WorksPhysics and Astronomy
Low-Rank Iteration for Efficient Calculations of Topological Surface States
Abstract
Surface states are hallmarks of topological phases, and their efficient computation is crucial for topology characterization. Conventional supercell methods (SCM) approximate semi-infinite systems as finite periodic ones, but suffer from high computational cost, artificial inter-surface coupling, and slow convergence near band edges. Recently, boundary problems for semi-infinite systems have been recast as block-tridiagonal equations and solved by the cyclic reduction method (CRM) and the transfer matrix method (TMM). However, these approaches remain resource-intensive, as fine-mesh discretization produces very large matrices. Recognizing the intrinsic sparsity and low rank of these matrices, we develop a low-rank iterative method (LIM) with stochastic trace estimation that substantially reduces complexity and memory without sacrificing accuracy, while remaining straightforwardly parallelizable. Benchmarks on 2D and 3D topological systems show orders-of-magnitude acceleration over CRM and TMM. LIM extends readily to heterostructures, defective interfaces, and non-Hermitian systems, providing an efficient platform for topological and plasmonic surface phenomena.
main_LIM-SDOS_NCS
*Correspondence: [email protected] (L.Z.) and [email protected] (S.M.)
Introduction
Artificially engineered materials, including photonic crystals, phononic crystals, and metamaterials, provide versatile platforms for exploring topological physics[1, 2, 3, 4, 5, 6, 7, 8, 9, 10]. As a central feature, topologically protected surface states enable robust wave transport in devices such as low-loss waveguides and unidirectional components[11, 12, 13], and their dispersion and propagation determine key device metrics including operational bandwidth and group velocity[14, 15, 16]. Efficient, accurate determination of the surface-state spectrum at a single boundary is therefore crucial for both topological physics and device engineering.
A common approach is the supercell method (SCM), which treats a semi-infinite system as a finite periodic slab and solves its eigenvalue problem for the surface-state dispersion. Suppressing coupling between the two artificial surfaces requires a thick supercell, incurring substantial cost, particularly for finely discretized 3D structures. However, the decay length diverges near band edges, so residual coupling persists at any thickness. Moreover, surface states constitute only a small fraction of the computed bands, embedded among numerous bulk states, so identifying them is nontrivial[17]. A more direct route is the surface Green’s function of a partially infinite system[18, 19, 20, 21, 22, 23], whose trace defines the surface density of states (SDOS) and links surface states to surface-source excitation spectra. For electronic systems, iterative[24, 25, 26], recursive[27, 28], and transfer-matrix[29, 30, 31, 32] methods compute the local SDOS directly on the block-tridiagonal Toeplitz matrices of tight-binding models, without assembling the full infinite matrix. In contrast, photonic and phononic systems generally exhibit more complex geometries and materials, and require finite-element discretization, producing far more degrees of freedom (DOF) than their electronic counterparts[33, 34, 35, 36]. Finite-element block-diagonal formulations recently extended the cyclic-reduction method (CRM) and transfer-matrix method (TMM) to such systems[37, 38]. Nevertheless, both costs grow as a power law in the DOFs per unit cell, posing a bottleneck for high-dimensional, finely discretized systems. Thus, applying mature electronic schemes directly to photonic and phononic systems is often prohibitive.
Here, recognizing that the discretization matrices are intrinsically sparse and that the inter-cell coupling matrices are low-rank, we propose a low-rank iterative method (LIM) that overcomes the main cost sources of CRM and TMM for complex photonic and phononic systems (see Figure 1). For a uniformly discretized unit cell with DOFs, the rank of the coupling matrix scales roughly as and in 2D and 3D, far smaller than . LIM reformulates the surface Green’s function as an iteration of a low-rank Riccati equation restricted to a subspace set by , avoiding operations on full matrices. The SDOS is then evaluated with Hutchinson’s stochastic trace estimator, without explicitly forming the dense Green’s function. These low-rank features substantially improve efficiency, enabling mildly super-linear scaling with finite-element DOFs. LIM also applies to systems formed by connecting multiple semi-infinite regions. We demonstrate this on bare and coated crystals, heterojunctions, sandwich structures, and 3D Weyl metamaterials, computing the SDOS and, where applicable, iso-frequency surfaces. Across these geometries, LIM resolves single-interface topological response, eliminates the mixing of opposite-surface states inherent to the SCM, and reduces memory and computation time. It extends directly to non-Hermitian systems, higher-order corner states, and inhomogeneous interfaces, thus offering an efficient, robust tool for topological device design.

Figure 1. The Low-Rank Iterative Method. (a) Schematic of a representable semi-infinity array structure. (b) Each cell is represented by a sparse block matrix coupled by low-rank sparse coupling matrices and in FEM, with blocks of the Green's function satisfying . (c) Schematic of low-rank QR factorization in LIM. The low-rank nature of ensures that the QR decomposition yields and containing only a finite number of orthonormal bases, thereby reducing the complexity of the iteration.
Results
Surface Green’s Function and Density of States
In the frequency domain, the source-free Maxwell equations yield an eigenvalue equation for the electric field , as does the fluid phononic wave equation for the acoustic pressure . Despite their distinct physical origins, both share a unified generalized-eigenvalue form with as the eigenvalue. The Green’s functions of the photonic Eq. (1) and phononic Eq. (2) systems are the responses of the corresponding wave operators to point-source excitations [38]:
where , , , and denote the permeability, permittivity, mass density, and bulk modulus at position . Denoting the differential operator and material terms by , both equations share the same form: . Physically, by introducing a small positive imaginary component into , the imaginary part of the Green’s function reflects the response intensity to an excitation at . Topological systems exhibit distinctive boundary properties, for which the SDOS spectral distribution is the key quantity. The SDOS follows from the surface Green’s function for a surface excitation:
Here, denotes the metric term, which is for photonic systems and for phononic systems.
The topological systems studied are periodic or quasiperiodic normal to the interface and semi-infinite along it, with identical bulk parameters and translationally invariant inter-cell coupling (see Figure 1a-Figure 1b). Eqs. (1) and (2) then take the same block-tridiagonal Toeplitz form after finite-element discretization:
Here, the diagonal blocks describe intra-cell coupling, and the off-diagonal blocks and represent inter-cell coupling. All blocks are of size for a unit cell discretized with DOFs and encode the geometry, materials, and periodicity information. The imaginary part of the trace of yields the SDOS for the boundary excitation . Note that can be any boundary excitation, not necessarily constant, to target a specific surface mode, e.g., one with a given polarization.
Boundary modes are conventionally obtained by truncating the semi-infinite to a finite number of cells in the SCM and diagonalizing the resulting finite matrix, which exploits only the sparsity of and tolerates cell-to-cell differences. Instead, CRM and TMM compute directly for a truly infinite system[37, 38]. CRM iteratively evaluates approximately by eliminating odd-layer Green’s functions via Gaussian elimination until the effective inter-layer coupling decays exponentially. By contrast, TMM relates adjacent-layer Green’s functions via a transfer matrix and, keeping only decaying eigenmodes, reduces the problem to a linear relation between surface and bulk Green’s functions that yields exactly. However, both approaches suffer from computational drawbacks that limit their efficiency: CRM destroys sparsity and iterates on dense matrices of order , whereas TMM obtains the transfer matrix in one step but expands the problem to order , making TMM slower.
Low-Rank Iterative Method and Stochastic Trace Estimation
As an illustration, we study a bare semi-infinite 2D photonic crystal: a square-lattice gyromagnetic crystal terminated at the top by a perfect electric conductor (PEC) boundary (see Figure 2a). The gyromagnetic term breaks time-reversal symmetry, yielding a Chern insulator with a nontrivial Chern number that guarantees unidirectional boundary states[39]. The unit cell is discretized as in Figure 2b. The governing equation is cast into matrix form , with nodal positions identifying the block rows and columns corresponding to , , and in Eq. (4). Once the FEM preprocessing is complete, LIM is applied to calculate the SDOS.

Figure 2. SDOS on a Semi-infinite Photonic Crystal. (a) Schematic of the semi-infinite photonic crystal with a PEC boundary at the top. (b) Discretized unit cell mesh by FEM, where nodes inside the unit cell and those shared between unit cells (highlighted in orange) contribute to the intra-cell and inter-cell terms, corresponding to the block matrix and the coupling matrices and in Figure 1b, respectively. (c) Band structure of a 9-cell supercell obtained by SCM, with surface states located near the upper and lower PEC boundaries highlighted in orange and blue, respectively. (d-f) SDOS calculated by CRM, TMM, and LIM, respectively. (g) The distribution of the eigenstate obtained by SCM and of the Green’s function solved by LIM, respectively.
Mathematically, solving for the Green’s function vector amounts to inverting the infinite block-tridiagonal Toeplitz matrix :
Here, and . The SDOS requires neither the full matrix nor itself, but only its trace. By translational invariance, the Green’s function blocks satisfy , where is the transfer matrix. With the spectral radius of , i.e., the largest eigenvalue modulus[31], the condition holds in the bandgap, thereby guaranteeing decay away from the boundary[38]. Thus, the transfer matrix is exactly the stabilizing solution of the block Riccati equation:
Given , the surface Green’s function satisfies:
Mathematically, the SDOS amounts rigorously to solving the Riccati equation rather than inverting . This formulation is fully equivalent to the TMM solution[29, 30, 31, 32, 38]. TMM linearizes the quadratic Riccati equation with auxiliary variables, converting it into a linear eigenvalue problem, but this doubles the eigenvalue-matrix dimension and raises the sparse-inversion cost to , thereby substantially reducing efficiency.
To improve efficiency, LIM exploits the low-rank and sparse structure of the finite-element blocks, using a low-rank factorization with an iterative scheme to solve the Riccati equation. In finite-element discretization, the coupling matrices and between adjacent cells are set by the mesh connectivity, specifically by the interface DOFs marked orange in Figure 2b. For shared DOFs, the coupling matrices involve only these DOFs, so their ranks are bounded by : . For a Hermitian system , the deviation arises only from the small imaginary part added to for SDOS calculations, hence . For the structure in Figure 2b with , the ranks are only , far smaller than . We factorize the coupling matrices through low-rank sparse QR decompositions:
where , with . The columns of form orthonormal bases satisfying , and satisfies (see Figure 1c).
Since the stable solution of the Riccati equation describes the exponentially decaying transfer driven by , its column space lies in the right singular subspace of . Physically, is governed by only a few interface-related modes, reducing to a lower-dimensional auxiliary factor : , where contains the interface-related components and is the projected auxiliary factor. Accordingly, satisfies the self-consistent equation:
All SDOS and field-distribution calculations depend only on the low-rank factor (derivation in Supplementary Section I). Setting , the self-consistent equation is solved by the iteration:
where and are constant matrices that can be precomputed. For a topological system, the iteration converges to a unique stabilizing solution and terminates when for a prescribed tolerance . During the iteration, all operators are restricted to subspaces of dimension and , avoiding explicit manipulation of dense matrices, which is the key to the mildly super-linear scaling of LIM.
Once is determined, the surface Green’s function can be explicitly derived from Eq. (7) using the Sherman–Morrison–Woodbury formula[40, 41]:
Compared with the single-layer solution , the second term acts as a correction accounting for the semi-infinite extension. Left-multiplying by and taking the imaginary part of its trace directly yields the SDOS. The poles are fully governed by , while acts only as a location-dependent weight and does not affect dispersion spectra. Thus, can be neglected when identifying boundary states, which is convenient when is a local matrix difficult to extract. A comparison with and without is given in Section II of the Supplementary Material.
A remaining challenge is computing exactly, which requires inverting . Although most matrices are already available from preceding iterations, is an sparse matrix, and explicitly forming its inverse generally destroys sparsity, with worst-case time and memory complexities of and . Explicit inversion is therefore prohibitive for large sparse problems. To improve efficiency further, we adopt a Hutchinson stochastic trace estimator[42, 43] for fast approximation of the SDOS:
where are independent Rademacher random vectors satisfying .
is approximated by averaging quadratic forms over random vectors (), instead of all diagonal entries, with statistical error decreasing as . This requires only the action of on random vectors: the product is obtained by solving via sparse LU back-substitution. Because is sparse, this back-substitution scales with the sparsity of its LU factors. The same mechanism computes . Since both and are much lower than , these operations are far cheaper than forming explicitly. With typically evaluations, the SDOS spectral lines are reproduced with high fidelity.
We validate LIM against SCM, CRM, and TMM, as shown in Figure 2. For SCM, the spectrum of a semi-infinite photonic crystal is approximated by diagonalizing a 9-cell supercell with PEC boundaries at both ends. Surface states protected by the nontrivial Chern number emerge and localize at both boundaries, with the corresponding bands highlighted in orange (upper) and blue (lower) in Figure 2c. Meanwhile, Figure 2d–Figure 2f show the SDOS of an ideal semi-infinite system from CRM, TMM, and LIM, respectively, with LIM agreeing with both while being faster. The SDOS renders bulk states continuous but surface states discrete. A forward-propagating surface state crosses the gap, while the backward-propagating modes coexisting in SCM vanish, underscoring the advantage of the ideal single interface.
Beyond the spectrum, LIM also yields field information that maps onto experimentally accessible boundary excitation fields. Unlike the spectrum, which requires only the trace of , the field calculation requires the full and the transfer matrix . Both are explicit in terms of . Once is available, all propagate layer by layer via , yielding the boundary-excited fields. Figure 2g compares the SCM eigenstate with the LIM Green’s function, showing excellent agreement and confirming field localization near the boundary.
Numerical Examples for Various Topological Systems
LIM is not restricted to 2D bare systems. For systems with cladding layers, heterojunctions, sandwich structures, or more complex 3D geometries, the mathematical structure is similar. The low-rank iteration solves the Riccati equation in each semi-infinite region in the same way. The interface coupling gives expressions for analogous to Eq. (7), and the stochastic trace estimator then yields the SDOS efficiently. We further demonstrate LIM with five representative examples: Figure 3 summarizes the dispersions of four distinct 2D topological systems, and Figure 4 presents a 3D Weyl system with its dispersion and iso-frequency surfaces. The formulation procedure and pseudocode are in Section III of the Supplementary Material.

Figure 3. Topological Surface States in different systems. (a-c) Semi-infinite photonic Chern insulator coated with a dielectric plate and terminated by a PEC boundary: (a) schematic, (b) band structure of a supercell obtained by SCM, and (c) SDOS spectrum calculated by LIM, with the representative field distribution shown in (a). (d–f) Heterojunction formed by joining two semi-infinite valley Chern insulator photonic crystals. (g–i) Hexagonal-lattice photonic crystal containing a single defect layer. (j–l) Phononic topological insulator formed by splicing two semi-infinite crystals. Panels (d–l) follow the same layout as (a–c). In the SCM results, modes at the target interface are marked in orange, while those at the other boundaries are marked in blue.

Figure 4. SDOS on a Semi-infinite 3D Weyl Metamaterial. (a) Schematic of the saddle-shaped metamaterial, which behaves as an ideal Weyl system, with PEC boundary conditions at both the top and the bottom. (b) Band structure of a 9-cell supercell obtained by the SCM. (c, d) SDOS spectra calculated by LIM, with PEC boundaries at the top and bottom, respectively. (e) Evolution of the surface Fermi arcs in the SDOS equifrequency contours on the top surface from 10 to 12 GHz.
The first 2D example is a coated Chern insulator, identical to Figure 2 except that a cladding layer is inserted between the upper interface and the PEC boundary, and every third cell along the boundary is enlarged to fold the Brillouin zone to one third of its size (see Figure 3a–Figure 3c). SCM bands for a 9-cell supercell and the LIM SDOS are shown in Figure 3b–Figure 3c. Bragg scattering complicates the bands and hinders mode selection in SCM. Nevertheless, the SDOS contains only clean single-interface topological states with positive group velocity, localized near the upper surface and matching the orange bands in Figure 3b.
The second 2D example is a heterojunction of two semi-infinite valley Chern insulator photonic crystals (see Figure 3d–Figure 3f). Opposite rod orientations give the two sides distinct valley Chern numbers, so valley boundary states propagate bidirectionally along the junction. Figure 3e–Figure 3f show SCM bands for a 19-cell supercell and the LIM SDOS. Two boundary bands with opposite group velocities appear within the valley gap, in excellent agreement with SCM.
The third 2D example is a hexagonal-lattice photonic crystal with a single defect layer, combining semi-infinite crystals and a sandwich layer (see Figure 3g–Figure 3i). Triangular air holes of varying size and orientation are etched in a dielectric slab. Unequal inclusion sizes in the upper and lower crystals break spatial symmetry and open a gap near the Dirac point, whereas equal-size inclusions in the middle layer restore the degeneracy locally, producing localized modes that propagate bidirectionally. SCM bands for a 19-cell supercell and the LIM SDOS are in Figure 3h–Figure 3i. The two agree well, and LIM confines the results to the single interface.
The last 2D example is a phononic topological insulator, formed by joining two semi-infinite hexagonal-lattice phononic crystals with different inclusion radii (see Figure 3j–Figure 3l). As in the photonic case, once the matrix is obtained by FEM, the SDOS follows the same procedure. SCM surface bands for a 9-cell supercell and the LIM SDOS are shown in Figure 3k–Figure 3l. Two pseudospin modes with opposite group velocities form interface-state bands traversing and connecting the band gap, again in excellent agreement.
The preceding examples are 2D, with scalar fields discretized by Lagrange finite elements with node-based DOFs. The final example extends LIM to the SDOS and constant-frequency surfaces of a 3D semi-infinite Weyl metamaterial (Figure 4). In 3D problems with curl operators, Nédélec edge elements are used, with low-order DOFs defined by tangential integrals along mesh edges[44]. Despite these differences, LIM still exploits the sublinear rank of the interface-related matrices. The 3D ideal Weyl metamaterial[45] consists of chiral metallic elements, with a PEC-terminated surface. Figure 4b shows SCM surface bands for a 9-cell supercell, with top and bottom surface states in orange and blue. Figure 4c-Figure 4d show the LIM SDOS for the two interfaces, examined separately through the boundary conditions, in excellent agreement with SCM. Figure 4e shows the SDOS equifrequency contours on the top surface from 10 to 12 GHz: the mode projections of the two oppositely chiral Weyl points are always linked by Fermi-arc surface states, in excellent agreement with Weyl-topology predictions.
Computing Efficiency and Accuracy
This section assesses the efficiency and accuracy of LIM, which mainly consists of two steps: solving by low-rank iteration and evaluating by a stochastic trace estimator.
We consider a Hermitian system with low-rank factors satisfying . The Riccati iteration for runs entirely in the low-rank subspace: each step involves only matrices and costs for iterations. Since , the main computational cost arises from initialization and applications of . The low-rank QR decompositions of and cost approximately , while the -related initialization comprises a sparse LU factorization with cost , where denotes the number of LU nonzeros. With optimal nested dissection, scale as in 2D and in 3D[46, 47]. Thus, the precomputation of cost , and the low-rank projections of cost , totaling for initialization. The same -related operations dominate stochastic trace estimation: each of the random vectors, typically , requires computing by sparse LU back-substitution costing plus matrix–vector products. Thus, the overall complexity is thus . Because -dependent terms have lower power in and small pre-factors, the cost is dominated by sparse -related operations, scaling as .
For concreteness, we measure the actual cost. For different DOFs, we compute the SDOS of the system in Figure 2a at 51 frequencies using LIM, LIM-invA, CRM, and TMM, recording peak memory and time. Here, LIM-invA solves by low-rank iteration and evaluates directly, without stochastic trace estimation. The comparison thus isolates the benefit of stochastic trace estimation over direct computation. The gain has a clear structural origin (see Figure 5a,Figure 5b): the boundary DOFs , the coupling rank , and the nonzeros of and all grow slowly with . Specifically, the -dependent terms and grow super-linearly with as in 2D, but their small pre-factors keep sparse operations dominant. The bottleneck is thus the LU factors, which grow as , only moderately steeper than the expected for optimal nested dissection. Consequently, the empirical memory and time complexities are only and (see Figure 5c,Figure 5d). By contrast, CRM scales as in memory and in time, whereas TMM requires memory and time. LIM is thus substantially more efficient for large-scale 2D systems.

Figure 5. Computing Efficiency and Accuracy of the LIM Algorithm. (a, b) Number of the and , and of nonzero matrix elements of and , in the 2D system, each fitted by a power function. (c, d) Average memory usage and computational time versus DOFs, obtained at 51 frequency points using LIM, LIM-invA, CRM, and TMM. For each algorithm, the growth rate with is fitted to a power law and marked to the right of the curve. (e) Error analysis of the stochastic trace estimator: RMSE versus the number of random vectors , decreasing as . Insets show the SDOS versus frequency at for different , illustrating the accuracy of the calculation.
For 3D systems, the LU factorization suffers far more fill-in, so each application of is more expensive and the gain of LIM is somewhat smaller than in 2D, though still substantial. A rigorous asymptotic analysis is not pursued because CRM and TMM become prohibitively costly. Instead, we benchmark a 3D structure with 4,643 DOFs, averaging over 20 independent runs at different frequencies (Table 1). Compared with CRM, LIM consumes about 1/17 of the memory and 1/15 of the time; relative to TMM, about 1/16 of the memory and only 1/184 of the time. LIM thus retains a clear 3D advantage, reducing both memory and time by more than one order of magnitude.
| Target Function | LIM | LIM-invA | CRM | TMM |
| Avg Time (s) | 2.5974 | 11.8334 | 38.1799 | 478.5181 |
| Time Ratio | 1 | 4.56 | 14.7 | 184.23 |
| Workspace Memory (MB) | 130.1375 | 582.0454 | 2214.2839 | 2032.0552 |
| Workspace Memory Ratio | 1 | 4.47 | 17.01 | 15.61 |
Table 1. Computational cost comparison of the four target functions in a 3D semi-infinite Weyl system with DOFs 4643. Time and workspace memory ratios are normalized to the LIM baseline.
Finally, we remark on the accuracy of LIM. Indeed, Figure 2–Figure 4 already demonstrate remarkable accuracy despite high efficiency. Once the low-rank iteration converges, accuracy is fully determined by the prescribed tolerance and can be regarded as numerically exact. For stochastic trace estimation, accuracy improves with the number of random vectors : as shown in Figure 5e, the root-mean-square error (RMSE) decreases as with increasing . The qualitative SDOS features are usually of primary interest, and already suffices for representative surface states. Increasing further reduces the statistical error, approaching explicit evaluation, but the cost grows approximately linearly with . A trade-off between accuracy and efficiency is therefore required.
Discussion
In summary, LIM efficiently computes the SDOS of semi-infinite topological photonic and phononic systems. Exploiting the low-rank and sparse structure of FEM discretization, it resolves the surface response by low-rank Riccati iteration and evaluates the SDOS by stochastic trace estimation without constructing large-scale matrices. Operating directly on the Green’s function, it avoids the boundary-coupling issues of SCM. LIM matches Green’s-function-based CRM and TMM in accuracy while substantially reducing memory and time, with advantages growing with DOFs and making it practical for large-scale topological systems.
Beyond photonic and phononic systems, LIM holds promise for electronic-structure calculations with localized bases such as numerical atomic orbitals (NAOs)[48] and maximally localized Wannier functions[49]. For complex structures that are challenging for conventional electronic-structure methods, such as large supercells and quasi-one-dimensional nanoribbons, the Hamiltonian retains the same sparse, low-rank block-tridiagonal form, so low-rank Riccati iteration could compute the embedding self-energies of semi-infinite leads without constructing the full device Hamiltonian, providing an efficient alternative to recursive Green’s function methods.
Methods
Simulation environment
All simulations reported in this work, both the numerical examples and the performance benchmarks, were carried out with COMSOL Multiphysics, controlled through a MATLAB scripting interface. COMSOL serves here primarily as a mesh generator; other open-source finite-element meshing packages could fulfill the same role. In a typical run, a supercell is constructed in COMSOL, and its finite-sized eigenmatrix is exported to MATLAB. There, is split into the diagonal self-coupling blocks and the off-diagonal inter-coupling blocks and , which LIM, CRM, and TMM all act on. Since every method is applied to the same geometry and the same mesh, the performance metrics, including computational time, memory usage, and accuracy, are compared under strictly identical conditions, so the comparison is rigorous and free of bias. Timing and memory were measured on a personal computer with an Intel Core i7-13700 CPU (16 cores, 24 threads, 2.10 GHz).
Details of numerical examples
All structures were discretized with the finite-element method. The specific structural and material settings are provided in the simulation files archived in the open data repository[50].
The bare photonic crystal in Figure 2 is based on a square lattice with lattice constant , and its unit cell contains a cylindrical inclusion of radius with relative permittivity and relative permeability , . This gyromagnetic response breaks time-reversal symmetry and yields the Chern-insulator phase. The coated crystal in Figure 3a–Figure 3c shares this geometry, except that an additional air layer of thickness separates the upper interface from the PEC boundary. Along this boundary, every third boundary unit cell is slightly enlarged, with its inclusion radius raised to , while the rest keep the original . This enlargement folds the Brillouin zone to one third of its original size.
The valley-Chern heterojunction in Figure 3d–Figure 3f joins two semi-infinite photonic crystals of triangular dielectric rods with side length , chamfer length , relative permittivity and relative permeability . Opposite rod orientations give the two sides distinct valley Chern numbers. The defect-layer structure in Figure 3g–Figure 3i consists of a dielectric slab with relative permittivity and permeability , perforated by triangular air holes of side length and in the upper and lower crystals, respectively, with equal-size holes in the middle layer. The phononic crystal in Figure 3j–Figure 3l joins two hexagonal-lattice acoustic crystals whose cylindrical inclusions have radii and .
The 3D Weyl metamaterial in Figure 4 consists of saddle-shaped chiral metallic elements embedded in a PCB substrate with relative permittivity , with in-plane periods and out-of-plane period . The element geometry is specified by , , and . Both the top and bottom surfaces are terminated by PEC boundaries.
Acknowledgements
Funding: This work was supported by National Key Research and Development Program of China (Grants No. 2023YFA1407700, 2023YFA1406901, and 2022YFA1404700), the Key Program of the National Natural Science Foundation of China (No. 12634017), National Natural Science Foundation China (Grants No. 12374343, 12221004, 62192771), and the start-up funding of Fudan University (JIH1232098Y). We also thank Dr. Yixin Sha for helpful discussions.
Author contributions: S.M. conceived the initial concept, K.L., C.H., Z.X. and S.M. performed numerical simulations and analytical calculations, S.S., L.Z. and S.M. participated in the analysis of the results, and S.M. and L.Z. supervised the project.
Competing interests: The authors declare no conflict of interest.
Data and materials availability: All data and source code that support the findings of this study are openly available in the Science Data Bank[50] and via GitHub at https://github.com/KaiqiLiu-LKQ/Low-Rank-Iteration-Method-for-Topological-Surface-States.
Supplementary Materials:
Correspondence: [email protected] (L.Z.) and [email protected] (S.M.).
SI_LIM-SDOS_NCS
This file includes:
Detailed Derivation of the LIM Algorithm
Discussion of the Metric in the SDOS calculation
Pseudocode of the LIM for representative partially infinite systems
Detailed Derivation of the LIM Algorithm
Starting from the block Riccati equation given in Eq. (6) of the main text, we have:
To proceed, we exploit the low-rank structure of the coupling matrices and . Specifically, these matrices can be factorized as:
Substituting Eq. (14) into Eq. (13) and rearranging the resulting terms yields:
Right-multiplying both sides of Eq. (15) by , we obtain:
Here, the common factor is invertible and can therefore be eliminated completely, yielding:
Next, left-multiplying Eq. (17) by and extracting , we arrive at:
From this, we can get:
Reversing the steps from Eq. (19) back to Eq. (17) and rearranging the results, we arrive at:
By invoking the following matrix identity:
With and , Eq. (20) can be recast into the following form:
Finally, by right-multiplying both sides of Eq. (22) by , we obtain the self-consistent equation corresponding to Eq. (9) in the main text.
In the computation, the following initialization matrices must be evaluated, along with the definition of :
Based on these definitions, the self-consistent equation for can be written as:
This self-consistent equation can be solved iteratively, with the iteration performed entirely in the low-rank subspace.
To obtain the explicit expression for efficiently, we invoke the Sherman–Morrison–Woodbury (SMW) formula, which reads:
The above form is valid provided that and are invertible, and that . The formula can be verified directly by matrix multiplication.
Substituting into Eq. (8) of the main text and using the low-rank decomposition of the coupling matrices, we obtain:
Here, using the SMW formula, we define and . Substituting the above definitions, we find that depends entirely on :
Likewise, the transfer matrix can also be expressed in terms of from Eq. (22):
When the dispersion relation of the surface states is the only quantity of interest, the SDOS spectrum can be obtained by applying the stochastic trace estimator to Eq. (28). If, however, the field distribution associated with the surface excitation must be computed rigorously, the inverse of must be evaluated exactly according to Eq. (28) to obtain , after which the transfer matrix is obtained via Eq. (29). The Green's function field distribution of the entire system is then given rigorously by .
Discussion of the Metric in the SDOS calculation
Strictly speaking, the SDOS result is given by Eq. (3), which depends not only on the surface Green's-function matrix but also on the metric term . Once the metric is included, Eq. (11) of the main text can still be evaluated with the stochastic trace estimator, so the leading-order computational cost is unchanged. In practice, however, the metric itself can be rather complicated: in materials containing gyromagnetic terms, for example, is a tensor at every node, which complicates both its extraction and its application. In this work we therefore neglect and directly use the imaginary part of the trace of to locate the modes. The metric-weighted and unweighted spectra are not equal in a strict mathematical sense.
The SDOS peaks, however, arise from the divergent response of the system at the mode frequencies under a given excitation, and the same divergence occurs at the mode frequencies for both and . In electromagnetic systems, the two correspond respectively to the electric field and the electric displacement field, whose divergences likewise occur at the mode frequencies. Hence, provided that is non-singular, the metric affects only the excitation strength and does not shift the mode positions.
To assess the influence of omitting , we calculated both the unweighted spectrum, , and the weighted spectrum, , for the 2D photonic crystal shown in Figure 2 of the main text. As shown in Figure S1-Figure S1), both approaches clearly identify the same surface state dispersion. Furthermore, Figure S1 compares the two spectra over the frequency range at . Although the two approaches yield noticeably different absolute spectral intensities, their peak positions and overall line shapes are nearly identical, indicating the same surface mode positions. Therefore, omitting does not affect the identification of the surface state position or its distinction from the bulk states in the present system.

Figure S1. Comparison of the SDOS with and without the metric . (a) Simplified SDOS spectrum, with the formula . (b) Rigorous SDOS spectrum, with the formula . (c) Comparison of the SDOS with and without at . Both spectra are calculated by LIM for the photonic crystal shown in Figure 2 of the main text.
This insensitivity can be understood from the pole structure of the Green’s function: weighting the spectrum in post-processing changes the spectral weight of each mode without changing its resonance poles. Since the Green’s function diverges at a mode in the limit , the peak position is fixed by its pole alone. Provided that no material parameter diverges at the same frequency, neither the spatial inhomogeneity nor the frequency dispersion of can shift it. The two definitions therefore give the same peak positions and generally similar line shapes, while their absolute intensities may differ noticeably. An exception arises in extreme cases: at the resonance frequency of a lossless material, the material parameters, and hence the metric itself, can diverge, so that the dispersion requires additional care. For instance, the frequency-dependent terms can be linearized by introducing auxiliary variables, so that the response at that frequency can be evaluated correctly.
In summary, the analysis in this work focuses on the spectral positions and dispersions of the surface states, as well as their distinction from the bulk states, rather than on the absolute calibration of the SDOS. We also note that the peak intensity of the numerical spectrum depends on the artificially introduced imaginary component . Therefore, the spectral intensities in this work are primarily used for relative comparison and surface-state identification.
Pseudocode of the LIM for Representative Partially Infinite Systems
In this chapter, we present the pseudocode of the LIM for four types of partially infinite systems. These four cases cover most scenarios involving partially infinite systems, and the LIM for more complex systems can also be derived by reference to these cases.
LIM for a Bare Semi-Infinite Crystal
This section presents the LIM for a bare semi-infinite crystal, which corresponds to the system in Fig. 2 and Fig. 4 of the main text. The operator matrix of this system can be expressed as:
A detailed discussion of this system is given in the main text.
The following is the detailed pseudo-code of the algorithm 1:
Algorithm 1: LIM for a bare semi-infinite crystal (Fig. 2 and Fig. 4 in the main text)
Initialization:
, ,
maximum iterations , error tolerance , number of random vectors
Output:
trZ00 (trace of Z00)
// Step 1: Low-rank truncation of and using QR decomposition
// Keep the columns and in and where the diagonal elements of and are non-zero, satisfying
// Step 2: Precompute quantities independent of the iteration
7. AinvUB = A⁻¹U_B
8. AinvUC = A⁻¹U_C
9. P = V_B† × AinvUC
10. Q = V_B† × AinvUB
11. R = V_C† × AinvUB
12. T = V_C† × AinvUC
13. // Step 3: Solve for the L̃ = V_B′L using low-rank Riccati iteration
14. Lᵗ = 0
15. for i = 1 to maximum iterations do:
16. Lᵗnew = −P + Q × (I_{r_b} + Lᵗ × R)⁻¹ × (Lᵗ × T)
17. if ‖Lᵗnew − Lᵗ‖_F/(‖Lᵗ‖_F + eps) < error tolerance then
18. break
19. end
20. Lᵗ = Lᵗnew
21. end
22. // Step 4: Precompute fixed matrices for Woodbury formula
23. VBLfinal = Lᵗ
24. K = I_{r_b} + VBLfinal · R
25. [LK, UK] = lu(K)
26. // Step 5: Estimate tr((A + B S)⁻¹) using Hutchinson's method
27. z = randomVector(m)
28. // Generate m Rademacher random vectors
29. w = A⁻¹z
30. s = UK⁻¹(LK⁻¹ × VBLfinal × V_C†w)
31. Z00z = w − AinvUB × s
32. sumtr = sum(z† × Z00z)
33. //Calculate each column's z_i† × Z00 × z_i
34. trZ00 = sumtr/m
35. return trZ00
Noted that the inverse matrix operator here is used only to express the mathematical formula clearly. In the calculation, LU decomposition is employed to obtain the inverse matrix and its products , and .
LIM for a Semi-Infinite Crystal with a Modified Boundary Block
This section presents the LIM for a semi-infinite crystal with a modified boundary block, which corresponds to the system in Figure 3(a) of the main text involving additional cladding layers. The operator matrix of this system can be expressed as:
The difference from algorithm 1 lies in the additional treatment required for the coupling of surface Green’s functions.
In this modified semi-infinity system, obtaining the final response requires considering two distinct sets of Green's functions simultaneously: the bulk surface Green's function and the boundary surface Green's function . They are defined respectively as:
Both Green's functions can be converted into explicit expressions via Woodbury transformations. The core expressions of the two Woodbury transformations are as follows:
and
where . The final SDOS can be expressed as a trace involving , which can likewise be obtained using the stochastic trace estimator:
The following is the detailed pseudo-code of the algorithm 2:
| Algorithm 2: LIM for a semi-infinite crystal with a boundary block (Fig. 3a in the main text) |
| Initialization: |
| A0, B0, C0, A1, B1, C1 |
| maximum iterations > 0, error tolerance > 0, number of random vectors > 0 |
| Output: |
| trZ00 (trace of Z00) |
| 1. // Step 1: Low-rank truncation of and using QR decomposition |
| 2. |
| 3. |
Table 2.
// Keep the columns in where the diagonal elements of are non-zero, satisfying
// Step 2: Precompute quantities independent of the iteration
// Step 3: Solve for the using low-rank Riccati iteration
for to maximum iterations do:
if then
break
end
end
// Step 4: Precompute fixed matrices for Woodbury formula
// Step 5: Construct the boundary Green's function using the Woodbury formula
// Step 6: Estimate using Hutchinson's method
// Generate Rademacher random vectors
//Calculate each column's
return
LIM for Two Semi-Infinite Crystals Interfaced with Each Other
This section presents the LIM for two semi-infinite crystals separated by another crystal slab, which corresponds to the photonic crystal in Figure 3(d) and the phononic crystal in Figure 3(j) of the main text. The operator matrix of both systems can be expressed as:
For these systems, the interface concatenation no longer includes an independent surface layer. Instead, it is composed of the last layer on the left and the first layer on the right.
The Green’s functions on both sides and respectively satisfy:
where and represent the transfer matrices. Therefore, the intermediate effective matrix should be defined as the block Green’s function on the two layers of the interface:
And the Green’s function in the middle of the interface is strictly defined as:
This expression can also be converted into an explicit form via Woodbury transformations. Specially, the Woodbury factor in the pseudo-code satisfies:
and the surface of Woodbury transformations is:
where .
The final SDOS can be expressed as a trace involving , which can likewise be obtained using the stochastic trace estimation:
The following is the detailed pseudo-code of the algorithm 3:
Algorithm 3: LIM for two semi-infinite crystals interfaced with each other (fig:3d in the main text)
Initialization: , , , , , , , maximum iterations , error tolerance , number of random vectors
Output: (trace of )
// Step 1: Low-rank truncation of and using QR decomposition
// Keep the columns in where the diagonal elements of are non-zero, satisfying ,
// Step 2: Precompute quantities independent of the iteration
14. A22invUB22 = A₂₂⁻¹U_B₂₂
15. A22invUC22 = A₂₂⁻¹U_C₂₂
16. Pleft = V_C₁₁ᵀ × A11invUB11
17. Qleft = V_C₁₁ᵀ × A11invUC11
18. Rleft = V_B₁₁ᵀ × A11invUC11
19. Tleft = V_B₁₁ᵀ × A11invUB11
20. Pright = V_B₂₂ᵀ × A22invUC22
21. Qright = V_B₂₂ᵀ × A22invUB22
22. Rright = V_C₂₂ᵀ × A22invUB22
23. Tright = V_C₂₂ᵀ × A22invUC22
24. // Step 3: Solve for the L using low-rank Riccati iteration
25. Ltleft = 0
26. Ltright = 0
27. for i = 1 to maximum iterations do:
28. Kleft = I + Ltleft × Rleft
29. LtLeftnew = −Pleft + Qleft × [Kleft⁻¹ × (Ltleft × Tleft)]
30. if ‖LtLeftnew − LtLeft‖F/(‖LtLeft‖F + eps) < error tolerance then
31. break
32. end
33. Ltleft = LtLeftnew
34. end
35. for i = 1 to maximum iterations do:
36. Kright = I + Ltright × Rright
37. Ltrightnew = −PlTright + Qright × [Kright⁻¹ × (Ltright × Tright)]
38. if ‖Ltrightnew − Ltright‖F/(‖Ltright‖F + eps) < error tolerance then
39. break
40. end
41. Ltright = Ltrightnew
42. end
43. // Step 4: Precompute fixed matrices for Woodbury formula
44. Ubd = [U_C₁₁, 0, U_B₂₂, 0; 0, U_B₂₂, 0, U_C₂₂]
45. Vbd = [V_B₁₁, 0, V_C₂₂, 0; 0, V_C₂₂, 0, V_B₁₂]
46. Ltint = − blkdiag(Ltleft, Ltright, I, I)
47. Aint = blkdiag(A₁₁, A₂₂)⁻¹ × Ubd
48. Kbd = I − Ltint × (Vbdᵀ × Aint)
49. // Step 6: Estimate tr(Z₀₀) using Hutchinson's method
50. z = randomVector(m)
51. // Generate m Rademacher random vectors
52. w = [A₁₁⁻¹ z(left), A₂₂⁻¹ z(right)]
53. s = Kbd⁻¹ × Ltint × (Vbdᵀ × w)
54. Zmidz = w + Aint × s
55. sumtr = sum(zᵀ × Z00z)
56. //Calculate each column's zᵢᵀ × Z00 × zᵢ
57. trZmid = sumtr/m
return
trZmid
LIM for Two Semi-Infinite Crystals Separated by Another Crystal Slab
This section presents the LIM for two semi-infinite crystals separated by another crystal slab, which corresponds to the system in Figure 3(g) of the main text. The operator matrix of this system can be expressed as:
In this system, the bulk Green's function in both sides can be defined as:
and the surface Green's function satisfies:
This expression can also be converted into an explicit form via Woodbury transformations. The core expressions are as follows:
and the final surface Green’s function is:
where .
The final SDOS can be expressed as a trace involving , which can likewise be obtained using the stochastic trace estimation:
The following is the detailed pseudo-code of the algorithm 4:
Algorithm 4: LIM for two semi-infinite crystals separated by another crystal slab
(fig:3g in the main text)
Initialization:
, , , , , , , , , ,
maximum iterations , error tolerance , number of random vectors
Output:
(trace of )
// Step 1: Low-rank truncation of and using QR decomposition
// Keep the columns in where the diagonal elements of are non-zero, satisfying
// Step 2: Precompute quantities independent of the iteration
// Step 3: Solve for the using low-rank Riccati iteration
for to maximum iterations do:
if then
break
end
end
for to maximum iterations do:
if then
break
end
end
// Step 4: Precompute fixed matrices for Woodbury formula
// Step 5: Construct the boundary Green's function using the Woodbury formula
// Step 6: Estimate using Hutchinson's method
// Generate Rademacher random vectors
65. sumtr = sum(z† × Z00 × z)
66. //Calculate each column's z_i† × Z00 × z_i
67. trZ00 = sumtr/m
68. return trZ00References
- [1]Bernevig, B. A. & Hughes, T. L. Topological insulators and topological superconductors. (Princeton university press, 2013).
- [2]Xue, H., Yang, Y. & Zhang, B. Topological Valley Photonics: Physics and Device Applications. Advanced Photonics Research 2, 2100013 (2021).DOI
- [3]Ma, S., Yang, B. & Zhang, S. Topological photonics in metamaterials. Photonics Insights 1, R02 (2022).DOI
- [4]Zhu, W. et al. Topological phononic metamaterials. Reports on Progress in Physics 86, 106501 (2023).arxiv.org/abs/2303.01426
- [5]Lu, L., Joannopoulos, J. D. & Soljačić, M. Topological photonics. Nature Photonics 8, 821–829 (2014).
- [6]Khanikaev, A. B. & Shvets, G. Two-dimensional topological photonics. Nature Photonics 11, 763–773 (2017).DOI
- [7]Leykam, D., Xue, H., Zhang, B. & Chong, Y. D. Limitations and possibilities of topological photonics. Nature Reviews Physics 8, 55–64 (2026).DOI
- [8]Zakeri, M. J., Perez-Leija, A. & Blanco-Redondo, A. High-dimensional topological photonic entanglement. Science 391, 1379–1382 (2026).
- [9]He, C., Zhao, L., Zhang, S., Zhou, L. & Ma, S. Intrinsic topological hinge states induced by boundary gauge fields in photonic metamaterials. eLight 5, 19 (2025).DOI
- [10]Ma, S. et al. Linked Weyl surfaces and Weyl arcs in photonic metamaterials. Science 373, 572–576 (2021).DOI
- [11]Yu, T.-J., Geng, Z.-G., Shen, Y.-X., Chen, Z. & Zhu, X.-F. Directional sound transport via coupled topological states in heterostructure sonic crystals with mirrored interfaces. Physical Review Applied 25, 024057 (2026).
- [12]Yuan, Y. et al. Supersymmetric Landau Levels in Subwavelength Type-I Dirac Metasurfaces. Physical Review Letters 136, 023802 (2026).DOI
- [13]Zhirihin, D. V. et al. Helical metasurfaces based on topological surface states in three-dimensional photonic topological insulators. Nature Materials 25, 762–766 (2026).DOI
- [14]Ma, Q. et al. Floquet topological states in time-varying metasurfaces. Science Advances 11, eadx9025 (2025).DOI
- [15]Qin, S. et al. Multifunctional Topological Metasurfaces with Modular Meta-Atoms for Multidimensional Wave Control. Advanced Optical Materials 13, 2500044 (2025).DOI
- [16]Ma, S. et al. Gauge Field Induced Chiral Zero Mode in Five-Dimensional Yang Monopole Metamaterials. Physical Review Letters 130, 243801 (2023).DOI
- [17]Rahachou, A. I. & Zozoulenko, I. V. Light propagation in finite and infinite photonic crystals: The recursive Green's function technique. Physical Review B 72, 155117 (2005).DOI
- [18]Sainidou, R., Stefanou, N. & Modinos, A. Green's function formalism for phononic crystals. Physical Review B 69, 064301 (2004).DOI
- [19]Rother, T. Green's functions in classical physics. Vol. 938 (Springer, 2017).DOI
- [20]Stakgold, I. & Holst, M. J. Green's functions and boundary value problems. Vol. 9 (Wiley New York, 1979).
- [21]Cederbaum, L. S. One-body Green's function for atoms and molecules: theory and application. Journal of Physics B: Atomic and Molecular Physics 8, 290 (1975).DOI
- [22]von Niessen, W., Schirmer, J. & Cederbaum, L. S. Computational methods for the one-particle green's function. Computer Physics Reports 1, 57–125 (1984).DOI
- [23]Economou, E. N. in Green’s Functions in Quantum Physics (ed Eleftherios N. Economou) 215–238 (Springer Berlin Heidelberg, 1983).
- [24]Sancho, M. P. L., Sancho, J. M. L. & Rubio, J. Quick iterative scheme for the calculation of transfer matrices: application to Mo (100). Journal of Physics F: Metal Physics 14, 1205 (1984).
- [25]Sancho, M. P. L., Sancho, J. M. L., Sancho, J. M. L. & Rubio, J. Highly convergent schemes for the calculation of bulk and surface Green functions. Journal of Physics F: Metal Physics 15, 851 (1985).
- [26]Metalidis, G. & Bruno, P. Green's function technique for studying electron flow in two-dimensional mesoscopic samples. Physical Review B 72, 235304 (2005).DOI
- [27]Haydock, R., Heine, V. & Kelly, M. J. Electronic structure based on the local atomic environment for tight-binding bands. Journal of Physics C: Solid State Physics 5, 2845 (1972).DOI
- [28]Haydock, R. The Recursive Solution of the Schrodinger Equation. Solid state physics 35, 215–294 (1980).
- [29]Lee, D. H. & Joannopoulos, J. D. Renormalization scheme for the transfer-matrix method and the surfaces of wurtzite ZnO. Physical Review B 24, 6899–6907 (1981).DOI
- [30]Lee, D. H. & Joannopoulos, J. D. A new theory of electronic surface states. Journal of Vacuum Science and Technology 19, 355–359 (1981).DOI
- [31]Lee, D. H. & Joannopoulos, J. D. Simple scheme for surface-band calculations. II. The Green's function. Physical Review B 23, 4997–5004 (1981).DOI
- [32]Lee, D. H. & Joannopoulos, J. D. Simple scheme for surface-band calculations. I. Physical Review B 23, 4988–4996 (1981).DOI
- [33]Boffi, D., Conforti, M. & Gastaldi, L. Modified edge finite elements for photonic crystals. Numerische Mathematik 105, 249–266 (2006).DOI
- [34]Hiett, B. P. et al. Application of finite element methods to photonic crystal modelling. IEE Proceedings - Science, Measurement and Technology 149, 293–296 (2002).DOI
- [35]Axmann, W. & Kuchment, P. An Efficient Finite Element Method for Computing Spectra of Photonic and Acoustic Band-Gap Materials: I. Scalar Case. Journal of Computational Physics 150, 468–481 (1999).DOI
- [36]Brechet, F., Marcou, J., Pagnoux, D. & Roy, P. Complete Analysis of the Characteristics of Propagation into Photonic Crystal Fibers, by the Finite Element Method. Optical Fiber Technology 6, 181–191 (2000).
- [37]Sha, Y.-X. et al. Surface density of states on semi-infinite topological photonic and acoustic crystals. Physical Review B 104, 115131 (2021).DOI
- [38]Sha, Y.-X., Xia, M.-Y., Lu, L. & Yang, Y. Efficient algorithms for the surface density of states in topological photonic and acoustic systems. Nature Computational Science 5, 1192–1201 (2025).DOI
- [39]Wang, Z., Chong, Y., Joannopoulos, J. D. & Soljačić, M. Observation of unidirectional backscattering-immune topological electromagnetic states. Nature 461, 772–775 (2009).DOI
- [40]Sherman, J. & Morrison, W. J. Adjustment of an inverse matrix corresponding to a change in one element of a given matrix. The Annals of Mathematical Statistics 21, 124–127 (1950).
- [41]Golub, G. H. & Van Loan, C. F. Matrix Computations - 4th Edition. (Johns Hopkins University Press, 2013).
- [42]Hutchinson, M. F. A stochastic estimator of the trace of the influence matrix for Laplacian smoothing splines. Communications in Statistics-Simulation and Computation 18, 1059–1076 (1989).
- [43]Chen, T. & Hallman, E. Krylov-Aware Stochastic Trace Estimation. SIAM Journal on Matrix Analysis and Applications 44, 1218–1244 (2023).DOI
- [44]Rylander, T. & Bondeson, A. Stable FEM-FDTD hybrid method for Maxwell's equations. Computer Physics Communications 125, 75–82 (2000).DOI
- [45]Yang, B. et al. Ideal Weyl points and helicoid surface states in artificial photonic crystal structures. Science 359, 1013–1016 (2018).DOI
- [46]George, A. Nested Dissection of a Regular Finite Element Mesh. SIAM Journal on Numerical Analysis 10, 345–363 (1973).DOI
- [47]Amestoy, P. et al. Combining Sparse Approximate Factorizations with Mixed-precision Iterative Refinement. ACM Trans. Math. Softw. 49, 4 (2023).DOI
- [48]Soler, J. M. et al. The SIESTA method for ab initio order-N materials simulation. Journal of Physics: Condensed Matter 14, 2745 (2002).DOI
- [49]Marzari, N. & Vanderbilt, D. Maximally localized generalized Wannier functions for composite energy bands. Physical Review B 56, 12847–12865 (1997).DOI
- [50]Liu, K. et al. Source Code and Data for the paper "Low-Rank Iteration for Efficient Calculations of Topological Surface States". Science Data Bank https://doi.org/10.57760/sciencedb.017v3 (2026).