WorksPhysics and Astronomy
Windowed Singularity Subtraction for the Spectral Integral Method
Abstract
In 2004, the spectral integral method (SIM) was proposed to solve TM scattering by homogeneous two-dimensional dielectric objects and perfect electrical conductors with smooth boundaries. Its error decreases exponentially as the number of boundary points increases linearly. Plasmonic resonances, however, require TE polarization and permittivities with a negative real part. The extension of SIM to this case shows that the original global singularity subtraction suffers catastrophic cancellation, with a round-off floor that grows exponentially with the attenuation accumulated across the object. To resolve this issue, this work introduces a windowed singularity subtraction that limits this cancellation amplification to an exponential factor that the user can set independently of the object size, provided the window is adequately resolved. Numerical results verify the accuracy of the proposed method.
Introduction
Boundary integral equations discretized with global trigonometric bases achieve exponential convergence for smooth, analytic boundaries, provided the logarithmic singularity of the two-dimensional (2D) Green’s function is treated analytically [1, 2, 3]. The spectral integral method (SIM) uses this idea for scattering by homogeneous dielectric and perfectly electric conductor (PEC) cylinders [4], where the cylinder’s cross section lies in the -plane, and the incident electric field is polarized along the direction. SIM splits the Green’s function into a smooth part and the product of a logarithmic kernel with an entire function, and evaluates both with fast Fourier transforms (FFTs). Later, the SIM was extended to objects embedded in layered media [5] and to three dimensions [6]. In all of these works, the scatterers were lossless dielectrics or PEC.
Metallic nanostructures support plasmonic resonances in both polarizations when the field can vary along all three directions, as in chains of metal nanoparticles [7, 8]. For a cylinder that is invariant along and long compared with the wavelength, however, a localized surface-plasmon resonance requires polarization (), because only then does the electric field have a component normal to the boundary that induces surface charges. Extending SIM to is therefore needed to study plasmonic resonances of long metallic scatterers and plasmonic waveguides [9].
Other high-order integral-equation methods have been developed for electromagnetic scattering by penetrable objects. For inhomogeneous two-dimensional scatterers, the Lippmann–Schwinger equation was solved with truncated Fourier expansions of the Green function, giving a high-order solver that runs in operations even for discontinuous refractive indices [10], and Bruno and Hyde proved that this Fourier-smoothing approach converges with higher order, with rates determined by the regularity of the scatterer [11]. For homogeneous objects, Balaban et al. discretized boundary integral equations, including the second-kind Müller equations for arbitrary dielectric cylinders, with Nyström quadratures adapted to smooth, logarithmic, and hypersingular kernels, and applied them to plasmon-assisted H-polarized scattering by silver nanostrips [12]. Sun et al. introduced a field-only, fully desingularized surface integral formulation for three-dimensional dielectric bodies [13]. Closest to the present problem, Helsing and Karlsson treated the two-dimensional transmission problem for the axial magnetic field of a cylinder with complex permittivity and sharp corners, combining second-kind Kleinman–Martin equations with Nyström discretization, local product integration, and recursively compressed inverse preconditioning, and obtained highly accurate fields even for real negative permittivities [14]. In these methods, the singular part of the kernel is handled either within a volume formulation or locally, on the boundary elements adjacent to the singularity, as in the partition-of-unity localization of high-order surface solvers [15]. Methods that instead remove the logarithmic singularity globally, as SIM does, face a different difficulty, which has not been addressed so far.
Extending SIM to TE scattering requires the careful treatment of two related details. First, for metals in the visible part of the spectrum, the interior wavenumber is complex, so the branches of square roots, logarithms, and Hankel functions must be chosen consistently. Second, when is large, where is the largest chord of the scatterer, the global singularity subtraction of [4] generates exponentially large intermediate terms that cancel catastrophically. This work proposes a windowed singularity subtraction that removes this limitation while preserving the structure of SIM: unknowns per field, FFT-based kernels, and no special quadrature. The cancellation mechanism and its round-off floor are analyzed, the non-analyticity of any localized splitting and its implications for super-algebraic convergence are established, and a signed-curvature self-term valid for non-convex boundaries is provided. An empirically calibrated rule for choosing is also proposed.
The paper is organized as follows. Section 2 presents a compact formulation of electromagnetic scattering that accommodates both TE and TM excitations, introduces the windowed singularity subtraction, and discusses its convergence properties. Section 3 validates the method against the Mie series for circular cylinders and an FDTD solver for a non-convex cylinder, investigates the window parameters and conditioning, evaluates the proposed rule for choosing , reports the computational cost, and then we conclude.
Formulation
A homogeneous object with relative permittivity and boundary is embedded in a lossless background . Both media are nonmagnetic, and the time dependence is . The scatterer is passive, , and . The wavenumbers inside the background and object are and , respectively, with the principal square root. The unit incident wave is
so that corresponds to a propagation along the direction.
The boundary is parametrized counter-clockwise, with . The Jacobian is , the unit tangent is , and the outward unit normal is
so that (Figure 1). All normal derivatives use this outward normal.

Figure 1. A 2D smooth object with relative permittivity in a background . The plane wave propagates in the direction , measured from the axis, and is the outward unit normal.
For TE polarization, and . Maxwell’s equations can be reduced to , , in each homogeneous region. No surface currents are present, so the tangential magnetic field is continuous, on , where and denote the exterior and interior sides. The tangential electric field is
Continuity of gives , where . For metals, and the normal derivative changes sign across the interface. This sign change supports surface plasmons, which, for a geometry invariant along , exist only under TE excitation.
Let denote for TE or for TM. The derivative of the Green’s function, with , is then
where and are Hankel functions of the first kind, of orders zero and one, respectively. For , Green’s representation of the radiating field in the exterior, with the background Green’s function and the outward normal, is
The incident field is regular in and satisfies the background Helmholtz equation there, so Green’s second identity in with the same gives, for ,
Adding the two equations removes from the integrand:
In the interior, Green’s representation of the total field , which satisfies the interior Helmholtz equation, uses the interior Green’s function :
As approaches , the single-layer potentials are continuous, and the double-layer potential tends to its principal value from outside and from inside. Equations (7) and (8) therefore give
for . With the unknowns and , and with from the continuity equation, we obtain
In TE, and are, up to sign and constant factors, the equivalent electric and magnetic surface currents and ; in TM the roles are interchanged. For these are the TM equations of [4].
On equispaced nodes and , the unknowns are expanded as truncated Fourier series, and . The small-argument expansions (9.1.13 and 9.1.11 of [16]; 10.8.2 and 10.8.1 of [17])
where , is the harmonic number, and is Euler’s constant (the bracket equals in terms of the digamma function), show that each Bessel function of the second kind consists of the logarithm times an entire function, , plus a remainder. The remainder is entire for and has a simple pole for . The pole of is canceled by geometry (because the boundary is smooth). On a smooth curve , where
is the signed curvature (positive for a convex boundary traversed counter-clockwise). Only the logarithm is genuinely non-smooth. Since , the kernels are split as
where in [4] and is the window introduced below. The logarithmic parts are discrete convolutions with the known Fourier coefficients of , namely and , and are evaluated with zero-padded FFTs.
The principal logarithm is applied only to in the kernels and to in the self-term, Eq. (18). The lossless limits are covered: a lossless dielectric () gives , and a lossless metal () gives , both inside the admissible range. The epsilon-near-zero limit () is excluded: there , the self-term diverges, and the interior equation degenerates into a Laplace problem that requires a separate limiting treatment. Active media () and lossy backgrounds are also outside the scope of this work.
As , and . With and Eq. (13),
With in Eq. (4),
because the logarithmic term of Eq. (17) is . Equation (19) is independent of . With signed as in Eq. (15), it holds on concave arcs. In the implementation, , and are evaluated analytically from and of the parametrization, not from finite differences of sampled points; the diagonal entry of the double-layer kernel is then simply Eq. (19) at .
Collocation at gives the system of [4], with the interior single-layer block scaled by . For each region , testing point and Fourier mode , let
and define and in the same way with and , where and are the factors multiplying the logarithm in Eqs. (16)–(17). The smooth integrals and are evaluated with the trapezoidal rule, and and are discrete convolutions of the Fourier coefficients of and with . Then and . Multiplying Eq. (11) by 2 and Eq. (12) by gives
with
where , , and and are the vectors of Fourier coefficients of and . The blocks have the same form for both regions, and only the sign of the jump term differs. Eliminating gives the Schur complement
For , the block changes sign relative to the dielectric case, so the two terms of can cancel. This cancellation is the discrete counterpart of a surface-plasmon resonance. In the quasi-static limit of a circular cylinder, the resonance is at . For the tested cases, peaks there at about and stays moderate otherwise (§ 3.6). The windowed subtraction addresses the accuracy of the matrix entries; it does not change this intrinsic sensitivity of the solution near a resonance.
In [4], is subtracted over the entire boundary. Since while , the smooth part and the convolution are both of size , where is the largest chord. They cancel to produce an matrix entry, so the relative error is bounded by
where is the machine precision. Equation (25) estimates the floating-point error of the assembled matrix entries. It is a mechanism-based estimate, not a rigorous backward-error bound for the complete algorithm. A relative perturbation of the entries of size changes the computed boundary solution by at most about in relative terms, and is – in our tests (§ 3.6). The error of the boundary solution additionally contains the discretization error. Equation (25) is therefore a floor, not a prediction of the total error, and it describes only the cancellation mechanism that the window controls. The relevant quantity is the attenuation accumulated across the object, , not the skin depth alone: the global splitting loses about significant digits. For gold at 800 nm, (skin depth nm). A cylinder of diameter gives , so the original method yields huge errors. However, the subtraction only needs to be exact near . Let be -periodic, with
where is the half-width of the window, for and . Then and , and vanishes to all orders at , although is not identically equal to one in any neighborhood of . The denominator makes and a smooth partition of unity on the transition region, and decreases monotonically from at to at . Hence is (see the Supplementary Materials and [18, 19, 20, 21, 22]), Eqs. (16)–(17) remain smooth, and the FFT part is unchanged. Outside the window, is evaluated directly and is exponentially small. The half-width, measured in the parameter , is
where the dimensionless parameter bounds the attenuation exponent inside the window: there , so and . The cancellation amplification inside the window is therefore at most about , and the window extends about skin depths along the boundary on each side of the source point. In Eq. (26), is the parameter difference reduced to , so that the window also covers source points across the parameter seam ; the implementation evaluates at the wrapped difference. Because the window is used only when , and every chord satisfies , Eq. (27) gives . For , the support of therefore lies inside one period, and its periodic extension is ; with the default values, . The window is applied to the interior kernel only, since the background is lossless, and only if , with the dimensionless switch-on threshold . Below , the round-off floor of the global splitting is acceptable, and its exponential convergence is retained.1
Inserting the large-argument form of (2D free-space Green’s function of the Helmholtz equation, evaluated with the background wavenumber) into Eq. (9) gives the scattered field
and the far-field amplitude
The bistatic scattering width and the cross sections per unit length are
With the unit incident amplitude of Eq. (1), the incident intensity is , with ; the cross sections are the scattered, extinguished, and absorbed powers per unit length divided by , and have the dimension of length. All integrals use the trapezoidal rule in . It is always nice to check the residual to see whether it is a very small number, but of course, it is not a sufficient accuracy check.
Numerical results
The method is implemented in MATLAB, and the simulations are done on a 2021 MacBook Pro with 32 GB of RAM. The error is calculated against independent references: the Mie series [23] for circular cylinders, a converged SIM solution for internal convergence, and FDTD for the non-convex case. Lengths are in nm. We consider two types of materials: BK7 glass is a lossless dielectric, and gold (Au) is a lossy conductor with a negative electrical permittivity in the visible part of the electromagnetic spectrum.
The permittivity of BK7 glass is calculated from its Sellmeier formula [24] ( at 633 nm). For gold, we use the analytic model of Etchegoin et al. [25], which was fitted to the Johnson–Christy data [26]. At 633 nm, this model gives , which corresponds to the complex refractive index with and . For the windowing, we set and .
Circular cylinders: comparison with the Mie series
Circular cylinders of radius nm are illuminated at nm with radian and unit amplitude, A/m. For these cases, for gold and 0 for BK7. Both are below , so the global splitting () is used. Figure 2 shows the boundary field and the scattering width. The maximum relative errors with respect to the Mie series are (for ) and (for ) for BK7 with , and and for gold with .

Figure 2. BK7 (left, ) and gold (right, ) cylinders, nm, nm, rad. (a,b) on the boundary versus polar angle . (c,d) Bistatic scattering width. Line: Mie series; circles: SIM.
Figure 3 shows the extinction and absorption efficiencies, , of the gold cylinder in vacuum from 400 to 800 nm. The SIM uses . Compared with the Mie series, the maximum deviations are in and in , and the energy residual is below . peaks at near 520–530 nm. Absorption dominates below about 500 nm, where interband transitions are active, and falls to at 800 nm.

Figure 3. Extinction and absorption efficiencies of a gold cylinder ( nm, vacuum, ). Curves: Mie series; markers: SIM, . Maximum deviations: () and ().
Figure 4 shows the maximum relative error of the boundary field, , against the number of points per wavelength (PPW), , for a BK7 and Au cylinder, where is the free-space wavelength. The error decreases exponentially, as expected for analytic kernels. For BK7 it reaches at 10 PPW and at 16 PPW. For gold, it reaches at 14 PPW and at 16 PPW. Gold needs more points per free-space wavelength because its interior field varies on shorter scales: and nm. Calculated per interior wavelength , the two materials behave similarly: about 6.7 (BK7) and 4.8 (gold) points give . Beyond 16 points per wavelength, the gold error saturates at . This plateau is the round-off floor, Eq. (25), of the global splitting: is an upper bound, and the observed floor lies about 20 times below it.

Figure 4. Convergence of the SIM (global splitting) for BK7 and gold cylinders, nm, nm: maximum relative error of the boundary with respect to the Mie series versus the number of boundary points per free-space wavelength.
The cases above do not activate the window. To test it, Figure 5 compares the two splittings for gold cylinders with , , and (, , and nm at nm). In each run, the window is either forced on or forced off. With the global splitting, the error reaches the round-off floor for nm, stalls at – for nm, and is for nm, for every . With the window, all three cases converge: the error reaches ( nm, ), ( nm, ) and ( nm, ). The cross sections converge at the same rate; for example, is accurate to and to for nm at .

Figure 5. Gold cylinders at nm, maximum relative error of the boundary with respect to the Mie series versus . Blue: nm (); red: nm (33.3); yellow: nm (66.6). Solid: windowed subtraction (); dashed: global subtraction.
In Figure 5, the windowed curves are almost straight with increasing negative slopes. We fitted by least squares against and against , using all points with error above , so that points dominated by the round-off floor are excluded. The slopes of the fits against , which equal , scale as , so that the coefficient is nearly the same in the three cases. For the three cases (, and ), the error follows the empirical law , where .
Non-convex boundary: gold rose cylinder
For our next numerical study, we chose a scatterer with a non-convex boundary, a five-lobed “rose” cylinder defined by
with nm and (inset of Figure 6). The radius varies between 112.5 and 187.5 nm, and the boundary is concave between neighboring lobes, where the radius of curvature is as small as 15 nm. The cylinder is gold, in vacuum, at nm, illuminated by a plane wave and . Here , so the global splitting is used. No analytical solution exists, so we compare with a commercial FDTD solver, Tidy3D [27]. FDTD is not an exact reference, so this comparison establishes agreement at the sub-percent level, not independent verification of the SIM’s internal ten-digit convergence.

Figure 6. Gold rose cylinder [Eq. (33), nm, ; inset] at nm, illumination along (). Magnitudes of the total fields and on a circle of radius nm versus polar angle ( is the illuminated side). Curves: SIM (). Circles: Tidy3D (uniform 2 nm grid).
Relative to a SIM solution with , the maximum error of the three cross sections, normalized by , is , , and for , 160, 256, respectively. The latter is ten significant digits by internal convergence; the independent comparison follows below. The rose needs about four times more points than a circle of the same perimeter because of the high curvature of its necks. The reference values below use . Fields outside the cylinder are evaluated from the exterior representation, Eq. (9), and its gradient, with .
The Tidy3D model is a 2D simulation (zero thickness and periodic boundaries along ) on a uniform 2 nm grid, in a m domain terminated by 24-layer perfectly matched layers (PML), with a margin of m () between the scattered-field flux box and the PML. The gold permittivity at 633 nm is represented by a single-pole Lorentz model that reproduces the same exactly at this wavelength. The plane wave is injected with a total-field/scattered-field source. The scattered power is the flux of the scattered field through a closed box outside the total-field region, and the absorbed power is the net inward flux through a box enclosing the cylinder. Both are normalized by the incident intensity from a second simulation without the scatterer. Fields on the probe circle are divided by those of the same empty simulation, which removes the source normalization and phase.
Figure 6 compares the magnitudes of the total magnetic and electric fields, normalized by the incident field, on a circle of radius nm. The SIM and FDTD curves are indistinguishable at the scale of the figure over the full angular range. This includes the interference structure on the illuminated side: an electric-field enhancement in the backscattering direction, and two near-zeros at and . Table 1 lists the cross sections. The extinction, scattering and absorption cross sections differ by , and of . Possible sources are the discretization of the FDTD model, since Tidy3D treats metallic media with staircasing by default, so that the 2 nm grid does not represent the curved gold surface exactly while the field varies over the skin depth nm, and the finite simulation domain and run time.
| SIM | Tidy3D (2 nm grid) | ||
| 1271.23 | 1276.41 | ||
| 1207.52 | 1212.01 | ||
| 63.71 | 64.40 |
Table 1. Cross sections per unit length (nm) of the gold rose cylinder, Eq. (33), at nm, illumination along ().
Computational cost
The assembly requires operations ( rows, each with FFT convolutions of length ), and the dense solve requires , so the solve dominates asymptotically. For the sizes considered here, however, the cost is set by the large constant of the assembly, because each matrix entry requires several complex Bessel and Hankel function evaluations, whereas the solve uses optimized dense linear algebra. Table 2 lists wall-clock times for the complete calculation, including assembly, solve, far field, and the three cross sections, with the windowed singularity subtraction for the gold cylinder with nm at nm. The far field and the cross sections take less than of the total time for . Over , the measured time grows approximately as . This growth is slower than the operation count suggests, because a fixed overhead per solution dominates at small and vectorized evaluation becomes more efficient as increases. The local exponent increases with , from about 1.3 at the smallest sizes to 1.8 between and 1536, but it remains below the value of about 2.1 that an cost would give in this range.
| 128 256 512 768 1024 1280 1536 | |
| time (s) | 0.06 0.16 0.46 0.88 1.48 2.13 3.43 |
Table 2. Wall-clock time of a complete SIM solution (assembly, solve, far field and cross sections), Au, nm, nm, window on. Medians of seven runs after one warup run.
Global vs. Windowed Singularity Subtraction
To demonstrate the need for the windowed singularity subtraction, we fix the sampling number to 1024 and calculate the relative error in the total magnetic field with global subtraction and windowed subtraction for circular gold with a radius ranging from 50 nm to 600 nm. At nm, this range corresponds to between 3.3 and 40. Figure 7 shows that the global error grows as predicted by Eq. (25), from at to at 33. The windowed error stays at for all values. For small , however, the global subtraction is more accurate than the windowed subtraction. For small , the global splitting is limited only by its exponentially convergent discretization error, because its cancellation floor is negligible; the window removes nothing there but adds a Gevrey-2 transition whose Fourier coefficients decay only root-exponentially, so at equal , it is less accurate.

Figure 7. Error at versus (gold, –600 nm, 633 nm). The global error follows the bound (dotted), and the windowed error stays at for all values.
Window parameters
For the previous examples, and were set to 8 and 15, respectively. Here, we run some additional test cases to investigate their influence on the accuracy. For the former, we assume a gold cylinder with nm at 633 nm, whose circumference is about . is fixed to 15 and we change from 2 to 16. Figure 8 (a) shows how the maximum relative error changes for and , 768, and 1024. For small , the window is narrow, its transition is steep, and the error is large: at it lies between about and . It decreases by one to two orders of magnitude up to and then varies by less than one order of magnitude for , with a shallow minimum near for and 1024. At it increases again, consistent with the cancellation factor inside the window, since . For the former studies, the default value of was 8; apparently, would have been a better choice.

Figure 8. (a) Maximum relative error of the boundary with respect to the Mie series versus for a gold cylinder, nm, nm, , 768 and 1024. (b) Worst-case maximum relative error of the boundary with respect to the Mie series versus the switch-on threshold , over gold cylinders with –600 nm (–40) at nm, , .
The threshold acts only as a switch: the window is used when and the global splitting otherwise. Its influence can therefore be seen only over a range of object sizes, not for a single object. We computed gold cylinders at nm with radii , 75, …, 600 nm, with and . For each cylinder, the scattering problem was solved once with the global splitting and once with the windowed splitting. For a given , the error that the method returns is the global result for the radii with and the windowed result for the others. Figure 8 (b) shows the largest of these errors over all radii as a function of . For , the worst-case error is constant at . It is set by the windowed result for the largest cylinder. For , the global splitting is used for cylinders whose round-off floor is no longer negligible, and the worst-case error grows with . With , any between 7 and 13 is therefore optimal for this range of sizes.
| group | case | at | ||||||
| shape | Au ellipse , 633 nm | 20.0 | 50 | 394 | 102 | 394 | 4.4 | |
| Au rose , 633 nm | 12.9 | 56 | – | 470 | 384 | 11.1 | ||
| Au ellipse , 633 nm | 40.0 | 100 | 786 | 102 | 1024 | 5.9 | ||
| resonance | Au circle , 506 nm | 2.3 | 10 | – | 16 | 16 | 2.7 | |
| circle , , 633 nm | 0.4 | 2 | – | 16 | 16 | 96.6 | ||
| material | circle , , 633 nm | 0.4 | 270 | – | 16 | 128 | 154 | |
| circle , , 633 nm | 26.7 | 198 | 524 | 16 | 768 | 29.6 |
Table 3. Validation of the sampling rule, Eq. (34). Dimensions are in nm. : maximum relative boundary-field error versus Mie or maximum cross-section error relative to versus SIM with ( for the large ellipse) (non-circular shapes). : smallest tested with . The column is evaluated at the predicted .
Conditioning
We also calculated the 2-norm condition numbers of the Schur complement, , Eq. (24), and reported them in Table 3. Without the window, the computed for nm fluctuates between and , which reflects corrupted matrix entries rather than intrinsic ill-conditioning. To probe the plasmon resonance, we swept from to for a small cylinder ( nm, 633 nm, ) with , and . peaks at , the quasi-static pole, with values of , and 20.1, i.e. . Even at the sharpest resonance, the error with respect to the Mie series stays below . For the additional cases of Table 3, the condition number ranges from 2.7 to 154. The near-resonant cylinder with gives 97, again close to .
How to choose for ultra-high accuracy
We run several other studies: a gold ellipse with 3:1 aspect ratio, a deeper-neck rose with (minimum neck radius of curvature 7.8 nm), and a larger 3:1 gold ellipse with the window on; two near-resonance cases: a gold cylinder with nm at its extinction peak (506 nm) and a small cylinder with close to the quasi-static pole; two other materials: a high-index, low-loss dielectric representative of silicon at 633 nm (, nm, with strong Mie resonances) and a strongly absorbing dielectric with (, nm, window on). Their numerical analyses are summarized in Table 4. Based on these studies, we recommend choosing as the largest of the three resolution requirements, rounded up to an even integer,
| case | |||||||
| BK7, nm, nm | 0 | 22 | – | 16 | 22 | ||
| Au, nm, nm | 13.3 | 48 | – | 16 | 48 | ||
| Au, nm, nm | 33.3 | 118 | 656 | 16 | 656 | ||
| Au, nm, nm | 38.3 | 134 | 752 | 16 | 752 | ||
| Au, nm, nm | 66.6 | 234 | 1310 | 16 | 1310 | ||
| BK7 rose, nm | 0 | 22 | – | 204 | 204 | – | |
| Au rose, nm | 12.0 | 48 | – | 204 | 204 | – |
Table 4. Sampling rule Eq. (34) versus observed error at the predicted ( in nm). : maximum relative boundary-field error versus Mie. : maximum cross-section error relative to , versus Mie or versus SIM with .
where
where is the number of points per interior wavelength, the number of nodes per window half-width, a curvature coefficient, and . The three terms resolve, respectively, the shortest wavelength, the window, and the boundary curvature; provides about nodes per smallest radius of curvature when the nodes are approximately equispaced in arc length. To find a good set of constants , a grid search is done on various cases, and it is observed that the constants , and lead to a maximum relative error of or so. Tables 3 and 4 list the , and values obtained using these constants for various cases, along with maximum relative errors. The error at the predicted ranges from to . Where the rule is off, it is conservative for the curvature- and wavelength-limited cases: the silicon cylinder reaches already at instead of 270, and the deeper-neck rose at instead of 470. It underestimates by 23–32% for the two large windowed cases, which reach at and 1024 instead of 524 and 786. The small resonant cylinders are resolved at the minimum , with errors of and . Proximity to a resonance therefore did not increase the required in these tests, although it does raise . Because this is an empirical guide, we recommend a doubling test on a quantity other than the energy residual. In all cases of Table 3, the energy residual at the predicted is below , irrespective of the actual error, which ranges over four orders of magnitude. The same root-exponential law describes the three windowed cases of Table 3, which include two non-circular boundaries and a dielectric with , with –1.60. Note that the constants and were chosen from a small benchmark set, and they should be regarded as an empirically calibrated criterion, not an error guarantee.
Conclusion
The spectral integral method has been extended from to scattering by homogeneous two-dimensional objects with complex permittivity, which makes it applicable to lossy dielectrics and to metals that support surface-plasmon resonances. The extension revealed that the original singularity subtraction, which is accurate for small or lossless objects, loses accuracy rapidly as the object becomes larger or more absorbing, because of numerical cancellation rather than insufficient resolution. The proposed windowed singularity subtraction removes this limitation while keeping the simplicity and FFT-based structure of the method, and it is applied only when the object is large enough to need it. Comparisons with the Mie series show that the method reaches errors below for gold cylinders that are several wavelengths in size, including cases in which the original subtraction returns completely wrong results. For a non-convex gold cylinder, the results agree with an independent FDTD solver to within 0.4%. An empirically calibrated rule for choosing the number of boundary points is also provided, so that users can obtain this level of accuracy without extensive trial runs and oversampling.
Data Availability
All the code to reproduce the results and figures shared in this manuscript can be found at https://github.com/simsekergun/SIM_TE.
Propositions
Proposition 1. With the convention and , let and be principal. Then for every value of , and , for all . Consequently, the expansions (13) and (14) of the manuscript, which hold for , never meet the branch cut, and the outgoing-wave factor of never grows: it decays exponentially when and is oscillatory in the lossless dielectric limit.
Proof. means , and the principal root halves the argument. Multiplying by does not change the argument.
Proposition 2. Let be real-analytic in a neighborhood of the diagonal and let be . Then near the diagonal.
Proof. Since , smoothness of requires to vanish to infinite order on the diagonal; otherwise the leading term is not . A real-analytic function that vanishes to infinite order at a point is identically zero in a neighborhood.
:::
A window that removes the logarithmic singularity and is real-analytic near the diagonal must therefore be identically one there, so a localized window is necessarily non-analytic. The proposition does not fix the convergence rate: windows of different regularity classes give different rates, and the analysis below applies to the Gevrey class-2 window of (26). Without the window (), the logarithmic singularity remains in the single-layer kernel, even for an analytic boundary. However, after analytically separating the universal logarithmic singularity, the remaining kernel is analytic for an analytic boundary. Fourier-based methods that treat the logarithmic part analytically can therefore achieve exponential convergence, provided the resulting discretization is stable [1, 3, 19]. The error then decays as until it reaches the floor (25). With the window, the regularized kernels are products of , and with functions that are analytic in for an analytic boundary. The individual factors and are not analytic across , because is not, but the combinations that occur in the kernels are. and are entire functions of , which is analytic in ; and are analytic, the latter because both numerator and denominator vanish to second order at ; the regular parts of and , including the pole term of , are therefore analytic after multiplication by the geometric factor of Eq. (4); and
is analytic because the last ratio is analytic and positive. The window built from belongs to the Gevrey class of order 2 [20], and Gevrey classes are closed under multiplication by analytic functions. The regularized kernels are therefore Gevrey-2, and their Fourier coefficients decay root-exponentially, as [20, 21, 22]. Equations (11)–(12) couple double-layer (second-kind) and single-layer (first-kind) operators with logarithmic kernels; they do not form a pure second-kind system. For such systems, trigonometric collocation and quadrature methods converge when the discrete system is uniformly stable, and the solution error is then bounded by a constant, related to the norm of the discrete inverse, times the kernel approximation error [18]. We assume this stability for the passive materials considered here. It degrades near the quasi-static plasmon pole, with vanishing loss, where the condition number grows as the loss decreases (§ 3.6). Under this assumption, the windowed method is super-algebraically convergent, and root-exponential convergence is expected.
Cross sections and energy balance
For , the field in the background is given by
Since for in-plane gradients, the time-averaged Poynting vector satisfies
The incident intensity is with , and .
The absorbed power per unit length is the net inward flux through , evaluated on the exterior side with the total field: . Dividing by with gives , which is Eq. (32). By Poynting's theorem in the interior, for passive media with the convention, so the sign is consistent: in all gold computations.
On a circle , . With the asymptotic form of , , so that
The extinguished power is minus the flux of the interference term, , through a large circle. Using and evaluating the angular integral by stationary phase at , where the forward contribution survives, gives the 2-D optical theorem . With the Mie coefficients one has , and Eqs. (31)–(32) reduce to the familiar
Because , and are computed from three different formulas, the balance is a consistency check of the convention. In Figure 3, and are additionally verified individually against the Mie series.
Implementation notes
The method is implemented in MATLAB (version R2024b) on a 2021 Apple MacBook Pro with 32 GB of RAM. The computations use double precision throughout.
, and are evaluated analytically from the parametrization on equispaced nodes, with even. , and the signed follow from Eqs. (15)–(19).
For each testing point , , , (set to 0 at ) and are computed. for the background and when , with the largest distance between two nodes. This sampled value can only underestimate the largest chord by at most about , and it is exact for a circle with even . It therefore only shifts the switch slightly near , where both splittings are accurate. , (MATLAB’s
besseljfunction) are evaluated only where , and , (MATLAB’sbesselhfunction) everywhere. The self-terms, Eqs. (18) and (19), are inserted at , with , there. No special treatment of small or large complex arguments is needed beyond the library routines, because the window keeps wherever is evaluated.
The products of with and are evaluated as discrete convolutions with , using zero-padded FFTs of length .
The two dense systems of Eq. (24) are solved with MATLAB’s backslash (LU with partial pivoting), without scaling or preconditioning.
The Mie series is truncated at , with .2 Internal references for non-circular shapes use –, as stated in each table.
The code and the scripts reproducing all figures and tables are available at https://github.com/simsekergun/SIM_TE.
References
- [1]R. Kress, “Boundary integral equations in time-harmonic acoustic scattering,” Math. Comput. Modelling, vol. 15, pp. 229–243, 1991.DOI
- [2]F. Q. Hu, “A spectral boundary integral equation method for the 2D Helmholtz equation,” J. Comput. Phys., vol. 120, pp. 340–347, 1995.DOI
- [3]L. N. Trefethen and J. A. C. Weideman, “The exponentially convergent trapezoidal rule,” SIAM Rev., vol. 56, no. 3, pp. 385–458, 2014.DOI
- [4]J. Liu and Q. H. Liu, “A spectral integral method for periodic and nonperiodic structures,” IEEE Microw. Wireless Compon. Lett., vol. 14, no. 3, pp. 97–99, 2004.DOI
- [5]E. Simsek, J. Liu, and Q. H. Liu, “A spectral integral method (SIM) for layered media,” IEEE Trans. Antennas Propag., vol. 54, no. 6, pp. 1742–1749, 2006.DOI
- [6]Q. H. Liu, Y. Lin, J. Liu, J. H. Lee, and E. Simsek, “A 3-D spectral integral method for surface integral equations,” IEEE Microw. Wireless Compon. Lett., vol. 19, no. 2, pp. 62–64, 2009.DOI
- [7]K. B. Crozier, E. Togan, E. Simsek, and T. Yang, “Experimental measurement of the dispersion relations of the surface plasmon modes of metal nanoparticle chains,” Opt. Express, vol. 15, pp. 17482–17493, 2007.pubmed.ncbi.nlm.nih.gov/19551041
- [8]E. Simsek, “Full analytical model for obtaining surface plasmon resonance modes of metal nanoparticle structures embedded in layered media,” Opt. Express, vol. 18, no. 2, pp. 1722–1733, 2010.DOI
- [9]R. F. Oulton, V. J. Sorger, D. A. Genov, D. F. P. Pile, and X. Zhang, “A hybrid plasmonic waveguide for subwavelength confinement and long-range propagation,” Nature Photon., vol. 2, pp. 496–500, 2008.
- [10]O. P. Bruno and A. Sei, “A fast high-order solver for EM scattering from complex penetrable bodies: TE case,” IEEE Trans. Antennas Propag., vol. 48, no. 12, pp. 1862–1864, Dec. 2000.DOI
- [11]O. P. Bruno and E. M. Hyde, “Higher-order Fourier approximation in scattering by two-dimensional, inhomogeneous media,” SIAM J. Numer. Anal., vol. 42, no. 6, pp. 2298–2319, 2005.DOI
- [12]M. V. Balaban, E. I. Smotrova, O. V. Shapoval, V. S. Bulygin, and A. I. Nosich, “Nystrom-type techniques for solving electromagnetics integral equations with smooth and singular kernels,” Int. J. Numer. Model., vol. 25, no. 5–6, pp. 490–511, 2012.DOI
- [13]Q. Sun, E. Klaseboer, A. J. Yuffa, and D. Y. C. Chan, “Field-only surface integral equations: scattering from a dielectric body,” J. Opt. Soc. Am. A, vol. 37, no. 2, pp. 284–293, 2020.DOI
- [14]J. Helsing and A. Karlsson, “On a Helmholtz transmission problem in planar domains with corners,” J. Comput. Phys., vol. 371, pp. 315–332, 2018.DOI
- [15]O. P. Bruno and L. A. Kun yansky, “A fast, high-order algorithm for the solution of surface scattering problems: basic implementation, tests, and applications,” J. Comput. Phys., vol. 169, pp. 80–110, 2001.DOI
- [16]M. Abramowitz and I. A. Stegun, Eds., Handbook of Mathematical Functions. Washington, DC: National Bureau of Standards, 1964.
- [17]“NIST digital library of mathematical functions,” https://dlmf.nist.gov, sec. 10.8.
- [18]R. Kress, Linear Integral Equations, 3rd ed. New York: Springer, 2014.
- [19]J. Galkowski and M. Zworski, “Outgoing solutions via Gevrey-2 properties,” Ann. PDE, vol. 7, no. 5, 2021.arxiv.org/abs/2004.07868
- [20]L. Rodino, Linear Partial Differential Operators in Gevrey Spaces. Singapore: World Scientific, 1993.DOI
- [21]S. G. Johnson, “Saddle-point integration of C∞ “bump” functions,” arXiv:1508.04376, 2015.arxiv.org/abs/1508.04376
- [22]J. P. Boyd, Chebyshev and Fourier Spectral Methods, 2nd ed. Mineola, NY: Dover, 2001.
- [23]C. F. Bohren and D. R. Huffman, Absorption and Scattering of Light by Small Particles. New York: Wiley, 1983, ch. 8.
- [24]M. N. Polyanskiy, “Refractiveindex.info database of optical constants,” Sci. Data, vol. 11, p. 94, 2024.DOI
- [25]P. G. Etchegoin, E. C. Le Ru, and M. Meyer, “An analytic model for the optical properties of gold,” J. Chem. Phys., vol. 125, p. 164705, 2006, erratum: vol. 127, 189901, 2007.pubmed.ncbi.nlm.nih.gov/17092118
- [26]P. B. Johnson and R. W. Christy, “Optical constants of the noble metals,” Phys. Rev. B, vol. 6, pp. 4370–4379, 1972.DOI
- [27]Flexcompute Inc., “Tidy3D FDTD solver, version 2.12.1,” https://www.flexcompute.com/tidy3d/.