathbf{t}}_v^{+}\right). \tag*{(58)} Δ θ v = atan2 ( z ^ ⋅ [ t ^ v − × t ^ v + ] , t ^ v − ⋅ t ^ v + ) . (58)
and the normalized total absolute turning is
T a b s = 1 2 π ∑ v ∣ Δ θ v ∣ . (59) T_{\mathrm{abs}}=\frac{1}{2\pi}\sum_v \lvert\Delta\theta_v\rvert. \tag*{(59)} T abs = 2 π 1 v ∑ ∣ Δ θ v ∣ . (59)
This quantity measures the accumulated absolute rotation of the boundary tangent, or equivalently of its normal. For a single simple closed polygonal boundary, the total-curvature theorem gives T a b s ≥ 1 T_{\mathrm{abs}}\ge1 T abs ≥ 1 , with equality for a convex boundary [68 ] . Reversals of the turning direction increase this quantity. For several boundary loops, their absolute-turning contributions are added.
The participation of order a a a at a vertex is measured by the mean squared charge-profile amplitude on its two adjacent edges:
S a ( v ; k t ) = 1 2 ∑ ± [ e ^ a ( k t ) ⋅ n ^ v ± ] 2 , a ∈ { 0 , p } . (60) S_a(v;\mathbf{k}_{\mathrm{t}})=\frac{1}{2}\sum_{\pm}\left[\hat{\mathbf{e}}_a(\mathbf{k}_{\mathrm{t}})\cdot\hat{\mathbf{n}}_v^{\pm}\right]^2,\qquad a\in\{0,p\}. \tag*{(60)} S a ( v ; k t ) = 2 1 ± ∑ [ e ^ a ( k t ) ⋅ n ^ v ± ] 2 , a ∈ { 0 , p } . (60)
This follows from the model profiles in Eq. (52) . Since the polarization vectors and boundary normals are unit vectors, 0 ≤ S a ≤ 1 0\le S_a\le1 0 ≤ S a ≤ 1 . The raw turning participation is
T p r a w ( k t ) = 1 2 π ∑ v ∣ Δ θ v ∣ 2 S 0 ( v ; k t ) S p ( v ; k t ) . (61) T_p^{\mathrm{raw}}(\mathbf{k}_{\mathrm{t}})=\frac{1}{2\pi}\sum_v\lvert\Delta\theta_v\rvert\,2\sqrt{S_0(v;\mathbf{k}_{\mathrm{t}})S_p(v;\mathbf{k}_{\mathrm{t}})}. \tag*{(61)} T p raw ( k t ) = 2 π 1 v ∑ ∣ Δ θ v ∣ 2 S 0 ( v ; k t ) S p ( v ; k t ) . (61)
Thus a vertex contributes strongly when its turn is pronounced and both profiles participate on the adjacent edges. This nonnegative weight measures jointly sampled boundary turning; cancellation from charge signs and reciprocal phases is retained in the complex overlap η 0 p \eta_{0p} η 0 p .
At each Bloch wavevector, we normalize the six raw participations and define the amplitude multipliers as
T p ( k t ) = T a b s T p r a w ( k t ) 1 6 ∑ r = 1 6 T r r a w ( k t ) , m ~ p ( k t ) = 2 T p ( k t ) . (62) T_p(\mathbf{k}_{\mathrm{t}})=\frac{T_{\mathrm{abs}}T_p^{\mathrm{raw}}(\mathbf{k}_{\mathrm{t}})}{\frac{1}{6}\sum_{r=1}^{6}T_r^{\mathrm{raw}}(\mathbf{k}_{\mathrm{t}})},\qquad
\widetilde{m}_p(\mathbf{k}_{\mathrm{t}})=\sqrt{2T_p(\mathbf{k}_{\mathrm{t}})}. \tag*{(62)} T p ( k t ) = 6 1 ∑ r = 1 6 T r raw ( k t ) T abs T p raw ( k t ) , m p ( k t ) = 2 T p ( k t ) . (62)
For a nonzero normalization denominator,
1 6 ∑ p = 1 6 T p ( k t ) = T a b s , (63) \frac{1}{6}\sum_{p=1}^{6}T_p(\mathbf{k}_{\mathrm{t}})=T_{\mathrm{abs}}, \tag*{(63)} 6 1 p = 1 ∑ 6 T p ( k t ) = T abs , (63)
1 6 ∑ p = 1 6 m ~ p 2 ( k t ) = 2 T a b s . \frac{1}{6}\sum_{p=1}^{6}\widetilde{m}_p^2(\mathbf{k}_{\mathrm{t}})=2T_{\mathrm{abs}}. 6 1 p = 1 ∑ 6 m p 2 ( k t ) = 2 T abs .
The normalization preserves the shell mean while allowing its distribution among the orders to vary with their polarization projections. The square root converts participation into an amplitude weight. The factor of two specifies the adopted amplitude normalization; no additional fitted scale multiplies Eq. (62) . The shell root-mean-square multiplier is therefore 2 T a b s \sqrt{2T_{\mathrm{abs}}} 2 T abs at every Bloch point.
Along the straight Γ \Gamma Γ –K segment,
k t ( s ) = s K , 0 ≤ s ≤ 1 , \mathbf{k}_{\mathrm{t}}(s)=s\mathbf{K},\qquad0\le s\le1, k t ( s ) = s K , 0 ≤ s ≤ 1 ,
K = 2 b 1 + b 2 3 . (64) \mathbf{K}=\frac{2\mathbf{b}_1+\mathbf{b}_2}{3}. \tag*{(64)} K = 3 2 b 1 + b 2 . (64)
where b 1 , b 2 \mathbf{b}_1,\mathbf{b}_2 b 1 , b 2 are the primitive reciprocal vectors and K \mathbf{K} K locates the selected Brillouin-zone corner. At every sampled point, the coupling uses m ~ p [ k t ( s ) ] \widetilde{m}_p[\mathbf{k}_{\mathrm{t}}(s)] m p [ k t ( s )] and η 0 p [ k t ( s ) ] \eta_{0p}[\mathbf{k}_{\mathrm{t}}(s)] η 0 p [ k t ( s )] evaluated at that same wavevector. The geometry remains fixed; the turning weights vary through the polarization projections in Eq. (60) . At Γ \Gamma Γ , the fundamental TE polarization and the multipliers are defined by the limit s → 0 + s\to0^+ s → 0 + along this path, with e ^ 0 = z ^ × K / ∣ K ∣ \hat{\mathbf{e}}_0=\hat{\mathbf{z}}\times\mathbf{K}/\lvert\mathbf{K}\rvert e ^ 0 = z ^ × K / ∣ K ∣ .
Figure 9 illustrates the vertex contributions at one representative point.
Uniform scaling of the boundary and lattice preserves these dimensionless turning quantities at corresponding normalized Bloch wavevectors. The prescription therefore introduces no additional length scale. The six multiplier traces along Γ \Gamma Γ –K are shown in Supplemental Fig. S3; the normalization and scaling properties are detailed in Sec. S6 of the Supplemental Material [54 ] .
Sideband self-energy and corrected dispersion
At each Bloch wavevector, the fundamental order couples to the six retained sidebands through
v p = m ~ p ( k t ) Γ ‾ T L Γ p η 0 p ( k t ) , D p = γ h p − Γ p . (65) v_p=\widetilde{m}_p(\mathbf{k}_t)\sqrt{\overline{\Gamma}_{\mathrm{TL}}\Gamma_p}\,\eta_{0p}(\mathbf{k}_t),\qquad
D_p=\gamma_{h_p}-\Gamma_p. \tag*{(65)} v p = m p ( k t ) Γ TL Γ p η 0 p ( k t ) , D p = γ h p − Γ p . (65)
Both quantities have units of m − 1 \mathrm{m}^{-1} m − 1 . The overlap and turning multiplier are evaluated at the current k t \mathbf{k}_t k t , while the decay quantities retain their dependence on ( k 0 , k t ) (k_0,\mathbf{k}_t) ( k 0 , k t ) . Each sideband detuning is determined by its own tangential wavevector and sheet loading.
Define v = ( v 1 , … , v 6 ) T \mathbf{v}=(v_1,\ldots,v_6)^{\mathsf{T}} v = ( v 1 , … , v 6 ) T , D = diag ( D 1 , … , D 6 ) \mathbf{D}=\operatorname{diag}(D_1,\ldots,D_6) D = diag ( D 1 , … , D 6 ) , and γ 0 = k t 2 − k 0 2 \gamma_0=\sqrt{k_t^2-k_0^2} γ 0 = k t 2 − k 0 2 . The adopted ( 1 + 6 ) (1+6) ( 1 + 6 ) -mode system is
\begin{pmatrix}
\gamma_0-\Gamma_{\mathrm{TL}} & -\mathbf{v}^{\dagger}\\
-\mathbf{v} & \mathbf{D}
\end{pmatrix}
\begin{pmatrix}
a_0\\
\mathbf{a}_{\mathrm{sb}}
\end{pmatrix}
=0, \tag*{(66)}\end{pmatrix}
where a 0 a_0 a 0 and a s b \mathbf{a}_{\mathrm{sb}} a sb are the fundamental and sideband amplitudes. The diagonal sideband block omits direct coupling between distinct sidebands.
For invertible D \mathbf{D} D , eliminating the sideband amplitudes by a Schur complement [54 , 69 ] gives
R ( k 0 , k t ) = γ 0 − Γ T L − Σ s b = 0. (67) R(k_0,\mathbf{k}_t)=\gamma_0-\Gamma_{\mathrm{TL}}-\Sigma_{\mathrm{sb}}=0. \tag*{(67)} R ( k 0 , k t ) = γ 0 − Γ TL − Σ sb = 0. (67)
Setting Σ s b = 0 \Sigma_{\mathrm{sb}}=0 Σ sb = 0 recovers the local condition in Eq. (48) . Without a symmetry projection, the self-energy is
Σ s b = v † D − 1 v = Γ T L ∑ p = 1 6 m ~ p 2 ( k t ) Γ p ∣ η 0 p ( k t ) ∣ 2 γ h p − Γ p . (68) \begin{aligned}
\Sigma_{\mathrm{sb}}=\mathbf{v}^{\dagger}\mathbf{D}^{-1}\mathbf{v} \\
&=\Gamma_{\mathrm{TL}}\sum_{p=1}^{6}\frac{\widetilde{m}_p^2(\mathbf{k}_t)\Gamma_p\lvert\eta_{0p}(\mathbf{k}_t)\rvert^2}{\gamma_{h_p}-\Gamma_p}. \tag*{(68)}
\end{aligned} Σ sb = v † D − 1 v = Γ TL p = 1 ∑ 6 γ h p − Γ p m p 2 ( k t ) Γ p ∣ η 0 p ( k t ) ∣ 2 . (68)
All quantities in this expression are evaluated at the same ( k 0 , k t ) (k_0,\mathbf{k}_t) ( k 0 , k t ) , with the full frequency dependence of the detunings retained. The turning factors and overlaps follow from the reconstructed boundary, and the sheet-loaded decays from the TE projections of the grid-length dyadic. No sideband coupling or detuning is fitted separately. At a singular sideband block, the eigencondition must be obtained from the coupled system (67).
When the propagation axis is a mirror of the complete unit-cell geometry, the sidebands can be organized into symmetry-adapted combinations. Reflection about u ^ = k ^ t \hat{\mathbf{u}}=\hat{\mathbf{k}}_t u ^ = k ^ t is represented by M = 2 u ^ u ^ T − I ˉ M=2\hat{\mathbf{u}}\hat{\mathbf{u}}^{\mathsf{T}}-\bar{I} M = 2 u ^ u ^ T − I ˉ and induces the permutation M G p = G π ( p ) M G_p=G_{\pi(p)} M G p = G π ( p ) . In the adopted TE basis, M e ^ p = − e ^ π ( p ) M\hat{\mathbf{e}}_p=-\hat{\mathbf{e}}_{\pi(p)} M e ^ p = − e ^ π ( p ) . The basis carries the odd parity of the fundamental TE field, so the allowed scalar amplitudes are equal within each mirror pair [51 , 52 , 70 , 71 ] . Their subspace is specified by
P T E = 1 2 ( I ˉ 6 + Π M ) , range ( U ) = range ( P T E ) , (69) P_{\mathrm{TE}}=\frac{1}{2}\left(\bar{I}_6+\Pi_M\right),\qquad
\operatorname{range}(U)=\operatorname{range}(P_{\mathrm{TE}}), \tag*{(69)} P TE = 2 1 ( I ˉ 6 + Π M ) , range ( U ) = range ( P TE ) , (69)
where Π M \Pi_M Π M implements the permutation and U † U = I ˉ r U^{\dagger}U=\bar{I}_r U † U = I ˉ r .
where Π M \Pi_M Π M implements the permutation and U † U = I ˉ r U^{\dagger}U=\bar{I}_r U † U = I ˉ r . Along Γ \Gamma Γ –K, the six orders form three mirror pairs, giving r = 3 r=3 r = 3 . The projected self-energy is
Σ s b p r o j = ( U † v ) † ( U † D U ) − 1 ( U † v ) . (70) \Sigma_{\mathrm{sb}}^{\mathrm{proj}}=(U^{\dagger}\mathbf{v})^{\dagger}(U^{\dagger}\mathbf{D}U)^{-1}(U^{\dagger}\mathbf{v}). \tag*{(70)} Σ sb proj = ( U † v ) † ( U † D U ) − 1 ( U † v ) . (70)
When mirror-symmetry reduction applies, Σ s b p r o j \Sigma_{\mathrm{sb}}^{\mathrm{proj}} Σ sb proj is used for Σ s b \Sigma_{\mathrm{sb}} Σ sb in Eq. (67) . Taking U = I ˉ 6 U=\bar{I}_6 U = I ˉ 6 recovers the unprojected expression.
The projection retains the complex coupling amplitudes, including the phases of η 0 p \eta_{0p} η 0 p . For an exactly mirror-symmetric geometry represented with its origin on the mirror axis, v π ( p ) = v p v_{\pi(p)}=v_p v π ( p ) = v p and D π ( p ) = D p D_{\pi(p)}=D_p D π ( p ) = D p . Each allowed pair combination therefore retains the full contribution of its two orders, while the complementary combination decouples. No additional multiplicity factor is introduced. A displaced geometry requires the corresponding Fourier phases in both the couplings and the mirror representation. The explicit permutation π = ( 1 3 ) ( 2 5 ) ( 4 6 ) \pi=(1\,3)(2\,5)(4\,6) π = ( 1 3 ) ( 2 5 ) ( 4 6 ) and construction of U U U are given in Sec. S6 of the Supplemental Material; the pairs are illustrated in Supplemental Fig. S2(b).
The pointwise prescription also recovers the local long-wavelength limit. For fixed geometry with finite, nondispersive loading, Γ T L , Γ p = O ( k 0 2 ) \Gamma_{\mathrm{TL}},\Gamma_p=O(k_0^2) Γ TL , Γ p = O ( k 0 2 ) . As k 0 , k t → 0 k_0,k_t\to0 k 0 , k t → 0 along the lowest branch, D p → ∣ G p ∣ > 0 D_p\to\lvert\mathbf{G}_p\rvert>0 D p → ∣ G p ∣ > 0 . The bounded turning multipliers and normalized overlaps then give Σ s b = O ( k 0 4 ) \Sigma_{\mathrm{sb}}=O(k_0^4) Σ sb = O ( k 0 4 ) and
Σ s b Γ T L ⟶ 0. (71) \frac{\Sigma_{\mathrm{sb}}}{\Gamma_{\mathrm{TL}}}\longrightarrow0. \tag*{(71)} Γ TL Σ sb ⟶ 0. (71)
The same limit holds after the mirror projection, whose sideband block has a finite, invertible limit. Local recovery therefore does not require the individual turning multipliers to approach unity or vanish.
The Schur elimination is exact within the adopted finite coupled system. The geometric coupling prescription and omission of direct sideband interactions remain approximations whose accuracy is assessed against the full-wave benchmarks in § 2.3 . With the lattice fixed, the construction follows
{ C ‾ , ∂ Σ } ⟶ { Γ T L , Γ p , m ~ p , η 0 p } \{\overline{\mathbf{C}},\partial\Sigma\}\longrightarrow\{\Gamma_{\mathrm{TL}},\Gamma_p,\widetilde{m}_p,\eta_{0p}\} { C , ∂ Σ } ⟶ { Γ TL , Γ p , m p , η 0 p }
⟶ Σ s b ⟶ R = 0. (72) \longrightarrow\Sigma_{\mathrm{sb}}\longrightarrow R=0. \tag*{(72)} ⟶ Σ sb ⟶ R = 0. (72)
The local branch depends on C ‾ \overline{\mathbf{C}} C through the projection A 0 ( k t ) A_0(\mathbf{k}_t) A 0 ( k t ) . The nonlocal correction additionally requires ∂ Σ \partial\Sigma ∂ Σ , which supplies the Fourier overlap and pointwise turning participation.
Numerical Results
The dispersion model of Sec. III is validated against full-wave eigenmode simulations for a set of geometries spanning the symmetry classes of Table 1 . Importantly, HFSS is not used to extract { C ‾ \{\overline{\mathbf{C}} { C or to fit any sideband parameter. For each geometry, the dyadic is computed once from the cell problem of Sec. II, the Floquet correction is evaluated from ∂ Σ \partial\Sigma ∂ Σ , and the HFSS eigenmode sweep serves solely as an independent benchmark of the resulting bands. Each structure is a hexagonal unit cell of circumradius R c e l l = 11.547 R_{\mathrm{cell}} = 11.547 R cell = 11.547 mm with a minimum edge-to-edge gap of 1 mm, with a random or canonical metal patch built from the honeycomb tessellation of Sec. II. The eigenfrequencies of the first bound TE surface-wave mode are computed with the HFSS eigenmode solver under the periodic boundary conditions, sweeping the tangential Bloch wavevector k t \mathbf{k}_{\mathrm{t}} k t along the Γ \Gamma Γ –K path of the hexagonal Brillouin zone. For each geometry, the capacitance dyadic C ‾ ‾ \overline{\overline{C}} C is computed once from the cell problem of Sec. II, projected onto the TE polarisation along the swept direction via (43), and inserted into the local (49), and sideband-corrected (68) dispersion relations. Since the sweep direction is fixed while the principal axes of C ‾ ‾ \overline{\overline{C}} C vary from structure to structure, the set probes the directional projection over a wide range of orientations relative to the anisotropy frame. The model accuracy is quantified by the relative frequency error at fixed k t k_{\mathrm{t}} k t , ϵ ( k t ) = ∣ f m o d e l − f H F S S ∣ / f H F S S \epsilon(k_{\mathrm{t}}) = \lvert f_{\mathrm{model}} - f_{\mathrm{HFSS}}\rvert/f_{\mathrm{HFSS}} ϵ ( k t ) = ∣ f model − f HFSS ∣ / f HFSS , evaluated at the simulated points.
Figure 11 complements the Γ \Gamma Γ –K cuts of Figure 10 by showing the model-predicted angular response throughout the first Brillouin zone. The reduced hybrid TE–TM calculation restores the longitudinal–transverse coupling within each retained Floquet order, while retaining the same geometric TE–TE coupling between orders as the quasi-TE calculation. The corresponding closure and the numerical construction of the full-Brillouin-zone isofrequency contours are detailed in Sec. S4 of the Supplemental Material [54 ] . The small contours surrounding Γ \Gamma Γ are nearly circular in both examples. At larger wavevectors, the Floquet correction depends on the propagation direction through the sideband decay constants γ h p \gamma_{h_p} γ h p , the detunings D p D_p D p , and the boundary coupling factors. Consequently, the same first-shell coupling that bends the Γ \Gamma Γ –K bands also reshapes the isofrequency contours in a direction-dependent manner. For the C 6 v C_{6v} C 6 v patch, the local capacitance dyadic is isotropic, and the contours remain
Figure 10. Surface-wave dispersion along the Γ \Gamma Γ –K path for six representative structures, one per symmetry class of Table I. The vertical dash-dotted marker indicates the onset of spatial dispersion, beyond which the Floquet-sideband coupling becomes significant and the local and sideband-corrected models separate. Below the onset, both models coincide and follow the quasi-linear branch of Eq. (38) ; beyond it the sideband-corrected model tracks the progressive flattening of the simulated band.
nearly circular over most of the plotted range; any finite-wavevector deformation must nevertheless preserve the sixfold symmetry and need not preserve exact circularity. For the C 1 C_1 C 1 patch, the anisotropic local dyadic combines with the direction-dependent Floquet correction to produce much stronger contour deformation towards the zone boundary. Thus, the low-frequency near-circular contours do not imply an isotropic C 1 C_1 C 1 dyadic. The quasi-TE and hybrid TE–TM calculations reproduce the same main directional features, although including polarisation coupling produces more visible shifts in the outer C 1 C_1 C 1 contours. Such shifts can become more apparent as the band flattens, because a smaller radial frequency slope makes the contour position more sensitive to changes in the modal frequency. These full-zone contours illustrate the model predictions; the independent HFSS validation reported here remains the Γ \Gamma Γ –K sweep.
Table 2 has the validation statistics for each symmetry class. The largest deviations occur in the transition region immediately beyond the spatial-dispersion onset, where the model can overestimate the modal frequency.
Table 2. Validation statistics by symmetry class: sweep-mean relative frequency error for the local dyadic model ϵ ˉ l o c a l d y a d i c \bar{\epsilon}_{\mathrm{local\ dyadic}} ϵ ˉ local dyadic , and the sweep-mean relative frequency error for the sideband-corrected model ϵ ˉ s i d e b a n d − c o r r e c t e d \bar{\epsilon}_{\mathrm{sideband-corrected}} ϵ ˉ sideband − corrected against HFSS eigenmode data along Γ \Gamma Γ –K.
Figure 11. Model-predicted isofrequency contours of the first TE-dominated surface-wave branch over the hexagonal first Brillouin zone for the C 6 v C_{6v} C 6 v (top row) and C 1 C_1 C 1 (bottom row) patches. The left and right columns show the quasi-TE approximation and the hybrid TE–TM model, respectively; both retain the six Floquet orders of the first reciprocal-shell. The colour maps give the modal frequency in GHz, the white curves mark selected isofrequencies, and the insets show the corresponding patches. The C 6 v C_{6v} C 6 v contours remain approximately circular, with weak deformation at larger wavevectors, whereas the C 1 C_1 C 1 contours develop a pronounced directional dependence. The two polarisation treatments give similar contours near Γ \Gamma Γ , with more visible differences in the outer C 1 C_1 C 1 contours.
This is the region where the six-sideband truncation of Eq. (67) is least accurate: the retained channels are no longer deeply evanescent.
Conclusion
We have developed a geometry-driven homogenisation framework that maps the unit-cell geometry of patch-type MTSs onto an effective surface-capacitance dyadic and its leading spatially dispersive correction. The resulting model predicts anisotropic surface-wave dispersion without per-geometry full-wave sheet retrieval, fitted sideband parameters, or a full-wave eigenmode band sweep for each new patch. The framework separates the three ingredients: Bravais periodicity defines the two-component macroscopic tangential loading space, gap connectivity ensures a non-degenerate capacitance response for the geometries considered here, and the point-group symmetry of the complete unit-cell geometry constrains the admissible form of the capacitance dyadic. A quasi-static boundary-value problem on a single periodic cell then determines its independent entries, scaled by the analytically calculated capacitance of the isotropic hexagonal reference patch. This construction provides the model’s quasi-static loading directly, avoiding the evaluation of frequency-dependent periodic Green’s functions for sheet-parameter retrieval and the extrapolation of finite-frequency or finite-Bloch-phase simulation data toward the static, long-wavelength limit. The resulting dyadic determines the local anisotropic dispersion through its TE projection along the propagation direction.
The leading spatially dispersive correction retains the six Floquet orders of the first reciprocal-shell. The coupling to the fundamental is determined by the complex boundary Fourier overlap and the directional turning multiplier, both evaluated at each sampled Bloch wavevector along the complete Γ \Gamma Γ –K path. Each order enters through the detuning D p = γ p − Γ p D_p = \gamma_p - \Gamma_p D p = γ p − Γ p , determined by its free-space evanescence and sheet loading. For propagation along a mirror axis of the complete unit-cell geometry, the complex coupling amplitudes are projected onto symmetry-allowed combinations of mirror-related orders. This projection retains their overlap phases and the full contribution of each pair. The correction becomes negligible relative to the local loading in the long-wavelength limit and captures the leading spatial-dispersion toward the Brillouin-zone boundary.
We validated against full-wave eigenmode simulations for six representative symmetry classes spanning C 6 v C_{6v} C 6 v to C 1 C_1 C 1 . The sideband-corrected model reduces the sweep-mean frequency error of the local dyadic model from 2.7–13.5% to 0.7–4.0%, with the largest improvements occurring for the low-symmetry geometries. A new gap-connected patch therefore yields a predicted band diagram without repeating a full-wave eigenmode campaign, while the same geometry-to-dispersion map provides a route toward inverse design of anisotropic surface-wave propagation. The framework also provides a basis for extensions to other propagation directions, including Γ \Gamma Γ –M, and to additional reciprocal shells. Substrate loading may be modeled using the effective permittivity of Eq. (17) , subject to separate validation.
Group-theoretic constraints on the surface-capacitance dyadic
This section provides the explicit derivation of the group-theoretic tensor forms summarized in Table I of Sec. II B in the main text. For a reciprocal lossless sheet, the in-plane surface-capacitance dyadic is real and symmetric, and may therefore be written in an arbitrary Cartesian basis as
C ‾ = ( C x x C x y C x y C y y ) . (73) \overline{\mathbf{C}} = \begin{pmatrix} C_{xx} & C_{xy} \\ C_{xy} & C_{yy} \end{pmatrix}. \tag*{(73)} C = ( C xx C x y C x y C y y ) . (73)
Let G G G denote the in-plane point group of the complete unit-cell geometry. If g ∈ G g \in G g ∈ G is represented in the tangential plane by the orthogonal matrix D ‾ ( g ) \overline{\mathbf{D}}(g) D ( g ) , invariance of the equivalent response requires
D ‾ ( g ) ⋅ C ‾ ⋅ D ‾ T ( g ) = C ‾ , ∀ g ∈ G . (74) \overline{\mathbf{D}}(g) \cdot\overline{\mathbf{C}} \cdot\overline{\mathbf{D}}^{T}(g) = \overline{\mathbf{C}}, \qquad\forall g \in G. \tag*{(74)} D ( g ) ⋅ C ⋅ D T ( g ) = C , ∀ g ∈ G . (74)
The forms derived below are the symmetry-allowed tensor forms obtained from Eq. (74) . Lower point symmetry permits anisotropy but does not require it; a particular geometry may still exhibit accidental isotropy.
Rotational symmetry
A counterclockwise in-plane rotation by an angle θ \theta θ is represented by
R ‾ ( θ ) = ( cos θ − sin θ sin θ cos θ ) . (75) \overline{\mathbf{R}}(\theta) = \begin{pmatrix} \cos\theta& -\sin\theta\\ \sin\theta& \cos\theta\end{pmatrix}. \tag*{(75)} R ( θ ) = ( cos θ sin θ − sin θ cos θ ) . (75)
The capacitance dyadic is separated into its isotropic and traceless parts by defining
C ‾ = C x x + C y y 2 , Δ C = C x x − C y y . (76) \overline{C} = \frac{C_{xx} + C_{yy}}{2}, \qquad\Delta C = C_{xx} - C_{yy}. \tag*{(76)} C = 2 C xx + C y y , Δ C = C xx − C y y . (76)
Equation (S1) can then be written as
C ‾ = C ‾ I ‾ + 1 2 ( Δ C 2 C x y 2 C x y − Δ C ) . (77) \overline{\mathbf{C}} = \overline{C}\,\overline{\mathbf{I}} + \frac{1}{2}\begin{pmatrix} \Delta C & 2C_{xy} \\ 2C_{xy} & -\Delta C \end{pmatrix}. \tag*{(77)} C = C I + 2 1 ( Δ C 2 C x y 2 C x y − Δ C ) . (77)
For convenience, we define the traceless anisotropy matrix
C ‾ a n = ( Δ C 2 C x y 2 C x y − Δ C ) , (78) \overline{\mathbf{C}}_{\mathrm{an}} = \begin{pmatrix} \Delta C & 2C_{xy} \\ 2C_{xy} & -\Delta C \end{pmatrix}, \tag*{(78)} C an = ( Δ C 2 C x y 2 C x y − Δ C ) , (78)
so that
C ‾ = C ‾ I ‾ + 1 2 C ‾ a n . (79) \overline{\mathbf{C}} = \overline{C}\,\overline{\mathbf{I}} + \frac{1}{2}\overline{\mathbf{C}}_{\mathrm{an}}. \tag*{(79)} C = C I + 2 1 C an . (79)
Under the rotation of Eq. (75) , the transformed capacitance dyadic is
C ‾ ′ = R ‾ ( θ ) ⋅ C ‾ ⋅ R ‾ T ( θ ) . (80) \overline{\mathbf{C}}' = \overline{\mathbf{R}}(\theta) \cdot\overline{\mathbf{C}} \cdot\overline{\mathbf{R}}^{T}(\theta). \tag*{(80)} C ′ = R ( θ ) ⋅ C ⋅ R T ( θ ) . (80)
Substituting Eq. (79) gives
C ‾ ′ = C ‾ R ‾ ( θ ) ⋅ I ‾ ⋅ R ‾ T ( θ ) + 1 2 R ‾ ( θ ) ⋅ C ‾ a n ⋅ R ‾ T ( θ ) = C ‾ I ‾ + 1 2 R ‾ ( θ ) ⋅ C ‾ a n ⋅ R ‾ T ( θ ) . (81) \overline{C}'=\overline{C}\overline{R}(\theta)\cdot\overline{I}\cdot\overline{R}^{T}(\theta)+\frac{1}{2}\overline{R}(\theta)\cdot\overline{C}_{\mathrm{an}}\cdot\overline{R}^{T}(\theta)
=\overline{C}\overline{I}+\frac{1}{2}\overline{R}(\theta)\cdot\overline{C}_{\mathrm{an}}\cdot\overline{R}^{T}(\theta). \tag*{(81)} C ′ = C R ( θ ) ⋅ I ⋅ R T ( θ ) + 2 1 R ( θ ) ⋅ C an ⋅ R T ( θ ) = C I + 2 1 R ( θ ) ⋅ C an ⋅ R T ( θ ) . (81)
where
R ‾ ( θ ) ⋅ R ‾ T ( θ ) = I ‾ . \overline{R}(\theta)\cdot\overline{R}^{T}(\theta)=\overline{I}. R ( θ ) ⋅ R T ( θ ) = I .
Thus, only the traceless part needs to be transformed. The first matrix multiplication gives
R ‾ ( θ ) ⋅ C ‾ a n = ( cos θ − sin θ sin θ cos θ ) ( Δ C 2 C x y 2 C x y − Δ C ) = ( Δ C cos θ − 2 C x y sin θ 2 C x y cos θ + Δ C sin θ Δ C sin θ + 2 C x y cos θ 2 C x y sin θ − Δ C cos θ ) . (82) \overline{R}(\theta)\cdot\overline{C}_{\mathrm{an}}=
\begin{pmatrix}
\cos\theta& -\sin\theta\\
\sin\theta& \cos\theta
\end{pmatrix}
\begin{pmatrix}
\Delta C & 2C_{xy}\\
2C_{xy} & -\Delta C
\end{pmatrix}
=
\begin{pmatrix}
\Delta C\cos\theta-2C_{xy}\sin\theta& 2C_{xy}\cos\theta+\Delta C\sin\theta\\
\Delta C\sin\theta+2C_{xy}\cos\theta& 2C_{xy}\sin\theta-\Delta C\cos\theta
\end{pmatrix}. \tag*{(82)} R ( θ ) ⋅ C an = ( cos θ sin θ − sin θ cos θ ) ( Δ C 2 C x y 2 C x y − Δ C ) = ( Δ C cos θ − 2 C x y sin θ Δ C sin θ + 2 C x y cos θ 2 C x y cos θ + Δ C sin θ 2 C x y sin θ − Δ C cos θ ) . (82)
Multiplication by R ‾ T ( θ ) \overline{R}^{T}(\theta) R T ( θ ) then yields
R ‾ ( θ ) ⋅ C ‾ a n ⋅ R ‾ T ( θ ) = ( Δ C ( cos 2 θ − sin 2 θ ) − 4 C x y sin θ cos θ 2 Δ C sin θ cos θ + 2 C x y ( cos 2 θ − sin 2 θ ) 2 Δ C sin θ cos θ + 2 C x y ( cos 2 θ − sin 2 θ ) − Δ C ( cos 2 θ − sin 2 θ ) + 4 C x y sin θ cos θ ) . (83) \overline{R}(\theta)\cdot\overline{C}_{\mathrm{an}}\cdot\overline{R}^{T}(\theta)
=
\begin{pmatrix}
\Delta C(\cos^{2}\theta-\sin^{2}\theta)-4C_{xy}\sin\theta\cos\theta& 2\Delta C\sin\theta\cos\theta+2C_{xy}(\cos^{2}\theta-\sin^{2}\theta)\\
2\Delta C\sin\theta\cos\theta+2C_{xy}(\cos^{2}\theta-\sin^{2}\theta) & -\Delta C(\cos^{2}\theta-\sin^{2}\theta)+4C_{xy}\sin\theta\cos\theta
\end{pmatrix}. \tag*{(83)} R ( θ ) ⋅ C an ⋅ R T ( θ ) = ( Δ C ( cos 2 θ − sin 2 θ ) − 4 C x y sin θ cos θ 2Δ C sin θ cos θ + 2 C x y ( cos 2 θ − sin 2 θ ) 2Δ C sin θ cos θ + 2 C x y ( cos 2 θ − sin 2 θ ) − Δ C ( cos 2 θ − sin 2 θ ) + 4 C x y sin θ cos θ ) . (83)
Using
cos 2 θ = cos 2 θ − sin 2 θ , sin 2 θ = 2 sin θ cos θ , (84) \cos2\theta=\cos^{2}\theta-\sin^{2}\theta,\qquad\sin2\theta=2\sin\theta\cos\theta, \tag*{(84)} cos 2 θ = cos 2 θ − sin 2 θ , sin 2 θ = 2 sin θ cos θ , (84)
this becomes
R ‾ ( θ ) ⋅ C ‾ a n ⋅ R ‾ T ( θ ) = ( Δ C cos 2 θ − 2 C x y sin 2 θ Δ C sin 2 θ + 2 C x y cos 2 θ Δ C sin 2 θ + 2 C x y cos 2 θ − Δ C cos 2 θ + 2 C x y sin 2 θ ) . (85) \overline{R}(\theta)\cdot\overline{C}_{\mathrm{an}}\cdot\overline{R}^{T}(\theta)=
\begin{pmatrix}
\Delta C\cos2\theta-2C_{xy}\sin2\theta& \Delta C\sin2\theta+2C_{xy}\cos2\theta\\
\Delta C\sin2\theta+2C_{xy}\cos2\theta& -\Delta C\cos2\theta+2C_{xy}\sin2\theta
\end{pmatrix}. \tag*{(85)} R ( θ ) ⋅ C an ⋅ R T ( θ ) = ( Δ C cos 2 θ − 2 C x y sin 2 θ Δ C sin 2 θ + 2 C x y cos 2 θ Δ C sin 2 θ + 2 C x y cos 2 θ − Δ C cos 2 θ + 2 C x y sin 2 θ ) . (85)
The transformed dyadic can also be written in the same isotropic–traceless form,
C ‾ ′ = C ‾ I ‾ + 1 2 ( Δ C ′ 2 C x y ′ 2 C x y ′ − Δ C ′ ) . (86) \overline{C}'=\overline{C}\overline{I}+\frac{1}{2}
\begin{pmatrix}
\Delta C' & 2C'_{xy}\\
2C'_{xy} & -\Delta C'
\end{pmatrix}. \tag*{(86)} C ′ = C I + 2 1 ( Δ C ′ 2 C x y ′ 2 C x y ′ − Δ C ′ ) . (86)
Comparison with Eq. (85) gives
Δ C ′ = Δ C cos 2 θ − 2 C x y sin 2 θ , (87) \Delta C'=\Delta C\cos2\theta-2C_{xy}\sin2\theta, \tag*{(87)} Δ C ′ = Δ C cos 2 θ − 2 C x y sin 2 θ , (87)
2 C x y ′ = Δ C sin 2 θ + 2 C x y cos 2 θ . (88) 2C'_{xy}=\Delta C\sin2\theta+2C_{xy}\cos2\theta. \tag*{(88)} 2 C x y ′ = Δ C sin 2 θ + 2 C x y cos 2 θ . (88)
Hence,
\begin{pmatrix}
\Delta C'\\
2C'_{xy}
\end{pmatrix}
=
\begin{pmatrix}
\cos2\theta& -\sin2\theta\\
\sin2\theta& \cos2\theta
\end{pmatrix}
\begin{pmatrix}
\Delta C\\
2C_{xy}
\end{pmatrix}. \tag*{(89)}\end{pmatrix}
Equation (S17) shows that the two independent components of the traceless part of the symmetric in-plane capacitance dyadic transform as
\begin{pmatrix}
\Delta C'\\
2C'_{xy}
\end{pmatrix}
=
\overline{R}(2\theta)\cdot
\begin{pmatrix}
\Delta C\\
2C_{xy}
\end{pmatrix}. \tag*{(90)}\end{pmatrix}
Thus, the traceless part transforms in the same way as a two-dimensional vector under a rotation through the doubled angle 2 θ 2\theta 2 θ .
For a rotation that belongs to the point group of the complete unit-cell geometry, the capacitance dyadic must be invariant, so that
C ‾ ‾ ′ = C ‾ ‾ . (91) \overline{\overline{C}}' = \overline{\overline{C}}. \tag*{(91)} C ′ = C . (91)
Equivalently,
Δ C ′ = Δ C , C x y ′ = C x y . (92) \Delta C' = \Delta C,\qquad C'_{xy} = C_{xy}. \tag*{(92)} Δ C ′ = Δ C , C x y ′ = C x y . (92)
Substituting these relations into Eq. (90) , the rotational invariance condition becomes
\begin{pmatrix} \Delta C \\ 2C_{xy} \end{pmatrix}
=
\overline{\overline{R}}(2\theta)\cdot
\begin{pmatrix} \Delta C \\ 2C_{xy} \end{pmatrix}. \tag*{(93)}\end{pmatrix}
Equation (S21) gives the rotational invariance condition for the anisotropic part of the surface-capacitance dyadic.
Three- and six-fold rotational symmetry
The three-fold and six-fold cyclic groups are written explicitly as
C 3 = { I ‾ ‾ , R ‾ ‾ ( 2 π 3 ) , R ‾ ‾ ( 4 π 3 ) } , (94) C_3 = \left\{ \overline{\overline{I}},\overline{\overline{R}}\left(\frac{2\pi}{3}\right),\overline{\overline{R}}\left(\frac{4\pi}{3}\right) \right\}, \tag*{(94)} C 3 = { I , R ( 3 2 π ) , R ( 3 4 π ) } , (94)
C 6 = { I ‾ ‾ , R ‾ ‾ ( π 3 ) , R ‾ ‾ ( 2 π 3 ) , R ‾ ‾ ( π ) , R ‾ ‾ ( 4 π 3 ) , R ‾ ‾ ( 5 π 3 ) } . (95) C_6 = \left\{ \overline{\overline{I}},\overline{\overline{R}}\left(\frac{\pi}{3}\right),\overline{\overline{R}}\left(\frac{2\pi}{3}\right),\overline{\overline{R}}(\pi),\overline{\overline{R}}\left(\frac{4\pi}{3}\right),\overline{\overline{R}}\left(\frac{5\pi}{3}\right) \right\}. \tag*{(95)} C 6 = { I , R ( 3 π ) , R ( 3 2 π ) , R ( π ) , R ( 3 4 π ) , R ( 3 5 π ) } . (95)
For a rotation belonging to the point group of the unit-cell geometry, Eq. (93) can be rearranged as
[ I ‾ ‾ − R ‾ ‾ ( 2 θ ) ] ( Δ C 2 C x y ) = ( 0 0 ) . (96) \left[\overline{\overline{I}}-\overline{\overline{R}}(2\theta)\right]
\begin{pmatrix} \Delta C \\ 2C_{xy} \end{pmatrix}
=
\begin{pmatrix} 0 \\ 0 \end{pmatrix}. \tag*{(96)} [ I − R ( 2 θ ) ] ( Δ C 2 C x y ) = ( 0 0 ) . (96)
A nonzero anisotropic part can exist only if the matrix in Eq. (96) is singular, i.e.,
det [ I ‾ ‾ − R ‾ ‾ ( 2 θ ) ] = 0. (97) \det\left[\overline{\overline{I}}-\overline{\overline{R}}(2\theta)\right] = 0. \tag*{(97)} det [ I − R ( 2 θ ) ] = 0. (97)
Using
I ‾ ‾ − R ‾ ‾ ( 2 θ ) = ( 1 − cos 2 θ sin 2 θ − sin 2 θ 1 − cos 2 θ ) , (98) \overline{\overline{I}}-\overline{\overline{R}}(2\theta)
=
\begin{pmatrix}
1-\cos2\theta& \sin2\theta\\
-\sin2\theta& 1-\cos2\theta
\end{pmatrix}, \tag*{(98)} I − R ( 2 θ ) = ( 1 − cos 2 θ − sin 2 θ sin 2 θ 1 − cos 2 θ ) , (98)
the determinant becomes
det [ I ‾ ‾ − R ‾ ‾ ( 2 θ ) ] = ( 1 − cos 2 θ ) 2 + sin 2 2 θ = 1 − 2 cos 2 θ + cos 2 2 θ + sin 2 2 θ = 2 − 2 cos 2 θ = 4 sin 2 θ . (99) \begin{aligned}
\det\left[\overline{\overline{I}}-\overline{\overline{R}}(2\theta)\right]
&= (1-\cos2\theta)^2+\sin^2 2\theta\\
&= 1-2\cos2\theta+\cos^2 2\theta+\sin^2 2\theta\\
&= 2-2\cos2\theta\\
&= 4\sin^2\theta. \tag*{(99)}
\end{aligned} det [ I − R ( 2 θ ) ] = ( 1 − cos 2 θ ) 2 + sin 2 2 θ = 1 − 2 cos 2 θ + cos 2 2 θ + sin 2 2 θ = 2 − 2 cos 2 θ = 4 sin 2 θ . (99)
For C 3 C_3 C 3 , the physical rotation angle is
θ = 2 π 3 . (100) \theta= \frac{2\pi}{3}. \tag*{(100)} θ = 3 2 π . (100)
Equation (S24) therefore takes the explicit form
[ I ‾ ‾ − R ‾ ‾ ( 4 π 3 ) ] ( Δ C 2 C x y ) = ( 0 0 ) , (101) \left[\overline{\overline{I}}-\overline{\overline{R}}\left(\frac{4\pi}{3}\right)\right]
\begin{pmatrix} \Delta C \\ 2C_{xy} \end{pmatrix}
=
\begin{pmatrix} 0 \\ 0 \end{pmatrix}, \tag*{(101)} [ I − R ( 3 4 π ) ] ( Δ C 2 C x y ) = ( 0 0 ) , (101)
where
I ‾ ‾ − R ‾ ‾ ( 4 π 3 ) = ( 3 2 − 3 2 3 2 3 2 ) . (102) \overline{\overline{I}}-\overline{\overline{R}}\left(\frac{4\pi}{3}\right)
=
\begin{pmatrix}
\frac{3}{2} & -\frac{\sqrt{3}}{2} \\
\frac{\sqrt{3}}{2} & \frac{3}{2}
\end{pmatrix}. \tag*{(102)} I − R ( 3 4 π ) = ( 2 3 2 3 − 2 3 2 3 ) . (102)
Its determinant is
det [ I ‾ ‾ − R ‾ ‾ ( 4 π 3 ) ] = 3 ≠ 0. (103) \det\left[\overline{\overline{I}}-\overline{\overline{R}}\left(\frac{4\pi}{3}\right)\right]=3\ne0. \tag*{(103)} det [ I − R ( 3 4 π ) ] = 3 = 0. (103)
Hence Eq. (101) has only the trivial solution,
Δ C = 0 , C x y = 0. (104) \Delta C=0,\qquad C_{xy}=0. \tag*{(104)} Δ C = 0 , C x y = 0. (104)
Since
Δ C = C x x − C y y , (105) \Delta C=C_{xx}-C_{yy}, \tag*{(105)} Δ C = C xx − C y y , (105)
Eq. (104) gives
C x x = C y y = C , C x y = 0 , (106) C_{xx}=C_{yy}=C,\qquad C_{xy}=0, \tag*{(106)} C xx = C y y = C , C x y = 0 , (106)
and therefore
C ‾ ‾ = C I ‾ ‾ . (107) \overline{\overline{C}}=C\overline{\overline{I}}. \tag*{(107)} C = C I . (107)
For C 6 C_{6} C 6 , the physical rotation angle is
θ = π 3 . (108) \theta=\frac{\pi}{3}. \tag*{(108)} θ = 3 π . (108)
(96) becomes
[ I ‾ ‾ − R ‾ ‾ ( 2 π 3 ) ] ( Δ C 2 C x y ) = ( 0 0 ) , (109) \left[\overline{\overline{I}}-\overline{\overline{R}}\left(\frac{2\pi}{3}\right)\right]\begin{pmatrix}\Delta C\\2C_{xy}\end{pmatrix}=\begin{pmatrix}0\\0\end{pmatrix}, \tag*{(109)} [ I − R ( 3 2 π ) ] ( Δ C 2 C x y ) = ( 0 0 ) , (109)
with
I ‾ ‾ − R ‾ ‾ ( 2 π 3 ) = ( 3 2 3 2 − 3 2 3 2 ) . (110) \overline{\overline{I}}-\overline{\overline{R}}\left(\frac{2\pi}{3}\right)=\begin{pmatrix}\frac{3}{2}&\frac{\sqrt{3}}{2}\\-\frac{\sqrt{3}}{2}&\frac{3}{2}\end{pmatrix}. \tag*{(110)} I − R ( 3 2 π ) = ( 2 3 − 2 3 2 3 2 3 ) . (110)
Again,
det [ I ‾ ‾ − R ‾ ‾ ( 2 π 3 ) ] = 3 ≠ 0 , (111) \det\left[\overline{\overline{I}}-\overline{\overline{R}}\left(\frac{2\pi}{3}\right)\right]=3\ne0, \tag*{(111)} det [ I − R ( 3 2 π ) ] = 3 = 0 , (111)
and hence
Δ C = 0 , C x y = 0. (112) \Delta C=0,\qquad C_{xy}=0. \tag*{(112)} Δ C = 0 , C x y = 0. (112)
Thus,
C 3 , C 6 : C ‾ ‾ = C I ‾ ‾ . (113) C_{3},\ C_{6}:\qquad\overline{\overline{C}}=C\overline{\overline{I}}. \tag*{(113)} C 3 , C 6 : C = C I . (113)
The mirror-containing groups additionally contain reflections. A mirror whose in-plane axis makes an angle ϕ \phi ϕ with the x x x -axis is represented by
M ‾ ‾ ( ϕ ) = R ‾ ‾ ( ϕ ) ( 1 0 0 − 1 ) = R ‾ ‾ T ( ϕ ) . (114) \overline{\overline{M}}(\phi)=\overline{\overline{R}}(\phi)\begin{pmatrix}1&0\\0&-1\end{pmatrix}=\overline{\overline{R}}^{T}(\phi). \tag*{(114)} M ( ϕ ) = R ( ϕ ) ( 1 0 0 − 1 ) = R T ( ϕ ) . (114)
Carrying out the multiplication gives
M ‾ ‾ ( ϕ ) = ( cos 2 ϕ sin 2 ϕ sin 2 ϕ − cos 2 ϕ ) . (115) \overline{\overline{M}}(\phi)=\begin{pmatrix}\cos2\phi&\sin2\phi\\\sin2\phi&-\cos2\phi\end{pmatrix}. \tag*{(115)} M ( ϕ ) = ( cos 2 ϕ sin 2 ϕ sin 2 ϕ − cos 2 ϕ ) . (115)
The three mirror operations of C 3 v C_{3v} C 3 v are therefore
C 3 v = C 3 ∪ { M ‾ ‾ ( 0 ) , M ‾ ‾ ( π 3 ) , M ‾ ‾ ( 2 π 3 ) } , (116) C_{3v}=C_{3}\cup\left\{\overline{\overline{M}}(0),\overline{\overline{M}}\left(\frac{\pi}{3}\right),\overline{\overline{M}}\left(\frac{2\pi}{3}\right)\right\}, \tag*{(116)} C 3 v = C 3 ∪ { M ( 0 ) , M ( 3 π ) , M ( 3 2 π ) } , (116)
whereas the six mirror operations of C 6 v C_{6v} C 6 v give
C 6 v = C 6 ∪ { M ‾ ( 0 ) , M ‾ ( π 6 ) , M ‾ ( π 3 ) , M ‾ ( π 2 ) , M ‾ ( 2 π 3 ) , M ‾ ( 5 π 6 ) } . (117) C_{6v}=C_6\cup\left\{\overline{M}(0),\overline{M}\left(\frac{\pi}{6}\right),\overline{M}\left(\frac{\pi}{3}\right),\overline{M}\left(\frac{\pi}{2}\right),\overline{M}\left(\frac{2\pi}{3}\right),\overline{M}\left(\frac{5\pi}{6}\right)\right\}. \tag*{(117)} C 6 v = C 6 ∪ { M ( 0 ) , M ( 6 π ) , M ( 3 π ) , M ( 2 π ) , M ( 3 2 π ) , M ( 6 5 π ) } . (117)
Since
C 3 ⊂ C 3 v , C 6 ⊂ C 6 v , (118) C_3\subset C_{3v},\qquad C_6\subset C_{6v}, \tag*{(118)} C 3 ⊂ C 3 v , C 6 ⊂ C 6 v , (118)
invariance under C 3 v C_{3v} C 3 v or C 6 v C_{6v} C 6 v necessarily includes invariance under the corresponding rotational subgroup. Therefore, the rotational constraints derived above already give
C 3 , C 3 v , C 6 , C 6 v : C ‾ = C ‾ I ‾ . (119) C_3,\ C_{3v},\ C_6,\ C_{6v}:\qquad\overline{C}=\overline{C}\,\overline{I}. \tag*{(119)} C 3 , C 3 v , C 6 , C 6 v : C = C I . (119)
Table S1. Rotational constraints on the anisotropic part of the surface-capacitance dyadic. The angle θ \theta θ denotes the physical rotation angle contained in the corresponding point group.
Two-fold rotational symmetry
The two-fold cyclic group is
C 2 = { I ‾ , R ‾ ( π ) } . (120) C_2=\left\{\overline{I},\overline{R}(\pi)\right\}. \tag*{(120)} C 2 = { I , R ( π ) } . (120)
For a rotation by π \pi π ,
R ‾ ( π ) = ( − 1 0 0 − 1 ) = − I ‾ . (121) \overline{R}(\pi)=\begin{pmatrix}-1&0\\0&-1\end{pmatrix}=-\overline{I}. \tag*{(121)} R ( π ) = ( − 1 0 0 − 1 ) = − I . (121)
The invariance condition therefore gives
R ‾ ( π ) ⋅ C ‾ ⋅ R ‾ T ( π ) = ( − I ‾ ) ⋅ C ‾ ⋅ ( − I ‾ ) = C ‾ . (122) \overline{R}(\pi)\cdot\overline{C}\cdot\overline{R}^{T}(\pi)=(-\overline{I})\cdot\overline{C}\cdot(-\overline{I})=\overline{C}. \tag*{(122)} R ( π ) ⋅ C ⋅ R T ( π ) = ( − I ) ⋅ C ⋅ ( − I ) = C . (122)
Thus the invariance condition is satisfied identically, independently of the values of C x x C_{xx} C xx , C y y C_{yy} C y y , and C x y C_{xy} C x y . Consequently, C 2 C_2 C 2 symmetry imposes no additional constraint beyond the symmetry of the capacitance dyadic:
C 2 : C ‾ = ( C x x C x y C x y C y y ) . (123) C_2:\qquad\overline{C}=\begin{pmatrix}C_{xx}&C_{xy}\\C_{xy}&C_{yy}\end{pmatrix}. \tag*{(123)} C 2 : C = ( C xx C x y C x y C y y ) . (123)
The two eigenvalues and the orientation of the principal axes therefore remain unrestricted by C 2 C_2 C 2 symmetry.
Mirror symmetry
We now determine the constraint imposed directly by a mirror operation on the capacitance dyadic. The general in-plane reflection M ‾ ( ϕ ) \overline{M}(\phi) M ( ϕ ) was introduced in Eq. (115) . Choosing a coordinate frame ( e ^ ∥ , e ^ ⊥ ) (\hat{\mathbf{e}}_{\parallel},\hat{\mathbf{e}}_{\perp}) ( e ^ ∥ , e ^ ⊥ ) aligned with the mirror axis corresponds to ϕ = 0 \phi=0 ϕ = 0 , in which case
M ‾ ( 0 ) = ( 1 0 0 − 1 ) (124) \overline{M}(0)=\begin{pmatrix}1&0\\0&-1\end{pmatrix} \tag*{(124)} M ( 0 ) = ( 1 0 0 − 1 ) (124)
In this basis, the most general symmetric capacitance dyadic is
C ‾ ‾ ∥ ⊥ = ( C ∥ ∥ C ∥ ⊥ C ∥ ⊥ C ⊥ ⊥ ) . (125) \overline{\overline{C}}_{\parallel\perp} =
\begin{pmatrix}
C_{\parallel\parallel} & C_{\parallel\perp} \\
C_{\parallel\perp} & C_{\perp\perp}
\end{pmatrix}. \tag*{(125)} C ∥⊥ = ( C ∥∥ C ∥⊥ C ∥⊥ C ⊥⊥ ) . (125)
Mirror invariance requires
M ‾ ‾ ( 0 ) ⋅ C ‾ ‾ ∥ ⊥ ⋅ M ‾ ‾ T ( 0 ) = ( 1 0 0 − 1 ) ( C ∥ ∥ C ∥ ⊥ C ∥ ⊥ C ⊥ ⊥ ) ( 1 0 0 − 1 ) = ( C ∥ ∥ − C ∥ ⊥ − C ∥ ⊥ C ⊥ ⊥ ) = C ‾ ‾ ∥ ⊥ . (126) \overline{\overline{M}}^{(0)} \cdot\overline{\overline{C}}_{\parallel\perp} \cdot\overline{\overline{M}}^{T(0)}
=
\begin{pmatrix}
1 & 0 \\
0 & -1
\end{pmatrix}
\begin{pmatrix}
C_{\parallel\parallel} & C_{\parallel\perp} \\
C_{\parallel\perp} & C_{\perp\perp}
\end{pmatrix}
\begin{pmatrix}
1 & 0 \\
0 & -1
\end{pmatrix}
=
\begin{pmatrix}
C_{\parallel\parallel} & -C_{\parallel\perp} \\
-C_{\parallel\perp} & C_{\perp\perp}
\end{pmatrix}
=
\overline{\overline{C}}_{\parallel\perp}. \tag*{(126)} M ( 0 ) ⋅ C ∥⊥ ⋅ M T ( 0 ) = ( 1 0 0 − 1 ) ( C ∥∥ C ∥⊥ C ∥⊥ C ⊥⊥ ) ( 1 0 0 − 1 ) = ( C ∥∥ − C ∥⊥ − C ∥⊥ C ⊥⊥ ) = C ∥⊥ . (126)
Therefore,
C ∥ ⊥ = 0 , (127) C_{\parallel\perp}=0, \tag*{(127)} C ∥⊥ = 0 , (127)
and the allowed dyadic form in the mirror-aligned basis is
C ‾ ‾ ∥ ⊥ = ( C ∥ 0 0 C ⊥ ) . (128) \overline{\overline{C}}_{\parallel\perp} =
\begin{pmatrix}
C_{\parallel} & 0 \\
0 & C_{\perp}
\end{pmatrix}. \tag*{(128)} C ∥⊥ = ( C ∥ 0 0 C ⊥ ) . (128)
Thus the directions parallel and perpendicular to the mirror are principal axes of the capacitance dyadic.
Cs symmetry
The group C s C_s C s contains the identity and a single mirror operation,
C s = { I ‾ ‾ , M ‾ ‾ ( ϕ ) } , (129) C_s = \left\{\overline{\overline{I}},\overline{\overline{M}}(\phi)\right\}, \tag*{(129)} C s = { I , M ( ϕ ) } , (129)
where the mirror axis makes an angle ϕ \phi ϕ with the fixed Cartesian x x x -axis. In the mirror-aligned basis, Eq. (128) gives
C ‾ ‾ ∥ ⊥ = ( C ∥ 0 0 C ⊥ ) . (130) \overline{\overline{C}}_{\parallel\perp} =
\begin{pmatrix}
C_{\parallel} & 0 \\
0 & C_{\perp}
\end{pmatrix}. \tag*{(130)} C ∥⊥ = ( C ∥ 0 0 C ⊥ ) . (130)
Transforming this dyadic to the fixed Cartesian frame gives
C ‾ ‾ = R ‾ ‾ ( ϕ ) ⋅ ( C ∥ 0 0 C ⊥ ) ⋅ R ‾ ‾ T ( ϕ ) . (131) \overline{\overline{C}} =
\overline{\overline{R}}(\phi) \cdot
\begin{pmatrix}
C_{\parallel} & 0 \\
0 & C_{\perp}
\end{pmatrix}
\cdot\overline{\overline{R}}^{T}(\phi). \tag*{(131)} C = R ( ϕ ) ⋅ ( C ∥ 0 0 C ⊥ ) ⋅ R T ( ϕ ) . (131)
Carrying out the multiplication,
C ‾ ‾ = ( C ∥ cos 2 ϕ + C ⊥ sin 2 ϕ ( C ∥ − C ⊥ ) sin ϕ cos ϕ ( C ∥ − C ⊥ ) sin ϕ cos ϕ C ∥ sin 2 ϕ + C ⊥ cos 2 ϕ ) . (132) \overline{\overline{C}} =
\begin{pmatrix}
C_{\parallel}\cos^2\phi+C_{\perp}\sin^2\phi&
(C_{\parallel}-C_{\perp})\sin\phi\cos\phi\\
(C_{\parallel}-C_{\perp})\sin\phi\cos\phi&
C_{\parallel}\sin^2\phi+C_{\perp}\cos^2\phi
\end{pmatrix}. \tag*{(132)} C = ( C ∥ cos 2 ϕ + C ⊥ sin 2 ϕ ( C ∥ − C ⊥ ) sin ϕ cos ϕ ( C ∥ − C ⊥ ) sin ϕ cos ϕ C ∥ sin 2 ϕ + C ⊥ cos 2 ϕ ) . (132)
Therefore,
C x x = C ∥ cos 2 ϕ + C ⊥ sin 2 ϕ , (133) C_{xx}=C_{\parallel}\cos^2\phi+C_{\perp}\sin^2\phi, \tag*{(133)} C xx = C ∥ cos 2 ϕ + C ⊥ sin 2 ϕ , (133)
C y y = C ∥ sin 2 ϕ + C ⊥ cos 2 ϕ , (134) C_{yy}=C_{\parallel}\sin^2\phi+C_{\perp}\cos^2\phi, \tag*{(134)} C y y = C ∥ sin 2 ϕ + C ⊥ cos 2 ϕ , (134)
C x y = ( C ∥ − C ⊥ ) sin ϕ cos ϕ . (135) C_{xy}=(C_{\parallel}-C_{\perp})\sin\phi\cos\phi. \tag*{(135)} C x y = ( C ∥ − C ⊥ ) sin ϕ cos ϕ . (135)
Hence the allowed tensor form may be summarized as
C s : C ‾ ‾ = R ‾ ‾ ( ϕ ) ⋅ ( C ∥ 0 0 C ⊥ ) ⋅ R ‾ ‾ T ( ϕ ) . (136) C_s:\quad\overline{\overline{C}} =
\overline{\overline{R}}(\phi)\cdot
\begin{pmatrix}
C_{\parallel} & 0 \\
0 & C_{\perp}
\end{pmatrix}
\cdot\overline{\overline{R}}^{T}(\phi). \tag*{(136)} C s : C = R ( ϕ ) ⋅ ( C ∥ 0 0 C ⊥ ) ⋅ R T ( ϕ ) . (136)
Thus the two principal capacitances remain independent, while their principal axes are fixed parallel and perpendicular to the mirror.
C2v symmetry
Choosing the Cartesian axes along the two mutually perpendicular mirror lines, the C 2 v C_{2v} C 2 v group can be written as
C 2 v = { I ‾ , R ‾ ( π ) , M ‾ ( 0 ) , M ‾ ( π 2 ) } . (137) C_{2v}=\left\{\overline{I},\overline{R}(\pi),\overline{M}(0),\overline{M}\left(\frac{\pi}{2}\right)\right\}. \tag*{(137)} C 2 v = { I , R ( π ) , M ( 0 ) , M ( 2 π ) } . (137)
The two mirror operations are
M ‾ ( 0 ) = ( 1 0 0 − 1 ) , M ‾ ( π 2 ) = ( − 1 0 0 1 ) . (138) \overline{M}(0)=\begin{pmatrix}1&0\\0&-1\end{pmatrix},\qquad\overline{M}\left(\frac{\pi}{2}\right)=\begin{pmatrix}-1&0\\0&1\end{pmatrix}. \tag*{(138)} M ( 0 ) = ( 1 0 0 − 1 ) , M ( 2 π ) = ( − 1 0 0 1 ) . (138)
Starting from the general symmetric dyadic
C ‾ = ( C x x C x y C x y C y y ) , (139) \overline{C}=\begin{pmatrix}C_{xx}&C_{xy}\\C_{xy}&C_{yy}\end{pmatrix}, \tag*{(139)} C = ( C xx C x y C x y C y y ) , (139)
invariance under the first mirror gives
M ‾ ( 0 ) ⋅ C ‾ ⋅ M ‾ T ( 0 ) = ( C x x − C x y − C x y C y y ) = C ‾ , (140) \overline{M}(0)\cdot\overline{C}\cdot\overline{M}^{T}(0)=\begin{pmatrix}C_{xx}&-C_{xy}\\-C_{xy}&C_{yy}\end{pmatrix}=\overline{C}, \tag*{(140)} M ( 0 ) ⋅ C ⋅ M T ( 0 ) = ( C xx − C x y − C x y C y y ) = C , (140)
and therefore
C x y = 0. (141) C_{xy}=0. \tag*{(141)} C x y = 0. (141)
The second mirror gives the same condition,
M ‾ ( π 2 ) ⋅ C ‾ ⋅ M ‾ T ( π 2 ) = C ‾ , C x y = 0. (142) \overline{M}\left(\frac{\pi}{2}\right)\cdot\overline{C}\cdot\overline{M}^{T}\left(\frac{\pi}{2}\right)=\overline{C},\qquad C_{xy}=0. \tag*{(142)} M ( 2 π ) ⋅ C ⋅ M T ( 2 π ) = C , C x y = 0. (142)
The two-fold rotation R ‾ ( π ) = − I ‾ \overline{R}(\pi)=-\overline{I} R ( π ) = − I introduces no additional constraint, as shown in Eq. (122) . The allowed form is therefore
C 2 v : C ‾ = ( C x x 0 0 C y y ) . (143) C_{2v}:\quad\overline{C}=\begin{pmatrix}C_{xx}&0\\0&C_{yy}\end{pmatrix}. \tag*{(143)} C 2 v : C = ( C xx 0 0 C y y ) . (143)
Thus C x x C_{xx} C xx and C y y C_{yy} C y y remain independent, while the two mirror axes fix the principal directions of the dyadic.
Trivial symmetry C1
The trivial point group contains only the identity operation,
C 1 = { I ‾ } . (144) C_{1}=\left\{\overline{I}\right\}. \tag*{(144)} C 1 = { I } . (144)
The invariance condition is therefore
I ‾ ⋅ C ‾ ⋅ I ‾ T = C ‾ , (145) \overline{I}\cdot\overline{C}\cdot\overline{I}^{T}=\overline{C}, \tag*{(145)} I ⋅ C ⋅ I T = C , (145)
which is identically satisfied. Hence C 1 C_{1} C 1 imposes no point-symmetry constraint on the capacitance dyadic. For a reciprocal lossless sheet, the most general allowed form is
C 1 : C ‾ = ( C x x C x y C x y C y y ) . (146) C_{1}:\quad\overline{C}=\begin{pmatrix}C_{xx}&C_{xy}\\C_{xy}&C_{yy}\end{pmatrix}. \tag*{(146)} C 1 : C = ( C xx C x y C x y C y y ) . (146)
The response therefore contains three independent real parameters, and the orientation of the principal axes is not fixed by point symmetry.
Let λ 1 \lambda_1 λ 1 and λ 2 \lambda_2 λ 2 denote the two principal values of the capacitance dyadic, and let the first principal axis make an angle ϕ C \phi_C ϕ C with the Cartesian x x x -axis. The dyadic can then be written as
C ‾ ‾ = R ‾ ( ϕ C ) ⋅ ( λ 1 0 0 λ 2 ) ⋅ R ‾ T ( ϕ C ) . (147) \overline{\overline{C}} = \overline{R}(\phi_C) \cdot\begin{pmatrix} \lambda_1 & 0 \\ 0 & \lambda_2 \end{pmatrix} \cdot\overline{R}^{T}(\phi_C). \tag*{(147)} C = R ( ϕ C ) ⋅ ( λ 1 0 0 λ 2 ) ⋅ R T ( ϕ C ) . (147)
The principal values and their orientation can be obtained directly from the Cartesian components of the dyadic. Defining the mean capacitance as
C ‾ = C x x + C y y 2 = 1 2 tr ( C ‾ ‾ ) , (148) \overline{C} = \frac{C_{xx}+C_{yy}}{2} = \frac{1}{2}\operatorname{tr}\left(\overline{\overline{C}}\right), \tag*{(148)} C = 2 C xx + C y y = 2 1 tr ( C ) , (148)
the two principal values are
λ 1 , 2 = C ‾ ± 1 2 ( C x x − C y y ) 2 + 4 C x y 2 , (149) \lambda_{1,2} = \overline{C} \pm\frac{1}{2}\sqrt{(C_{xx}-C_{yy})^2 + 4C_{xy}^2}, \tag*{(149)} λ 1 , 2 = C ± 2 1 ( C xx − C y y ) 2 + 4 C x y 2 , (149)
where the upper sign corresponds to λ 1 \lambda_1 λ 1 . Their orientation is
ϕ C = 1 2 atan2 ( 2 C x y , C x x − C y y ) ( m o d π ) . (150) \phi_C = \frac{1}{2}\operatorname{atan2}(2C_{xy}, C_{xx}-C_{yy}) \pmod{\pi}. \tag*{(150)} ϕ C = 2 1 atan2 ( 2 C x y , C xx − C y y ) ( mod π ) . (150)
When λ 1 = λ 2 \lambda_1 = \lambda_2 λ 1 = λ 2 , the response is isotropic and ϕ C \phi_C ϕ C is arbitrary. For an arbitrary in-plane direction
u ^ ( θ ) = cos θ x ^ + sin θ y ^ , (151) \hat{\mathbf{u}}(\theta) = \cos\theta\,\hat{\mathbf{x}} + \sin\theta\,\hat{\mathbf{y}}, \tag*{(151)} u ^ ( θ ) = cos θ x ^ + sin θ y ^ , (151)
the directional capacitance is
C ( u ^ ) = u ^ ⋅ C ‾ ‾ ⋅ u ^ . (152) C(\hat{\mathbf{u}}) = \hat{\mathbf{u}} \cdot\overline{\overline{C}} \cdot\hat{\mathbf{u}}. \tag*{(152)} C ( u ^ ) = u ^ ⋅ C ⋅ u ^ . (152)
Substitution of Eq. (147) gives
C ( u ^ ) = λ 1 cos 2 ( θ − ϕ C ) + λ 2 sin 2 ( θ − ϕ C ) = λ 1 + λ 2 2 + λ 1 − λ 2 2 cos [ 2 ( θ − ϕ C ) ] . (153) \begin{aligned}
C(\hat{\mathbf{u}}) &= \lambda_1\cos^2(\theta-\phi_C) + \lambda_2\sin^2(\theta-\phi_C) \\
&= \frac{\lambda_1+\lambda_2}{2} + \frac{\lambda_1-\lambda_2}{2}\cos[2(\theta-\phi_C)]. \tag*{(153)}
\end{aligned} C ( u ^ ) = λ 1 cos 2 ( θ − ϕ C ) + λ 2 sin 2 ( θ − ϕ C ) = 2 λ 1 + λ 2 + 2 λ 1 − λ 2 cos [ 2 ( θ − ϕ C )] . (153)
The normalized directional-capacitance indicator plotted in Figure 3 of the main text becomes
C ( u ^ ) C ‾ = 1 + λ 1 − λ 2 λ 1 + λ 2 cos [ 2 ( θ − ϕ C ) ] . (154) \frac{C(\hat{\mathbf{u}})}{\overline{C}} = 1 + \frac{\lambda_1-\lambda_2}{\lambda_1+\lambda_2}\cos[2(\theta-\phi_C)]. \tag*{(154)} C C ( u ^ ) = 1 + λ 1 + λ 2 λ 1 − λ 2 cos [ 2 ( θ − ϕ C )] . (154)
Quasi-static corrector problem and numerical evaluation of the capacitance dyadic
This section gives the mathematical and numerical details of the quasi-static calculation introduced in Sec. II C of the main text. We first define the affine-periodic electrostatic cell problem and its two basis correctors. We then derive the corresponding minimum-energy and Euler–Lagrange formulations, describe the periodic triangular-graph approximation used for arbitrary patch geometries, and finally identify the entries of the auxiliary capacitance dyadic. The dyadic obtained here has units of F/m. Empty-cell subtraction and conversion to the equivalent sheet-capacitance dyadic, with units of F, are separate operations described in Sec. II D of the main text.
Affine-periodic electrostatic cell problem
Let U ⊂ R 2 U \subset\mathbb{R}^{2} U ⊂ R 2 be a primitive cell with lattice translations a 1 \mathbf{a}_{1} a 1 and a 2 \mathbf{a}_{2} a 2 . The metallic region is the union of N c N_{\mathrm{c}} N c electrically disconnected perfect-electric-conductor (PEC) components,
Σ = ⋃ m = 1 N c Σ m , Ω = U ∖ Σ , (155) \Sigma= \bigcup_{m=1}^{N_{\mathrm{c}}} \Sigma_{m}, \qquad\Omega= U \setminus\Sigma, \tag*{(155)} Σ = m = 1 ⋃ N c Σ m , Ω = U ∖ Σ , (155)
where Ω \Omega Ω is the dielectric gap. Connectivity is understood on the periodically identified cell; consequently, portions meeting through opposite cell boundaries belong to the same physical component. We consider the gap-connected geometries described in Sec. II A of the main text, for which independent winding paths exist along both primitive translations.
For a dimensionless imposed in-plane direction
v = v x e ^ x + v y e ^ y , (156) \mathbf{v} = v_{x}\hat{\mathbf{e}}_{x} + v_{y}\hat{\mathbf{e}}_{y}, \tag*{(156)} v = v x e ^ x + v y e ^ y , (156)
let E 0 E_{0} E 0 be the equivalent field amplitude, in V/m. Introduce the normalized total potential-like function
u v ( r ) = v ⋅ r + w v ( r ) , Φ v ( r ) = − E 0 u v ( r ) . (157) u_{\mathbf{v}}(\mathbf{r}) = \mathbf{v} \cdot\mathbf{r} + w_{\mathbf{v}}(\mathbf{r}), \qquad\Phi_{\mathbf{v}}(\mathbf{r}) = -E_{0}u_{\mathbf{v}}(\mathbf{r}). \tag*{(157)} u v ( r ) = v ⋅ r + w v ( r ) , Φ v ( r ) = − E 0 u v ( r ) . (157)
The associated microscopic electric field is
E v c e l l ( r ) = − ∇ Φ v = E 0 ( v + ∇ w v ) . (158) \mathbf{E}^{\mathrm{cell}}_{\mathbf{v}}(\mathbf{r}) = -\nabla\Phi_{\mathbf{v}} = E_{0}\left(\mathbf{v} + \nabla w_{\mathbf{v}}\right). \tag*{(158)} E v cell ( r ) = − ∇ Φ v = E 0 ( v + ∇ w v ) . (158)
Thus u v u_{\mathbf{v}} u v and w v w_{\mathbf{v}} w v have units of length, whereas v + ∇ w v \mathbf{v} + \nabla w_{\mathbf{v}} v + ∇ w v is dimensionless.
Only the microscopic correction is periodic:
w v ( r + a i ) = w v ( r ) , (159) w_{\mathbf{v}}(\mathbf{r}+\mathbf{a}_{i}) = w_{\mathbf{v}}(\mathbf{r}), \tag*{(159)} w v ( r + a i ) = w v ( r ) , (159)
u v ( r + a i ) = u v ( r ) + v ⋅ a i , i = 1 , 2. (160) u_{\mathbf{v}}(\mathbf{r}+\mathbf{a}_{i}) = u_{\mathbf{v}}(\mathbf{r}) + \mathbf{v}\cdot\mathbf{a}_{i}, \qquad i=1,2. \tag*{(160)} u v ( r + a i ) = u v ( r ) + v ⋅ a i , i = 1 , 2. (160)
Accordingly, the total potential is affine-periodic, while its gradient and the microscopic cell field are periodic.
The tangential electric field vanishes on each PEC component. Therefore it is the total potential, rather than the corrector alone, that is constant on a conductor:
u v ( r ) = v ⋅ r + w v ( r ) = c m ( v ) , r ∈ ∂ Σ m . (161) u_{\mathbf{v}}(\mathbf{r}) = \mathbf{v}\cdot\mathbf{r} + w_{\mathbf{v}}(\mathbf{r}) = c_{m}(\mathbf{v}), \qquad\mathbf{r}\in\partial\Sigma_{m}. \tag*{(161)} u v ( r ) = v ⋅ r + w v ( r ) = c m ( v ) , r ∈ ∂ Σ m . (161)
The constants c m c_{m} c m are not prescribed because the conductors are floating; their equivalent potentials are − E 0 c m -E_{0}c_{m} − E 0 c m .
Let H p e r 1 ( Ω ) H^{1}_{\mathrm{per}}(\Omega) H per 1 ( Ω ) denote the Sobolev space whose traces agree on paired cell boundaries. For a prescribed loading v \mathbf{v} v , define the affine admissible set
A v = { w ∈ H p e r 1 ( Ω ) : v ⋅ r + w = c m on ∂ Σ m for some c m ∈ R } . (162) \mathcal{A}_{\mathbf{v}} = \left\{w\in H^{1}_{\mathrm{per}}(\Omega):\ \mathbf{v}\cdot\mathbf{r}+w=c_{m}\ \text{on }\partial\Sigma_{m}\ \text{for some }c_{m}\in\mathbb{R}\right\}. \tag*{(162)} A v = { w ∈ H per 1 ( Ω ) : v ⋅ r + w = c m on ∂ Σ m for some c m ∈ R } . (162)
The field is unchanged under w v ↦ w v + C w_{\mathbf{v}}\mapsto w_{\mathbf{v}}+C w v ↦ w v + C . A convenient continuum gauge is
∫ Ω w v d A = 0. (163) \int_{\Omega} w_{\mathbf{v}}\,\mathrm{d}A = 0. \tag*{(163)} ∫ Ω w v d A = 0. (163)
This fixes only the arbitrary potential zero and has no effect on the field or on the capacitance dyadic.
For a trial corrector w ∈ A v w \in\mathcal{A}_{\mathbf{v}} w ∈ A v , the cell-averaged electrostatic energy density of the two-dimensional surrogate is
W e [ w ; v ] = ϵ 0 E 0 2 2 ∣ U ∣ ∫ Ω ∣ v + ∇ w ∣ 2 d A . (164) \mathcal{W}_{e}[w;\mathbf{v}] = \frac{\epsilon_0 E_0^2}{2|U|} \int_{\Omega} \left|\mathbf{v}+\nabla w\right|^2\,\mathrm{d}A. \tag*{(164)} W e [ w ; v ] = 2∣ U ∣ ϵ 0 E 0 2 ∫ Ω ∣ v + ∇ w ∣ 2 d A . (164)
Here ϵ 0 E 0 2 \epsilon_0 E_0^2 ϵ 0 E 0 2 has units of J/m3 ^3 3 , while the area integral divided by ∣ U ∣ |U| ∣ U ∣ is dimensionless. Equation (164) is therefore the raw energy-density quantity associated with the auxiliary two-dimensional problem, not yet the energy of the final sheet model. For a substrate-backed array, ϵ 0 \epsilon_0 ϵ 0 is replaced by the quasi-static effective permittivity defined in Eq. (17) of the main text.
Since the positive factor ϵ 0 E 0 2 \epsilon_0 E_0^2 ϵ 0 E 0 2 does not affect the minimizer, define the normalized Dirichlet functional
J v [ w ] = 1 2 ∣ U ∣ ∫ Ω ∣ v + ∇ w ∣ 2 d A . (165) \mathcal{J}_{\mathbf{v}}[w] = \frac{1}{2|U|}\int_{\Omega}\left|\mathbf{v}+\nabla w\right|^2\,\mathrm{d}A. \tag*{(165)} J v [ w ] = 2∣ U ∣ 1 ∫ Ω ∣ v + ∇ w ∣ 2 d A . (165)
The corrector is the minimum-energy solution
w v ∈ arg min w ∈ A v / R J v [ w ] , W e ( v ) = ϵ 0 E 0 2 J v [ w v ] . (166) w_{\mathbf{v}} \in\underset{w\in\mathcal{A}_{\mathbf{v}}/\mathbb{R}}{\arg\min}\mathcal{J}_{\mathbf{v}}[w],\qquad\mathcal{W}_{e}(\mathbf{v})=\epsilon_0 E_0^2\mathcal{J}_{\mathbf{v}}[w_{\mathbf{v}}]. \tag*{(166)} w v ∈ w ∈ A v / R arg min J v [ w ] , W e ( v ) = ϵ 0 E 0 2 J v [ w v ] . (166)
The admissible variations form the space
V = { φ ∈ H p e r 1 ( Ω ) : φ = d m on ∂ Σ m for some d m ∈ R } . (167) \mathcal{V}=\left\{\varphi\in H^1_{\mathrm{per}}(\Omega):\varphi=d_m\ \text{on }\partial\Sigma_m\ \text{for some }d_m\in\mathbb{R}\right\}. \tag*{(167)} V = { φ ∈ H per 1 ( Ω ) : φ = d m on ∂ Σ m for some d m ∈ R } . (167)
The constants d m d_m d m may vary independently because the corresponding conductor potentials are floating. For every φ ∈ V \varphi\in\mathcal{V} φ ∈ V , the first variation is
0 = d d t J v [ w v + t φ ] ∣ t = 0 = 1 ∣ U ∣ ∫ Ω ( v + ∇ w v ) ⋅ ∇ φ d A . (168) 0=\left.\frac{\mathrm{d}}{\mathrm{d}t}\mathcal{J}_{\mathbf{v}}[w_{\mathbf{v}}+t\varphi]\right|_{t=0}
=\frac{1}{|U|}\int_{\Omega}(\mathbf{v}+\nabla w_{\mathbf{v}})\cdot\nabla\varphi\,\mathrm{d}A. \tag*{(168)} 0 = d t d J v [ w v + tφ ] t = 0 = ∣ U ∣ 1 ∫ Ω ( v + ∇ w v ) ⋅ ∇ φ d A . (168)
Hence the weak cell problem is
∫ Ω ( v + ∇ w v ) ⋅ ∇ φ d A = 0 , ∀ φ ∈ V . (169) \int_{\Omega}(\mathbf{v}+\nabla w_{\mathbf{v}})\cdot\nabla\varphi\,\mathrm{d}A=0,\qquad\forall\varphi\in\mathcal{V}. \tag*{(169)} ∫ Ω ( v + ∇ w v ) ⋅ ∇ φ d A = 0 , ∀ φ ∈ V . (169)
Compactly supported variations give Laplace’s equation in the gap. The contributions from paired cell boundaries cancel because the field and the test function are periodic. Independent variation of each d m d_m d m gives zero total flux on every floating conductor. The equivalent strong problem is
∇ ⋅ ( v + ∇ w v ) = 0 in Ω , v ⋅ r + w v = c m ( v ) on ∂ Σ m , ∫ ∂ Σ m n ^ m ⋅ ( v + ∇ w v ) d ℓ = 0 m = 1 , … , N c , w v ( r + a i ) = w v ( r ) i = 1 , 2. (170) \begin{aligned}
\nabla\cdot(\mathbf{v}+\nabla w_{\mathbf{v}})&=0 &&\text{in }\Omega,\\
\mathbf{v}\cdot\mathbf{r}+w_{\mathbf{v}}&=c_m(\mathbf{v}) &&\text{on }\partial\Sigma_m,\\
\int_{\partial\Sigma_m}\hat{\mathbf{n}}_m\cdot(\mathbf{v}+\nabla w_{\mathbf{v}})\,\mathrm{d}\ell&=0 &&m=1,\ldots,N_c,\\
w_{\mathbf{v}}(\mathbf{r}+\mathbf{a}_i)&=w_{\mathbf{v}}(\mathbf{r}) &&i=1,2. \tag*{(170)}
\end{aligned} ∇ ⋅ ( v + ∇ w v ) v ⋅ r + w v ∫ ∂ Σ m n ^ m ⋅ ( v + ∇ w v ) d ℓ w v ( r + a i ) = 0 = c m ( v ) = 0 = w v ( r ) in Ω , on ∂ Σ m , m = 1 , … , N c , i = 1 , 2. (170)
together with the gauge in Eq. (163) . Here n ^ m \hat{\mathbf{n}}_m n ^ m points from the metal into the gap. Multiplication of Eq. (S97c) by ϵ 0 E 0 \epsilon_0 E_0 ϵ 0 E 0 gives the zero-net-charge condition on the m m m th floating conductor.
The cell problem is linear in the imposed loading. With the gauge fixed,
w v = v x w x + v y w y , c m ( v ) = v x c m , x + v y c m , y , (171) w_{\mathbf{v}}=v_xw_x+v_yw_y,\qquad c_m(\mathbf{v})=v_xc_{m,x}+v_yc_{m,y}, \tag*{(171)} w v = v x w x + v y w y , c m ( v ) = v x c m , x + v y c m , y , (171)
where w x w_x w x and w y w_y w y are generated by v = e ^ x \mathbf{v}=\hat{\mathbf{e}}_x v = e ^ x and v = e ^ y \mathbf{v}=\hat{\mathbf{e}}_y v = e ^ y , respectively. Only these two basis problems must therefore be solved.
Periodic triangular-graph approximation
The numerical implementation replaces the continuum problem by a periodic triangular graph. The primitive parallelogram is sampled at
r i j = s i a 1 + t j a 2 , s i = i − 1 2 N − 1 2 , t j = j − 1 2 N − 1 2 , (172) \mathbf{r}_{ij}=s_i\mathbf{a}_1+t_j\mathbf{a}_2,\qquad s_i=\frac{i-\frac{1}{2}}{N}-\frac{1}{2},\qquad t_j=\frac{j-\frac{1}{2}}{N}-\frac{1}{2}, \tag*{(172)} r ij = s i a 1 + t j a 2 , s i = N i − 2 1 − 2 1 , t j = N j − 2 1 − 2 1 , (172)
for i , j = 1 , … , N i,j=1,\ldots,N i , j = 1 , … , N . The value of N N N is chosen odd so that sampling points do not lie systematically on symmetry or polygon-boundary lines. For the metal-membership test, each point is folded into the Wigner–Seitz representative of the same periodic cell; the associated lattice translation is retained when affine edge offsets are formed. Each sample is connected to its periodic neighbours in the three positive triangular-grid directions
a 1 N , a 2 N , − a 1 + a 2 N . (173) \frac{\mathbf{a}_1}{N},\qquad\frac{\mathbf{a}_2}{N},\qquad\frac{-\mathbf{a}_1+\mathbf{a}_2}{N}. \tag*{(173)} N a 1 , N a 2 , N − a 1 + a 2 . (173)
Thus, each undirected physical bond is stored exactly once. Before contraction, the periodic graph contains N 2 N^2 N 2 samples and 3 N 2 3N^2 3 N 2 oriented bond representatives.
Let
Q N = { r q } q = 1 N 2 = G N ∪ ˙ ( ⋃ m = 1 N c Σ m , N ) (174) Q_N=\{\mathbf{r}_q\}_{q=1}^{N^2}=\mathcal{G}_N\mathbin{\dot\cup}\left(\bigcup_{m=1}^{N_c}\Sigma_{m,N}\right) \tag*{(174)} Q N = { r q } q = 1 N 2 = G N ∪ ˙ ( m = 1 ⋃ N c Σ m , N ) (174)
be the sampled points, partitioned into the gap samples G N \mathcal{G}_N G N and the samples belonging to each periodically connected PEC component. Every gap sample carries one independent degree of freedom, whereas all samples in the same PEC component are contracted to one metal supernode. If N d N_d N d is the number of gap samples, the number of graph unknowns is
N v = N d + N c . (175) N_v=N_d+N_c. \tag*{(175)} N v = N d + N c . (175)
Introduce the point-to-unknown map
ν ( q ) = { m , r q ∈ Σ m , N , N c + j ( q ) , r q is the j ( q ) th gap sample , (176) \nu(q)=
\begin{cases}
m, & \mathbf{r}_q\in\Sigma_{m,N},\\
N_c+j(q), & \mathbf{r}_q\text{ is the }j(q)\text{th gap sample},
\end{cases} \tag*{(176)} ν ( q ) = { m , N c + j ( q ) , r q ∈ Σ m , N , r q is the j ( q ) th gap sample , (176)
and the graph coordinate functions
ψ p ( r q ) = δ p , ν ( q ) , p = 1 , … , N v . (177) \psi_p(\mathbf{r}_q)=\delta_{p,\nu(q)},\qquad p=1,\ldots,N_v. \tag*{(177)} ψ p ( r q ) = δ p , ν ( q ) , p = 1 , … , N v . (177)
The resulting discrete potential space is
X N = span { ψ 1 , … , ψ N v } . (178) \mathcal{X}_N=\operatorname{span}\{\psi_1,\ldots,\psi_{N_v}\}. \tag*{(178)} X N = span { ψ 1 , … , ψ N v } . (178)
For a gap degree of freedom, the corresponding basis function is nonzero at only one sampled gap point. By contrast, the basis function associated with the m m m th metal supernode is equal to unity at every sampled point belonging to Σ m , N \Sigma_{m,N} Σ m , N and vanishes elsewhere. Thus, the contraction of a connected PEC component to one basis function imposes its equipotential condition directly.
These are coordinate functions on the sampled graph and are not being identified with conforming H 1 H^1 H 1 finite-element basis functions.
For loading direction ξ ∈ { x , y } \xi\in\{x,y\} ξ ∈ { x , y } , let z ξ ∈ R N v \mathbf{z}_{\xi}\in\mathbb{R}^{N_v} z ξ ∈ R N v contain the graph values of the normalized total affine potential. At a representative sample,
u ξ , N ( r q ) = ∑ p = 1 N v z ξ , p ψ p ( r q ) = z ξ , ν ( q ) , (179) u_{\xi,N}(\mathbf{r}_q)=\sum_{p=1}^{N_v}z_{\xi,p}\psi_p(\mathbf{r}_q)=z_{\xi,\nu(q)}, \tag*{(179)} u ξ , N ( r q ) = p = 1 ∑ N v z ξ , p ψ p ( r q ) = z ξ , ν ( q ) , (179)
up to the known affine lattice translation attached to that representative. In particular, z ξ , m = c m , ξ z_{\xi,m}=c_{m,\xi} z ξ , m = c m , ξ is the constant total potential of the m m m th metal component; it is not a constant value of the periodic corrector. In continuum notation the corresponding corrector trace is
w ξ ( r ) = c m , ξ − e ^ ξ ⋅ r , r ∈ ∂ Σ m . (180) w_{\xi}(\mathbf{r})=c_{m,\xi}-\widehat{\mathbf{e}}_{\xi}\cdot\mathbf{r},\qquad\mathbf{r}\in\partial\Sigma_m. \tag*{(180)} w ξ ( r ) = c m , ξ − e ξ ⋅ r , r ∈ ∂ Σ m . (180)
Choose one orientation for each stored bond. An edge e = ( q , r ; R e ) e=(q,r;\mathbf{R}_e) e = ( q , r ; R e ) joins sample q q q to the appropriate periodic image of sample r r r . The vector R e \mathbf{R}_e R e collects the lattice translation produced by a periodic wrap and, when a metal sample has been folded to the Wigner–Seitz representative, the corresponding change of representative. After metal contraction, define the incidence matrix by
B e , ν ( q ) = − 1 , B e , ν ( r ) = + 1 , (181) B_{e,\nu(q)}=-1,\qquad B_{e,\nu(r)}=+1, \tag*{(181)} B e , ν ( q ) = − 1 , B e , ν ( r ) = + 1 , (181)
with all other entries in row e e e equal to zero. A bond whose two endpoints map to the same contracted unknown has a zero incidence row and is discarded from the active solve-edge set
E N a c t = { e : ν ( q e ) ≠ ν ( r e ) } . (182) \mathcal{E}^{\mathrm{act}}_N=\{e:\nu(q_e)\ne\nu(r_e)\}. \tag*{(182)} E N act = { e : ν ( q e ) = ν ( r e )} . (182)
For the ξ \xi ξ -directed basis problem, affine periodicity supplies the known offset
( g ξ ) e = e ^ ξ ⋅ R e . (183) (g_\xi)_e=\hat{\mathbf{e}}_\xi\cdot\mathbf{R}_e. \tag*{(183)} ( g ξ ) e = e ^ ξ ⋅ R e . (183)
The oriented total-potential drop is therefore
d ξ , e = z ξ , ν ( r e ) − z ξ , ν ( q e ) + e ^ ξ ⋅ R e = ( B z ξ + g ξ ) e . (184) \begin{aligned}
d_{\xi,e}=z_{\xi,\nu(r_e)}-z_{\xi,\nu(q_e)}+\hat{\mathbf{e}}_\xi\cdot\mathbf{R}_e \\
&=(B\mathbf{z}_\xi+\mathbf{g}_\xi)_e. \tag*{(184)}
\end{aligned} d ξ , e = z ξ , ν ( r e ) − z ξ , ν ( q e ) + e ^ ξ ⋅ R e = ( B z ξ + g ξ ) e . (184)
The entries of z ξ \mathbf{z}_\xi z ξ , g ξ \mathbf{g}_\xi g ξ , and d ξ \mathbf{d}_\xi d ξ all have units of length.
For an equilateral triangular grid, the cotangent weight of every periodic bond is
κ △ = 1 2 ( cot π 3 + cot π 3 ) = 1 3 . (185) \kappa_\triangle=\frac{1}{2}\left(\cot\frac{\pi}{3}+\cot\frac{\pi}{3}\right)=\frac{1}{\sqrt{3}}. \tag*{(185)} κ △ = 2 1 ( cot 3 π + cot 3 π ) = 3 1 . (185)
Because every undirected bond is stored once, the discrete energy used to assemble the graph problem is
J ξ , N a c t ( z ξ ) = κ △ 2 ∣ U ∣ ∑ e ∈ E N a c t d ξ , e 2 . (186) \mathcal{J}^{\mathrm{act}}_{\xi,N}(\mathbf{z}_\xi)=\frac{\kappa_\triangle}{2\lvert U\rvert}\sum_{e\in\mathcal{E}^{\mathrm{act}}_N}d_{\xi,e}^{2}. \tag*{(186)} J ξ , N act ( z ξ ) = 2 ∣ U ∣ κ △ e ∈ E N act ∑ d ξ , e 2 . (186)
There is no additional factor associated with duplicated edge orientations, because the implementation does not store the reverse copy of each bond.
Writing
W = κ △ I , L = B T W B , b ξ = − B T W g ξ , (187) \mathbf{W}=\kappa_\triangle\mathbf{I},\qquad\mathbf{L}=\mathbf{B}^{T}\mathbf{W}\mathbf{B},\qquad\mathbf{b}_\xi=-\mathbf{B}^{T}\mathbf{W}\mathbf{g}_\xi, \tag*{(187)} W = κ △ I , L = B T WB , b ξ = − B T W g ξ , (187)
stationarity of Eq. (186) gives
L z ξ = b ξ . (188) \mathbf{L}\mathbf{z}_\xi=\mathbf{b}_\xi. \tag*{(188)} L z ξ = b ξ . (188)
The equation associated with a freely varying metal supernode is the discrete counterpart of the component-neutrality condition in Eq. (S97c).
The Laplacian has one constant null vector on every connected component of the active graph. The implementation removes these gauge freedoms by selecting one index f c f_c f c in every graph component c c c and imposing
z ξ , f c = 0. (189) z_{\xi,f_c}=0. \tag*{(189)} z ξ , f c = 0. (189)
Equivalently, if R \mathcal{R} R denotes the unpinned indices, the computed solution satisfies
L R R z ξ , R = b ξ , R , z ξ , f c = 0. (190) \mathbf{L}_{\mathcal{R}\mathcal{R}}\mathbf{z}_{\xi,\mathcal{R}}=\mathbf{b}_{\xi,\mathcal{R}},\qquad z_{\xi,f_c}=0. \tag*{(190)} L RR z ξ , R = b ξ , R , z ξ , f c = 0. (190)
The full-size form used in the code may be written by defining a diagonal mask D \mathbf{D} D , whose entries are zero at the pinned indices and unity elsewhere, and P = I − D \mathbf{P}=\mathbf{I}-\mathbf{D} P = I − D :
( D L D + P ) z ξ = D b ξ . (191) (\mathbf{D}\mathbf{L}\mathbf{D}+\mathbf{P})\mathbf{z}_\xi=\mathbf{D}\mathbf{b}_\xi. \tag*{(191)} ( DLD + P ) z ξ = D b ξ . (191)
This pinned-node convention chooses a different representative from the mean-zero gauge in Eq. (163) , but both give the same edge drops and hence the same capacitance dyadic.
Identification and numerical evaluation of the capacitance dyadic
After solving the two basis problems, define
F ξ = e ^ ξ + ∇ w ξ , ξ ∈ { x , y } . (192) \mathbf{F}_{\xi}=\hat{\mathbf{e}}_{\xi}+\nabla w_{\xi},\qquad\xi\in\{x,y\}. \tag*{(192)} F ξ = e ^ ξ + ∇ w ξ , ξ ∈ { x , y } . (192)
Equation (171) gives
v + ∇ w v = v x F x + v y F y . (193) \mathbf{v}+\nabla w_v=v_x\mathbf{F}_x+v_y\mathbf{F}_y. \tag*{(193)} v + ∇ w v = v x F x + v y F y . (193)
Substitution into the minimum energy yields
W e ( v ) = ϵ 0 E 0 2 2 ∣ U ∣ ∫ Ω ( v x 2 F x ⋅ F x + 2 v x v y F x ⋅ F y + v y 2 F y ⋅ F y ) d A . (194) W_e(\mathbf{v})=\frac{\epsilon_0 E_0^2}{2|U|}\int_{\Omega}\left(v_x^2\mathbf{F}_x\cdot\mathbf{F}_x+2v_xv_y\mathbf{F}_x\cdot\mathbf{F}_y+v_y^2\mathbf{F}_y\cdot\mathbf{F}_y\right)\,dA. \tag*{(194)} W e ( v ) = 2∣ U ∣ ϵ 0 E 0 2 ∫ Ω ( v x 2 F x ⋅ F x + 2 v x v y F x ⋅ F y + v y 2 F y ⋅ F y ) d A . (194)
The minimum energy is a quadratic form,
W e ( v ) = E 0 2 2 v ⋅ C ‾ [ Σ ] ⋅ v , (195) W_e(\mathbf{v})=\frac{E_0^2}{2}\,\mathbf{v}\cdot\overline{\mathbf{C}}[\Sigma]\cdot\mathbf{v}, \tag*{(195)} W e ( v ) = 2 E 0 2 v ⋅ C [ Σ ] ⋅ v , (195)
with entries
C ξ η [ Σ ] = ϵ 0 ∣ U ∣ ∫ Ω ( e ^ ξ + ∇ w ξ ) ⋅ ( e ^ η + ∇ w η ) d A , ξ , η ∈ { x , y } . (196) C_{\xi\eta}[\Sigma]=\frac{\epsilon_0}{|U|}\int_{\Omega}\left(\hat{\mathbf{e}}_{\xi}+\nabla w_{\xi}\right)\cdot\left(\hat{\mathbf{e}}_{\eta}+\nabla w_{\eta}\right)\,dA,\qquad\xi,\eta\in\{x,y\}. \tag*{(196)} C ξ η [ Σ ] = ∣ U ∣ ϵ 0 ∫ Ω ( e ^ ξ + ∇ w ξ ) ⋅ ( e ^ η + ∇ w η ) d A , ξ , η ∈ { x , y } . (196)
The area integral divided by ∣ U ∣ |U| ∣ U ∣ is dimensionless, so [ C ξ η ] = F / m [C_{\xi\eta}]={\rm F}/{\rm m} [ C ξ η ] = F / m . The off-diagonal entry is the signed overlap of the two basis fields and cannot, in general, be reconstructed from the two self-energies alone. Symmetry of the bilinear form gives
C x y = C y x . (197) C_{xy}=C_{yx}. \tag*{(197)} C x y = C y x . (197)
At finite resolution, the implementation distinguishes the edge set used to solve the graph problem from the edge set used to approximate the integral over the open dielectric region Ω \Omega Ω . Equation (186) is assembled on all active bonds, including gap–gap and gap–metal bonds. The latter couple the gap unknowns to the equipotential metal supernodes. The dyadic quadrature, however, uses the subset
E N g g = { e ∈ E N a c t : q e ∈ G N , r e ∈ G N } , (198) \mathcal{E}_{N}^{\mathrm{gg}}=\{e\in\mathcal{E}_{N}^{\mathrm{act}}:q_e\in\mathcal{G}_{N},\ r_e\in\mathcal{G}_{N}\}, \tag*{(198)} E N gg = { e ∈ E N act : q e ∈ G N , r e ∈ G N } , (198)
containing only bonds whose two endpoints lie in the sampled gap. This avoids assigning a wholly dielectric energy contribution to bonds cut by the PEC boundary. With the current numerical option, the computed entries are therefore
C ξ η ( N ) [ Σ ] = ϵ 0 κ △ ∣ U ∣ ∑ e ∈ E N g g d ξ , e d η , e , ξ , η ∈ { x , y } . (199) C_{\xi\eta}^{(N)}[\Sigma]=\frac{\epsilon_0\kappa\triangle}{|U|}\sum_{e\in\mathcal{E}_{N}^{\mathrm{gg}}}d_{\xi,e}d_{\eta,e},\qquad\xi,\eta\in\{x,y\}. \tag*{(199)} C ξ η ( N ) [ Σ ] = ∣ U ∣ ϵ 0 κ △ e ∈ E N gg ∑ d ξ , e d η , e , ξ , η ∈ { x , y } . (199)
For a substrate-backed calculation, ϵ 0 \epsilon_0 ϵ 0 in Eq. (199) is replaced by the same effective permittivity used in the continuum cell problem. The distinction between E N a c t \mathcal{E}_{N}^{\mathrm{act}} E N act and E N g g \mathcal{E}_{N}^{\mathrm{gg}} E N gg is a finite-resolution boundary treatment: interface bonds participate in the solution but not in the final gap-domain quadrature. Its numerical influence is consequently assessed by the resolution comparison described below.
Equation (199) is a Gram form. In particular, its off-diagonal entries are algebraically symmetric, and for any
q = q x e ^ x + q y e ^ y , \mathbf{q}=q_x\hat{\mathbf{e}}_x+q_y\hat{\mathbf{e}}_y, q = q x e ^ x + q y e ^ y ,
q ⋅ C ‾ ( N ) ⋅ q = ϵ 0 κ △ ∣ U ∣ ∑ e ∈ E N g g ( q x d x , e + q y d y , e ) 2 ≥ 0. (200) \mathbf{q}\cdot\overline{\mathbf{C}}^{(N)}\cdot\mathbf{q}=\frac{\epsilon_0\kappa\triangle}{|U|}\sum_{e\in\mathcal{E}_{N}^{\mathrm{gg}}}\left(q_xd_{x,e}+q_yd_{y,e}\right)^2\geq0. \tag*{(200)} q ⋅ C ( N ) ⋅ q = ∣ U ∣ ϵ 0 κ △ e ∈ E N gg ∑ ( q x d x , e + q y d y , e ) 2 ≥ 0. (200)
The continuum counterpart is
q ⋅ C ‾ ⋅ q = ϵ 0 ∣ U ∣ ∫ Ω ∣ q + ∇ w q ∣ 2 d A ≥ 0. (201) \mathbf{q}\cdot\overline{\mathbf{C}}\cdot\mathbf{q}=\frac{\epsilon_0}{|U|}\int_{\Omega}\left|\mathbf{q}+\nabla w_q\right|^2\,dA\geq0. \tag*{(201)} q ⋅ C ⋅ q = ∣ U ∣ ϵ 0 ∫ Ω ∣ q + ∇ w q ∣ 2 d A ≥ 0. (201)
If the right-hand side vanished, integration of the zero field along the two independent gap-winding cycles γ i \gamma_i γ i would give
0 = ∫ γ i ( q + ∇ w q ) ⋅ d ℓ = q ⋅ a i , i = 1 , 2. (202) 0 = \int_{\gamma_i} \left(\mathbf{q} + \nabla w_{\mathbf{q}}\right) \cdot\mathrm{d}\ell= \mathbf{q} \cdot\mathbf{a}_i, \qquad i = 1,2. \tag*{(202)} 0 = ∫ γ i ( q + ∇ w q ) ⋅ d ℓ = q ⋅ a i , i = 1 , 2. (202)
Because a 1 \mathbf{a}_1 a 1 and a 2 \mathbf{a}_2 a 2 are linearly independent, this implies q = 0 \mathbf{q} = 0 q = 0 . Hence the continuum dyadic is positive definite for the gap-connected geometries considered here. In the empty-cell limit, w x = w y = 0 w_x = w_y = 0 w x = w y = 0 , so the analytic background response is
C ‾ ‾ [ ∅ ] = ϵ 0 I . (203) \overline{\overline{C}}[\emptyset] = \epsilon_0 I. \tag*{(203)} C [ ∅ ] = ϵ 0 I . (203)
To monitor spatial resolution, the raw dyadic is recomputed on a second grid. If the first calculation uses N N N samples along each primitive direction, the second uses
N ′ = odd ( round ( 1.5 N ) ) , (204) N' = \operatorname{odd}(\operatorname{round}(1.5N)), \tag*{(204)} N ′ = odd ( round ( 1.5 N )) , (204)
where odd ( n ) = n \operatorname{odd}(n) = n odd ( n ) = n for odd n n n and odd ( n ) = n + 1 \operatorname{odd}(n) = n + 1 odd ( n ) = n + 1 for even n n n . The recorded relative change is
δ N , N ′ = ∥ C ‾ ‾ ( N ) − C ‾ ‾ ( N ′ ) ∥ F ∥ C ‾ ‾ ( N ′ ) ∥ F . (205) \delta_{N,N'} = \frac{\left\lVert\overline{\overline{C}}^{(N)} - \overline{\overline{C}}^{(N')} \right\rVert_F}{\left\lVert\overline{\overline{C}}^{(N')} \right\rVert_F}. \tag*{(205)} δ N , N ′ = C ( N ′ ) F C ( N ) − C ( N ′ ) F . (205)
The calculations reported in the present implementation use N = 151 N = 151 N = 151 , for which N ′ = 227 N' = 227 N ′ = 227 . A value δ N , N ′ > 10 − 2 \delta_{N,N'} > 10^{-2} δ N , N ′ > 1 0 − 2 produces a numerical warning and indicates that N N N should be increased. This comparison is a two-resolution consistency indicator rather than an a posteriori error bound. No equivalent sheet scaling or dispersion information enters this auxiliary corrector calculation.
Mathematical derivation of the reference-calibrated sheet capacitance
Starting from the auxiliary dyadic defined in Eq. (196) , we derive the excess response, its reference normalization, and the equivalent sheet capacitance. We use the energy interpretation and empty-cell limit established in § 3 . We distinguish the resulting continuum identities from the additional assumption relating auxiliary volume energy to sheet energy. Throughout this section, we consider a uniform free-space background of permittivity ϵ 0 \epsilon_{0} ϵ 0 .
Full-cell extension and the excess capacitance dyadic
We retain the correctors w x , w y w_{x},w_{y} w x , w y , the normalized fields F x , F y \mathbf{F}_{x},\mathbf{F}_{y} F x , F y , and the gap-domain gauge of Eq. (163) . For an isolated PEC component, we use Eq. (161) to extend the correctors as
w ξ ( r ) = c m , ξ − e ^ ξ ⋅ r in Σ m , ∇ w ξ = − e ^ ξ in Σ , (206) w_{\xi}(\mathbf{r})=c_{m,\xi}-\widehat{\mathbf{e}}_{\xi}\cdot\mathbf{r}\ \text{ in }\Sigma_{m},\qquad\nabla w_{\xi}=-\widehat{\mathbf{e}}_{\xi}\ \text{ in }\Sigma, \tag*{(206)} w ξ ( r ) = c m , ξ − e ξ ⋅ r in Σ m , ∇ w ξ = − e ξ in Σ , (206)
for ξ ∈ { x , y } \xi\in\{x,y\} ξ ∈ { x , y } . We interpret the constant on a continuous physical representative of the conductor. If that conductor is cut by the chosen cell boundary, its translated representatives follow the affine periodicity of Eq. (S87b). In particular, for a lattice translation R ∈ Λ \mathbf{R}\in\Lambda R ∈ Λ ,
c m , ξ ( R ) = c m , ξ + e ^ ξ ⋅ R , w ξ ( r + R ) = w ξ ( r ) . (207) c_{m,\xi}^{(\mathbf{R})}=c_{m,\xi}+\widehat{\mathbf{e}}_{\xi}\cdot\mathbf{R},\qquad w_{\xi}(\mathbf{r}+\mathbf{R})=w_{\xi}(\mathbf{r}). \tag*{(207)} c m , ξ ( R ) = c m , ξ + e ξ ⋅ R , w ξ ( r + R ) = w ξ ( r ) . (207)
Thus the extension respects periodic identification even when a physical patch appears in several pieces in the chosen primitive cell. The argument requires isolated conductors for which both imposed basis loadings are admissible; it does not extend unchanged to a conductor winding around the periodic cell.
For Lipschitz PEC boundaries, the matching traces of the gap and metal expressions give an extended function in H p e r 1 ( U ) H^{1}_{\mathrm{per}}(U) H per 1 ( U ) . We retain the notation w ξ w_{\xi} w ξ for this extension. Its mean over U U U need not vanish: the original gauge on Ω \Omega Ω is sufficient. Using periodicity, we obtain
∫ U ∇ w ξ d A = ∫ ∂ U w ξ n ^ d ℓ = 0. (208) \int_{U}\nabla w_{\xi}\,\mathrm{d}A=\int_{\partial U}w_{\xi}\widehat{\mathbf{n}}\,\mathrm{d}\ell=0. \tag*{(208)} ∫ U ∇ w ξ d A = ∫ ∂ U w ξ n d ℓ = 0. (208)
For smooth periodic functions the boundary integrals cancel between paired faces because the traces agree and the normals are opposite. We obtain the same identity for H p e r 1 ( U ) H^{1}_{\mathrm{per}}(U) H per 1 ( U ) by approximation with smooth periodic functions.
The normalized total field vanishes in the metal, F ξ = e ^ ξ + ∇ w ξ = 0 \mathbf{F}_{\xi}=\widehat{\mathbf{e}}_{\xi}+\nabla w_{\xi}=0 F ξ = e ξ + ∇ w ξ = 0 in Σ \Sigma Σ . We therefore extend the integral in Eq. (196) from Ω \Omega Ω to U U U and expand:
C ξ η [ Σ ] = ϵ 0 ∣ U ∣ ∫ U ( e ^ ξ + ∇ w ξ ) ⋅ ( e ^ η + ∇ w η ) d A = ϵ 0 δ ξ η + ϵ 0 ∣ U ∣ e ^ ξ ⋅ ∫ U ∇ w η d A + ϵ 0 ∣ U ∣ e ^ η ⋅ ∫ U ∇ w ξ d A + ϵ 0 ∣ U ∣ ∫ U ∇ w ξ ⋅ ∇ w η d A . (209) \begin{aligned}
C_{\xi\eta}[\Sigma]&=\frac{\epsilon_{0}}{|U|}\int_{U}\left(\widehat{\mathbf{e}}_{\xi}+\nabla w_{\xi}\right)\cdot\left(\widehat{\mathbf{e}}_{\eta}+\nabla w_{\eta}\right)\,\mathrm{d}A\\
&=\epsilon_{0}\delta_{\xi\eta}+\frac{\epsilon_{0}}{|U|}\widehat{\mathbf{e}}_{\xi}\cdot\int_{U}\nabla w_{\eta}\,\mathrm{d}A+\frac{\epsilon_{0}}{|U|}\widehat{\mathbf{e}}_{\eta}\cdot\int_{U}\nabla w_{\xi}\,\mathrm{d}A\\
&\qquad+\frac{\epsilon_{0}}{|U|}\int_{U}\nabla w_{\xi}\cdot\nabla w_{\eta}\,\mathrm{d}A. \tag*{(209)}
\end{aligned} C ξ η [ Σ ] = ∣ U ∣ ϵ 0 ∫ U ( e ξ + ∇ w ξ ) ⋅ ( e η + ∇ w η ) d A = ϵ 0 δ ξ η + ∣ U ∣ ϵ 0 e ξ ⋅ ∫ U ∇ w η d A + ∣ U ∣ ϵ 0 e η ⋅ ∫ U ∇ w ξ d A + ∣ U ∣ ϵ 0 ∫ U ∇ w ξ ⋅ ∇ w η d A . (209)
The mixed terms vanish by Eq. (208) . Using the empty-cell value already established in Eq. (203) , we define
C ‾ ‾ e x c [ Σ ] = C ‾ ‾ [ Σ ] − C ‾ ‾ [ ∅ ] = C ‾ ‾ [ Σ ] − ϵ 0 I ‾ ‾ . (210) \overline{\overline{C}}_{\mathrm{exc}}[\Sigma]=\overline{\overline{C}}[\Sigma]-\overline{\overline{C}}[\varnothing]=\overline{\overline{C}}[\Sigma]-\epsilon_{0}\overline{\overline{I}}. \tag*{(210)} C exc [ Σ ] = C [ Σ ] − C [ ∅ ] = C [ Σ ] − ϵ 0 I . (210)
From Eq. (209) , we obtain
( C ‾ ‾ e x c [ Σ ] ) ξ η = ϵ 0 ∣ U ∣ ∫ U ∇ w ξ ⋅ ∇ w η d A . (211) \left(\overline{\overline{C}}_{\mathrm{exc}}[\Sigma]\right)_{\xi\eta}=\frac{\epsilon_{0}}{|U|}\int_{U}\nabla w_{\xi}\cdot\nabla w_{\eta}\,\mathrm{d}A. \tag*{(211)} ( C exc [ Σ ] ) ξ η = ∣ U ∣ ϵ 0 ∫ U ∇ w ξ ⋅ ∇ w η d A . (211)
The full-cell integration is essential. Indeed, the metal extension contributes
∫ Σ ∇ w ξ ⋅ ∇ w η d A = ∣ Σ ∣ δ ξ η , (212) \int_{\Sigma}\nabla w_{\xi}\cdot\nabla w_{\eta}\,\mathrm{d}A=|\Sigma|\delta_{\xi\eta}, \tag*{(212)} ∫ Σ ∇ w ξ ⋅ ∇ w η d A = ∣Σ∣ δ ξ η , (212)
which would be omitted if U U U were replaced by Ω \Omega Ω in Eq. (211) .
Using the loading linearity in Eq. (171) , we obtain, for every real in-plane vector v \mathbf{v} v ,
v ⋅ C ‾ e x c ⋅ v = ε 0 ∣ U ∣ ∫ U ∣ v x ∇ w x + v y ∇ w y ∣ 2 d A = ε 0 ∣ U ∣ ∫ U ∣ ∇ w v ∣ 2 d A ≥ 0. (213) \mathbf{v}\cdot\overline{\mathbf{C}}_{\mathrm{exc}}\cdot\mathbf{v}
=\frac{\varepsilon_0}{\lvert U\rvert}\int_U\lvert v_x\nabla w_x+v_y\nabla w_y\rvert^2\,\mathrm{d}A
=\frac{\varepsilon_0}{\lvert U\rvert}\int_U\lvert\nabla w_{\mathbf{v}}\rvert^2\,\mathrm{d}A\ge0. \tag*{(213)} v ⋅ C exc ⋅ v = ∣ U ∣ ε 0 ∫ U ∣ v x ∇ w x + v y ∇ w y ∣ 2 d A = ∣ U ∣ ε 0 ∫ U ∣ ∇ w v ∣ 2 d A ≥ 0. (213)
Thus the excess is symmetric and positive semidefinite. For a nonempty positive-area isolated patch, we obtain the stronger estimate
v ⋅ C ‾ e x c ⋅ v ≥ ε 0 ∣ U ∣ ∫ Σ ∣ ∇ w v ∣ 2 d A = ε 0 ∣ Σ ∣ ∣ U ∣ ∣ v ∣ 2 (214) \mathbf{v}\cdot\overline{\mathbf{C}}_{\mathrm{exc}}\cdot\mathbf{v}
\ge\frac{\varepsilon_0}{\lvert U\rvert}\int_\Sigma\lvert\nabla w_{\mathbf{v}}\rvert^2\,\mathrm{d}A
=\varepsilon_0\frac{\lvert\Sigma\rvert}{\lvert U\rvert}\lvert\mathbf{v}\rvert^2 \tag*{(214)} v ⋅ C exc ⋅ v ≥ ∣ U ∣ ε 0 ∫ Σ ∣ ∇ w v ∣ 2 d A = ε 0 ∣ U ∣ ∣ Σ ∣ ∣ v ∣ 2 (214)
which proves positive definiteness, since ∇ w v = − v \nabla w_{\mathbf{v}}=-\mathbf{v} ∇ w v = − v in the metal. These statements concern the continuum cell problem. The empty geometry has zero excess.
We subtract the two energies at the same imposed macroscopic field. Using Eq. (195) , we write
Δ W e [ Σ ; v ] = W e [ Σ ; v ] − W e [ ∅ ; v ] = E 0 2 2 v ⋅ C ‾ e x c [ Σ ] ⋅ v . (215) \Delta W_e[\Sigma;\mathbf{v}]=W_e[\Sigma;\mathbf{v}]-W_e[\varnothing;\mathbf{v}]
=\frac{E_0^2}{2}\mathbf{v}\cdot\overline{\mathbf{C}}_{\mathrm{exc}}[\Sigma]\cdot\mathbf{v}. \tag*{(215)} Δ W e [ Σ ; v ] = W e [ Σ ; v ] − W e [ ∅ ; v ] = 2 E 0 2 v ⋅ C exc [ Σ ] ⋅ v . (215)
This nonnegative quantity has units of J/m3 ^3 3 . Its quadratic-form interpretation follows the variational energy framework of Milton [46 Sec. 13.1] .
We also use the extension to establish the equivalence of closed cavities. Let H ⊂ Ω H\subset\Omega H ⊂ Ω be a charge-free cavity whose entire boundary belongs to one connected PEC component, with no additional conductor inside. The total normalized potential u v u_{\mathbf{v}} u v is harmonic in H H H and has the constant trace c m ( v ) c_m(\mathbf{v}) c m ( v ) on ∂ H \partial H ∂ H . We test its weak harmonic equation with u v − c m ( v ) ∈ H 0 1 ( H ) u_{\mathbf{v}}-c_m(\mathbf{v})\in H_0^1(H) u v − c m ( v ) ∈ H 0 1 ( H ) to obtain
∫ H ∣ ∇ ( u v − c m ( v ) ) ∣ 2 d A = 0. (216) \int_H\lvert\nabla(u_{\mathbf{v}}-c_m(\mathbf{v}))\rvert^2\,\mathrm{d}A=0. \tag*{(216)} ∫ H ∣ ∇ ( u v − c m ( v )) ∣ 2 d A = 0. (216)
Hence the total field vanishes in H H H . Filling H H H with metal adds a constraint already satisfied by the old minimizer. The new admissible set is a subset of the old one, so its minimum energy cannot be smaller; the old minimizer remains admissible and attains the same energy. From equality for every loading, we conclude that
C ‾ [ Σ ∪ H ] = C ‾ [ Σ ] , C ‾ e x c [ Σ ∪ H ] = C ‾ e x c [ Σ ] . (217) \overline{\mathbf{C}}[\Sigma\cup H]=\overline{\mathbf{C}}[\Sigma],
\qquad
\overline{\mathbf{C}}_{\mathrm{exc}}[\Sigma\cup H]=\overline{\mathbf{C}}_{\mathrm{exc}}[\Sigma]. \tag*{(217)} C [ Σ ∪ H ] = C [ Σ ] , C exc [ Σ ∪ H ] = C exc [ Σ ] . (217)
The conclusion extends to any finite collection of such cavities. It applies only to this auxiliary two-dimensional problem: open slots are not filled, and an aperture in a physical thin sheet is not thereby equivalent to solid metal.
Reference normalization and directional representation
We use the same lattice, permittivity, field convention, and boundary conditions for the target and reference cell problems. For the centered regular hexagonal reference, the complete problem is invariant under the rotation R \mathbf{R} R through π / 3 \pi/3 π /3 . Rotating an admissible trial field preserves its energy and maps the admissible set for loading v \mathbf{v} v bijectively onto that for loading R v \mathbf{R}\mathbf{v} Rv . Consequently, W e [ Σ r e f ; R v ] = W e [ Σ r e f ; v ] W_e[\Sigma_{\mathrm{ref}};\mathbf{R}\mathbf{v}]=W_e[\Sigma_{\mathrm{ref}};\mathbf{v}] W e [ Σ ref ; Rv ] = W e [ Σ ref ; v ] , and
R T C ‾ e x c [ Σ r e f ] R = C ‾ e x c [ Σ r e f ] . (218) \mathbf{R}^{T}\overline{\mathbf{C}}_{\mathrm{exc}}[\Sigma_{\mathrm{ref}}]\mathbf{R}
=\overline{\mathbf{C}}_{\mathrm{exc}}[\Sigma_{\mathrm{ref}}]. \tag*{(218)} R T C exc [ Σ ref ] R = C exc [ Σ ref ] . (218)
To establish isotropy explicitly, we write the symmetric tensor as ( a b b d ) \begin{pmatrix}a&b\\b&d\end{pmatrix} ( a b b d ) and use R = cos ( π / 3 ) I + sin ( π / 3 ) J \mathbf{R}=\cos(\pi/3)\mathbf{I}+\sin(\pi/3)\mathbf{J} R = cos ( π /3 ) I + sin ( π /3 ) J , where
J = ( 0 − 1 1 0 ) . \mathbf{J}=\begin{pmatrix}0&-1\\1&0\end{pmatrix}. J = ( 0 1 − 1 0 ) .
(218) implies commutation with J \mathbf{J} J , which gives b = 0 b=0 b = 0 and a = d a=d a = d . Therefore,
C ‾ e x c [ Σ r e f ] = p r e f I , p r e f = 1 2 tr ( C ‾ e x c [ Σ r e f ] ) > 0. (219) \overline{\mathbf{C}}_{\mathrm{exc}}[\Sigma_{\mathrm{ref}}]=p_{\mathrm{ref}}\mathbf{I},
\qquad
p_{\mathrm{ref}}=\frac{1}{2}\operatorname{tr}\left(\overline{\mathbf{C}}_{\mathrm{exc}}[\Sigma_{\mathrm{ref}}]\right)>0. \tag*{(219)} C exc [ Σ ref ] = p ref I , p ref = 2 1 tr ( C exc [ Σ ref ] ) > 0. (219)
Using Eq. (214) , we obtain p r e f ≥ ε 0 ∣ Σ r e f ∣ / ∣ U ∣ > 0 p_{\mathrm{ref}}\ge\varepsilon_0\lvert\Sigma_{\mathrm{ref}}\rvert/\lvert U\rvert>0 p ref ≥ ε 0 ∣ Σ ref ∣ / ∣ U ∣ > 0 . The symmetry argument concerns the complete cell, including its lattice.
We define the dimensionless normalized excess by
S ‾ [ Σ ] = C ‾ e x c [ Σ ] p r e f , [ p r e f ] = F / m , [ S ‾ ] = 1. (220) \overline{\mathcal{S}}[\Sigma]=\frac{\overline{\mathbf{C}}_{\mathrm{exc}}[\Sigma]}{p_{\mathrm{ref}}},
\qquad
[p_{\mathrm{ref}}]=\mathrm{F/m},
\qquad
[\overline{\mathcal{S}}]=1. \tag*{(220)} S [ Σ ] = p ref C exc [ Σ ] , [ p ref ] = F/m , [ S ] = 1. (220)
The empty-cell subtraction and isotropic reference give
S ‾ ~ [ ∅ ] = 0 ‾ , S ‾ ~ [ Σ r e f ] = I ‾ . (221) \widetilde{\overline{S}}[\emptyset]=\overline{0},\qquad\widetilde{\overline{S}}[\Sigma_{\mathrm{ref}}]=\overline{I}. \tag*{(221)} S [ ∅ ] = 0 , S [ Σ ref ] = I . (221)
For the figures, we separately normalize each tensor by half its own trace. For a nonzero symmetric positive-semidefinite tensor M ‾ \overline{M} M , we define
M ‾ ~ = M ‾ 1 2 tr ( M ‾ ) , r M ( θ ) = u ^ ( θ ) ⋅ M ‾ ~ ⋅ u ^ ( θ ) , (222) \widetilde{\overline{M}}=\frac{\overline{M}}{\frac{1}{2}\operatorname{tr}(\overline{M})},\qquad r_M(\theta)=\hat{\mathbf{u}}(\theta)\cdot\widetilde{\overline{M}}\cdot\hat{\mathbf{u}}(\theta), \tag*{(222)} M = 2 1 tr ( M ) M , r M ( θ ) = u ^ ( θ ) ⋅ M ⋅ u ^ ( θ ) , (222)
where u ^ ( θ ) = cos θ e ^ x + sin θ e ^ y \hat{\mathbf{u}}(\theta)=\cos\theta\,\hat{\mathbf{e}}_x+\sin\theta\,\hat{\mathbf{e}}_y u ^ ( θ ) = cos θ e ^ x + sin θ e ^ y . Let m 1 ≥ m 2 ≥ 0 m_1\ge m_2\ge0 m 1 ≥ m 2 ≥ 0 be the eigenvalues, and let an eigenvector for m 1 m_1 m 1 make angle φ \varphi φ with the x x x axis. Evaluating in the orthonormal principal basis [59 App. B.5] , we obtain
r M ( θ ) = 2 [ m 1 cos 2 ( θ − φ ) + m 2 sin 2 ( θ − φ ) ] m 1 + m 2 = 1 + m 1 − m 2 m 1 + m 2 cos ( 2 ( θ − φ ) ) . (223) r_M(\theta)=\frac{2\left[m_1\cos^2(\theta-\varphi)+m_2\sin^2(\theta-\varphi)\right]}{m_1+m_2}
=1+\frac{m_1-m_2}{m_1+m_2}\cos\left(2(\theta-\varphi)\right). \tag*{(223)} r M ( θ ) = m 1 + m 2 2 [ m 1 cos 2 ( θ − φ ) + m 2 sin 2 ( θ − φ ) ] = 1 + m 1 + m 2 m 1 − m 2 cos ( 2 ( θ − φ ) ) . (223)
The denominator is positive because the tensor is nonzero and positive semidefinite. Thus this normalization preserves the principal basis and eigenvalue ratio when defined, and gives the unit circle for an isotropic tensor. The radial curve is a directional quadratic-form indicator, not generally an ellipse. The zero tensor has no such trace-normalized representation.
For the excess, we perform the subtraction before the plotting normalization:
C ‾ ~ e x c = C ‾ − ϵ 0 I ‾ 1 2 tr ( C ‾ ) − ϵ 0 , (224) \widetilde{\overline{C}}_{\mathrm{exc}}=\frac{\overline{C}-\epsilon_0\overline{I}}{\frac{1}{2}\operatorname{tr}(\overline{C})-\epsilon_0}, \tag*{(224)} C exc = 2 1 tr ( C ) − ϵ 0 C − ϵ 0 I , (224)
provided the excess is nonzero. The trace normalization in the figures does not enter the sheet-capacitance calculation.
Reference-energy matching and the equivalent sheet capacitance
We now relate the auxiliary excess energy to the equivalent sheet energy. For the same real macroscopic field E 0 v \mathbf{E}_0 v E 0 v , we write
Δ W e [ Σ ; v ] = E 0 2 2 v ⋅ C ‾ ~ e x c [ Σ ] ⋅ v , [ Δ W e ] = J / m 3 , \Delta W_e[\Sigma;\mathbf{v}]=\frac{E_0^2}{2}\mathbf{v}\cdot\widetilde{\overline{C}}_{\mathrm{exc}}[\Sigma]\cdot\mathbf{v},\qquad[\Delta W_e]=\mathrm{J/m^3}, Δ W e [ Σ ; v ] = 2 E 0 2 v ⋅ C exc [ Σ ] ⋅ v , [ Δ W e ] = J/ m 3 ,
U s m o d e l [ Σ ; v ] = E 0 2 2 v ⋅ C ‾ ~ e q v [ Σ ] ⋅ v , [ U s m o d e l ] = J / m 2 . (225) U_s^{\mathrm{model}}[\Sigma;\mathbf{v}]=\frac{E_0^2}{2}\mathbf{v}\cdot\widetilde{\overline{C}}_{\mathrm{eqv}}[\Sigma]\cdot\mathbf{v},\qquad[U_s^{\mathrm{model}}]=\mathrm{J/m^2}. \tag*{(225)} U s model [ Σ ; v ] = 2 E 0 2 v ⋅ C eqv [ Σ ] ⋅ v , [ U s model ] = J/ m 2 . (225)
We use a reciprocal electric-sheet constitutive model [57 ] . For a selected geometry family F \mathcal{F} F , we introduce the energy-matching assumption
∃ ℓ r e f > 0 : U s m o d e l [ Σ ; v ] = ℓ r e f Δ W e [ Σ ; v ] , ∀ Σ ∈ F , ∀ v ∈ R 2 . (226) \exists\,\ell_{\mathrm{ref}}>0:\quad U_s^{\mathrm{model}}[\Sigma;\mathbf{v}]=\ell_{\mathrm{ref}}\Delta W_e[\Sigma;\mathbf{v}],\qquad\forall\Sigma\in\mathcal{F},\quad\forall\mathbf{v}\in\mathbb{R}^2. \tag*{(226)} ∃ ℓ ref > 0 : U s model [ Σ ; v ] = ℓ ref Δ W e [ Σ ; v ] , ∀Σ ∈ F , ∀ v ∈ R 2 . (226)
We therefore assume one conversion length for all geometries and both in-plane polarizations, under the same macroscopic-field convention.
For E 0 ≠ 0 E_0\ne0 E 0 = 0 , substitution gives
v T D v = 0 ∀ v , D = C ‾ ~ e q v − ℓ r e f C ‾ ~ e x c , D T = D . (227) \mathbf{v}^{T}\mathbf{D}\mathbf{v}=0\quad\forall\mathbf{v},\qquad\mathbf{D}=\widetilde{\overline{C}}_{\mathrm{eqv}}-\ell_{\mathrm{ref}}\widetilde{\overline{C}}_{\mathrm{exc}},\qquad\mathbf{D}^{T}=\mathbf{D}. \tag*{(227)} v T Dv = 0 ∀ v , D = C eqv − ℓ ref C exc , D T = D . (227)
We evaluate this quadratic form on three loadings:
v = e ^ x ⟹ D x x = 0 , \mathbf{v}=\hat{\mathbf{e}}_x\quad\Longrightarrow\quad D_{xx}=0, v = e ^ x ⟹ D xx = 0 ,
v = e ^ y ⟹ D y y = 0 , \mathbf{v}=\hat{\mathbf{e}}_y\quad\Longrightarrow\quad D_{yy}=0, v = e ^ y ⟹ D y y = 0 ,
v = e ^ x + e ^ y ⟹ D x x + 2 D x y + D y y = 0. (228) \mathbf{v}=\hat{\mathbf{e}}_x+\hat{\mathbf{e}}_y\quad\Longrightarrow\quad D_{xx}+2D_{xy}+D_{yy}=0. \tag*{(228)} v = e ^ x + e ^ y ⟹ D xx + 2 D x y + D y y = 0. (228)
Consequently,
D x x = D y y = D x y = D y x = 0 ⟹ C ‾ ~ e q v [ Σ ] = ℓ r e f C ‾ ~ e x c [ Σ ] . (229) D_{xx}=D_{yy}=D_{xy}=D_{yx}=0\quad\Longrightarrow\quad\widetilde{\overline{C}}_{\mathrm{eqv}}[\Sigma]=\ell_{\mathrm{ref}}\widetilde{\overline{C}}_{\mathrm{exc}}[\Sigma]. \tag*{(229)} D xx = D y y = D x y = D y x = 0 ⟹ C eqv [ Σ ] = ℓ ref C exc [ Σ ] . (229)
For the isotropic reference, we impose the independently prescribed sheet value C ‾ r e f e q v > 0 \overline{C}^{\mathrm{eqv}}_{\mathrm{ref}}>0 C ref eqv > 0 :
C ‾ e x c [ Σ r e f ] = p r e f I ‾ , p r e f > 0 , \overline{\mathbf C}_{\mathrm{exc}}[\Sigma_{\mathrm{ref}}]=p_{\mathrm{ref}}\overline{\mathbf I},\qquad p_{\mathrm{ref}}>0, C exc [ Σ ref ] = p ref I , p ref > 0 ,
C ‾ e q v [ Σ r e f ] = C ‾ r e f e q v I ‾ . (230) \overline{\mathbf C}_{\mathrm{eqv}}[\Sigma_{\mathrm{ref}}]=\overline{C}^{\mathrm{eqv}}_{\mathrm{ref}}\overline{\mathbf I}. \tag*{(230)} C eqv [ Σ ref ] = C ref eqv I . (230)
Combining Eqs. (229) and (230) , we obtain
ℓ r e f p r e f I ‾ = C ‾ r e f e q v I ‾ ⟹ ℓ r e f = C ‾ r e f e q v p r e f . (231) \ell_{\mathrm{ref}}p_{\mathrm{ref}}\overline{\mathbf I}=\overline{C}^{\mathrm{eqv}}_{\mathrm{ref}}\overline{\mathbf I}\quad\Longrightarrow\quad\ell_{\mathrm{ref}}=\frac{\overline{C}^{\mathrm{eqv}}_{\mathrm{ref}}}{p_{\mathrm{ref}}}. \tag*{(231)} ℓ ref p ref I = C ref eqv I ⟹ ℓ ref = p ref C ref eqv . (231)
Therefore,
C ‾ e q v [ Σ ] = C ‾ r e f e q v p r e f C ‾ e x c [ Σ ] = C ‾ r e f e q v S ‾ [ Σ ] = C ‾ r e f e q v p r e f ( C ‾ [ Σ ] − ϵ 0 I ‾ ) . (232) \begin{aligned}
\overline{\mathbf C}_{\mathrm{eqv}}[\Sigma]&=\frac{\overline{C}^{\mathrm{eqv}}_{\mathrm{ref}}}{p_{\mathrm{ref}}}\overline{\mathbf C}_{\mathrm{exc}}[\Sigma]\\
&=\overline{C}^{\mathrm{eqv}}_{\mathrm{ref}}\overline{\mathbf S}[\Sigma]\\
&=\frac{\overline{C}^{\mathrm{eqv}}_{\mathrm{ref}}}{p_{\mathrm{ref}}}\left(\overline{\mathbf C}[\Sigma]-\epsilon_0\overline{\mathbf I}\right). \tag*{(232)}
\end{aligned} C eqv [ Σ ] = p ref C ref eqv C exc [ Σ ] = C ref eqv S [ Σ ] = p ref C ref eqv ( C [ Σ ] − ϵ 0 I ) . (232)
We verify the units and the two reference conditions directly:
[ ℓ r e f ] = F F / m = m , [ C ‾ e q v ] = m F m = F , (233) [\ell_{\mathrm{ref}}]=\frac{\mathrm F}{\mathrm{F}/\mathrm m}=\mathrm m,\qquad[\overline{\mathbf C}_{\mathrm{eqv}}]=\mathrm m\frac{\mathrm F}{\mathrm m}=\mathrm F, \tag*{(233)} [ ℓ ref ] = F / m F = m , [ C eqv ] = m m F = F , (233)
C ‾ e q v [ ∅ ] = 0 ‾ , C ‾ e q v [ Σ r e f ] = C ‾ r e f e q v I ‾ . (234) \overline{\mathbf C}_{\mathrm{eqv}}[\varnothing]=\overline{\mathbf0},\qquad\overline{\mathbf C}_{\mathrm{eqv}}[\Sigma_{\mathrm{ref}}]=\overline{C}^{\mathrm{eqv}}_{\mathrm{ref}}\overline{\mathbf I}. \tag*{(234)} C eqv [ ∅ ] = 0 , C eqv [ Σ ref ] = C ref eqv I . (234)
We can also derive this normalization as the unique map within the scalar-affine calibration class
C ‾ e q v [ Σ ] = a C ‾ [ Σ ] + b I ‾ , [ a ] = m , [ b ] = F , (235) \overline{\mathbf C}_{\mathrm{eqv}}[\Sigma]=a\overline{\mathbf C}[\Sigma]+b\overline{\mathbf I},\qquad[a]=\mathrm m,\qquad[b]=\mathrm F, \tag*{(235)} C eqv [ Σ ] = a C [ Σ ] + b I , [ a ] = m , [ b ] = F , (235)
where a a a , b b b are independent of Σ \Sigma Σ . Indeed,
C ‾ [ ∅ ] = ϵ 0 I ‾ , \overline{\mathbf C}[\varnothing]=\epsilon_0\overline{\mathbf I}, C [ ∅ ] = ϵ 0 I ,
C ‾ [ Σ r e f ] = ( ϵ 0 + p r e f ) I ‾ . (236) \overline{\mathbf C}[\Sigma_{\mathrm{ref}}]=(\epsilon_0+p_{\mathrm{ref}})\overline{\mathbf I}. \tag*{(236)} C [ Σ ref ] = ( ϵ 0 + p ref ) I . (236)
so we write the two matching conditions as
\begin{pmatrix}
\epsilon_0 & 1\\
\epsilon_0+p_{\mathrm{ref}} & 1
\end{pmatrix}
\begin{pmatrix}
a\\
b
\end{pmatrix}
=
\begin{pmatrix}
0\\
\overline{C}^{\mathrm{eqv}}_{\mathrm{ref}}
\end{pmatrix},
\qquad
\det\begin{pmatrix}
\epsilon_0 & 1\\
\epsilon_0+p_{\mathrm{ref}} & 1
\end{pmatrix}
=-p_{\mathrm{ref}}\ne0. \tag*{(237)}\end{pmatrix}
Solving this system, we obtain
a = C ‾ r e f e q v p r e f = ℓ r e f , b = − ϵ 0 C ‾ r e f e q v p r e f = − ϵ 0 ℓ r e f . (238) a=\frac{\overline{C}^{\mathrm{eqv}}_{\mathrm{ref}}}{p_{\mathrm{ref}}}=\ell_{\mathrm{ref}},\qquad b=-\epsilon_0\frac{\overline{C}^{\mathrm{eqv}}_{\mathrm{ref}}}{p_{\mathrm{ref}}}=-\epsilon_0\ell_{\mathrm{ref}}. \tag*{(238)} a = p ref C ref eqv = ℓ ref , b = − ϵ 0 p ref C ref eqv = − ϵ 0 ℓ ref . (238)
Thus the empty-cell condition fixes the offset and the reference fixes the slope. Uniqueness is conditional on the class (S162).
To identify the spectral consequences, we write [59 App. B.5]
C ‾ e x c = ∑ i = 1 2 λ i v ^ i ⊗ v ^ i , λ i ≥ 0 , v ^ i ⋅ v ^ j = δ i j . (239) \overline{\mathbf C}_{\mathrm{exc}}=\sum_{i=1}^{2}\lambda_i\widehat{\mathbf v}_i\otimes\widehat{\mathbf v}_i,\qquad\lambda_i\ge0,\qquad\widehat{\mathbf v}_i\cdot\widehat{\mathbf v}_j=\delta_{ij}. \tag*{(239)} C exc = i = 1 ∑ 2 λ i v i ⊗ v i , λ i ≥ 0 , v i ⋅ v j = δ ij . (239)
Then
C ‾ = ∑ i = 1 2 ( ϵ 0 + λ i ) v ^ i ⊗ v ^ i , S ‾ = ∑ i = 1 2 λ i p r e f v ^ i ⊗ v ^ i , C ‾ e q v = ∑ i = 1 2 C ‾ r e f e q v λ i p r e f v ^ i ⊗ v ^ i . (240) \begin{aligned}
\overline{\mathbf C}&=\sum_{i=1}^{2}(\epsilon_0+\lambda_i)\widehat{\mathbf v}_i\otimes\widehat{\mathbf v}_i,\\
\overline{\mathbf S}&=\sum_{i=1}^{2}\frac{\lambda_i}{p_{\mathrm{ref}}}\widehat{\mathbf v}_i\otimes\widehat{\mathbf v}_i,\\
\overline{\mathbf C}_{\mathrm{eqv}}&=\sum_{i=1}^{2}\frac{\overline{C}^{\mathrm{eqv}}_{\mathrm{ref}}\lambda_i}{p_{\mathrm{ref}}}\widehat{\mathbf v}_i\otimes\widehat{\mathbf v}_i. \tag*{(240)}
\end{aligned} C S C eqv = i = 1 ∑ 2 ( ϵ 0 + λ i ) v i ⊗ v i , = i = 1 ∑ 2 p ref λ i v i ⊗ v i , = i = 1 ∑ 2 p ref C ref eqv λ i v i ⊗ v i . (240)
In particular,
s i = λ i p r e f , C ‾ i e q v = C ‾ r e f e q v s i , C ‾ 1 e q v C ‾ 2 e q v = λ 1 λ 2 ( λ 2 > 0 ) , (241) s_i = \frac{\lambda_i}{p_{\mathrm{ref}}}, \qquad\overline{C}^{\mathrm{eqv}}_i = \overline{C}^{\mathrm{eqv}}_{\mathrm{ref}}s_i, \qquad\frac{\overline{C}^{\mathrm{eqv}}_1}{\overline{C}^{\mathrm{eqv}}_2} = \frac{\lambda_1}{\lambda_2} \quad(\lambda_2 > 0), \tag*{(241)} s i = p ref λ i , C i eqv = C ref eqv s i , C 2 eqv C 1 eqv = λ 2 λ 1 ( λ 2 > 0 ) , (241)
ker C ‾ e q v = ker C ‾ e x c , rank C ‾ e q v = rank C ‾ e x c , (242) \ker\overline{C}^{\mathrm{eqv}} = \ker\overline{C}_{\mathrm{exc}}, \qquad\operatorname{rank} \overline{C}^{\mathrm{eqv}} = \operatorname{rank} \overline{C}_{\mathrm{exc}}, \tag*{(242)} ker C eqv = ker C exc , rank C eqv = rank C exc , (242)
v ⋅ C ‾ e q v ⋅ v = ℓ r e f ∑ i = 1 2 λ i ( v ⋅ v ^ i ) 2 ≥ 0. (243) \mathbf{v} \cdot\overline{C}_{\mathrm{eqv}} \cdot\mathbf{v} = \ell_{\mathrm{ref}} \sum_{i=1}^{2} \lambda_i(\mathbf{v} \cdot\hat{\mathbf{v}}_i)^2 \ge0. \tag*{(243)} v ⋅ C eqv ⋅ v = ℓ ref i = 1 ∑ 2 λ i ( v ⋅ v ^ i ) 2 ≥ 0. (243)
The principal basis is shared; it is not unique for an isotropic tensor. For an orthogonal symmetry R \mathbf{R} R , we also obtain
R C ‾ R T = C ‾ ⟹ R C ‾ e q v R T = ℓ r e f ( R C ‾ R T − ϵ 0 I ‾ ) = C ‾ e q v . (244) \mathbf{R}\overline{C}\mathbf{R}^{T} = \overline{C} \quad\Longrightarrow\quad\mathbf{R}\overline{C}_{\mathrm{eqv}}\mathbf{R}^{T} = \ell_{\mathrm{ref}}\left(\mathbf{R}\overline{C}\mathbf{R}^{T} - \epsilon_0\overline{I}\right) = \overline{C}_{\mathrm{eqv}}. \tag*{(244)} R C R T = C ⟹ R C eqv R T = ℓ ref ( R C R T − ϵ 0 I ) = C eqv . (244)
We prescribe the reference value independently of the auxiliary cell problem. For the free-standing regular hexagonal reference, we use
C ‾ r e f e q v = 2 ϵ 0 A i s o , \overline{C}^{\mathrm{eqv}}_{\mathrm{ref}} = 2\epsilon_0 A_{\mathrm{iso}}, C ref eqv = 2 ϵ 0 A iso ,
A i s o = 3 D h 2 π ln [ csc ( π g 2 D h ) ] , 0 < g < D h . (245) A_{\mathrm{iso}} = \frac{\sqrt{3}D_h}{2\pi}\ln\left[\csc\left(\frac{\pi g}{2D_h}\right)\right], \qquad0 < g < D_h. \tag*{(245)} A iso = 2 π 3 D h ln [ csc ( 2 D h π g ) ] , 0 < g < D h . (245)
Here [ A i s o ] = m [A_{\mathrm{iso}}] = \mathrm{m} [ A iso ] = m and C ‾ r e f e q v > 0 \overline{C}^{\mathrm{eqv}}_{\mathrm{ref}} > 0 C ref eqv > 0 . We motivate the logarithmic dependence by the strip and square-patch models of Luukkonen et al. [10 ] and treat the hexagonal prefactor as an independently testable approximation. We may replace this expression with a computed reference capacitance.
Finally, we identify what the isotropic reference leaves undetermined. We decompose the excess into its isotropic and traceless parts:
C ‾ e x c = t I ‾ + C ‾ d e v , t = 1 2 tr C ‾ e x c , tr C ‾ d e v = 0. (246) \overline{C}_{\mathrm{exc}} = t\overline{I} + \overline{C}_{\mathrm{dev}}, \qquad t = \frac{1}{2}\operatorname{tr}\overline{C}_{\mathrm{exc}}, \qquad\operatorname{tr}\overline{C}_{\mathrm{dev}} = 0. \tag*{(246)} C exc = t I + C dev , t = 2 1 tr C exc , tr C dev = 0. (246)
For comparison, we consider the more general rotation-covariant linear calibration
L ( C ‾ e x c ) = ℓ i s o t I ‾ + ℓ d e v C ‾ d e v , [ ℓ i s o ] = [ ℓ d e v ] = m . (247) \mathcal{L}\left(\overline{C}_{\mathrm{exc}}\right) = \ell_{\mathrm{iso}}t\overline{I} + \ell_{\mathrm{dev}}\overline{C}_{\mathrm{dev}}, \qquad[\ell_{\mathrm{iso}}] = [\ell_{\mathrm{dev}}] = \mathrm{m}. \tag*{(247)} L ( C exc ) = ℓ iso t I + ℓ dev C dev , [ ℓ iso ] = [ ℓ dev ] = m . (247)
For the reference,
t r e f = p r e f , C ‾ d e v [ Σ r e f ] = 0 ‾ , (248) t_{\mathrm{ref}} = p_{\mathrm{ref}}, \qquad\overline{C}_{\mathrm{dev}}[\Sigma_{\mathrm{ref}}] = \overline{0}, \tag*{(248)} t ref = p ref , C dev [ Σ ref ] = 0 , (248)
and therefore
L ( p r e f I ‾ ) = ℓ i s o p r e f I ‾ = C ‾ r e f e q v I ‾ ⟹ ℓ i s o = C ‾ r e f e q v p r e f , (249) \mathcal{L}(p_{\mathrm{ref}}\overline{I}) = \ell_{\mathrm{iso}}p_{\mathrm{ref}}\overline{I} = \overline{C}^{\mathrm{eqv}}_{\mathrm{ref}}\overline{I} \quad\Longrightarrow\quad\ell_{\mathrm{iso}} = \frac{\overline{C}^{\mathrm{eqv}}_{\mathrm{ref}}}{p_{\mathrm{ref}}}, \tag*{(249)} L ( p ref I ) = ℓ iso p ref I = C ref eqv I ⟹ ℓ iso = p ref C ref eqv , (249)
with no condition on ℓ d e v \ell_{\mathrm{dev}} ℓ dev . Positivity alone does not force equality: the eigenvalues of the output in Eq. (247) are
μ i = ℓ d e v λ i + ( ℓ i s o − ℓ d e v ) t ≥ 0 if 0 ≤ ℓ d e v ≤ ℓ i s o . (250) \mu_i = \ell_{\mathrm{dev}}\lambda_i + (\ell_{\mathrm{iso}} - \ell_{\mathrm{dev}})t \ge0 \qquad\text{if} \qquad0 \le\ell_{\mathrm{dev}} \le\ell_{\mathrm{iso}}. \tag*{(250)} μ i = ℓ dev λ i + ( ℓ iso − ℓ dev ) t ≥ 0 if 0 ≤ ℓ dev ≤ ℓ iso . (250)
In the present model, we choose
ℓ d e v = ℓ i s o = ℓ r e f = C ‾ r e f e q v p r e f . (251) \ell_{\mathrm{dev}} = \ell_{\mathrm{iso}} = \ell_{\mathrm{ref}} = \frac{\overline{C}^{\mathrm{eqv}}_{\mathrm{ref}}}{p_{\mathrm{ref}}}. \tag*{(251)} ℓ dev = ℓ iso = ℓ ref = p ref C ref eqv . (251)
We thus use the reference to fix the overall scale and assume equal scaling of the isotropic and anisotropic parts throughout the selected geometry family. This is the physical hypothesis in Eq. (226) . Its validity requires electromagnetic validation or an interface-cell reduction; the two reference conditions alone do not establish it.
Continuum identities and numerical consistency
We have derived the excess Gram identity and its positivity for the continuum problem. At finite resolution, we solve on the active graph bonds described in § 3 and evaluate the raw dyadic using the gap–gap subset in Eq. (199) . The latter is a Gram form and therefore gives a positive-semidefinite raw tensor. That fact alone does not imply positivity after empty-cell subtraction:
C ‾ ( N ) ⪰ 0 ⇏ C ‾ ( N ) − ϵ 0 I ⪰ 0. (252) \overline{\mathcal{C}}^{(N)} \succeq0 \not\Rightarrow\overline{\mathcal{C}}^{(N)}-\epsilon_0 I \succeq0. \tag*{(252)} C ( N ) ⪰ 0 ⇒ C ( N ) − ϵ 0 I ⪰ 0. (252)
The same distinction applies when a numerically computed empty-cell tensor is subtracted. The gap–gap quadrature does not, by itself, provide a discrete counterpart of the full-cell identity (S138).
We therefore require a consistent background-subtraction convention and discretization for the target and reference calculations. Numerical checks should examine the excess eigenvalues, the positivity of the computed p r e f p_{\mathrm{ref}} p ref , and the stability of the normalized tensor under refinement. Small relative changes of the raw tensor, as measured by Eq. (205) , need not imply equally small relative changes of the excess when the background subtraction causes cancellation. These checks assess numerical recovery of the continuum identities; we do not assert a discrete convergence theorem here.
For the physical calibration, we obtain Eq. (232) uniquely once we assume the common conversion length in Eq. (226) and prescribe a positive independent reference capacitance. Applying that same length across geometries and principal directions requires electromagnetic validation or a further interface-cell derivation. The auxiliary normalization alone does not establish this applicability.
Full local anisotropic transverse-resonance condition and quasi-TE reduction
This section derives the local anisotropic eigencondition underlying the quasi-TE dispersion model used in the main text. We first express the geometry-derived grid-length dyadic in a basis fixed to the in-plane propagation direction, then derive the complete coupled TE–TM transverse-resonance condition, and finally obtain the TE-dominant reduction used for the dispersion calculations.
The equivalent sheet-capacitance dyadic is related to the grid-length dyadic by
A ‾ = C ‾ e q v 2 ϵ 0 , (253) \overline{A} = \frac{\overline{C}_{\mathrm{eqv}}}{2\epsilon_0}, \tag*{(253)} A = 2 ϵ 0 C eqv , (253)
so that, in a fixed Cartesian basis,
A ‾ = ( A x x A x y A x y A y y ) . (254) \overline{A} =
\begin{pmatrix}
A_{xx} & A_{xy} \\
A_{xy} & A_{yy}
\end{pmatrix}. \tag*{(254)} A = ( A xx A x y A x y A y y ) . (254)
The symmetry of A ‾ \overline{A} A follows from the reciprocity and symmetry of the capacitance dyadic established in the main text.
Propagation-fixed representation of the grid-length dyadic
For a nonzero in-plane Bloch wavevector
k t = k x x ^ + k y y ^ , k t = k x 2 + k y 2 , (255) \mathbf{k}_{t} = k_x\hat{\mathbf{x}} + k_y\hat{\mathbf{y}}, \qquad k_t = \sqrt{k_x^2 + k_y^2}, \tag*{(255)} k t = k x x ^ + k y y ^ , k t = k x 2 + k y 2 , (255)
we define the longitudinal and transverse unit vectors
e ^ ∥ = k t k t = 1 k t ( k x k y ) , e ^ ⊥ = e ^ T E = z ^ × k t k t = 1 k t ( − k y k x ) . (256) \hat{\mathbf{e}}_{\parallel} = \frac{\mathbf{k}_t}{k_t} = \frac{1}{k_t}
\begin{pmatrix}
k_x \\
k_y
\end{pmatrix},
\qquad
\hat{\mathbf{e}}_{\perp} = \hat{\mathbf{e}}_{\mathrm{TE}} = \frac{\hat{\mathbf{z}}\times\mathbf{k}_t}{k_t} = \frac{1}{k_t}
\begin{pmatrix}
-k_y \\
k_x
\end{pmatrix}. \tag*{(256)} e ^ ∥ = k t k t = k t 1 ( k x k y ) , e ^ ⊥ = e ^ TE = k t z ^ × k t = k t 1 ( − k y k x ) . (256)
The pair { e ^ ∥ , e ^ ⊥ } \{\hat{\mathbf{e}}_{\parallel},\hat{\mathbf{e}}_{\perp}\} { e ^ ∥ , e ^ ⊥ } forms an orthonormal basis of the tangential plane. The associated orthogonal transformation is
Q ( k t ) = ( e ^ ∥ e ^ ⊥ ) = 1 k t ( k x − k y k y k x ) , (257) Q(\mathbf{k}_t) = \left(\hat{\mathbf{e}}_{\parallel}\ \hat{\mathbf{e}}_{\perp}\right)
= \frac{1}{k_t}
\begin{pmatrix}
k_x & -k_y \\
k_y & k_x
\end{pmatrix}, \tag*{(257)} Q ( k t ) = ( e ^ ∥ e ^ ⊥ ) = k t 1 ( k x k y − k y k x ) , (257)
for which
Q T Q = Q Q T = I . (258) Q^{T}Q = QQ^{T} = I. \tag*{(258)} Q T Q = Q Q T = I . (258)
The representation of the grid-length dyadic in the propagation-fixed basis is therefore
Q T A ‾ Q = ( A ∥ A × A × A ⊥ ) , (259) Q^{T}\overline{A}Q =
\begin{pmatrix}
A_{\parallel} & A_{\times} \\
A_{\times} & A_{\perp}
\end{pmatrix}, \tag*{(259)} Q T A Q = ( A ∥ A × A × A ⊥ ) , (259)
where the diagonal entries are the longitudinal and transverse quadratic projections and the off-diagonal entry is the longitudinal–transverse cross projection.
Using Eqs. (254) and (256) ,
A ∥ = e ^ ∥ ⋅ A ‾ ⋅ e ^ ∥ = A x x k x 2 + 2 A x y k x k y + A y y k y 2 k t 2 , (260) A_{\parallel} = \hat{\mathbf{e}}_{\parallel}\cdot\overline{A}\cdot\hat{\mathbf{e}}_{\parallel}
= \frac{A_{xx}k_x^2 + 2A_{xy}k_xk_y + A_{yy}k_y^2}{k_t^2}, \tag*{(260)} A ∥ = e ^ ∥ ⋅ A ⋅ e ^ ∥ = k t 2 A xx k x 2 + 2 A x y k x k y + A y y k y 2 , (260)
A ⊥ = e ^ ⊥ ⋅ A ‾ ⋅ e ^ ⊥ = A x x k y 2 − 2 A x y k x k y + A y y k x 2 k t 2 , (261) A_{\perp} = \hat{\mathbf{e}}_{\perp}\cdot\overline{A}\cdot\hat{\mathbf{e}}_{\perp}
= \frac{A_{xx}k_y^2 - 2A_{xy}k_xk_y + A_{yy}k_x^2}{k_t^2}, \tag*{(261)} A ⊥ = e ^ ⊥ ⋅ A ⋅ e ^ ⊥ = k t 2 A xx k y 2 − 2 A x y k x k y + A y y k x 2 , (261)
A × = e ^ ∥ ⋅ A ‾ ⋅ e ^ ⊥ = A x y ( k x 2 − k y 2 ) + ( A y y − A x x ) k x k y k t 2 . (262) A_{\times} = \hat{\mathbf{e}}_{\parallel}\cdot\overline{A}\cdot\hat{\mathbf{e}}_{\perp}
= \frac{A_{xy}(k_x^2-k_y^2) + (A_{yy}-A_{xx})k_xk_y}{k_t^2}. \tag*{(262)} A × = e ^ ∥ ⋅ A ⋅ e ^ ⊥ = k t 2 A x y ( k x 2 − k y 2 ) + ( A y y − A xx ) k x k y . (262)
Consequently,
A ‾ = A ∥ e ^ ∥ e ^ ∥ + A ⊥ e ^ ⊥ e ^ ⊥ + A × ( e ^ ∥ e ^ ⊥ + e ^ ⊥ e ^ ∥ ) . (263) \overline{\mathbf{A}}=A_{\parallel}\hat{\mathbf{e}}_{\parallel}\hat{\mathbf{e}}_{\parallel}+A_{\perp}\hat{\mathbf{e}}_{\perp}\hat{\mathbf{e}}_{\perp}
+A_{\times}\left(\hat{\mathbf{e}}_{\parallel}\hat{\mathbf{e}}_{\perp}+\hat{\mathbf{e}}_{\perp}\hat{\mathbf{e}}_{\parallel}\right). \tag*{(263)} A = A ∥ e ^ ∥ e ^ ∥ + A ⊥ e ^ ⊥ e ^ ⊥ + A × ( e ^ ∥ e ^ ⊥ + e ^ ⊥ e ^ ∥ ) . (263)
Juxtaposition of two unit vectors denotes their dyadic product.
The physical meaning of these three quantities is immediate: A ∥ A_{\parallel} A ∥ is the response sampled by a tangential electric field parallel to k t \mathbf{k}_{t} k t , A ⊥ A_{\perp} A ⊥ is that sampled by the TE-polarised electric field, and A × A_{\times} A × couples these two field components.
The condition for exact TE–TM decoupling can also be seen directly in the principal frame of A ‾ \overline{\mathbf{A}} A . Let its eigenvalues be A 1 A_{1} A 1 and A 2 A_{2} A 2 , and let ϑ \vartheta ϑ denote the angle between k t \mathbf{k}_{t} k t and the first principal axis. In this frame,
A ∥ = A 1 cos 2 ϑ + A 2 sin 2 ϑ , (264) A_{\parallel}=A_{1}\cos^{2}\vartheta+A_{2}\sin^{2}\vartheta, \tag*{(264)} A ∥ = A 1 cos 2 ϑ + A 2 sin 2 ϑ , (264)
A ⊥ = A 1 sin 2 ϑ + A 2 cos 2 ϑ , (265) A_{\perp}=A_{1}\sin^{2}\vartheta+A_{2}\cos^{2}\vartheta, \tag*{(265)} A ⊥ = A 1 sin 2 ϑ + A 2 cos 2 ϑ , (265)
A × = A 2 − A 1 2 sin 2 ϑ . (266) A_{\times}=\frac{A_{2}-A_{1}}{2}\sin2\vartheta. \tag*{(266)} A × = 2 A 2 − A 1 sin 2 ϑ . (266)
Thus A × = 0 A_{\times}=0 A × = 0 identically for an isotropic sheet, A 1 = A 2 A_{1}=A_{2} A 1 = A 2 , and also whenever the propagation direction is aligned with either principal axis, ϑ = 0 \vartheta=0 ϑ = 0 or π / 2 \pi/2 π /2 . Away from these directions, an anisotropic sheet generally has A × ≠ 0 A_{\times}\ne0 A × = 0 .
At the Γ \Gamma Γ point, k t = 0 \mathbf{k}_{t}=0 k t = 0 , the propagation-fixed basis itself is undefined. Direction-dependent quantities at Γ \Gamma Γ are therefore understood as one-sided limits along the selected path in reciprocal space.
Exact local TE–TM transverse-resonance condition
The tangential electric field of the most general local eigenmode is expanded in the propagation-fixed basis as
E t = E ∥ e ^ ∥ + E ⊥ e ^ ⊥ . (267) \mathbf{E}_{t}=E_{\parallel}\hat{\mathbf{e}}_{\parallel}+E_{\perp}\hat{\mathbf{e}}_{\perp}. \tag*{(267)} E t = E ∥ e ^ ∥ + E ⊥ e ^ ⊥ . (267)
The local constitutive relation of the capacitive sheet is
Y ‾ s = j ω C ‾ e q v = j 2 k 0 η 0 A ‾ . (268) \overline{\mathbf{Y}}_{s}=j\omega\overline{\mathbf{C}}_{\mathrm{eqv}}=j\frac{2k_{0}}{\eta_{0}}\overline{\mathbf{A}}. \tag*{(268)} Y s = j ω C eqv = j η 0 2 k 0 A . (268)
The exterior free-space admittance dyadic is diagonal in the propagation-fixed basis because the TE and TM plane-wave polarisations are eigenpolarisations of homogeneous free space. It may therefore be written as
Y ‾ 0 = Y 0 T M e ^ ∥ e ^ ∥ + Y 0 T E e ^ ⊥ e ^ ⊥ . (269) \overline{\mathbf{Y}}_{0}=Y_{0}^{\mathrm{TM}}\hat{\mathbf{e}}_{\parallel}\hat{\mathbf{e}}_{\parallel}+Y_{0}^{\mathrm{TE}}\hat{\mathbf{e}}_{\perp}\hat{\mathbf{e}}_{\perp}. \tag*{(269)} Y 0 = Y 0 TM e ^ ∥ e ^ ∥ + Y 0 TE e ^ ⊥ e ^ ⊥ . (269)
For a free-standing sheet in the same medium on both sides, the two half-spaces contribute equally, so that the source-free transverse-resonance condition is
( 2 Y ‾ 0 + Y ‾ s ) ⋅ E t = 0 t . (270) \left(2\overline{\mathbf{Y}}_{0}+\overline{\mathbf{Y}}_{s}\right)\cdot\mathbf{E}_{t}=\mathbf{0}_{t}. \tag*{(270)} ( 2 Y 0 + Y s ) ⋅ E t = 0 t . (270)
Defining
T ‾ 0 = 2 Y ‾ 0 + Y ‾ s , (271) \overline{\mathbf{T}}_{0}=2\overline{\mathbf{Y}}_{0}+\overline{\mathbf{Y}}_{s}, \tag*{(271)} T 0 = 2 Y 0 + Y s , (271)
Eq. (270) becomes
T ‾ 0 ⋅ E t = 0 t . (272) \overline{\mathbf{T}}_{0}\cdot\mathbf{E}_{t}=\mathbf{0}_{t}. \tag*{(272)} T 0 ⋅ E t = 0 t . (272)
Using Eq. (259) , the matrix representation of the transverse-resonance dyadic is
[ T ‾ 0 ] ∥ ⊥ = ( T T M T × T × T T E ) . (273) \left[\overline{\mathbf{T}}_{0}\right]_{\parallel\perp}=
\begin{pmatrix}
\mathcal{T}_{\mathrm{TM}} & \mathcal{T}_{\times}\\
\mathcal{T}_{\times} & \mathcal{T}_{\mathrm{TE}}
\end{pmatrix}. \tag*{(273)} [ T 0 ] ∥⊥ = ( T TM T × T × T TE ) . (273)
with
T T M = 2 Y 0 T M + j 2 k 0 η 0 A ∥ , (274) \mathcal{T}_{\mathrm{TM}}=2Y_{0}^{\mathrm{TM}}+j\frac{2k_{0}}{\eta_{0}}A_{\parallel}, \tag*{(274)} T TM = 2 Y 0 TM + j η 0 2 k 0 A ∥ , (274)
T T E = 2 Y 0 T E + j 2 k 0 η 0 A ⊥ , (275) \mathcal{T}_{\mathrm{TE}}=2Y_{0}^{\mathrm{TE}}+j\frac{2k_{0}}{\eta_{0}}A_{\perp}, \tag*{(275)} T TE = 2 Y 0 TE + j η 0 2 k 0 A ⊥ , (275)
T × = j 2 k 0 η 0 A × . (276) \mathcal{T}_{\times}=j\frac{2k_{0}}{\eta_{0}}A_{\times}. \tag*{(276)} T × = j η 0 2 k 0 A × . (276)
(272) is therefore equivalent to the pair
T T M E ∥ + T × E ⊥ = 0 , (277) \mathcal{T}_{\mathrm{TM}} E_{\parallel}+\mathcal{T}_{\times}E_{\perp}=0, \tag*{(277)} T TM E ∥ + T × E ⊥ = 0 , (277)
T × E ∥ + T T E E ⊥ = 0. (278) \mathcal{T}_{\times}E_{\parallel}+\mathcal{T}_{\mathrm{TE}}E_{\perp}=0. \tag*{(278)} T × E ∥ + T TE E ⊥ = 0. (278)
A nontrivial field requires the determinant of Eq. (273) to vanish:
det ( T ‾ ‾ 0 ) = T T M T T E − T × 2 = 0. (279) \det\left(\overline{\overline{\mathcal{T}}}_0\right)=\mathcal{T}_{\mathrm{TM}}\mathcal{T}_{\mathrm{TE}}-\mathcal{T}_{\times}^{2}=0. \tag*{(279)} det ( T 0 ) = T TM T TE − T × 2 = 0. (279)
Explicitly,
( 2 Y 0 T M + j 2 k 0 η 0 A ∥ ) ( 2 Y 0 T E + j 2 k 0 η 0 A ⊥ ) − ( j 2 k 0 η 0 A × ) 2 = 0. (280) \left(2Y_0^{\mathrm{TM}}+j\frac{2k_0}{\eta_0}A_{\parallel}\right)\left(2Y_0^{\mathrm{TE}}+j\frac{2k_0}{\eta_0}A_{\perp}\right)-\left(j\frac{2k_0}{\eta_0}A_{\times}\right)^2=0. \tag*{(280)} ( 2 Y 0 TM + j η 0 2 k 0 A ∥ ) ( 2 Y 0 TE + j η 0 2 k 0 A ⊥ ) − ( j η 0 2 k 0 A × ) 2 = 0. (280)
This is the complete local anisotropic TE–TM eigencondition. The cross term is generated entirely by the representation of the anisotropic sheet response in a propagation-fixed frame. It vanishes for an isotropic sheet and along the principal directions of an anisotropic sheet. Otherwise, the eigenpolarisation is generally hybrid TE–TM, consistently with the behaviour of anisotropic electromagnetic sheets [25 ] .
Schur-complement reduction and hybridisation measure
For the positive-definite capacitive patch arrays considered in this work, the TM diagonal factor does not independently vanish on the bound branch of interest. In the bound-wave convention used in the main text,
Y 0 T M = j ω ε 0 γ , γ = k t 2 − k 0 2 > 0. (281) Y_0^{\mathrm{TM}}=j\frac{\omega\varepsilon_0}{\gamma},\qquad\gamma=\sqrt{k_t^2-k_0^2}>0. \tag*{(281)} Y 0 TM = j γ ω ε 0 , γ = k t 2 − k 0 2 > 0. (281)
Since positive definiteness of A ‾ \overline{A} A implies A ∥ > 0 A_{\parallel}>0 A ∥ > 0 for every nonzero direction,
T T M = j ( 2 ω ε 0 γ + 2 k 0 η 0 A ∥ ) ≠ 0. (282) \mathcal{T}_{\mathrm{TM}}=j\left(\frac{2\omega\varepsilon_0}{\gamma}+\frac{2k_0}{\eta_0}A_{\parallel}\right)\ne0. \tag*{(282)} T TM = j ( γ 2 ω ε 0 + η 0 2 k 0 A ∥ ) = 0. (282)
Thus no independent bound TM root is produced by the uncoupled capacitive sheet. (277) can therefore be solved for the longitudinal field component:
E ∥ = − T × T T M E ⊥ . (283) E_{\parallel}=-\frac{\mathcal{T}_{\times}}{\mathcal{T}_{\mathrm{TM}}}E_{\perp}. \tag*{(283)} E ∥ = − T TM T × E ⊥ . (283)
Substitution into Eq. (278) yields
[ T T E − T × 2 T T M ] E ⊥ = 0. (284) \left[\mathcal{T}_{\mathrm{TE}}-\frac{\mathcal{T}_{\times}^{2}}{\mathcal{T}_{\mathrm{TM}}}\right]E_{\perp}=0. \tag*{(284)} [ T TE − T TM T × 2 ] E ⊥ = 0. (284)
Hence the exact hybrid eigencondition may equivalently be written in Schur-complement form as
T T E − T × 2 T T M = 0. (285) \mathcal{T}_{\mathrm{TE}}-\frac{\mathcal{T}_{\times}^{2}}{\mathcal{T}_{\mathrm{TM}}}=0. \tag*{(285)} T TE − T TM T × 2 = 0. (285)
The second term is the correction to the uncoupled TE transverse-resonance condition produced by anisotropic TE–TM hybridisation. (283) also provides a direct measure of the hybrid character of the mode. We define
ζ h y b = ∣ E ∥ E ⊥ ∣ = ∣ T × T T M ∣ . (286) \zeta_{\mathrm{hyb}}=\left|\frac{E_{\parallel}}{E_{\perp}}\right|=\left|\frac{\mathcal{T}_{\times}}{\mathcal{T}_{\mathrm{TM}}}\right|. \tag*{(286)} ζ hyb = E ⊥ E ∥ = T TM T × . (286)
When A × = 0 A_{\times}=0 A × = 0 , T × = 0 \mathcal{T}_{\times}=0 T × = 0 and therefore ζ h y b = 0 \zeta_{\mathrm{hyb}}=0 ζ hyb = 0 ; the TE and TM sectors decouple exactly. When ζ h y b ≪ 1 \zeta_{\mathrm{hyb}}\ll1 ζ hyb ≪ 1 , the longitudinal component is perturbatively small and the eigenmode is TE dominant.
For the positive-definite capacitive patch arrays considered here, T T M ≠ 0 \mathcal{T}_{\mathrm{TM}}\ne0 T TM = 0 on the bound branch of interest, so the uncoupled TM sector does not support an independent bound solution. The hybrid branch is therefore continuously connected to the pure TE solution in the limit A × → 0 A_{\times} \to0 A × → 0 , while the longitudinal component E ∥ E_{\parallel} E ∥ is induced by the anisotropic cross coupling. Indeed, from Eq. (283) ,
E ∥ E ⊥ = − T × T T M , (287) \frac{E_{\parallel}}{E_{\perp}}=-\frac{T_{\times}}{T_{\mathrm{TM}}}, \tag*{(287)} E ⊥ E ∥ = − T TM T × , (287)
so that the longitudinal field is first order in the coupling ratio. By contrast, the correction to the transverse-resonance condition enters through the Schur complement as
− T × 2 T T M , (288) -\frac{T_{\times}^{2}}{T_{\mathrm{TM}}}, \tag*{(288)} − T TM T × 2 , (288)
and is therefore quadratic in the cross-coupling term. Hence, when ζ h y b ≪ 1 \zeta_{\mathrm{hyb}} \ll1 ζ hyb ≪ 1 , the leading-order dispersion of the hybrid mode is governed by its transverse component, with the longitudinal field providing only a perturbative correction. The reduced dispersion model used in the main text therefore follows the dominant transverse component of this TE-dominant branch. Its effective capacitive response is the transverse quadratic projection A 0 ( k t ) = A ⊥ A_{0}(\mathbf{k}_{t})=A_{\perp} A 0 ( k t ) = A ⊥ , while the explicit Schur-complement correction is omitted in the reduced closure.
Quasi-TE reduction used in the dispersion model
The transverse response sampled by the TE-dominant field is
A 0 ( k t ) ≡ A ⊥ = e ^ T E ⋅ A ‾ ⋅ e ^ T E , (289) A_{0}(\mathbf{k}_{t}) \equiv A_{\perp} = \hat{\mathbf{e}}_{\mathrm{TE}} \cdot\overline{\mathbf{A}} \cdot\hat{\mathbf{e}}_{\mathrm{TE}}, \tag*{(289)} A 0 ( k t ) ≡ A ⊥ = e ^ TE ⋅ A ⋅ e ^ TE , (289)
or, in Cartesian components,
A 0 ( k t ) = A x x k y 2 − 2 A x y k x k y + A y y k x 2 k t 2 . (290) A_{0}(\mathbf{k}_{t}) = \frac{A_{xx}k_{y}^{2}-2A_{xy}k_{x}k_{y}+A_{yy}k_{x}^{2}}{k_{t}^{2}}. \tag*{(290)} A 0 ( k t ) = k t 2 A xx k y 2 − 2 A x y k x k y + A y y k x 2 . (290)
For an isotropic sheet, A ‾ = A g I ‾ \overline{\mathbf{A}}=A_{g}\overline{\mathbf{I}} A = A g I and hence
A 0 ( k t ) = A g (291) A_{0}(\mathbf{k}_{t}) = A_{g} \tag*{(291)} A 0 ( k t ) = A g (291)
for every propagation direction. Neglecting the explicit Schur-complement term in Eq. (285) gives the dominant transverse condition
T T E = 0. (292) T_{\mathrm{TE}}=0. \tag*{(292)} T TE = 0. (292)
This reduction is exact when A × = 0 A_{\times}=0 A × = 0 , and therefore whenever the propagation direction coincides with a principal axis of A ‾ \overline{\mathbf{A}} A . Away from a principal direction it is the quasi-TE approximation to the exact hybrid eigenproblem, with its accuracy characterised by Eq. (286) .
The local dyadic constitutive law itself is independent of k t \mathbf{k}_{t} k t . To retain the same k t \mathbf{k}_{t} k t -dependent correction used for the scalar patch-array model, the Luukkonen correction factor is applied only to the transverse scalar projection A 0 A_{0} A 0 :
2 Y 0 T E + j 2 k 0 η 0 A 0 ( k t ) F L ( k t ) = 0 , (293) 2Y_{0}^{\mathrm{TE}}+j\frac{2k_{0}}{\eta_{0}}A_{0}(\mathbf{k}_{t})F_{L}(k_{t})=0, \tag*{(293)} 2 Y 0 TE + j η 0 2 k 0 A 0 ( k t ) F L ( k t ) = 0 , (293)
where
F L ( k t ) = 1 − k t 2 2 k 0 2 . (294) F_{L}(k_{t})=1-\frac{k_{t}^{2}}{2k_{0}^{2}}. \tag*{(294)} F L ( k t ) = 1 − 2 k 0 2 k t 2 . (294)
No tensorial extension of F L F_{L} F L to the off-diagonal hybridisation term is introduced. Equation (293) is therefore a reduced closure: the complete anisotropy of the quasi-static sheet is retained through the directional quadratic projection A 0 ( k t ) A_{0}(\mathbf{k}_{t}) A 0 ( k t ) , whereas the explicit TE–TM Schur-complement correction remains omitted.
Closed-form evanescent decay constant
For a bound surface wave,
k t 2 = k 0 2 + γ 2 , γ > 0. (295) k_t^2 = k_0^2 + \gamma^2,\qquad\gamma> 0. \tag*{(295)} k t 2 = k 0 2 + γ 2 , γ > 0. (295)
Using the scalar TE transverse-resonance relation corresponding to Eq. (293) , the common admittance factors cancel to give
γ = k 0 2 A 0 ( k t ) F L ( k t ) . (296) \gamma= k_0^2 A_0(k_t) F_L(k_t). \tag*{(296)} γ = k 0 2 A 0 ( k t ) F L ( k t ) . (296)
The k t k_t k t -dependent correction factor can be expressed in terms of γ \gamma γ through Eq. (295) :
F L ( k t ) = 1 − k 0 2 + γ 2 2 k 0 2 = 1 2 ( 1 − γ 2 k 0 2 ) . (297) \begin{aligned}
F_L(k_t) &= 1 - \frac{k_0^2+\gamma^2}{2k_0^2}\\
&= \frac{1}{2}\left(1-\frac{\gamma^2}{k_0^2}\right). \tag*{(297)}
\end{aligned} F L ( k t ) = 1 − 2 k 0 2 k 0 2 + γ 2 = 2 1 ( 1 − k 0 2 γ 2 ) . (297)
Substituting Eq. (297) into Eq. (296) gives
γ = A 0 2 ( k 0 2 − γ 2 ) , (298) \gamma= \frac{A_0}{2}\left(k_0^2-\gamma^2\right), \tag*{(298)} γ = 2 A 0 ( k 0 2 − γ 2 ) , (298)
where the argument k t k_t k t of A 0 A_0 A 0 has been suppressed for compactness. Rearrangement yields the quadratic equation
A 0 γ 2 + 2 γ − A 0 k 0 2 = 0. (299) A_0\gamma^2+2\gamma-A_0k_0^2=0. \tag*{(299)} A 0 γ 2 + 2 γ − A 0 k 0 2 = 0. (299)
Its two algebraic roots are
γ ± = − 1 ± 1 + ( k 0 A 0 ) 2 A 0 . (300) \gamma_{\pm}=\frac{-1\pm\sqrt{1+(k_0A_0)^2}}{A_0}. \tag*{(300)} γ ± = A 0 − 1 ± 1 + ( k 0 A 0 ) 2 . (300)
For the positive-definite capacitive sheets considered here, A 0 > 0 A_0>0 A 0 > 0 . The root with the minus sign is negative and is therefore incompatible with exponential decay away from the sheet. The physical root is consequently
Γ T L ( k 0 , k t ) = 1 + [ k 0 A 0 ( k t ) ] 2 − 1 A 0 ( k t ) . (301) \Gamma_{\mathrm{TL}}(k_0,k_t)=\frac{\sqrt{1+[k_0A_0(k_t)]^2}-1}{A_0(k_t)}. \tag*{(301)} Γ TL ( k 0 , k t ) = A 0 ( k t ) 1 + [ k 0 A 0 ( k t ) ] 2 − 1 . (301)
Because Eq. (301) was obtained after explicit substitution of F L ( k t ) F_L(k_t) F L ( k t ) , the k t k_t k t -dependent correction is already embedded in the closed-form root and must not be applied a second time. The corresponding quasi-TE residual is
R 0 ( k 0 , k t ) = k t 2 − k 0 2 − Γ T L ( k 0 , k t ) , (302) R_0(k_0,k_t)=\sqrt{k_t^2-k_0^2}-\Gamma_{\mathrm{TL}}(k_0,k_t), \tag*{(302)} R 0 ( k 0 , k t ) = k t 2 − k 0 2 − Γ TL ( k 0 , k t ) , (302)
and the reduced anisotropic dispersion condition is
R 0 ( k 0 , k t ) = 0. (303) R_0(k_0,k_t)=0. \tag*{(303)} R 0 ( k 0 , k t ) = 0. (303)
In the isotropic limit, A 0 = A g A_0=A_g A 0 = A g , Eq. (301) reduces identically to the positive root of the scalar patch-array condition derived in the main text.
Directional dependence and principal-axis limit
The directional character of the reduced model follows directly from Eq. (265) . In the principal frame,
A 0 ( ϑ ) = A 1 sin 2 ϑ + A 2 cos 2 ϑ . (304) A_0(\vartheta)=A_1\sin^2\vartheta+A_2\cos^2\vartheta. \tag*{(304)} A 0 ( ϑ ) = A 1 sin 2 ϑ + A 2 cos 2 ϑ . (304)
Thus the magnitude of the effective capacitive response sampled by the TE field changes continuously as the propagation direction rotates with respect to the principal axes.
For the positive-definite capacitance dyadics considered in this work,
A 1 > 0 , A 2 > 0 , (305) A_1 > 0,\qquad A_2 > 0, \tag*{(305)} A 1 > 0 , A 2 > 0 , (305)
and therefore
A 0 ( ϑ ) > 0 for all ϑ . (306) A_0(\vartheta) > 0\qquad\text{for all }\vartheta. \tag*{(306)} A 0 ( ϑ ) > 0 for all ϑ . (306)
The resulting quasi-TE isofrequency contours consequently remain closed. In the weak-anisotropy regime they approach elliptical contours whose principal directions coincide with the eigenvectors of A ‾ \overline{\mathbf{A}} A . This is the positive-definite counterpart of the sign-indefinite anisotropic-sheet regime, in which one principal response changes sign and open hyperbolic isofrequency contours may occur [23 , 24 , 60 ] .
Equations (S206) and (S212) provide the exact local anisotropic eigencondition for the quasi-static sheet constitutive law. Equation (S220), together with Eqs. (S228) and (S229), specifies the reduced quasi-TE closure with the prescribed k t k_t k t -dependent transverse correction. Local TE–TM decoupling is exact along a principal axis and in the isotropic limit. The hybrid calculation described below retains the same transverse correction and restores the cross-polarisation response within every retained Floquet order.
Hybrid TE–TM closure and full-zone isofrequency calculation
For the isofrequency comparison in the main text, we use a reduced hybrid extension of the quasi-TE model. The factor F L F_L F L continues to act only on the TE sheet entry, while the TM and cross entries retain their quasi-static values. This specifies the polarisation closure used in the calculation; it is an additional modelling assumption beyond the strictly local constitutive law of Eq. (S195).
We retain the fundamental order and the six orders of the first reciprocal shell. For p = 0 , … , 6 p=0,\ldots,6 p = 0 , … , 6 , let
q p = k t + G p , G 0 = 0 , γ p = ∣ q p ∣ 2 − k 0 2 > 0. (307) \mathbf{q}_p=\mathbf{k}_t+\mathbf{G}_p,\qquad\mathbf{G}_0=0,\qquad\gamma_p=\sqrt{\lvert\mathbf{q}_p\rvert^2-k_0^2}>0. \tag*{(307)} q p = k t + G p , G 0 = 0 , γ p = ∣ q p ∣ 2 − k 0 2 > 0. (307)
Here G 1 , … , G 6 \mathbf{G}_1,\ldots,\mathbf{G}_6 G 1 , … , G 6 are the six first-shell reciprocal vectors. The projections A ∥ , p A_{\parallel,p} A ∥ , p , A ⊥ , p A_{\perp,p} A ⊥ , p , and A × , p A_{\times,p} A × , p are obtained from Eqs. (S187)–(S189) by replacing k t \mathbf{k}_t k t with q p \mathbf{q}_p q p . Thus every retained harmonic samples the same dyadic in its own propagation-fixed basis.
Using the bound-wave admittance Y 0 , p T E = − j γ p / ( k 0 η 0 ) Y_{0,p}^{\mathrm{TE}}=-j\gamma_p/(k_0\eta_0) Y 0 , p TE = − j γ p / ( k 0 η 0 ) , the corrected TE diagonal is
T T E , p ( L ) = j 2 k 0 η 0 [ A ⊥ , p F L ( ∣ q p ∣ ) − γ p k 0 2 ] , (308) \mathcal{T}_{\mathrm{TE},p}^{(L)}=j\frac{2k_0}{\eta_0}\left[A_{\perp,p}F_L(\lvert\mathbf{q}_p\rvert)-\frac{\gamma_p}{k_0^2}\right], \tag*{(308)} T TE , p ( L ) = j η 0 2 k 0 [ A ⊥ , p F L (∣ q p ∣) − k 0 2 γ p ] , (308)
whereas T T M , p \mathcal{T}_{\mathrm{TM},p} T TM , p and T × , p \mathcal{T}_{\times,p} T × , p follow Eqs. (S201) and (S203) with the order-dependent projections. Substituting these entries into the Schur complement of Eq. (S212), and using ∣ q p ∣ 2 = k 0 2 + γ p 2 \lvert\mathbf{q}_p\rvert^2=k_0^2+\gamma_p^2 ∣ q p ∣ 2 = k 0 2 + γ p 2 , gives the uncoupled hybrid residual
F p h y b = γ p − A ⊥ , p 2 ( k 0 2 − γ p 2 ) + k 0 2 A × , p 2 γ p 1 + A ∥ , p γ p . (309) \mathcal{F}_p^{\mathrm{hyb}}=\gamma_p-\frac{A_{\perp,p}}{2}\left(k_0^2-\gamma_p^2\right)+\frac{k_0^2A_{\times,p}^2\gamma_p}{1+A_{\parallel,p}\gamma_p}. \tag*{(309)} F p hyb = γ p − 2 A ⊥ , p ( k 0 2 − γ p 2 ) + 1 + A ∥ , p γ p k 0 2 A × , p 2 γ p . (309)
The local hybrid condition for order p p p is F p h y b = 0 \mathcal{F}_p^{\mathrm{hyb}}=0 F p hyb = 0 . The last term restores the longitudinal–transverse coupling; omitting it gives the corrected quasi-TE condition.
To use the same decay-detuning normalisation as the first-shell model, define
Γ p = Γ T L ( k 0 , q p ) = k 0 2 A ⊥ , p 1 + 1 + ( k 0 A ⊥ , p ) 2 . (310) \Gamma_p=\Gamma_{\mathrm{TL}}(k_0,\mathbf{q}_p)=\frac{k_0^2A_{\perp,p}}{1+\sqrt{1+(k_0A_{\perp,p})^2}}. \tag*{(310)} Γ p = Γ TL ( k 0 , q p ) = 1 + 1 + ( k 0 A ⊥ , p ) 2 k 0 2 A ⊥ , p . (310)
N p = 1 + A ⊥ , p 2 ( γ p + Γ p ) > 0. (311) \mathcal{N}_p=1+\frac{A_{\perp,p}}{2}\left(\gamma_p+\Gamma_p\right)>0. \tag*{(311)} N p = 1 + 2 A ⊥ , p ( γ p + Γ p ) > 0. (311)
The quantity Γ p \Gamma_p Γ p remains the quasi-TE loaded decay of Eq. (S228) in both calculations. Its defining quadratic gives the identity
γ p − A ⊥ , p 2 ( k 0 2 − γ p 2 ) = N p ( γ p − Γ p ) . (312) \gamma_p-\frac{A_{\perp,p}}{2}\left(k_0^2-\gamma_p^2\right)=\mathcal{N}_p\left(\gamma_p-\Gamma_p\right). \tag*{(312)} γ p − 2 A ⊥ , p ( k 0 2 − γ p 2 ) = N p ( γ p − Γ p ) . (312)
Consequently, the normalised detunings used in the two models are
D p q T E = γ p − Γ p , (313) D^{\mathrm{qTE}}_{p}=\gamma_{p}-\Gamma_{p}, \tag*{(313)} D p qTE = γ p − Γ p , (313)
D p h y b = F p h y b N p = γ p − Γ p + k 0 2 A × , p 2 γ p N p ( 1 + A ∥ , p γ p ) . (314) D^{\mathrm{hyb}}_{p}=\frac{\mathcal{F}^{\mathrm{hyb}}_{p}}{\mathcal{N}_{p}}=\gamma_{p}-\Gamma_{p}+\frac{k_{0}^{2}A_{\times,p}^{2}\gamma_{p}}{\mathcal{N}_{p}(1+A_{\parallel,p}\gamma_{p})}. \tag*{(314)} D p hyb = N p F p hyb = γ p − Γ p + N p ( 1 + A ∥ , p γ p ) k 0 2 A × , p 2 γ p . (314)
Both detunings have units of inverse length. The division by N p \mathcal{N}_{p} N p preserves the uncoupled roots and puts the hybrid residual in the normalisation of the existing Floquet coupling ansatz.
We retain that same geometric coupling in both models:
v p = m ~ p ( k t ) Γ 0 Γ p η 0 p ( k t ) , p = 1 , … , 6 , (315) v_{p}=\widetilde{m}_{p}(\mathbf{k}_{t})\sqrt{\Gamma_{0}\Gamma_{p}}\,\eta_{0p}(\mathbf{k}_{t}),\qquad p=1,\ldots,6, \tag*{(315)} v p = m p ( k t ) Γ 0 Γ p η 0 p ( k t ) , p = 1 , … , 6 , (315)
With v = ( v 1 , … , v 6 ) T \mathbf{v}=(v_{1},\ldots,v_{6})^{\mathrm{T}} v = ( v 1 , … , v 6 ) T , the TM-eliminated system is
H ( m ) ( k 0 , k t ) = ( D 0 ( m ) − v † − v diag ( D 1 ( m ) , … , D 6 ( m ) ) ) , m ∈ { q T E , h y b } . (316) \mathbf{H}^{(m)}(k_{0},\mathbf{k}_{t})=
\begin{pmatrix}
D_{0}^{(m)} & -\mathbf{v}^{\dagger}\\
-\mathbf{v} & \operatorname{diag}(D_{1}^{(m)},\ldots,D_{6}^{(m)})
\end{pmatrix},
\qquad m\in\{\mathrm{qTE},\mathrm{hyb}\}. \tag*{(316)} H ( m ) ( k 0 , k t ) = ( D 0 ( m ) − v − v † diag ( D 1 ( m ) , … , D 6 ( m ) ) ) , m ∈ { qTE , hyb } . (316)
Inter-order coupling acts on the TE amplitudes, while the TM response enters through the local Schur complement of each order. The sideband block remains diagonal, as in the reduced first-shell model. All six sidebands are retained separately at a general k t \mathbf{k}_{t} k t .
Whenever the sideband detunings are nonzero, their elimination gives
D 0 ( m ) − ∑ p = 1 6 ∣ v p ∣ 2 D p ( m ) = 0. (317) D_{0}^{(m)}-\sum_{p=1}^{6}\frac{\lvert v_{p}\rvert^{2}}{D_{p}^{(m)}}=0. \tag*{(317)} D 0 ( m ) − p = 1 ∑ 6 D p ( m ) ∣ v p ∣ 2 = 0. (317)
For numerical evaluation we use the Hermitian matrix in Eq. (316) directly, including near a vanishing sideband detuning. The frequency maps displayed in the main text are obtained from the lowest eigenvalue condition
λ min [ H ( m ) ( k 0 , k t ) ] = 0 , 0 < k 0 < ∣ k t ∣ , f ( m ) ( k t ) = c 0 k 0 2 π . (318) \lambda_{\min}\left[\mathbf{H}^{(m)}(k_{0},\mathbf{k}_{t})\right]=0,\qquad0<k_{0}<\lvert\mathbf{k}_{t}\rvert,\qquad f^{(m)}(\mathbf{k}_{t})=\frac{c_{0}k_{0}}{2\pi}. \tag*{(318)} λ m i n [ H ( m ) ( k 0 , k t ) ] = 0 , 0 < k 0 < ∣ k t ∣ , f ( m ) ( k t ) = 2 π c 0 k 0 . (318)
We solve this condition throughout the first Brillouin zone and plot the level sets f ( m ) ( k x , k y ) = f ∗ f^{(m)}(k_{x},k_{y})=f_{\ast} f ( m ) ( k x , k y ) = f ∗ . At Γ \Gamma Γ , the frequency map uses the continuous lowest-branch limit f → 0 f\to0 f → 0 .
The fundamental polarisation ratio follows from the eliminated TM row:
E ∥ , 0 E ⊥ , 0 = − A × , 0 γ 0 1 + A ∥ , 0 γ 0 , ζ h y b , 0 = ∣ A × , 0 ∣ γ 0 1 + A ∥ , 0 γ 0 . (319) \frac{E_{\parallel,0}}{E_{\perp,0}}=-\frac{A_{\times,0}\gamma_{0}}{1+A_{\parallel,0}\gamma_{0}},\qquad\zeta_{\mathrm{hyb},0}=\frac{\lvert A_{\times,0}\rvert\gamma_{0}}{1+A_{\parallel,0}\gamma_{0}}. \tag*{(319)} E ⊥ , 0 E ∥ , 0 = − 1 + A ∥ , 0 γ 0 A × , 0 γ 0 , ζ hyb , 0 = 1 + A ∥ , 0 γ 0 ∣ A × , 0 ∣ γ 0 . (319)
For an isotropic dyadic, A × , p = 0 A_{\times,p}=0 A × , p = 0 for every order, so the quasi-TE and hybrid matrices coincide exactly. This explains their agreement for the C 6 v C_{6v} C 6 v example within this reduced model. For a C 1 C_{1} C 1 dyadic, the cross projections are generally nonzero and modify both the fundamental and sideband detunings. Even when k t \mathbf{k}_{t} k t lies along a principal axis and A × , 0 = 0 A_{\times,0}=0 A × , 0 = 0 , the shifted vectors q p \mathbf{q}_{p} q p can have A × , p ≠ 0 A_{\times,p}\ne0 A × , p = 0 ; the full first-shell comparison can therefore retain a hybrid correction along that direction.
Reciprocal-shell hierarchy and Floquet truncation
The nonlocal correction introduced in Sec. IV C of the main text is constructed by retaining the first nonzero shell of Floquet harmonics. Here we first identify the Floquet orders belonging to the first three reciprocal shells and then compare their evanescent character at the high-symmetry points Γ \Gamma Γ , M M M , and K K K . This provides a simple reciprocal-space justification for retaining the complete six-member first shell while neglecting the more strongly evanescent higher shells.
The triangular Bravais lattice is associated with the regular hexagonal unit cell of circumradius
R h = 11.547 m m . (320) R_h = 11.547\ \mathrm{mm}. \tag*{(320)} R h = 11.547 mm . (320)
For the Wigner–Seitz hexagon of the triangular lattice, the circumradius and lattice constant satisfy
R h = a 3 , a = 3 R h = 20.000 m m . (321) R_h = \frac{a}{\sqrt{3}}, \qquad a = \sqrt{3}R_h = 20.000\ \mathrm{mm}. \tag*{(321)} R h = 3 a , a = 3 R h = 20.000 mm . (321)
The lattice constant a a a determines the reciprocal-lattice scale, whereas R h R_h R h is used below only as a characteristic geometrical length against which the evanescent decay lengths may be compared.
First three reciprocal shells
We choose the primitive direct-lattice vectors as
a 1 = a ( 1 0 ) , a 2 = a ( 1 / 2 3 / 2 ) . (322) \mathbf{a}_1 = a\begin{pmatrix}1\\0\end{pmatrix}, \qquad\mathbf{a}_2 = a\begin{pmatrix}1/2\\\sqrt{3}/2\end{pmatrix}. \tag*{(322)} a 1 = a ( 1 0 ) , a 2 = a ( 1/2 3 /2 ) . (322)
The primitive reciprocal vectors are defined by
a i ⋅ b j = 2 π δ i j , (323) \mathbf{a}_i\cdot\mathbf{b}_j = 2\pi\delta_{ij}, \tag*{(323)} a i ⋅ b j = 2 π δ ij , (323)
which gives
b 1 = 2 π a ( 1 − 1 / 3 ) , b 2 = 2 π a ( 0 2 / 3 ) . (324) \mathbf{b}_1 = \frac{2\pi}{a}\begin{pmatrix}1\\-1/\sqrt{3}\end{pmatrix}, \qquad\mathbf{b}_2 = \frac{2\pi}{a}\begin{pmatrix}0\\2/\sqrt{3}\end{pmatrix}. \tag*{(324)} b 1 = a 2 π ( 1 − 1/ 3 ) , b 2 = a 2 π ( 0 2/ 3 ) . (324)
Their common magnitude is
G 0 = ∣ b 1 ∣ = ∣ b 2 ∣ = 4 π 3 a = 0.36275987 m m − 1 . (325) G_0 = |\mathbf{b}_1| = |\mathbf{b}_2| = \frac{4\pi}{\sqrt{3}a} = 0.36275987\ \mathrm{mm}^{-1}. \tag*{(325)} G 0 = ∣ b 1 ∣ = ∣ b 2 ∣ = 3 a 4 π = 0.36275987 mm − 1 . (325)
An arbitrary reciprocal-lattice vector is
G m n = m b 1 + n b 2 , m , n ∈ Z . (326) \mathbf{G}_{mn} = m\mathbf{b}_1+n\mathbf{b}_2,\qquad m,n\in\mathbb{Z}. \tag*{(326)} G mn = m b 1 + n b 2 , m , n ∈ Z . (326)
Using Eq. (324) ,
∣ G m n ∣ 2 = G 0 2 ( m 2 − m n + n 2 ) . (327) |\mathbf{G}_{mn}|^2 = G_0^2\left(m^2-mn+n^2\right). \tag*{(327)} ∣ G mn ∣ 2 = G 0 2 ( m 2 − mn + n 2 ) . (327)
Defining
s m n = m 2 − m n + n 2 , (328) s_{mn}=m^2-mn+n^2, \tag*{(328)} s mn = m 2 − mn + n 2 , (328)
the reciprocal-shell radius is therefore
∣ G m n ∣ = G 0 s m n . (329) |\mathbf{G}_{mn}|=G_0\sqrt{s_{mn}}. \tag*{(329)} ∣ G mn ∣ = G 0 s mn . (329)
The first nonzero shell is obtained for s m n = 1 s_{mn}=1 s mn = 1 . Its six Floquet orders are
( m , n ) = ( 1 , 0 ) , ( 0 , 1 ) , ( 1 , 1 ) , ( − 1 , 0 ) , ( 0 , − 1 ) , ( − 1 , − 1 ) , (330) (m,n)=(1,0),\ (0,1),\ (1,1),\ (-1,0),\ (0,-1),\ (-1,-1), \tag*{(330)} ( m , n ) = ( 1 , 0 ) , ( 0 , 1 ) , ( 1 , 1 ) , ( − 1 , 0 ) , ( 0 , − 1 ) , ( − 1 , − 1 ) , (330)
corresponding to
G 1 = b 1 , G 2 = b 2 , G 3 = b 1 + b 2 , G 4 = − b 1 , G 5 = − b 2 , G 6 = − ( b 1 + b 2 ) . (331) \mathbf{G}_1=\mathbf{b}_1,\qquad\mathbf{G}_2=\mathbf{b}_2,\qquad\mathbf{G}_3=\mathbf{b}_1+\mathbf{b}_2,
\\
\mathbf{G}_4=-\mathbf{b}_1,\qquad\mathbf{G}_5=-\mathbf{b}_2,\qquad\mathbf{G}_6=-(\mathbf{b}_1+\mathbf{b}_2). \tag*{(331)} G 1 = b 1 , G 2 = b 2 , G 3 = b 1 + b 2 , G 4 = − b 1 , G 5 = − b 2 , G 6 = − ( b 1 + b 2 ) . (331)
All six have the same reciprocal-space magnitude,
∣ G p ∣ = G 0 , p = 1 , … , 6. (332) |\mathbf{G}_p|=G_0,\qquad p=1,\ldots,6. \tag*{(332)} ∣ G p ∣ = G 0 , p = 1 , … , 6. (332)
The second shell corresponds to s m n = 3 s_{mn}=3 s mn = 3 and contains
( m , n ) = ( 1 , − 1 ) , ( − 1 , 1 ) , ( 2 , 1 ) , ( 1 , 2 ) , ( − 2 , − 1 ) , ( − 1 , − 2 ) , (333) (m,n)=(1,-1),\ (-1,1),\ (2,1),\ (1,2),\ (-2,-1),\ (-1,-2), \tag*{(333)} ( m , n ) = ( 1 , − 1 ) , ( − 1 , 1 ) , ( 2 , 1 ) , ( 1 , 2 ) , ( − 2 , − 1 ) , ( − 1 , − 2 ) , (333)
with
∣ G ∣ I I = 3 G 0 . (334) |\mathbf{G}|_{\mathrm{II}}=\sqrt{3}G_0. \tag*{(334)} ∣ G ∣ II = 3 G 0 . (334)
The third shell corresponds to s m n = 4 s_{mn}=4 s mn = 4 and contains
( m , n ) = ( 2 , 0 ) , ( 0 , 2 ) , ( 2 , 2 ) , ( − 2 , 0 ) , ( 0 , − 2 ) , ( − 2 , − 2 ) , (335) (m,n)=(2,0),\ (0,2),\ (2,2),\ (-2,0),\ (0,-2),\ (-2,-2), \tag*{(335)} ( m , n ) = ( 2 , 0 ) , ( 0 , 2 ) , ( 2 , 2 ) , ( − 2 , 0 ) , ( 0 , − 2 ) , ( − 2 , − 2 ) , (335)
with
∣ G ∣ I I I = 2 G 0 . (336) |\mathbf{G}|_{\mathrm{III}}=2G_0. \tag*{(336)} ∣ G ∣ III = 2 G 0 . (336)
The first three shell radii are consequently
∣ G ∣ I = G 0 , ∣ G ∣ I I = 3 G 0 , ∣ G ∣ I I I = 2 G 0 . (337) |\mathbf{G}|_{\mathrm{I}}=G_0,\qquad|\mathbf{G}|_{\mathrm{II}}=\sqrt{3}G_0,\qquad|\mathbf{G}|_{\mathrm{III}}=2G_0. \tag*{(337)} ∣ G ∣ I = G 0 , ∣ G ∣ II = 3 G 0 , ∣ G ∣ III = 2 G 0 . (337)
The corresponding reciprocal-space construction is shown in Figure S1 .
Figure S1. First three reciprocal shells of the triangular Bravais lattice, normalized by G 0 = ∣ b 1 ∣ = ∣ b 2 ∣ G_0 = |\mathbf{b}_1| = |\mathbf{b}_2| G 0 = ∣ b 1 ∣ = ∣ b 2 ∣ . The solid blue, dashed gray, and dotted green contours correspond respectively to the first, second, and third shells. The six vectors G 1 , … , G 6 \mathbf{G}_1,\ldots,\mathbf{G}_6 G 1 , … , G 6 form the complete first nonzero reciprocal shell retained in the Floquet model.
Selection of the six first-shell Floquet harmonics
A Bloch field of the periodic sheet can be expanded in Floquet harmonics as
E t ( ρ , z ) = ∑ m , n ∈ Z E m n ( z ) exp [ j ( k t + G m n ) ⋅ ρ ] . (338) \mathbf{E}_{t}(\boldsymbol{\rho},z)=\sum_{m,n\in\mathbb{Z}}\mathbf{E}_{mn}(z)\exp\left[j\left(\mathbf{k}_{t}+\mathbf{G}_{mn}\right)\cdot\boldsymbol{\rho}\right]. \tag*{(338)} E t ( ρ , z ) = m , n ∈ Z ∑ E mn ( z ) exp [ j ( k t + G mn ) ⋅ ρ ] . (338)
The fundamental Floquet harmonic corresponds to G 00 = 0 \mathbf{G}_{00}=0 G 00 = 0 , while the nonzero reciprocal vectors generate the higher Floquet harmonics of the same Bloch field.
For a Bloch wavevector k t \mathbf{k}_{t} k t , the transverse wavenumber of the harmonic associated with G m n \mathbf{G}_{mn} G mn is
q m n = k t + G m n , q m n = ∣ q m n ∣ . (339) \mathbf{q}_{mn}=\mathbf{k}_{t}+\mathbf{G}_{mn},\qquad q_{mn}=|\mathbf{q}_{mn}|. \tag*{(339)} q mn = k t + G mn , q mn = ∣ q mn ∣. (339)
Its free-space normal wavenumber is
k z , m n = k 0 2 − q m n 2 , k 0 = 2 π f c 0 . (340) k_{z,mn}=\sqrt{k_0^2-q_{mn}^2},\qquad k_0=\frac{2\pi f}{c_0}. \tag*{(340)} k z , mn = k 0 2 − q mn 2 , k 0 = c 0 2 π f . (340)
For q m n > k 0 q_{mn}>k_0 q mn > k 0 , the harmonic is evanescent. Defining
γ m n = q m n 2 − k 0 2 , (341) \gamma_{mn}=\sqrt{q_{mn}^2-k_0^2}, \tag*{(341)} γ mn = q mn 2 − k 0 2 , (341)
the normal wavenumber can be written as
k z , m n = ± j γ m n , (342) k_{z,mn}=\pm j\gamma_{mn}, \tag*{(342)} k z , mn = ± j γ mn , (342)
where the appropriate sign is chosen to enforce decay away from the sheet. The field envelope therefore varies as
e − γ m n ∣ z ∣ , (343) e^{-\gamma_{mn}|z|}, \tag*{(343)} e − γ mn ∣ z ∣ , (343)
with 1 / e 1/e 1/ e decay length
ℓ m n = 1 γ m n . (344) \ell_{mn} = \frac{1}{\gamma_{mn}}. \tag*{(344)} ℓ mn = γ mn 1 . (344)
A large γ m n \gamma_{mn} γ mn , or equivalently a small ℓ m n \ell_{mn} ℓ mn , corresponds to a deeply evanescent and strongly confined Floquet harmonic. Conversely, as
q m n → k 0 + , (345) q_{mn} \to k_0^+, \tag*{(345)} q mn → k 0 + , (345)
one has
γ m n → 0 + , ℓ m n → ∞ , (346) \gamma_{mn} \to0^+,\qquad\ell_{mn} \to\infty, \tag*{(346)} γ mn → 0 + , ℓ mn → ∞ , (346)
and the harmonic becomes progressively more weakly evanescent. The equality
q m n = k 0 (347) q_{mn} = k_0 \tag*{(347)} q mn = k 0 (347)
defines the free-space propagation threshold of that Floquet harmonic,
f t h , m n = c 0 q m n 2 π . (348) f_{\mathrm{th},mn}=\frac{c_0q_{mn}}{2\pi}. \tag*{(348)} f th , mn = 2 π c 0 q mn . (348)
This frequency is not a resonance of the metasurface; it only marks the transition of the corresponding Floquet harmonic from evanescent to propagating.
To compare the first three reciprocal shells, we evaluate the least-confined Floquet harmonic from each shell at the three high-symmetry points Γ \Gamma Γ , M M M , and K K K .
At Γ \Gamma Γ ,
k t = 0 , (349) \mathbf{k}_{t}=0, \tag*{(349)} k t = 0 , (349)
so that q m n = ∣ G m n ∣ q_{mn}=\lvert\mathbf{G}_{mn}\rvert q mn = ∣ G mn ∣ . Therefore
q I ( Γ ) = G 0 , q I I ( Γ ) = 3 G 0 , q I I I ( Γ ) = 2 G 0 . (350) q_{\mathrm{I}}(\Gamma)=G_0,\qquad q_{\mathrm{II}}(\Gamma)=\sqrt{3}G_0,\qquad q_{\mathrm{III}}(\Gamma)=2G_0. \tag*{(350)} q I ( Γ ) = G 0 , q II ( Γ ) = 3 G 0 , q III ( Γ ) = 2 G 0 . (350)
The corresponding propagation thresholds are
f t h , I ( Γ ) = 17.31 G H z , f t h , I I ( Γ ) = 29.98 G H z , f t h , I I I ( Γ ) = 34.62 G H z . (351) f_{\mathrm{th,I}}(\Gamma)=17.31\ \mathrm{GHz},\qquad f_{\mathrm{th,II}}(\Gamma)=29.98\ \mathrm{GHz},\qquad f_{\mathrm{th,III}}(\Gamma)=34.62\ \mathrm{GHz}. \tag*{(351)} f th , I ( Γ ) = 17.31 GHz , f th , II ( Γ ) = 29.98 GHz , f th , III ( Γ ) = 34.62 GHz . (351)
In particular, all six members of the first shell have the same transverse wavenumber q p = G 0 q_{\mathrm{p}}=G_0 q p = G 0 at Γ \Gamma Γ and therefore the same free-space evanescent decay constant.
For a representative M M M point,
M = b 1 2 , (352) \mathbf{M}=\frac{\mathbf{b}_1}{2}, \tag*{(352)} M = 2 b 1 , (352)
the first-shell vector G 4 = − b 1 \mathbf{G}_4=-\mathbf{b}_1 G 4 = − b 1 gives
∣ M + G 4 ∣ = G 0 2 . (353) \lvert\mathbf{M}+\mathbf{G}_4\rvert=\frac{G_0}{2}. \tag*{(353)} ∣ M + G 4 ∣ = 2 G 0 . (353)
The minimum values obtained from the second and third shells are
q I I , min ( M ) = 7 2 G 0 , q I I I , min ( M ) = 3 2 G 0 . (354) q_{\mathrm{II},\min}(M)=\frac{\sqrt{7}}{2}G_0,\qquad q_{\mathrm{III},\min}(M)=\frac{3}{2}G_0. \tag*{(354)} q II , m i n ( M ) = 2 7 G 0 , q III , m i n ( M ) = 2 3 G 0 . (354)
Hence
f t h , I ( M ) = 8.65 G H z , f t h , I I ( M ) = 22.9 G H z , f t h , I I I ( M ) = 25.96 G H z . (355) f_{\mathrm{th,I}}(M)=8.65\ \mathrm{GHz},\qquad f_{\mathrm{th,II}}(M)=22.9\ \mathrm{GHz},\qquad f_{\mathrm{th,III}}(M)=25.96\ \mathrm{GHz}. \tag*{(355)} f th , I ( M ) = 8.65 GHz , f th , II ( M ) = 22.9 GHz , f th , III ( M ) = 25.96 GHz . (355)
For a representative K K K point,
K = 2 b 1 + b 2 3 , (356) \mathbf{K}=\frac{2\mathbf{b}_1+\mathbf{b}_2}{3}, \tag*{(356)} K = 3 2 b 1 + b 2 , (356)
the two first-shell vectors
G 4 = − b 1 , G 6 = − ( b 1 + b 2 ) (357) \mathbf{G}_4=-\mathbf{b}_1,\qquad\mathbf{G}_6=-(\mathbf{b}_1+\mathbf{b}_2) \tag*{(357)} G 4 = − b 1 , G 6 = − ( b 1 + b 2 ) (357)
give
∣ K + G 4 ∣ = ∣ K + G 6 ∣ = G 0 3 . (358) \lvert\mathbf{K}+\mathbf{G}_4\rvert=\lvert\mathbf{K}+\mathbf{G}_6\rvert=\frac{G_0}{\sqrt{3}}. \tag*{(358)} ∣ K + G 4 ∣ = ∣ K + G 6 ∣ = 3 G 0 . (358)
For the second and third shells,
q I I , min ( K ) = 2 G 0 3 , q I I I , min ( K ) = G 0 7 3 . (359) q_{\mathrm{II},\min}(K)=\frac{2G_0}{\sqrt{3}},\qquad q_{\mathrm{III},\min}(K)=G_0\sqrt{\frac{7}{3}}. \tag*{(359)} q II , m i n ( K ) = 3 2 G 0 , q III , m i n ( K ) = G 0 3 7 . (359)
The corresponding thresholds are
f t h , I ( K ) = 9.99 G H z , f t h , I I ( K ) = 19.99 G H z , f t h , I I I ( K ) = 26.44 G H z . (360) f_{\mathrm{th,I}}(K)=9.99\ \mathrm{GHz},\qquad f_{\mathrm{th,II}}(K)=19.99\ \mathrm{GHz},\qquad f_{\mathrm{th,III}}(K)=26.44\ \mathrm{GHz}. \tag*{(360)} f th , I ( K ) = 9.99 GHz , f th , II ( K ) = 19.99 GHz , f th , III ( K ) = 26.44 GHz . (360)
The comparison is summarized in Table S2 .
Table S2. Propagation threshold of the least-confined Floquet harmonic from each of the first three reciprocal shells at the high-symmetry points, calculated for a = 20 a = 20 a = 20 mm.
The first-shell harmonics are therefore the first higher Floquet orders to approach weak evanescence near the Brillouin-zone boundary. The second and third shells remain substantially farther from their propagation thresholds and consequently remain more strongly evanescent over the same frequency range.
The difference can also be seen directly from the decay lengths. For example, at f = 9 f = 9 f = 9 GHz the least-confined harmonic of each shell has
Γ M K first shell 3.23 mm propagating 10.99 mm second shell 1.67 mm 2.27 mm 2.67 mm third shell 1.43 mm 1.96 mm 1.92 mm (361) \begin{array}{c|ccc}
& \Gamma& M & K \\
\text{first shell} & 3.23\ \text{mm} & \text{propagating} & 10.99\ \text{mm} \\
\text{second shell} & 1.67\ \text{mm} & 2.27\ \text{mm} & 2.67\ \text{mm} \\
\text{third shell} & 1.43\ \text{mm} & 1.96\ \text{mm} & 1.92\ \text{mm} \tag*{(361)}
\end{array} first shell second shell third shell Γ 3.23 mm 1.67 mm 1.43 mm M propagating 2.27 mm 1.96 mm K 10.99 mm 2.67 mm 1.92 mm (361)
Compared with the characteristic hexagonal radius R h = 11.547 R_h = 11.547 R h = 11.547 mm, the first-shell harmonics near M M M and K K K possess considerably longer decay lengths than those of the second and third shells. They therefore provide the leading higher-harmonic correction to the fundamental G = 0 \mathbf{G} = 0 G = 0 channel.
The evanescent hierarchy is used here as a spectral truncation criterion; it does not imply that the higher-shell harmonics vanish identically. Their actual contribution also depends on the geometry-dependent coupling to the fundamental harmonic. Nevertheless, the substantially larger transverse wavenumbers and shorter decay lengths of the second and higher shells identify them as higher-order corrections over the frequency range considered here.
Finally, the complete first shell is retained rather than only the harmonic that is least evanescent for one chosen propagation direction. At Γ \Gamma Γ , all six first-shell vectors have the same q p = G 0 q_p = G_0 q p = G 0 . At the representative M M M point of Eq. (352) , G 4 \mathbf{G}_4 G 4 is the least-confined first-shell harmonic. Moving to the other geometrically related M M M points of the hexagonal Brillouin zone selects the other members of the same first shell. Likewise, at the representative K K K point of Eq. (356) , the pair G 4 \mathbf{G}_4 G 4 and G 6 \mathbf{G}_6 G 6 has the smallest shifted transverse wavenumber, while the corresponding pairs are permuted among the first-shell vectors at the other K K K vertices.
Accordingly, the truncation is performed by reciprocal-shell radius, rather than by selecting only those harmonics that are least evanescent along a particular propagation direction. The six vectors
{ ± b 1 , ± b 2 , ± ( b 1 + b 2 ) } (362) \{\pm\mathbf{b}_1,\ \pm\mathbf{b}_2,\ \pm(\mathbf{b}_1+\mathbf{b}_2)\} \tag*{(362)} { ± b 1 , ± b 2 , ± ( b 1 + b 2 )} (362)
form the complete nearest nonzero reciprocal shell. Retaining this complete six-member set therefore provides a magnitude-complete first-shell truncation without artificially privileging a particular propagation direction, while the more deeply evanescent second and higher shells are neglected as higher-order spectral corrections.
Geometry-derived first-shell Floquet correction
This section gives the detailed construction of the nonlocal correction used in the main text. The local sheet response is supplied by the equivalent surface-capacitance dyadic C ‾ e q v \overline{\mathbf{C}}_{\mathrm{eqv}} C eqv , whereas the leading spatially dispersive correction additionally uses the reconstructed metal boundary ∂ Σ \partial\Sigma ∂ Σ . The complete first reciprocal shell is retained as six separate Floquet orders. No positive-half-shell multiplicity and no fitted per-order coupling coefficient are introduced.
The derivation proceeds in four stages. First, the same local dyadic is projected onto the TE direction of every shifted Floquet wavevector. Second, the complex coupling between the fundamental order and each sideband is computed by an exact Fourier integral over the polygonal boundary. Third, a directional-turning functional assigns a geometry-derived amplitude to each of the six orders. Finally, the six amplitudes are eliminated by a Schur complement, with a little-group projection applied when the propagation path is a mirror axis of the complete unit-cell geometry.
Complete first shell and TE-projected local quantities
Let the triangular Bravais lattice be generated by
a 1 = a ( 1 0 0 ) , a 2 = a ( 1 / 2 3 / 2 0 ) . (363) \mathbf{a}_1=a\begin{pmatrix}1\\0\\0\end{pmatrix},\qquad
\mathbf{a}_2=a\begin{pmatrix}1/2\\\sqrt{3}/2\\0\end{pmatrix}. \tag*{(363)} a 1 = a 1 0 0 , a 2 = a 1/2 3 /2 0 . (363)
Writing
A l a t = ( a 1 a 2 ) , B l a t = 2 π A l a t − T = ( b 1 b 2 ) , (364) \mathbf{A}_{\mathrm{lat}}=(\mathbf{a}_1\ \mathbf{a}_2),\qquad
\mathbf{B}_{\mathrm{lat}}=2\pi\mathbf{A}_{\mathrm{lat}}^{-\mathrm{T}}=(\mathbf{b}_1\ \mathbf{b}_2), \tag*{(364)} A lat = ( a 1 a 2 ) , B lat = 2 π A lat − T = ( b 1 b 2 ) , (364)
gives
b 1 = 2 π a ( 1 − 1 / 3 ) , b 2 = 2 π a ( 0 2 / 3 ) , (365) \mathbf{b}_1=\frac{2\pi}{a}\begin{pmatrix}1\\-1/\sqrt{3}\end{pmatrix},\qquad
\mathbf{b}_2=\frac{2\pi}{a}\begin{pmatrix}0\\2/\sqrt{3}\end{pmatrix}, \tag*{(365)} b 1 = a 2 π ( 1 − 1/ 3 ) , b 2 = a 2 π ( 0 2/ 3 ) , (365)
so that a i ⋅ b j = 2 π δ i j \mathbf{a}_i\cdot\mathbf{b}_j=2\pi\delta_{ij} a i ⋅ b j = 2 π δ ij . Their magnitudes and mutual scalar product are
∣ b 1 ∣ = ∣ b 2 ∣ = ∣ b 1 + b 2 ∣ = G 0 , G 0 = 4 π 3 a , b 1 ⋅ b 2 = − G 0 2 2 . (366) |\mathbf{b}_1|=|\mathbf{b}_2|=|\mathbf{b}_1+\mathbf{b}_2|=G_0,\qquad
G_0=\frac{4\pi}{\sqrt{3}a},\qquad
\mathbf{b}_1\cdot\mathbf{b}_2=-\frac{G_0^2}{2}. \tag*{(366)} ∣ b 1 ∣ = ∣ b 2 ∣ = ∣ b 1 + b 2 ∣ = G 0 , G 0 = 3 a 4 π , b 1 ⋅ b 2 = − 2 G 0 2 . (366)
Consequently, the six shortest nonzero reciprocal vectors form the first shell [61 , 62 ] . We use the fixed ordering
G 1 = + b 1 , G 2 = + b 2 , G 3 = + ( b 1 + b 2 ) , \mathbf{G}_1=+\mathbf{b}_1,\qquad
\mathbf{G}_2=+\mathbf{b}_2,\qquad
\mathbf{G}_3=+(\mathbf{b}_1+\mathbf{b}_2), G 1 = + b 1 , G 2 = + b 2 , G 3 = + ( b 1 + b 2 ) ,
G 4 = − b 1 , G 5 = − b 2 , G 6 = − ( b 1 + b 2 ) . (367) \mathbf{G}_4=-\mathbf{b}_1,\qquad
\mathbf{G}_5=-\mathbf{b}_2,\qquad
\mathbf{G}_6=-(\mathbf{b}_1+\mathbf{b}_2). \tag*{(367)} G 4 = − b 1 , G 5 = − b 2 , G 6 = − ( b 1 + b 2 ) . (367)
The selected K K K point and the straight Γ \Gamma Γ –K K K path are
K = 2 b 1 + b 2 3 = 4 π 3 a x ^ , k t ( s ) = s K , 0 ≤ s ≤ 1. (368) \mathbf{K}=\frac{2\mathbf{b}_1+\mathbf{b}_2}{3}=\frac{4\pi}{3a}\hat{\mathbf{x}},\qquad
\mathbf{k}_t(s)=s\mathbf{K},\qquad
0\le s\le1. \tag*{(368)} K = 3 2 b 1 + b 2 = 3 a 4 π x ^ , k t ( s ) = s K , 0 ≤ s ≤ 1. (368)
For notational uniformity, introduce the fundamental order G 0 = 0 \mathbf{G}_0=\mathbf{0} G 0 = 0 . The transverse wavevector of order a ∈ { 0 , 1 , … , 6 } a\in\{0,1,\ldots,6\} a ∈ { 0 , 1 , … , 6 } is then
q a = k t + G a , q 0 = k t . (369) \mathbf{q}_a=\mathbf{k}_t+\mathbf{G}_a,\qquad
\mathbf{q}_0=\mathbf{k}_t. \tag*{(369)} q a = k t + G a , q 0 = k t . (369)
The complete first shell and the corresponding shifted Floquet wavevectors are summarized in Figure S2 . Panel (b) also previews the mirror pairing used later in Sec. S6E.
For an evanescent harmonic, the free-space normal decay is evaluated on the branch with positive real part,
γ h a = ∣ q a ∣ 2 − k 0 2 , Re γ h a > 0. (370) \gamma_{h a}=\sqrt{|\mathbf{q}_a|^2-k_0^2},\qquad
\operatorname{Re}\gamma_{h a}>0. \tag*{(370)} γ ha = ∣ q a ∣ 2 − k 0 2 , Re γ ha > 0. (370)
Thus γ h 0 = k t 2 − k 0 2 \gamma_{h0}=\sqrt{k_t^2-k_0^2} γ h 0 = k t 2 − k 0 2 on the bound fundamental branch.
Every order samples the same local sheet dyadic through its own TE polarization. Specifically,
e ^ a = e ^ T E ( q a ) = z ^ × q a ∣ q a ∣ = 1 ∣ q a ∣ ( − q a , y q a , x ) . (371) \widehat{\mathbf{e}}_a=\widehat{\mathbf{e}}_{\mathrm{TE}}(\mathbf{q}_a)
=\frac{\widehat{\mathbf{z}}\times\mathbf{q}_a}{|\mathbf{q}_a|}
=\frac{1}{|\mathbf{q}_a|}
\begin{pmatrix}
-q_{a,y}\\
q_{a,x}
\end{pmatrix}. \tag*{(371)} e a = e TE ( q a ) = ∣ q a ∣ z × q a = ∣ q a ∣ 1 ( − q a , y q a , x ) . (371)
and
A ‾ = C ‾ e q v 2 ε 0 , A a = e ^ a ⋅ A ‾ ⋅ e ^ a . (372) \overline{\mathbf{A}}=\frac{\overline{\mathbf{C}}_{\mathrm{eqv}}}{2\varepsilon_0},\qquad A_a=\widehat{\mathbf{e}}_a\cdot\overline{\mathbf{A}}\cdot\widehat{\mathbf{e}}_a. \tag*{(372)} A = 2 ε 0 C eqv , A a = e a ⋅ A ⋅ e a . (372)
Because A ‾ \overline{\mathbf{A}} A is real symmetric and positive definite, A a > 0 A_a>0 A a > 0 for every nonzero q a \mathbf{q}_a q a . At Γ \Gamma Γ , the direction of e ^ 0 \widehat{\mathbf{e}}_0 e 0 is defined by the one-sided path limit
e ^ 0 ( Γ ) = lim s → 0 z ^ × k t ( s ) ∣ k t ( s ) ∣ = z ^ × K ^ . (373) \widehat{\mathbf{e}}_0(\Gamma)=\lim_{s\to0}\frac{\widehat{\mathbf{z}}\times\mathbf{k}_t(s)}{\lvert\mathbf{k}_t(s)\rvert}=\widehat{\mathbf{z}}\times\widehat{\mathbf{K}}. \tag*{(373)} e 0 ( Γ ) = s → 0 lim ∣ k t ( s )∣ z × k t ( s ) = z × K . (373)
This removes the apparent 0/0 ambiguity without assigning an arbitrary polarization at Γ \Gamma Γ .
Let Γ T L ( k 0 , A ) \Gamma_{\mathrm{TL}}(k_0,A) Γ TL ( k 0 , A ) denote the sheet-loaded decay obtained from the local transverse-resonance condition in the main text. With the Luukkonen obliquity factor used there, it is the positive solution of
A Γ 2 + 2 Γ − A k 0 2 = 0. (374) A\Gamma^2+2\Gamma-Ak_0^2=0. \tag*{(374)} A Γ 2 + 2Γ − A k 0 2 = 0. (374)
For A > 0 A>0 A > 0 , the two algebraic roots are
Γ ± = − 1 ± 1 + ( k 0 A ) 2 A . (375) \Gamma_{\pm}=\frac{-1\pm\sqrt{1+(k_0A)^2}}{A}. \tag*{(375)} Γ ± = A − 1 ± 1 + ( k 0 A ) 2 . (375)
The decaying bound branch requires Γ > 0 \Gamma>0 Γ > 0 , and therefore
Γ T L ( k 0 , A ) = 1 + ( k 0 A ) 2 − 1 A = k 0 2 A 1 + 1 + ( k 0 A ) 2 . (376) \Gamma_{\mathrm{TL}}(k_0,A)=\frac{\sqrt{1+(k_0A)^2}-1}{A}=\frac{k_0^2A}{1+\sqrt{1+(k_0A)^2}}. \tag*{(376)} Γ TL ( k 0 , A ) = A 1 + ( k 0 A ) 2 − 1 = 1 + 1 + ( k 0 A ) 2 k 0 2 A . (376)
The second form follows by rationalizing the numerator and is numerically stable when k 0 A ≪ 1 k_0A\ll1 k 0 A ≪ 1 . In particular,
Γ T L ( k 0 , A ) = k 0 2 A 2 − k 0 4 A 3 8 + O ( k 0 6 A 5 ) . (377) \Gamma_{\mathrm{TL}}(k_0,A)=\frac{k_0^2A}{2}-\frac{k_0^4A^3}{8}+O(k_0^6A^5). \tag*{(377)} Γ TL ( k 0 , A ) = 2 k 0 2 A − 8 k 0 4 A 3 + O ( k 0 6 A 5 ) . (377)
We henceforth use
Γ 0 ≡ Γ T L ( k 0 , A 0 ) , Γ p ≡ Γ T L ( k 0 , A p ) , p = 1 , … , 6. (378) \Gamma_0\equiv\Gamma_{\mathrm{TL}}(k_0,A_0),\qquad\Gamma_p\equiv\Gamma_{\mathrm{TL}}(k_0,A_p),\qquad p=1,\ldots,6. \tag*{(378)} Γ 0 ≡ Γ TL ( k 0 , A 0 ) , Γ p ≡ Γ TL ( k 0 , A p ) , p = 1 , … , 6. (378)
The local fundamental residual is
R 0 ( k 0 , k t ) ≡ γ h 0 − Γ 0 = k t 2 − k 0 2 − Γ 0 . (379) R_0(k_0,\mathbf{k}_t)\equiv\gamma_{h0}-\Gamma_0=\sqrt{k_t^2-k_0^2}-\Gamma_0. \tag*{(379)} R 0 ( k 0 , k t ) ≡ γ h 0 − Γ 0 = k t 2 − k 0 2 − Γ 0 . (379)
Exact complex boundary Fourier overlap
After unioning the metallic tessellation cells and clipping the result to the unit cell, the metal–gap interface is reconstructed as a collection of closed polygonal loops. Duplicate consecutive vertices and vertices lying on a straight continuation are removed. The resulting set of open straight edges is denoted by E ( ∂ Σ ) \mathcal{E}(\partial\Sigma) E ( ∂ Σ ) . For each edge e ∈ E ( ∂ Σ ) e \in\mathcal{E}(\partial\Sigma) e ∈ E ( ∂ Σ ) , let r e − , r e + \mathbf{r}_e^{-}, \mathbf{r}_e^{+} r e − , r e + be its endpoints,
d e = r e + − r e − , L e = ∣ d e ∣ , (380) \mathbf{d}_e=\mathbf{r}_e^{+}-\mathbf{r}_e^{-}, \qquad L_e=\lvert\mathbf{d}_e\rvert, \tag*{(380)} d e = r e + − r e − , L e = ∣ d e ∣ , (380)
and define
t ^ e = d e L e , n ^ e = ( t e , y − t e , x ) . (381) \hat{\mathbf{t}}_e=\frac{\mathbf{d}_e}{L_e}, \qquad\hat{\mathbf{n}}_e=\begin{pmatrix} t_{e,y} \\ -t_{e,x} \end{pmatrix}. \tag*{(381)} t ^ e = L e d e , n ^ e = ( t e , y − t e , x ) . (381)
The normal is constant on the interior of a polygon edge. Reversing the orientation of a complete loop changes both factors of n ^ e \hat{\mathbf{n}}_e n ^ e below and consequently leaves the overlap unchanged.
The electromagnetic field behaviour near conducting edges and the associated concentration of induced charge motivate a boundary-based description of the capacitive response [50 , 58 ] . Within the present reduced model, the normal projections specify geometric weights for the assumed edge-charge profiles.
Let ℓ ∈ [ 0 , L ] \ell\in[0,L] ℓ ∈ [ 0 , L ] parameterize the reconstructed boundary, where L L L is its total length and r ( ℓ ) ∈ ∂ Σ \mathbf{r}(\ell) \in\partial\Sigma r ( ℓ ) ∈ ∂ Σ . For several boundary loops, their arclength parameterizations are concatenated in a fixed order. Consistently with the main text, we define the dimensionless model profiles
ρ ~ a ( ℓ ; k t ) = [ e ^ a ⋅ n ^ ( ℓ ) ] e − i ( k t + G a ) ⋅ r ( ℓ ) , a = 0 , … , 6 , (382) \widetilde{\rho}_{a}(\ell;\mathbf{k}_t)=\left[\hat{\mathbf{e}}_a\cdot\hat{\mathbf{n}}(\ell)\right]e^{-i(\mathbf{k}_t+\mathbf{G}_a)\cdot\mathbf{r}(\ell)}, \qquad a=0,\ldots,6, \tag*{(382)} ρ a ( ℓ ; k t ) = [ e ^ a ⋅ n ^ ( ℓ ) ] e − i ( k t + G a ) ⋅ r ( ℓ ) , a = 0 , … , 6 , (382)
where G 0 = 0 \mathbf{G}_0=\mathbf{0} G 0 = 0 and G p \mathbf{G}_p G p , p = 1 , … , 6 p=1,\ldots,6 p = 1 , … , 6 , are the retained first-shell reciprocal vectors. Here n ^ \hat{\mathbf{n}} n ^ is the in-plane boundary normal, e ^ 0 ≡ e ^ T E \hat{\mathbf{e}}_0\equiv\hat{\mathbf{e}}_{\mathrm{TE}} e ^ 0 ≡ e ^ TE , and e ^ p \hat{\mathbf{e}}_p e ^ p is the TE polarization relative to k t + G p \mathbf{k}_t+\mathbf{G}_p k t + G p . The polarization vectors and k t \mathbf{k}_t k t are real. The modal amplitudes are factored out; the profile shapes constitute the adopted edge-projection approximation.
The full-boundary cross moment between the fundamental order and sideband p p p is
B 0 p = ∫ 0 L ρ ~ 0 ∗ ( ℓ ; k t ) ρ ~ p ( ℓ ; k t ) d ℓ = ∮ ∂ Σ ( e ^ 0 ⋅ n ^ ) ( e ^ p ⋅ n ^ ) e − i G p ⋅ r d ℓ , (383) \begin{aligned}
B_{0p}=\int_{0}^{L}\widetilde{\rho}_0^{*}(\ell;\mathbf{k}_t)\widetilde{\rho}_p(\ell;\mathbf{k}_t)\,\mathrm{d}\ell\\
&=\oint_{\partial\Sigma}(\hat{\mathbf{e}}_0\cdot\hat{\mathbf{n}})(\hat{\mathbf{e}}_p\cdot\hat{\mathbf{n}})e^{-i\mathbf{G}_p\cdot\mathbf{r}}\,\mathrm{d}\ell, \tag*{(383)}
\end{aligned} B 0 p = ∫ 0 L ρ 0 ∗ ( ℓ ; k t ) ρ p ( ℓ ; k t ) d ℓ = ∮ ∂ Σ ( e ^ 0 ⋅ n ^ ) ( e ^ p ⋅ n ^ ) e − i G p ⋅ r d ℓ , (383)
and the squared boundary-profile norms are
N a = ∫ 0 L ∣ ρ ~ a ( ℓ ; k t ) ∣ 2 d ℓ = ∮ ∂ Σ ( e ^ a ⋅ n ^ ) 2 d ℓ , a ∈ { 0 , p } . (384) \begin{aligned}
N_a=\int_{0}^{L}\left\lvert\widetilde{\rho}_a(\ell;\mathbf{k}_t)\right\rvert^2\,\mathrm{d}\ell\\
&=\oint_{\partial\Sigma}(\hat{\mathbf{e}}_a\cdot\hat{\mathbf{n}})^2\,\mathrm{d}\ell,\qquad a\in\{0,p\}. \tag*{(384)}
\end{aligned} N a = ∫ 0 L ∣ ρ a ( ℓ ; k t ) ∣ 2 d ℓ = ∮ ∂ Σ ( e ^ a ⋅ n ^ ) 2 d ℓ , a ∈ { 0 , p } . (384)
The common Bloch phase cancels in the cross moment, leaving the reciprocal phase e − i G p ⋅ r e^{-i\mathbf{G}_p\cdot\mathbf{r}} e − i G p ⋅ r . The phase factors do not contribute to the squared norms.
For nonzero profile norms, the normalized overlap is
η 0 p ( k t ) = B 0 p N 0 N p , (385) \eta_{0p}(\mathbf{k}_t)=\frac{B_{0p}}{\sqrt{N_0N_p}}, \tag*{(385)} η 0 p ( k t ) = N 0 N p B 0 p , (385)
which measures the complex correlation between the two assumed charge-profile shapes. The full-boundary quantities B 0 p B_{0p} B 0 p and N a N_a N a correspond to B 0 p ( L ) B_{0p}(L) B 0 p ( L ) and N a ( L ) N_a(L) N a ( L ) in the cumulative arclength notation of the main text.
Floquet orthogonality and boundary-weighted overlap
We distinguish the orthogonality of Floquet harmonics over a complete primitive cell from the boundary-weighted overlap used in the present coupling model.
For a real tangential Bloch wavevector k t \mathbf{k}_t k t , the fundamental and sideband harmonic functions at the sheet plane are
ψ 0 ( r ) = e ^ 0 e − i k t ⋅ r , ψ p ( r ) = e ^ p e − i ( k t + G p ) ⋅ r . (386) \psi_0(\mathbf{r})=\hat{\mathbf{e}}_0e^{-i\mathbf{k}_t\cdot\mathbf{r}}, \qquad\psi_p(\mathbf{r})=\hat{\mathbf{e}}_pe^{-i(\mathbf{k}_t+\mathbf{G}_p)\cdot\mathbf{r}}. \tag*{(386)} ψ 0 ( r ) = e ^ 0 e − i k t ⋅ r , ψ p ( r ) = e ^ p e − i ( k t + G p ) ⋅ r . (386)
Here the TE polarization vectors are real. Their unweighted area inner product over a complete primitive cell U U U is
⟨ ψ 0 , ψ p ⟩ U = ∫ U ψ 0 ∗ ( r ) ⋅ ψ p ( r ) d A = ( e ^ 0 ⋅ e ^ p ) ∫ U e − i G p ⋅ r d A . (387) \begin{aligned}
\langle\psi_0,\psi_p\rangle_U=\int_U \psi_0^*(\mathbf r)\cdot\psi_p(\mathbf r)\,\mathrm{d}A \\
&=(\widehat{\mathbf e}_0\cdot\widehat{\mathbf e}_p)\int_U e^{-i\mathbf G_p\cdot\mathbf r}\,\mathrm{d}A. \tag*{(387)}
\end{aligned} ⟨ ψ 0 , ψ p ⟩ U = ∫ U ψ 0 ∗ ( r ) ⋅ ψ p ( r ) d A = ( e 0 ⋅ e p ) ∫ U e − i G p ⋅ r d A . (387)
The common Bloch phase cancels in this product.
Let a 1 , a 2 \mathbf a_1,\mathbf a_2 a 1 , a 2 be the primitive lattice vectors and b 1 , b 2 \mathbf b_1,\mathbf b_2 b 1 , b 2 their reciprocal vectors, with a i ⋅ b j = 2 π δ i j \mathbf a_i\cdot\mathbf b_j=2\pi\delta_{ij} a i ⋅ b j = 2 π δ ij . We write
G p = m 1 b 1 + m 2 b 2 , r = t 1 a 1 + t 2 a 2 , 0 ≤ t 1 , t 2 < 1 , (388) \mathbf G_p=m_1\mathbf b_1+m_2\mathbf b_2,\qquad\mathbf r=t_1\mathbf a_1+t_2\mathbf a_2,\qquad0\le t_1,t_2<1, \tag*{(388)} G p = m 1 b 1 + m 2 b 2 , r = t 1 a 1 + t 2 a 2 , 0 ≤ t 1 , t 2 < 1 , (388)
where m 1 , m 2 m_1,m_2 m 1 , m 2 are integers. Consequently,
∫ U e − i G p ⋅ r d A = ∣ U ∣ ∫ 0 1 e − i 2 π m 1 t 1 d t 1 ∫ 0 1 e − i 2 π m 2 t 2 d t 2 = ∣ U ∣ δ m 1 , 0 δ m 2 , 0 . (389) \begin{aligned}
\int_U e^{-i\mathbf G_p\cdot\mathbf r}\,\mathrm{d}A=|U|\int_0^1e^{-i2\pi m_1t_1}\,\mathrm{d}t_1\int_0^1e^{-i2\pi m_2t_2}\,\mathrm{d}t_2 \\
&=|U|\delta_{m_1,0}\delta_{m_2,0}. \tag*{(389)}
\end{aligned} ∫ U e − i G p ⋅ r d A = ∣ U ∣ ∫ 0 1 e − i 2 π m 1 t 1 d t 1 ∫ 0 1 e − i 2 π m 2 t 2 d t 2 = ∣ U ∣ δ m 1 , 0 δ m 2 , 0 . (389)
For every nonzero integer m m m ,
∫ 0 1 e − i 2 π m t d t = e − i 2 π m − 1 − i 2 π m = 0. (390) \int_0^1e^{-i2\pi mt}\,\mathrm{d}t=\frac{e^{-i2\pi m}-1}{-i2\pi m}=0. \tag*{(390)} ∫ 0 1 e − i 2 π m t d t = − i 2 π m e − i 2 π m − 1 = 0. (390)
Thus, for G p ≠ 0 \mathbf G_p\ne0 G p = 0 ,
⟨ ψ 0 , ψ p ⟩ U = 0. (391) \langle\psi_0,\psi_p\rangle_U=0. \tag*{(391)} ⟨ ψ 0 , ψ p ⟩ U = 0. (391)
Although this derivation uses a primitive parallelogram, the integrand is lattice-periodic, so the same result holds over the hexagonal Wigner–Seitz cell.
The boundary cross moment uses a different inner product. Let ℓ ∈ [ 0 , L ] \ell\in[0,L] ℓ ∈ [ 0 , L ] parameterize the reconstructed boundary, with r ( ℓ ) ∈ ∂ Σ \mathbf r(\ell)\in\partial\Sigma r ( ℓ ) ∈ ∂ Σ , and define the normal traces
g a ( ℓ ) = n ^ ( ℓ ) ⋅ ψ a ( r ( ℓ ) ) , a ∈ { 0 , p } . (392) g_a(\ell)=\widehat{\mathbf n}(\ell)\cdot\psi_a(\mathbf r(\ell)),\qquad a\in\{0,p\}. \tag*{(392)} g a ( ℓ ) = n ( ℓ ) ⋅ ψ a ( r ( ℓ )) , a ∈ { 0 , p } . (392)
These normal traces coincide with the dimensionless model edge-charge profiles defined above:
g a ( ℓ ) = ρ ~ a ( ℓ ; k t ) , a ∈ { 0 , p } . (393) g_a(\ell)=\widetilde{\rho}_{a}(\ell;\mathbf k_t),\qquad a\in\{0,p\}. \tag*{(393)} g a ( ℓ ) = ρ a ( ℓ ; k t ) , a ∈ { 0 , p } . (393)
Thus the boundary inner product used below is precisely the model charge-profile overlap introduced in the main text.
For several boundary loops, the following integrals denote the sum over their arclength parameterizations. Substituting Eq. (386) gives
B 0 p = ∫ 0 L g 0 ∗ ( ℓ ′ ) g p ( ℓ ′ ) d ℓ ′ = ∫ 0 L ( e ^ 0 ⋅ n ^ ( ℓ ′ ) ) ( e ^ p ⋅ n ^ ( ℓ ′ ) ) e − i G p ⋅ r ( ℓ ′ ) d ℓ ′ . (394) \begin{aligned}
B_{0p}=\int_0^L g_0^*(\ell')g_p(\ell')\,\mathrm{d}\ell' \\
&=\int_0^L\left(\widehat{\mathbf e}_0\cdot\widehat{\mathbf n}(\ell')\right)\left(\widehat{\mathbf e}_p\cdot\widehat{\mathbf n}(\ell')\right)e^{-i\mathbf G_p\cdot\mathbf r(\ell')}\,\mathrm{d}\ell'. \tag*{(394)}
\end{aligned} B 0 p = ∫ 0 L g 0 ∗ ( ℓ ′ ) g p ( ℓ ′ ) d ℓ ′ = ∫ 0 L ( e 0 ⋅ n ( ℓ ′ ) ) ( e p ⋅ n ( ℓ ′ ) ) e − i G p ⋅ r ( ℓ ′ ) d ℓ ′ . (394)
Likewise,
N a = ∫ 0 L ∣ g a ( ℓ ′ ) ∣ 2 d ℓ ′ , η 0 p = ⟨ g 0 , g p ⟩ ∂ Σ ∥ g 0 ∥ ∂ Σ ∥ g p ∥ ∂ Σ . (395) N_a=\int_0^L|g_a(\ell')|^2\,\mathrm{d}\ell',\qquad\eta_{0p}=\frac{\langle g_0,g_p\rangle_{\partial\Sigma}}{\|g_0\|_{\partial\Sigma}\|g_p\|_{\partial\Sigma}}. \tag*{(395)} N a = ∫ 0 L ∣ g a ( ℓ ′ ) ∣ 2 d ℓ ′ , η 0 p = ∥ g 0 ∥ ∂ Σ ∥ g p ∥ ∂ Σ ⟨ g 0 , g p ⟩ ∂ Σ . (395)
The boundary integral follows the metal–gap interface and contains its local normal projections. The complete-cell cancellation in Eq. (389) therefore does not apply to this integral.
To display the geometric weighting explicitly, introduce the boundary Fourier tensor
F ‾ ∂ Σ ( G ) = ∮ ∂ Σ n ^ n ^ T e − i G ⋅ r d ℓ . (396) \overline{\mathbf F}_{\partial\Sigma}(\mathbf G)=\oint_{\partial\Sigma}\widehat{\mathbf n}\widehat{\mathbf n}^{T}e^{-i\mathbf G\cdot\mathbf r}\,\mathrm{d}\ell. \tag*{(396)} F ∂ Σ ( G ) = ∮ ∂ Σ n n T e − i G ⋅ r d ℓ . (396)
Then
B 0 p = e ^ 0 T F ‾ ∂ Σ ( G p ) e ^ p . (397) B_{0p}=\widehat{\mathbf e}_0^{T}\overline{\mathbf F}_{\partial\Sigma}(\mathbf G_p)\widehat{\mathbf e}_p. \tag*{(397)} B 0 p = e 0 T F ∂ Σ ( G p ) e p . (397)
This tensor contains the spatial distribution and orientation of the metal edges. Its coefficients at nonzero reciprocal vectors need not vanish, although particular geometries, polarizations, or symmetry selection rules can make an individual contraction zero.
Orthogonality of the Floquet basis consequently does not require the patterned interaction operator to be diagonal in that basis. The fundamental and sideband harmonics can contribute to the same Bloch eigenmode through nonzero inter-order matrix elements.
Within the present reduced closure, η 0 p \eta_{0p} η 0 p is a boundary-weighted geometric measure used to approximate fundamental-to-sideband coupling. Its possible nonzero value is consistent with complete-cell Floquet orthogonality. The integral is evaluated exactly for the prescribed polygonal boundary, while its use as an electromagnetic coupling measure remains part of the adopted edge-projection approximation.
Analytical evaluation on polygon edges
For completeness, Eq. (383) can be evaluated analytically on every straight edge. Parameterize edge e e e by
r e ( u ) = r e − + u d e , 0 ≤ u ≤ 1 , d ℓ = L e d u . (398) \mathbf r_e(u)=\mathbf r_e^{-}+u\mathbf d_e,\qquad0\leq u\leq1,\qquad d\ell=L_e\,du. \tag*{(398)} r e ( u ) = r e − + u d e , 0 ≤ u ≤ 1 , d ℓ = L e d u . (398)
Writing x e p = G p ⋅ d e x_{ep}=\mathbf G_p\cdot\mathbf d_e x e p = G p ⋅ d e , its exponential integral is
I e ( G p ) = ∫ e e − i G p ⋅ r d ℓ = L e e − i G p ⋅ r e − ∫ 0 1 e − i u x e p d u = L e e − i G p ⋅ r e − 1 − e − i x e p i x e p . (399) \begin{aligned}
I_e(\mathbf G_p)&=\int_e e^{-i\mathbf G_p\cdot\mathbf r}\,d\ell\\
&=L_e e^{-i\mathbf G_p\cdot\mathbf r_e^-}\int_0^1e^{-iux_{ep}}\,du\\
&=L_e e^{-i\mathbf G_p\cdot\mathbf r_e^-}\frac{1-e^{-ix_{ep}}}{ix_{ep}}. \tag*{(399)}
\end{aligned} I e ( G p ) = ∫ e e − i G p ⋅ r d ℓ = L e e − i G p ⋅ r e − ∫ 0 1 e − i u x e p d u = L e e − i G p ⋅ r e − i x e p 1 − e − i x e p . (399)
Introducing the edge midpoint r e c = ( r e − + r e + ) / 2 \mathbf r_e^c=(\mathbf r_e^-+\mathbf r_e^+)/2 r e c = ( r e − + r e + ) /2 and using
1 − e − i x = 2 i e − i x / 2 sin ( x / 2 ) , (400) 1-e^{-ix}=2ie^{-ix/2}\sin(x/2), \tag*{(400)} 1 − e − i x = 2 i e − i x /2 sin ( x /2 ) , (400)
gives the equivalent midpoint form
I e ( G p ) = L e e − i G p ⋅ r e c sinc ( x e p 2 ) , sinc ( y ) = sin y y . (401) I_e(\mathbf G_p)=L_e e^{-i\mathbf G_p\cdot\mathbf r_e^c}\operatorname{sinc}\left(\frac{x_{ep}}{2}\right),\qquad\operatorname{sinc}(y)=\frac{\sin y}{y}. \tag*{(401)} I e ( G p ) = L e e − i G p ⋅ r e c sinc ( 2 x e p ) , sinc ( y ) = y sin y . (401)
The removable singularity is fixed by sinc ( 0 ) = 1 \operatorname{sinc}(0)=1 sinc ( 0 ) = 1 , so that I e ( G p ) = L e e − i G p ⋅ r e c I_e(\mathbf G_p)=L_e e^{-i\mathbf G_p\cdot\mathbf r_e^c} I e ( G p ) = L e e − i G p ⋅ r e c whenever G p ⋅ d e = 0 \mathbf G_p\cdot\mathbf d_e=0 G p ⋅ d e = 0 . This form avoids numerical quadrature and the cancellation that would otherwise occur when ∣ x e p ∣ ≪ 1 \lvert x_{ep}\rvert\ll1 ∣ x e p ∣ ≪ 1 .
Since the polarization and normal vectors are constant on each edge, Eqs. (383) and (384) reduce to the finite sums
B 0 p = ∑ e ∈ E ( ∂ Σ ) ( e ^ 0 ⋅ n ^ e ) ( e ^ p ⋅ n ^ e ) I e ( G p ) , (402) B_{0p}=\sum_{e\in\mathcal E(\partial\Sigma)}(\widehat{\mathbf e}_0\cdot\widehat{\mathbf n}_e)(\widehat{\mathbf e}_p\cdot\widehat{\mathbf n}_e)I_e(\mathbf G_p), \tag*{(402)} B 0 p = e ∈ E ( ∂ Σ ) ∑ ( e 0 ⋅ n e ) ( e p ⋅ n e ) I e ( G p ) , (402)
N a = ∑ e ∈ E ( ∂ Σ ) L e ( e ^ a ⋅ n ^ e ) 2 . (403) N_a=\sum_{e\in\mathcal E(\partial\Sigma)}L_e(\widehat{\mathbf e}_a\cdot\widehat{\mathbf n}_e)^2. \tag*{(403)} N a = e ∈ E ( ∂ Σ ) ∑ L e ( e a ⋅ n e ) 2 . (403)
The normalized contribution of a single edge is therefore
d η 0 p ( e ) = ( e ^ 0 ⋅ n ^ e ) ( e ^ p ⋅ n ^ e ) I e ( G p ) N 0 N p , η 0 p = ∑ e d η 0 p ( e ) . (404) d\eta_{0p}^{(e)}=\frac{(\widehat{\mathbf e}_0\cdot\widehat{\mathbf n}_e)(\widehat{\mathbf e}_p\cdot\widehat{\mathbf n}_e)I_e(\mathbf G_p)}{\sqrt{N_0N_p}},\qquad\eta_{0p}=\sum_e d\eta_{0p}^{(e)}. \tag*{(404)} d η 0 p ( e ) = N 0 N p ( e 0 ⋅ n e ) ( e p ⋅ n e ) I e ( G p ) , η 0 p = e ∑ d η 0 p ( e ) . (404)
This is the quantity whose phase colors the boundary in the six overlap maps of the main text. If the edges are traversed in the order e 1 , e 2 , … , e N e e_1,e_2,\ldots,e_{N_e} e 1 , e 2 , … , e N e , the displayed cumulative phasor is
z j ( p ) = ∑ n = 1 j d η 0 p ( e n ) , z 0 ( p ) = 0 , z N e ( p ) = η 0 p . (405) z_j^{(p)}=\sum_{n=1}^{j}d\eta_{0p}^{(e_n)},\qquad z_0^{(p)}=0,\qquad z_{N_e}^{(p)}=\eta_{0p}. \tag*{(405)} z j ( p ) = n = 1 ∑ j d η 0 p ( e n ) , z 0 ( p ) = 0 , z N e ( p ) = η 0 p . (405)
The bound ∣ η 0 p ∣ ≤ 1 |\eta_{0p}| \le1 ∣ η 0 p ∣ ≤ 1 follows directly from Cauchy–Schwarz. Indeed, define the boundary functions
f 0 ( r ) = e ^ 0 ⋅ n ^ ( r ) , f p ( r ) = ( e ^ p ⋅ n ^ ( r ) ) e − i G p ⋅ r . (406) f_0(\mathbf{r}) = \hat{\mathbf{e}}_0 \cdot\hat{\mathbf{n}}(\mathbf{r}), \qquad f_p(\mathbf{r}) = \left(\hat{\mathbf{e}}_p \cdot\hat{\mathbf{n}}(\mathbf{r})\right)e^{-i\mathbf{G}_p \cdot\mathbf{r}}. \tag*{(406)} f 0 ( r ) = e ^ 0 ⋅ n ^ ( r ) , f p ( r ) = ( e ^ p ⋅ n ^ ( r ) ) e − i G p ⋅ r . (406)
Because ∣ e − i G p ⋅ r ∣ = 1 |e^{-i\mathbf{G}_p \cdot\mathbf{r}}| = 1 ∣ e − i G p ⋅ r ∣ = 1 ,
∥ f 0 ∥ L 2 ( ∂ Σ ) = N 0 , ∥ f p ∥ L 2 ( ∂ Σ ) = N p , B 0 p = ⟨ f 0 , f p ⟩ . (407) \|f_0\|_{L^2(\partial\Sigma)} = N_0, \qquad\|f_p\|_{L^2(\partial\Sigma)} = N_p, \qquad B_{0p} = \langle f_0, f_p\rangle. \tag*{(407)} ∥ f 0 ∥ L 2 ( ∂ Σ ) = N 0 , ∥ f p ∥ L 2 ( ∂ Σ ) = N p , B 0 p = ⟨ f 0 , f p ⟩ . (407)
Hence
∣ B 0 p ∣ 2 ≤ ∥ f 0 ∥ L 2 2 ∥ f p ∥ L 2 2 = N 0 N p , ∣ η 0 p ∣ ≤ 1. (408) |B_{0p}|^2 \le\|f_0\|^2_{L^2}\|f_p\|^2_{L^2} = N_0N_p, \qquad|\eta_{0p}| \le1. \tag*{(408)} ∣ B 0 p ∣ 2 ≤ ∥ f 0 ∥ L 2 2 ∥ f p ∥ L 2 2 = N 0 N p , ∣ η 0 p ∣ ≤ 1. (408)
The phase in Eq. (385) also records the choice of spatial origin. If the complete patch is translated by d \mathbf{d} d , then the change of variables r ′ = r + d \mathbf{r}' = \mathbf{r} + \mathbf{d} r ′ = r + d gives
B 0 p [ Σ + d ] = e − i G p ⋅ d B 0 p [ Σ ] , N a [ Σ + d ] = N a [ Σ ] , (409) B_{0p}[\Sigma+ \mathbf{d}] = e^{-i\mathbf{G}_p \cdot\mathbf{d}}B_{0p}[\Sigma], \qquad N_a[\Sigma+ \mathbf{d}] = N_a[\Sigma], \tag*{(409)} B 0 p [ Σ + d ] = e − i G p ⋅ d B 0 p [ Σ ] , N a [ Σ + d ] = N a [ Σ ] , (409)
and consequently
η 0 p [ Σ + d ] = e − i G p ⋅ d η 0 p [ Σ ] . (410) \eta_{0p}[\Sigma+ \mathbf{d}] = e^{-i\mathbf{G}_p \cdot\mathbf{d}}\eta_{0p}[\Sigma]. \tag*{(410)} η 0 p [ Σ + d ] = e − i G p ⋅ d η 0 p [ Σ ] . (410)
Its magnitude is translation invariant, but its phase is not. The patch is therefore centered at the symmetry origin before the mirror-related sidebands are combined. The complex edge contributions are summed before taking the modulus, retaining constructive and destructive interference along the boundary in each overlap. The resulting complex overlaps are used to construct the symmetry-adapted coupling in Sec. S6E.
Directional turning and the six geometry-derived multipliers
Let r v \mathbf{r}_v r v be a vertex of a reconstructed polygonal boundary loop, with adjacent vertices r v − 1 \mathbf{r}_{v-1} r v − 1 and r v + 1 \mathbf{r}_{v+1} r v + 1 . The one-sided unit tangents are
t ^ v − = r v − r v − 1 ∣ r v − r v − 1 ∣ , t ^ v + = r v + 1 − r v ∣ r v + 1 − r v ∣ , (411) \hat{\mathbf{t}}_v^- = \frac{\mathbf{r}_v-\mathbf{r}_{v-1}}{|\mathbf{r}_v-\mathbf{r}_{v-1}|}, \qquad\hat{\mathbf{t}}_v^+ = \frac{\mathbf{r}_{v+1}-\mathbf{r}_v}{|\mathbf{r}_{v+1}-\mathbf{r}_v|}, \tag*{(411)} t ^ v − = ∣ r v − r v − 1 ∣ r v − r v − 1 , t ^ v + = ∣ r v + 1 − r v ∣ r v + 1 − r v , (411)
and the corresponding one-sided normals can be taken as
n ^ v ± = ( t ^ v , y ± − t ^ v , x ± ) . (412) \hat{\mathbf{n}}_v^\pm= \begin{pmatrix} \hat{t}_{v,y}^\pm\\ -\hat{t}_{v,x}^\pm\end{pmatrix}. \tag*{(412)} n ^ v ± = ( t ^ v , y ± − t ^ v , x ± ) . (412)
The exterior signed turning from the incoming to the outgoing edge is
Δ θ v = atan2 [ z ^ ⋅ ( t ^ v − × t ^ v + ) , t ^ v − ⋅ t ^ v + ] , − π < Δ θ v ≤ π . (413) \Delta\theta_v = \operatorname{atan2}\left[\hat{\mathbf{z}}\cdot\left(\hat{\mathbf{t}}_v^- \times\hat{\mathbf{t}}_v^+\right),\hat{\mathbf{t}}_v^- \cdot\hat{\mathbf{t}}_v^+\right], \qquad-\pi< \Delta\theta_v \le\pi. \tag*{(413)} Δ θ v = atan2 [ z ^ ⋅ ( t ^ v − × t ^ v + ) , t ^ v − ⋅ t ^ v + ] , − π < Δ θ v ≤ π . (413)
For a counterclockwise-oriented simple closed polygonal loop, the signed turning theorem gives ∑ v Δ θ v = 2 π \sum_v \Delta\theta_v = 2\pi ∑ v Δ θ v = 2 π . Moreover, ∑ v ∣ Δ θ v ∣ \sum_v |\Delta\theta_v| ∑ v ∣Δ θ v ∣ is the polygonal total curvature of the loop. Milnor’s total-curvature theorem states that this quantity is at least 2 π 2\pi 2 π , with equality if and only if the loop is convex [68 ] . Hence, for a simple convex loop,
∑ v Δ θ v = 2 π , ∑ v ∣ Δ θ v ∣ = 2 π . (414) \sum_v \Delta\theta_v = 2\pi, \qquad\sum_v |\Delta\theta_v| = 2\pi. \tag*{(414)} v ∑ Δ θ v = 2 π , v ∑ ∣Δ θ v ∣ = 2 π . (414)
For a nonconvex or multicomponent boundary, positive and negative signed turns can cancel in the first sum but not in the second. The normalised total absolute turning is therefore defined by
T a b s = 1 2 π ∑ v ∣ Δ θ v ∣ . (415) T_{\mathrm{abs}} = \frac{1}{2\pi}\sum_v |\Delta\theta_v|. \tag*{(415)} T abs = 2 π 1 v ∑ ∣Δ θ v ∣. (415)
Consequently, T a b s ≥ 1 T_{\mathrm{abs}} \ge1 T abs ≥ 1 for a single simple closed boundary loop, with equality if and only if the loop is convex. Additional boundary components contribute their own positive total curvature. Unlike the signed rotation index, T a b s T_{\mathrm{abs}} T abs is a shape-sensitive geometric quantity rather than a topological invariant.
The turning at a vertex contributes to sideband p p p only to the extent that both the fundamental and sideband TE directions load its adjacent edges. For a ∈ { 0 , p } a \in\{0,p\} a ∈ { 0 , p } , define
S a ( v ; k t ) = 1 2 [ ( e ^ a ⋅ n ^ v ) 2 + ( e ^ a ⋅ n ^ v + ) 2 ] . (416) S_a(\mathbf{v};\mathbf{k}_t)=\frac{1}{2}\left[(\hat{\mathbf{e}}_a\cdot\hat{\mathbf{n}}_v)^2+(\hat{\mathbf{e}}_a\cdot\hat{\mathbf{n}}_v^{+})^2\right]. \tag*{(416)} S a ( v ; k t ) = 2 1 [ ( e ^ a ⋅ n ^ v ) 2 + ( e ^ a ⋅ n ^ v + ) 2 ] . (416)
Since all vectors entering a scalar product are unit vectors,
0 ≤ S a ( v ; k t ) ≤ 1. (417) 0\le S_a(\mathbf{v};\mathbf{k}_t)\le1. \tag*{(417)} 0 ≤ S a ( v ; k t ) ≤ 1. (417)
The symmetric geometric-mean participation of the two TE directions is
w p ( v ; k t ) = 2 S 0 ( v ; k t ) S p ( v ; k t ) . (418) w_p(\mathbf{v};\mathbf{k}_t)=2\sqrt{S_0(\mathbf{v};\mathbf{k}_t)S_p(\mathbf{v};\mathbf{k}_t)}. \tag*{(418)} w p ( v ; k t ) = 2 S 0 ( v ; k t ) S p ( v ; k t ) . (418)
It vanishes if either TE direction is tangent to both adjacent edges, it is unchanged by reversing a boundary normal, and it treats the two coupled orders symmetrically. The vertex contribution displayed by the color and size of the markers in the main-text turning maps is
c v ( p ) ( k t ) = ∣ Δ θ v ∣ 2 π w p ( v ; k t ) = ∣ Δ θ v ∣ 2 π 2 S 0 ( v ; k t ) S p ( v ; k t ) . (419) c_v^{(p)}(\mathbf{k}_t)=\frac{|\Delta\theta_v|}{2\pi}w_p(\mathbf{v};\mathbf{k}_t)=\frac{|\Delta\theta_v|}{2\pi}2\sqrt{S_0(\mathbf{v};\mathbf{k}_t)S_p(\mathbf{v};\mathbf{k}_t)}. \tag*{(419)} c v ( p ) ( k t ) = 2 π ∣Δ θ v ∣ w p ( v ; k t ) = 2 π ∣Δ θ v ∣ 2 S 0 ( v ; k t ) S p ( v ; k t ) . (419)
Summing all vertex contributions gives the raw directional turning
T p r a w ( k t ) = ∑ v c v ( p ) ( k t ) = 1 2 π ∑ v ∣ Δ θ v ∣ 2 S 0 ( v ; k t ) S p ( v ; k t ) . (420) T_p^{\mathrm{raw}}(\mathbf{k}_t)=\sum_v c_v^{(p)}(\mathbf{k}_t)=\frac{1}{2\pi}\sum_v|\Delta\theta_v|2\sqrt{S_0(\mathbf{v};\mathbf{k}_t)S_p(\mathbf{v};\mathbf{k}_t)}. \tag*{(420)} T p raw ( k t ) = v ∑ c v ( p ) ( k t ) = 2 π 1 v ∑ ∣Δ θ v ∣2 S 0 ( v ; k t ) S p ( v ; k t ) . (420)
The six raw values distribute the boundary turning anisotropically, but their mean can vary with the instantaneous set of TE directions. We therefore define
⟨ T r a w ⟩ 6 = 1 6 ∑ r = 1 6 T r r a w (421) \langle T^{\mathrm{raw}}\rangle_6=\frac{1}{6}\sum_{r=1}^{6}T_r^{\mathrm{raw}} \tag*{(421)} ⟨ T raw ⟩ 6 = 6 1 r = 1 ∑ 6 T r raw (421)
and, for ⟨ T r a w ( k t ) ⟩ 6 > 0 \langle T^{\mathrm{raw}}(\mathbf{k}_t)\rangle_6>0 ⟨ T raw ( k t ) ⟩ 6 > 0 , use the mean-preserving normalization
T p ( k t ) = T a b s T p r a w ( k t ) ⟨ T r a w ( k t ) ⟩ 6 . (422) T_p(\mathbf{k}_t)=T_{\mathrm{abs}}\frac{T_p^{\mathrm{raw}}(\mathbf{k}_t)}{\langle T^{\mathrm{raw}}(\mathbf{k}_t)\rangle_6}. \tag*{(422)} T p ( k t ) = T abs ⟨ T raw ( k t ) ⟩ 6 T p raw ( k t ) . (422)
The normalization fixes the shell-average scale exactly, because
1 6 ∑ p = 1 6 T p = 1 6 ∑ p = 1 6 T a b s T p r a w ⟨ T r a w ⟩ 6 = T a b s ⟨ T r a w ⟩ 6 ( 1 6 ∑ p = 1 6 T p r a w ) = T a b s . (423) \begin{aligned}
\frac{1}{6}\sum_{p=1}^{6}T_p
&=\frac{1}{6}\sum_{p=1}^{6}T_{\mathrm{abs}}\frac{T_p^{\mathrm{raw}}}{\langle T^{\mathrm{raw}}\rangle_6}\\
&=\frac{T_{\mathrm{abs}}}{\langle T^{\mathrm{raw}}\rangle_6}\left(\frac{1}{6}\sum_{p=1}^{6}T_p^{\mathrm{raw}}\right)\\
&=T_{\mathrm{abs}}. \tag*{(423)}
\end{aligned} 6 1 p = 1 ∑ 6 T p = 6 1 p = 1 ∑ 6 T abs ⟨ T raw ⟩ 6 T p raw = ⟨ T raw ⟩ 6 T abs ( 6 1 p = 1 ∑ 6 T p raw ) = T abs . (423)
Hence T a b s T_{\mathrm{abs}} T abs fixes the total geometric-complexity scale, while T p r a w / ⟨ T r a w ⟩ 6 T_p^{\mathrm{raw}}/\langle T^{\mathrm{raw}}\rangle_6 T p raw / ⟨ T raw ⟩ 6 assigns that scale among the six reciprocal directions.
The pointwise amplitude multiplier is
m ~ p ( k t ) = 2 T p ( k t ) . (424) \widetilde{m}_p(\mathbf{k}_t)=\sqrt{2T_p(\mathbf{k}_t)}. \tag*{(424)} m p ( k t ) = 2 T p ( k t ) . (424)
The square root converts the normalized participation weight into a coupling-amplitude multiplier. The factor of two is the fixed normalization adopted for the present six-order closure. Together with Eq. (S350), it gives
1 6 ∑ p = 1 6 m ~ p 2 ( k t ) = 2 T a b s . (425) \frac{1}{6}\sum_{p=1}^{6}\widetilde{m}_p^2(\mathbf{k}_t)=2T_{\mathrm{abs}}. \tag*{(425)} 6 1 p = 1 ∑ 6 m p 2 ( k t ) = 2 T abs . (425)
This normalization is applied once through Eq. (424) ; the six-order sum and subsequent mirror projection use these amplitudes directly. No adjustable scale multiplies the turning multiplier.
For the dispersion calculation along k t ( s ) = s K \mathbf{k}_t(s)=s\mathbf{K} k t ( s ) = s K , the directional multiplier m ~ p [ k t ( s ) ] \widetilde{m}_p[\mathbf{k}_t(s)] m p [ k t ( s )] is evaluated at every sampled Bloch wavevector and used directly in the fundamental-to-sideband coupling. The turning multiplier and boundary Fourier overlap η 0 p [ k t ( s ) ] \eta_{0p}[\mathbf{k}_t(s)] η 0 p [ k t ( s )] are evaluated at the same wavevector throughout the complete Γ \Gamma Γ –K K K path.
The reconstructed geometry remains fixed. The wavevector dependence of the turning multipliers arises from the TE polarization projections entering the vertex participation factors. The shell-mean normalization therefore holds pointwise along the path. At Γ \Gamma Γ , the fundamental TE polarization and the resulting multipliers are defined by the one-sided limit s → 0 + s\to0^{+} s → 0 + along the selected propagation direction.
For the representative unit cell used in the main-text edgewise maps, the six pointwise multipliers are shown in Figure S3 . The point s = 0.82 s=0.82 s = 0.82 identifies the representative wavevector used for those visualizations; the dispersion calculation uses the corresponding pointwise quantities at every sampled wavevector.
Figure S3. Directional turning multipliers for the representative unit cell used in the main-text edgewise maps, along k t ( s ) = s K \mathbf{k}_t(s)=s\mathbf{K} k t ( s ) = s K . Solid and dashed curves show the pointwise quantities m ~ p [ k t ( s ) ] \widetilde{m}_p[\mathbf{k}_t(s)] m p [ k t ( s )] for the positive and negative first-shell orders, respectively. For an optional wavevector-independent approximation, one constant per order may be defined by m ‾ p = median s ∈ [ 1 / 2 , 1 ] m ~ p [ k t ( s ) ] \overline{m}_p=\operatorname{median}_{s\in[1/2,1]}\widetilde{m}_p[\mathbf{k}_t(s)] m p = median s ∈ [ 1/2 , 1 ] m p [ k t ( s )] . The shaded region marks this prescribed sampling interval; horizontal dotted segments and the displayed vector give the resulting constants. Open markers and the vertical dash-dotted line identify s = 0.82 s=0.82 s = 0.82 , used for the representative main-text visualizations. The dispersion results reported here use the full pointwise multipliers.
Finally, consider a positive uniform dilation of the reconstructed boundary about an arbitrary fixed point r c \mathbf{r}_c r c ,
r v ( λ ) = r c + λ ( r v − r c ) , λ > 0 , (426) \mathbf{r}_v^{(\lambda)}=\mathbf{r}_c+\lambda(\mathbf{r}_v-\mathbf{r}_c),\qquad\lambda>0, \tag*{(426)} r v ( λ ) = r c + λ ( r v − r c ) , λ > 0 , (426)
while keeping the reciprocal lattice and the dimensionless Bloch-path coordinate s s s fixed. The adjacent edge vectors consequently transform as
d v ± , ( λ ) = λ d v ± , t ^ v ± , ( λ ) = λ d v ± ∣ λ d v ± ∣ = t ^ v ± . (427) \mathbf{d}_v^{\pm,(\lambda)}=\lambda\mathbf{d}_v^{\pm},\qquad
\widehat{\mathbf{t}}_v^{\pm,(\lambda)}
=\frac{\lambda\mathbf{d}_v^{\pm}}{\lvert\lambda\mathbf{d}_v^{\pm}\rvert}
=\widehat{\mathbf{t}}_v^{\pm}. \tag*{(427)} d v ± , ( λ ) = λ d v ± , t v ± , ( λ ) = ∣ λ d v ± ∣ λ d v ± = t v ± . (427)
The unit normals and turning angles are therefore unchanged. Moreover, the TE directions e ^ a \widehat{\mathbf{e}}_a e a are fixed by the reciprocal lattice and the selected Bloch point and do not depend on the lengths of the polygon edges. Hence
n ^ v ± , ( λ ) = n ^ v ± , Δ θ v ( λ ) = Δ θ v , S a ( λ ) ( v ; k t ) = S a ( v ; k t ) . (428) \widehat{\mathbf{n}}_v^{\pm,(\lambda)}=\widehat{\mathbf{n}}_v^{\pm},\qquad
\Delta\theta_v^{(\lambda)}=\Delta\theta_v,\qquad
S_a^{(\lambda)}(\mathbf{v};\mathbf{k}_t)=S_a(\mathbf{v};\mathbf{k}_t). \tag*{(428)} n v ± , ( λ ) = n v ± , Δ θ v ( λ ) = Δ θ v , S a ( λ ) ( v ; k t ) = S a ( v ; k t ) . (428)
It follows successively that
T a b s ( λ ) = T a b s , T p r a w , ( λ ) = T p r a w , T p ( λ ) = T p , m ~ p ( λ ) = m ~ p . (429) T_{\mathrm{abs}}^{(\lambda)}=T_{\mathrm{abs}},\qquad
T_p^{\mathrm{raw},(\lambda)}=T_p^{\mathrm{raw}},\qquad
T_p^{(\lambda)}=T_p,\qquad
\widetilde{m}_p^{(\lambda)}=\widetilde{m}_p. \tag*{(429)} T abs ( λ ) = T abs , T p raw , ( λ ) = T p raw , T p ( λ ) = T p , m p ( λ ) = m p . (429)
Equation (S356) holds at every Bloch wavevector along the selected path and therefore applies directly to the pointwise multipliers used in the dispersion calculation. The optional median constants illustrated in Figure S3 inherit the same invariance for a fixed sampling interval.
Thus, the directional-turning multiplier contains no metric edge-length scale. A dilation of the patch at fixed lattice still changes the boundary Fourier overlap through G p ⋅ r \mathbf{G}_p \cdot\mathbf{r} G p ⋅ r , and it also changes C ‾ e q v \overline{C}_{\mathrm{eqv}} C eqv and the associated decay constants. The scale invariance established here therefore applies only to the turning multiplier, not to the complete electromagnetic dispersion model.
Six-order Schur reduction
Let a 0 a_0 a 0 denote the fundamental TE amplitude and a p a_p a p , p = 1 , … , 6 p=1,\ldots,6 p = 1 , … , 6 , the six first-shell amplitudes. For each retained order define the detuning
D p = γ h p − Γ p (430) D_p = \gamma_{h p} - \Gamma_p \tag*{(430)} D p = γ h p − Γ p (430)
and the complex fundamental-to-sideband coupling
v p ( k 0 , k t ) = m ~ p ( k t ) Γ 0 Γ p η 0 p ( k t ) . (431) v_p(k_0,\mathbf{k}_t) = \widetilde{m}_p(\mathbf{k}_t)\sqrt{\Gamma_0\Gamma_p}\,\eta_{0p}(\mathbf{k}_t). \tag*{(431)} v p ( k 0 , k t ) = m p ( k t ) Γ 0 Γ p η 0 p ( k t ) . (431)
Here Γ 0 ≡ Γ T L \Gamma_0 \equiv\Gamma_{\mathrm{TL}} Γ 0 ≡ Γ TL is the fundamental sheet-loaded decay used in the main text. At each Bloch wavevector, m ~ p ( k t ) \widetilde{m}_p(\mathbf{k}_t) m p ( k t ) follows from Eq. (424) , and η 0 p ( k t ) \eta_{0p}(\mathbf{k}_t) η 0 p ( k t ) follows from the boundary Fourier integral. The decays Γ 0 \Gamma_0 Γ 0 and Γ p \Gamma_p Γ p are obtained from the corresponding TE projections of the same local dyadic. All quantities are evaluated at the same k t \mathbf{k}_t k t , with their frequency dependence retained where applicable.
Collect the sideband amplitudes and couplings into
a h = ( a 1 ⋯ a 6 ) T , v = ( v 1 ⋯ v 6 ) T , (432) \mathbf{a}_h = (a_1\ \cdots\ a_6)^{\mathsf{T}}, \qquad\mathbf{v} = (v_1\ \cdots\ v_6)^{\mathsf{T}}, \tag*{(432)} a h = ( a 1 ⋯ a 6 ) T , v = ( v 1 ⋯ v 6 ) T , (432)
and introduce the diagonal retained sideband block
D = diag ( D 1 , … , D 6 ) . (433) \mathbf{D} = \operatorname{diag}(D_1,\ldots,D_6). \tag*{(433)} D = diag ( D 1 , … , D 6 ) . (433)
The scalar TE-projected fundamental-plus-six-order system is
\begin{pmatrix}
R_0 & -\mathbf{v}^{\mathsf{T}} \\
-\mathbf{v} & \mathbf{D}
\end{pmatrix}
\begin{pmatrix}
a_0 \\
\mathbf{a}_h
\end{pmatrix}
= 0. \tag*{(434)}\end{pmatrix}
Written row by row, Eq. (434) is
R 0 a 0 − ∑ p = 1 6 v p ∗ a p = 0 , (435) R_0a_0-\sum_{p=1}^{6}v_p^*a_p=0, \tag*{(435)} R 0 a 0 − p = 1 ∑ 6 v p ∗ a p = 0 , (435)
D p a p − v p a 0 = 0 , p = 1 , … , 6. (436) D_pa_p-v_pa_0=0,\qquad p=1,\ldots,6. \tag*{(436)} D p a p − v p a 0 = 0 , p = 1 , … , 6. (436)
If D p ≠ 0 D_p\ne0 D p = 0 , each sideband amplitude follows directly from Eq. (436) ,
a p = v p D p a 0 . (437) a_p=\frac{v_p}{D_p}a_0. \tag*{(437)} a p = D p v p a 0 . (437)
Substitution into Eq. (435) gives
0 = R 0 a 0 − ∑ p = 1 6 v p ∗ ( v p D p a 0 ) = [ R 0 − ∑ p = 1 6 ∣ v p ∣ 2 D p ] a 0 . (438) \begin{aligned}
0=R_0a_0-\sum_{p=1}^{6}v_p^*\left(\frac{v_p}{D_p}a_0\right) \\
&=\left[R_0-\sum_{p=1}^{6}\frac{\lvert v_p\rvert^2}{D_p}\right]a_0. \tag*{(438)}
\end{aligned} 0 = R 0 a 0 − p = 1 ∑ 6 v p ∗ ( D p v p a 0 ) = [ R 0 − p = 1 ∑ 6 D p ∣ v p ∣ 2 ] a 0 . (438)
A nontrivial fundamental amplitude therefore requires
R 0 − Σ s b = 0 , Σ s b = v T D − 1 v = ∑ p = 1 6 ∣ v p ∣ 2 D p . (439) R_0-\Sigma_{\mathrm{sb}}=0,\qquad\Sigma_{\mathrm{sb}}=\mathbf{v}^{\mathsf{T}}\mathbf{D}^{-1}\mathbf{v}=\sum_{p=1}^{6}\frac{\lvert v_p\rvert^2}{D_p}. \tag*{(439)} R 0 − Σ sb = 0 , Σ sb = v T D − 1 v = p = 1 ∑ 6 D p ∣ v p ∣ 2 . (439)
Using Eqs. (S357) and (S358), the unprojected six-order self-energy becomes
Σ s b ( k 0 , k t ) = Γ 0 ∑ p = 1 6 m ~ p 2 ( k t ) Γ p ∣ η 0 p ( k t ) ∣ 2 γ h p − Γ p . (440) \Sigma_{\mathrm{sb}}(k_0,k_t)=\Gamma_0\sum_{p=1}^{6}\frac{\widetilde{m}_p^2(k_t)\Gamma_p\left|\eta_{0p}(k_t)\right|^2}{\gamma_{hp}-\Gamma_p}. \tag*{(440)} Σ sb ( k 0 , k t ) = Γ 0 p = 1 ∑ 6 γ h p − Γ p m p 2 ( k t ) Γ p ∣ η 0 p ( k t ) ∣ 2 . (440)
All decay quantities in this expression are evaluated at the same ( k 0 , k t ) (k_0,k_t) ( k 0 , k t ) . The corrected unprojected dispersion equation is
k t 2 − k 0 2 − Γ 0 ( k 0 , k t ) − Σ s b ( k 0 , k t ) = 0. (441) \sqrt{k_t^2-k_0^2}-\Gamma_0(k_0,k_t)-\Sigma_{\mathrm{sb}}(k_0,k_t)=0. \tag*{(441)} k t 2 − k 0 2 − Γ 0 ( k 0 , k t ) − Σ sb ( k 0 , k t ) = 0. (441)
The dimensions are consistent: [ Γ 0 ] = [ Γ p ] = [ γ h p ] = m − 1 [\Gamma_0]=[\Gamma_p]=[\gamma_{hp}]=\mathrm{m}^{-1} [ Γ 0 ] = [ Γ p ] = [ γ h p ] = m − 1 , [ v p ] = m − 1 [v_p]=\mathrm{m}^{-1} [ v p ] = m − 1 , and hence [ ∣ v p ∣ 2 / D p ] = [ Σ s b ] = m − 1 [|v_p|^2/D_p]=[\Sigma_{\mathrm{sb}}]=\mathrm{m}^{-1} [ ∣ v p ∣ 2 / D p ] = [ Σ sb ] = m − 1 . Setting v = 0 \mathbf{v}=0 v = 0 recovers the local residual R 0 = 0 R_0=0 R 0 = 0 .
Equation (S360) is the defining reduced-closure approximation. Before this approximation, a general first-shell operator would have the form
D h h e x a c t = ( D 11 D 12 ⋯ D 16 D 21 D 22 ⋯ D 26 ⋮ ⋮ ⋱ ⋮ D 61 D 62 ⋯ D 66 ) , (442) \mathbf{D}_{hh}^{\mathrm{exact}}=
\begin{pmatrix}
D_{11} & D_{12} & \cdots& D_{16}\\
D_{21} & D_{22} & \cdots& D_{26}\\
\vdots& \vdots& \ddots& \vdots\\
D_{61} & D_{62} & \cdots& D_{66}
\end{pmatrix}, \tag*{(442)} D hh exact = D 11 D 21 ⋮ D 61 D 12 D 22 ⋮ D 62 ⋯ ⋯ ⋱ ⋯ D 16 D 26 ⋮ D 66 , (442)
and its Schur self-energy would be
Σ s b e x a c t = v † ( D h h e x a c t ) − 1 v . (443) \Sigma_{\mathrm{sb}}^{\mathrm{exact}}=\mathbf{v}^{\dagger}\left(\mathbf{D}_{hh}^{\mathrm{exact}}\right)^{-1}\mathbf{v}. \tag*{(443)} Σ sb exact = v † ( D hh exact ) − 1 v . (443)
The present model replaces Eq. (S369) by
D h h e x a c t ⟶ D = diag ( γ h 1 − Γ 1 , … , γ h 6 − Γ 6 ) . (444) \mathbf{D}_{hh}^{\mathrm{exact}}\longrightarrow\mathbf{D}=\operatorname{diag}(\gamma_{h1}-\Gamma_1,\ldots,\gamma_{h6}-\Gamma_6). \tag*{(444)} D hh exact ⟶ D = diag ( γ h 1 − Γ 1 , … , γ h 6 − Γ 6 ) . (444)
Thus, fundamental-to-sideband coupling is retained, whereas explicit sideband-to-sideband matrix elements are not calculated. The resulting model is a geometry-derived reduced closure, not an exact reduction of the full periodic Maxwell operator.
Mirror projection on the Γ \Gamma Γ –K K K path
When reflection about the propagation direction is not a symmetry of the complete patch-plus-lattice geometry, the six orders are treated as the independent channels of Eq. (S367). When the Γ \Gamma Γ –K K K path is a mirror axis of the complete unit-cell geometry, the six retained orders can be organized into three symmetry-adapted pairs associated with the mirror subgroup C s = { E , M } C_s=\{E,M\} C s = { E , M } .
Along the selected Γ \Gamma Γ –K K K path, define
u ^ = K ∣ K ∣ , M = 2 u ^ u ^ T − I 2 . (445) \widehat{\mathbf{u}}=\frac{\mathbf{K}}{|\mathbf{K}|},\qquad\mathbf{M}=2\widehat{\mathbf{u}}\widehat{\mathbf{u}}^{\mathsf{T}}-\mathbf{I}_2. \tag*{(445)} u = ∣ K ∣ K , M = 2 u u T − I 2 . (445)
This gives the in-plane reflection about the propagation axis, including its one-sided limit at Γ \Gamma Γ . To determine its action on a TE vector, introduce the 90° rotation matrix
J = ( 0 − 1 1 0 ) , e ^ T E ( q ) = J q ∣ q ∣ . (446) \mathbf{J}=
\begin{pmatrix}
0 & -1\\
1 & 0
\end{pmatrix},
\qquad
\widehat{\mathbf{e}}_{\mathrm{TE}}(\mathbf{q})=\frac{\mathbf{J}\mathbf{q}}{|\mathbf{q}|}. \tag*{(446)} J = ( 0 1 − 1 0 ) , e TE ( q ) = ∣ q ∣ Jq . (446)
For any two-dimensional orthogonal matrix R \mathbf{R} R ,
R J = det ( R ) J R . (447) \mathbf{R}\mathbf{J}=\det(\mathbf{R})\mathbf{J}\mathbf{R}. \tag*{(447)} RJ = det ( R ) JR . (447)
Using ∣ R q ∣ = ∣ q ∣ |\mathbf{R}\mathbf{q}|=|\mathbf{q}| ∣ Rq ∣ = ∣ q ∣ , it follows that
R e ^ T E ( q ) = R J q ∣ q ∣ = det ( R ) J R q ∣ R q ∣ = det ( R ) e ^ T E ( R q ) . (448) \begin{aligned}
\mathbf{R}\widehat{\mathbf{e}}_{\mathrm{TE}}(\mathbf{q})
&=\frac{\mathbf{R}\mathbf{J}\mathbf{q}}{|\mathbf{q}|}\\
&=\det(\mathbf{R})\frac{\mathbf{J}\mathbf{R}\mathbf{q}}{|\mathbf{R}\mathbf{q}|}\\
&=\det(\mathbf{R})\widehat{\mathbf{e}}_{\mathrm{TE}}(\mathbf{R}\mathbf{q}). \tag*{(448)}
\end{aligned} R e TE ( q ) = ∣ q ∣ RJq = det ( R ) ∣ Rq ∣ JRq = det ( R ) e TE ( Rq ) . (448)
Since det ( M ) = − 1 \det(\mathbf{M})=-1 det ( M ) = − 1 ,
M e ^ T E ( q ) = − e ^ T E ( M q ) . (449) \mathbf{M}\hat{\mathbf{e}}_{\mathrm{TE}}(\mathbf{q})=-\hat{\mathbf{e}}_{\mathrm{TE}}(\mathbf{M}\mathbf{q}). \tag*{(449)} M e ^ TE ( q ) = − e ^ TE ( Mq ) . (449)
The minus sign is essential: the TE polarization is mirror odd when its wavevector lies in the mirror plane. The first reciprocal shell is closed under M \mathbf{M} M . Hence there is a permutation π \pi π satisfying
M G p = G π ( p ) . (450) \mathbf{M}\mathbf{G}_{p}=\mathbf{G}_{\pi(p)}. \tag*{(450)} M G p = G π ( p ) . (450)
For the coordinate choice of Eqs. (365) and (368) , M = diag ( 1 , − 1 ) \mathbf{M}=\operatorname{diag}(1,-1) M = diag ( 1 , − 1 ) . Direct application gives
M b 1 = b 1 + b 2 , M ( b 1 + b 2 ) = b 1 , M b 2 = − b 2 , M ( − b 2 ) = b 2 , M ( − b 1 ) = − ( b 1 + b 2 ) , M [ − ( b 1 + b 2 ) ] = − b 1 . (451) \begin{aligned}
\mathbf{M}\mathbf{b}_{1}&=\mathbf{b}_{1}+\mathbf{b}_{2}, & \mathbf{M}(\mathbf{b}_{1}+\mathbf{b}_{2})&=\mathbf{b}_{1},\\
\mathbf{M}\mathbf{b}_{2}&=-\mathbf{b}_{2}, & \mathbf{M}(-\mathbf{b}_{2})&=\mathbf{b}_{2},\\
\mathbf{M}(-\mathbf{b}_{1})&=-(\mathbf{b}_{1}+\mathbf{b}_{2}), & \mathbf{M}[-(\mathbf{b}_{1}+\mathbf{b}_{2})]&=-\mathbf{b}_{1}. \tag*{(451)}
\end{aligned} M b 1 M b 2 M ( − b 1 ) = b 1 + b 2 , = − b 2 , = − ( b 1 + b 2 ) , M ( b 1 + b 2 ) M ( − b 2 ) M [ − ( b 1 + b 2 )] = b 1 , = b 2 , = − b 1 . (451)
With the ordering in Eq. (367) , the permutation is therefore
π = ( 1 3 ) ( 2 5 ) ( 4 6 ) . (452) \pi=(1\ 3)(2\ 5)(4\ 6). \tag*{(452)} π = ( 1 3 ) ( 2 5 ) ( 4 6 ) . (452)
These three mirror-related pairs are illustrated geometrically in Figure S2 . Its matrix representation is
Π M = ( 0 0 1 0 0 0 0 0 0 0 1 0 1 0 0 0 0 0 0 0 0 0 0 1 0 1 0 0 0 0 0 0 0 1 0 0 ) , Π M 2 = I 6 . (453) \mathbf{\Pi}_{M}=
\begin{pmatrix}
0&0&1&0&0&0\\
0&0&0&0&1&0\\
1&0&0&0&0&0\\
0&0&0&0&0&1\\
0&1&0&0&0&0\\
0&0&0&1&0&0
\end{pmatrix},
\qquad\mathbf{\Pi}_{M}^{2}=\mathbf{I}_{6}. \tag*{(453)} Π M = 0 0 1 0 0 0 0 0 0 0 1 0 1 0 0 0 0 0 0 0 0 0 0 1 0 1 0 0 0 0 0 0 0 1 0 0 , Π M 2 = I 6 . (453)
Because the polarization sign in Eq. (449) is part of the modal basis, the representation of the mirror on the six TE coefficients is
D T E ( M ) = − Π M . (454) \mathbf{D}_{\mathrm{TE}}(\mathbf{M})=-\mathbf{\Pi}_{M}. \tag*{(454)} D TE ( M ) = − Π M . (454)
The fundamental TE mode is odd under the same reflection, so its C s C_s C s character is
χ T E ( E ) = 1 , χ T E ( M ) = − 1. (455) \chi_{\mathrm{TE}}(E)=1,\qquad\chi_{\mathrm{TE}}(M)=-1. \tag*{(455)} χ TE ( E ) = 1 , χ TE ( M ) = − 1. (455)
The character projector onto the sideband sector transforming like the fundamental is consequently
P T E = 1 ∣ C s ∣ ∑ g ∈ C s χ T E ∗ ( g ) D T E ( g ) = 1 2 [ I 6 + ( − 1 ) ( − Π M ) ] = 1 2 ( I 6 + Π M ) . (456) \begin{aligned}
\mathbf{P}_{\mathrm{TE}}&=\frac{1}{\lvert C_s\rvert}\sum_{g\in C_s}\chi_{\mathrm{TE}}^{*}(g)\mathbf{D}_{\mathrm{TE}}(g)\\
&=\frac{1}{2}\left[\mathbf{I}_{6}+(-1)(-\mathbf{\Pi}_{M})\right]\\
&=\frac{1}{2}\left(\mathbf{I}_{6}+\mathbf{\Pi}_{M}\right). \tag*{(456)}
\end{aligned} P TE = ∣ C s ∣ 1 g ∈ C s ∑ χ TE ∗ ( g ) D TE ( g ) = 2 1 [ I 6 + ( − 1 ) ( − Π M ) ] = 2 1 ( I 6 + Π M ) . (456)
Thus the apparently symmetric coefficient combinations represent the mirror-odd physical TE sector; the physical minus sign has already been carried by the transformed TE basis vectors. An orthonormal basis for range ( P T E ) \operatorname{range}(\mathbf{P}_{\mathrm{TE}}) range ( P TE ) is
U = 1 2 ( 1 0 0 0 1 0 1 0 0 0 0 1 0 1 0 0 0 1 ) , U † U = I 3 , U U † = P T E . (457) \mathbf{U}=\frac{1}{\sqrt{2}}
\begin{pmatrix}
1&0&0\\
0&1&0\\
1&0&0\\
0&0&1\\
0&1&0\\
0&0&1
\end{pmatrix},
\qquad\mathbf{U}^{\dagger}\mathbf{U}=\mathbf{I}_{3},\qquad\mathbf{U}\mathbf{U}^{\dagger}=\mathbf{P}_{\mathrm{TE}}. \tag*{(457)} U = 2 1 1 0 1 0 0 0 0 1 0 0 1 0 0 0 0 1 0 1 , U † U = I 3 , U U † = P TE . (457)
Hence the six first-shell orders reduce to three TE-compatible mirror channels. Write
a h = U b , b ∈ C 3 . (458) \mathbf{a}_{h}=\mathbf{U}\mathbf{b},\qquad\mathbf{b}\in\mathbb{C}^{3}. \tag*{(458)} a h = Ub , b ∈ C 3 . (458)
Substitution into Eq. (434) , followed by left multiplication of the sideband rows by U † \mathbf{U}^{\dagger} U † , gives the reduced system
\begin{pmatrix}
R_0 & -(\mathbf{U}^{\dagger}\mathbf{v})^{\dagger} \\
-\mathbf{U}^{\dagger}\mathbf{v} & \mathbf{U}^{\dagger}\mathbf{D}\mathbf{U}
\end{pmatrix}
\begin{pmatrix}
a_0 \\
\mathbf{b}
\end{pmatrix}
=0. \tag*{(459)}\end{pmatrix}
Define
v r = U † v , D r = U † D U . (460) \mathbf{v}_r=\mathbf{U}^{\dagger}\mathbf{v}, \qquad\mathbf{D}_r=\mathbf{U}^{\dagger}\mathbf{D}\mathbf{U}. \tag*{(460)} v r = U † v , D r = U † DU . (460)
The reduced sideband equation is
D r b = v r a 0 , b = D r − 1 v r a 0 , (461) \mathbf{D}_r\mathbf{b}=\mathbf{v}_r a_0,\qquad\mathbf{b}=\mathbf{D}_r^{-1}\mathbf{v}_r a_0, \tag*{(461)} D r b = v r a 0 , b = D r − 1 v r a 0 , (461)
and the corresponding Schur complement is
Σ s b p r o j = v r † D r − 1 v r = ( U † v ) † ( U † D U ) − 1 ( U † v ) . (462) \Sigma_{\mathrm{sb}}^{\mathrm{proj}}=\mathbf{v}_r^{\dagger}\mathbf{D}_r^{-1}\mathbf{v}_r=(\mathbf{U}^{\dagger}\mathbf{v})^{\dagger}(\mathbf{U}^{\dagger}\mathbf{D}\mathbf{U})^{-1}(\mathbf{U}^{\dagger}\mathbf{v}). \tag*{(462)} Σ sb proj = v r † D r − 1 v r = ( U † v ) † ( U † DU ) − 1 ( U † v ) . (462)
The projected dispersion equation is
k t 2 − k 0 2 − Γ 0 − Σ s b p r o j = 0. (463) \sqrt{k_t^2-k_0^2}-\Gamma_0-\Sigma_{\mathrm{sb}}^{\mathrm{proj}}=0. \tag*{(463)} k t 2 − k 0 2 − Γ 0 − Σ sb proj = 0. (463)
For the explicit basis in Eq. (457) ,
v r = 1 2 ( v 1 + v 3 v 2 + v 5 v 4 + v 6 ) , D r = 1 2 ( D 1 + D 3 0 0 0 D 2 + D 5 0 0 0 D 4 + D 6 ) . (464) \mathbf{v}_r=\frac{1}{\sqrt{2}}
\begin{pmatrix}
v_1+v_3\\
v_2+v_5\\
v_4+v_6
\end{pmatrix},
\qquad
\mathbf{D}_r=\frac{1}{2}
\begin{pmatrix}
D_1+D_3 & 0 & 0\\
0 & D_2+D_5 & 0\\
0 & 0 & D_4+D_6
\end{pmatrix}. \tag*{(464)} v r = 2 1 v 1 + v 3 v 2 + v 5 v 4 + v 6 , D r = 2 1 D 1 + D 3 0 0 0 D 2 + D 5 0 0 0 D 4 + D 6 . (464)
Therefore Eq. (462) can also be written as
Σ s b p r o j = ∣ v 1 + v 3 ∣ 2 D 1 + D 3 + ∣ v 2 + v 5 ∣ 2 D 2 + D 5 + ∣ v 4 + v 6 ∣ 2 D 4 + D 6 . (465) \Sigma_{\mathrm{sb}}^{\mathrm{proj}}=\frac{\lvert v_1+v_3\rvert^2}{D_1+D_3}+\frac{\lvert v_2+v_5\rvert^2}{D_2+D_5}+\frac{\lvert v_4+v_6\rvert^2}{D_4+D_6}. \tag*{(465)} Σ sb proj = D 1 + D 3 ∣ v 1 + v 3 ∣ 2 + D 2 + D 5 ∣ v 2 + v 5 ∣ 2 + D 4 + D 6 ∣ v 4 + v 6 ∣ 2 . (465)
For an exactly mirror-symmetric geometry, with the patch centred at the symmetry origin, D 1 = D 3 D_1=D_3 D 1 = D 3 , D 2 = D 5 D_2=D_5 D 2 = D 5 , and D 4 = D 6 D_4=D_6 D 4 = D 6 . In the adopted TE convention, mirror symmetry also gives v 1 = v 3 v_1=v_3 v 1 = v 3 , v 2 = v 5 v_2=v_5 v 2 = v 5 , and v 4 = v 6 v_4=v_6 v 4 = v 6 . For each mirror-related pair ( p , π ( p ) ) (p,\pi(p)) ( p , π ( p )) , the coupling to the normalised symmetric coefficient combination is therefore ( v p + v π ( p ) ) / 2 = 2 v p (v_p+v_{\pi(p)})/\sqrt{2}=\sqrt{2}v_p ( v p + v π ( p ) ) / 2 = 2 v p , whereas the coupling to the complementary antisymmetric combination vanishes: ( v p − v π ( p ) ) / 2 = 0 (v_p-v_{\pi(p)})/\sqrt{2}=0 ( v p − v π ( p ) ) / 2 = 0 . (465) thus retains the full self-energy contribution 2 ∣ v p ∣ 2 / D p 2\lvert v_p\rvert^2/D_p 2 ∣ v p ∣ 2 / D p of each original pair, while the antisymmetric combination is decoupled from the fundamental mode.
Whenever D r \mathbf{D}_r D r is nonsingular, the ordinary inverse in Eq. (462) applies. To examine numerical rank, let
D r = Q diag ( d 1 , d 2 , d 3 ) Q † . (466) \mathbf{D}_r=\mathbf{Q}\operatorname{diag}(d_1,d_2,d_3)\mathbf{Q}^{\dagger}. \tag*{(466)} D r = Q diag ( d 1 , d 2 , d 3 ) Q † . (466)
A Moore–Penrose inverse with a fixed machine-level cutoff can be defined by
D r + = Q diag ( d 1 + , d 2 + , d 3 + ) Q † , d j + = { d j − 1 , ∣ d j ∣ > τ D D r e f , 0 , ∣ d j ∣ ≤ τ D D r e f , (467) \mathbf{D}_r^{+}=\mathbf{Q}\operatorname{diag}(d_1^{+},d_2^{+},d_3^{+})\mathbf{Q}^{\dagger},\qquad
d_j^{+}=
\begin{cases}
d_j^{-1}, & \lvert d_j\rvert>\tau_DD_{\mathrm{ref}},\\
0, & \lvert d_j\rvert\le\tau_DD_{\mathrm{ref}},
\end{cases} \tag*{(467)} D r + = Q diag ( d 1 + , d 2 + , d 3 + ) Q † , d j + = { d j − 1 , 0 , ∣ d j ∣ > τ D D ref , ∣ d j ∣ ≤ τ D D ref , (467)
where D r e f = max j ∣ d j ∣ D_{\mathrm{ref}}=\max_j\lvert d_j\rvert D ref = max j ∣ d j ∣ and τ D \tau_D τ D is fixed independently of geometry. The cutoff provides a diagnostic of numerical rank loss. At a true singularity of D r \mathbf{D}_r D r , the inverse-based Schur expression is not applicable; the dispersion must be assessed from the coupled system in Eq. (459) . A singular channel must not be discarded through the pseudoinverse and interpreted as an equivalent dispersion condition.
Combining the preceding results, the complete calculation is
{ C ‾ e q v , ∂ Σ } ⟶ { Γ 0 , Γ p , η 0 p , T p r a w , T p , m ~ p ( k t ) } ⟶ Σ s b ⟶ R 0 − Σ s b = 0. (468) \left\{\overline{C}_{\mathrm{eqv}},\partial\Sigma\right\}
\longrightarrow
\left\{\Gamma_0,\Gamma_p,\eta_{0p},T_p^{\mathrm{raw}},T_p,\widetilde{m}_p(k_t)\right\}
\longrightarrow\Sigma_{\mathrm{sb}}\longrightarrow R_0-\Sigma_{\mathrm{sb}}=0. \tag*{(468)} { C eqv , ∂ Σ } ⟶ { Γ 0 , Γ p , η 0 p , T p raw , T p , m p ( k t ) } ⟶ Σ sb ⟶ R 0 − Σ sb = 0. (468)
The geometric factors and polarization projections are evaluated at the same Bloch wavevector, with the frequency dependence of the decay quantities retained. The unprojected expression in Eq. (440) is used when the propagation-axis reflection is not a symmetry of the complete patch-plus-lattice geometry. When that reflection belongs to the little group of the complete geometry, the projected self-energy in Eq. (462) is used for Σ s b \Sigma_{\mathrm{sb}} Σ sb in the dispersion residual. The projected expression retains the contributions of all six original orders through three symmetry-adapted channels. The symmetry reduction introduces no additional multiplicity factor or fitted coupling parameter.
Long-wavelength consistency
For fixed geometry with finite, nondispersive loading, the projections A 0 A_0 A 0 and A p A_p A p remain finite and positive. The expansion in Eq. (377) gives
Γ 0 ≃ A 0 k 0 2 2 , Γ p ≃ A p k 0 2 2 , k 0 A 0 , k 0 A p ≪ 1. (469) \Gamma_0 \simeq\frac{A_0 k_0^2}{2}, \qquad\Gamma_p \simeq\frac{A_p k_0^2}{2}, \qquad k_0 A_0, k_0 A_p \ll1. \tag*{(469)} Γ 0 ≃ 2 A 0 k 0 2 , Γ p ≃ 2 A p k 0 2 , k 0 A 0 , k 0 A p ≪ 1. (469)
Thus both sheet-loaded decays vanish quadratically with frequency. As k 0 , k t → 0 k_0, k_t \to0 k 0 , k t → 0 along the lowest bound branch, each retained nonzero reciprocal order satisfies
D p = γ h p − Γ p ⟶ ∣ G p ∣ = G 0 > 0. (470) D_p = \gamma_{h p} - \Gamma_p \longrightarrow\lvert\mathbf{G}_p \rvert= G_0 > 0. \tag*{(470)} D p = γ h p − Γ p ⟶ ∣ G p ∣ = G 0 > 0. (470)
The normalized overlaps obey ∣ η 0 p ∣ ≤ 1 \lvert\eta_{0p} \rvert\le1 ∣ η 0 p ∣ ≤ 1 by Eq. (408) , and Eq. (425) keeps the six turning multipliers bounded. The unprojected self-energy therefore satisfies
Σ s b Γ 0 = ∑ p = 1 6 m ~ p 2 ( k t ) Γ p ∣ η 0 p ( k t ) ∣ 2 D p ⟶ 0. (471) \frac{\Sigma_{\mathrm{sb}}}{\Gamma_0} = \sum_{p=1}^{6} \frac{\widetilde{m}_p^2(k_t)\,\Gamma_p\,\lvert\eta_{0p}(k_t) \rvert^2}{D_p} \longrightarrow0. \tag*{(471)} Γ 0 Σ sb = p = 1 ∑ 6 D p m p 2 ( k t ) Γ p ∣ η 0 p ( k t ) ∣ 2 ⟶ 0. (471)
After mirror projection, D r → G 0 I 3 \mathbf{D}_{r} \to G_0 \mathbf{I}_3 D r → G 0 I 3 , so the same relative suppression holds for Σ s b p r o j \Sigma_{\mathrm{sb}}^{\mathrm{proj}} Σ sb proj . The sideband correction consequently becomes negligible relative to the local loading, recovering the leading local long-wavelength dispersion. The individual turning multipliers need not approach unity or vanish at Γ \Gamma Γ .
References