WorksMathematics
A Hierarchical direct solver for Lippmann-Schwinger equations based on boundary traces
Abstract
A high-order hierarchical direct solver for variable-media scattering in two dimensions is presented. Standard discretizations for these problems produce dense linear systems that are commonly solved iteratively, requiring a new global solve for each incident field. The method proposed here instead constructs a reusable scattering operator for the inhomogeneous medium, based on a hierarchical subdivision of the computational domain. First, high-accuracy leaf scattering operators are built from discretized local integral equations. Up the hierarchy, neighboring boxes are merged to build the scattering operators for successively larger boxes. A key feature of the method is that each scattering operator is represented as a map of incoming traces to outgoing traces. Consequently, volume unknowns appear only in the leaf computations; all higher levels of the hierarchy operate entirely on boundary traces. The hierarchical solver is presented first at the continuum level and discretized only afterward using high-order local volume quadrature, polynomial trace representations, and boundary integral techniques. For a discretization involving degrees of freedom, the method requires work and storage to build all scattering operators. Once these are available, any incoming field can be processed at cost if only the outgoing field is required, and cost if the full field is required. Numerical experiments demonstrate high-order accuracy on a range of challenging scattering problems including cavities, back-scattering, limited regularity, and high wavenumber, that pose significant challenges to iterative methods.
Introduction
This paper describes a method for numerically solving a variable coefficient Helmholtz equation of the form
Here is the wave-number, is a prescribed incoming field satisfying the free-space Helmholtz equation in a neighborhood of the scatterer, and is a compactly supported scattering potential. We let denote a rectangular domain that encloses the support of , and further assume for simplicity that satisfies the free-space Helmholtz equation throughout . When , an acoustic interpretation is obtained by relating to the local wave speed and the free-space value through
The scattered field is required to satisfy the Sommerfeld radiation condition. Figure 1 illustrates the geometry.
Figure 1. Geometry of the scattering problem. Waves propagate in a medium with constant wave speed everywhere except in the shaded areas, where the wave speed is given by , with a given, smooth, compactly supported “scattering potential.” An incident field hits the scattering potential and induces the scattered field . The dashed line marks the artificial domain , which encloses the support of the scattering potential.
A particularly attractive formulation of (1) is obtained by introducing the free-space fundamental solution
where is the Hankel function of the first kind of order zero. This is the outgoing free-space fundamental solution for . We then write
The source density then satisfies the Lippmann–Schwinger equation
The formulation (3) has several well-known advantages: the computational domain is finite, the radiation condition is enforced automatically, and the equation has the structure of a Fredholm equation of the second kind [12]. High-order discretizations are available for smooth media [1, 15, 2], with related high-order volume-integral discretizations also developed for more general acoustic and electromagnetic media [3, 7].
The main computational drawback is that a discretization of (3) produces a dense linear system involving all degrees of freedom in the scattering region. A standard approach to overcome this problem is to deploy an iterative solver, where each application of the operator is accelerated using a fast method such as the FFT or the Fast Multipole Method [1, 15]. Considerable work has been devoted to making such iterations effective at high frequency and for strongly scattering media, including sparsifying and sweeping preconditioners [37, 39, 28, 22]. For many problems these techniques are highly effective and are the natural methods of choice. There remain, however, important regimes in which the global iteration can be difficult, and applications involving many incoming fields require repeated solution of essentially the same scattering problem. These considerations motivate the construction of a reusable direct solver.
The method developed in this paper is based on a simple observation. Consider any square or rectangular subdomain . The field incident on from sources outside the box satisfies the homogeneous Helmholtz equation inside the box, while the field generated by sources inside satisfies the homogeneous Helmholtz equation outside it. We represent both fields by their traces on the boundary . The response of the entire volume can then be summarized by a boundary-to-boundary scattering operator
Once has been constructed, the volume degrees of freedom inside no longer need to participate in interactions with the rest of the problem.
The method is hierarchical. The computational domain is first partitioned into a collection of small leaf boxes, which are organized into a binary tree by recursively pairing neighboring boxes. Each parent box is the union of its two children, and the root of the tree is the full computational domain . On each leaf, a scattering operator is constructed by solving local Lippmann–Schwinger problems in the volume. The scattering operators are then merged pairwise in an upward pass through the tree: given the scattering operators of two children, a boundary-only calculation produces the scattering operator of their parent. Repeating this process yields a scattering operator for the entire inhomogeneous medium.
A key feature of the method is that volume unknowns appear only in the leaf computations. At all higher levels, the hierarchy operates entirely on boundary traces. This dimensional reduction is particularly attractive at high accuracy, since inter-box interactions are represented directly by high-order boundary fields rather than by increasingly rich subsets of a volume discretization.
The method naturally separates into three computational stages. First, a set of geometry- and frequency-dependent Helmholtz continuation and evaluation operators is constructed. These operators depend only on the box sizes and on the wave-number , not on the scattering potential, and can therefore be precomputed and reused for different media. Second, once the medium is specified, local scattering operators are built and merged in an upward pass. Third, after this build stage, each new incoming field can be processed rapidly through the stored hierarchy.
For a discretization with degrees of freedom, the method requires work to construct the hierarchical scattering representation. Retaining the operators needed for subsequent recovery of the full volume solution requires storage, while a variant that constructs only the root scattering operator requires only storage. Once the hierarchy has been built, new incident fields can be processed without any further volume solves. (Let us note that the matrices representing the scattering operators appear to have internal rank structure that could in principle be exploited to further reduce asymptotic complexities, see Remark 4.)
In this work, the reduction to boundary traces and the hierarchical structure of scattering operators are derived at the continuum level, before discretization. To realize the resulting algorithm numerically, we use high-order discretizations of the local Lippmann–Schwinger equation and high-order polynomial representations of the boundary traces (with or being representative). The continuation and evaluation operators are constructed using standard high-order boundary integral equation techniques.
The present approach is related to several existing lines of work. Fast iterative methods for the Lippmann–Schwinger equation form the main computational baseline [1, 15, 37, 39, 28]. Hierarchical construction of scattering operators for volume integral equations has important precedents, notably the work of Chen [8]. T-matrix and multiple-scattering methods likewise characterize individual scatterers by maps from incident to scattered fields [17, 19, 18]. There is also a substantial literature on direct solvers that first discretize the volume integral equation and then exploit rank structure in the resulting dense matrix [2, 14]. A particularly close predecessor is the direct solver of [22], which likewise uses a hierarchical construction based on scattering operators for the Lippmann–Schwinger equation. In [22], the volume integral equation is first discretized, and rank structure in the resulting dense matrix is then exploited to construct and merge discrete scattering operators. The present construction reverses the order of these operations: the hierarchical reduction is formulated at the continuum level, so that all inter-box interactions above the leaf level are represented directly by boundary traces, with discretization introduced only afterward. This distinction becomes particularly important in the high-accuracy regime, where the interaction ranks in the discrete formulation of [22] can become large.
The organization of the present method is reminiscent of hierarchical Poincaré–Steklov (HPS) solvers for differential equations [29, 20]: local problems are condensed into boundary response operators and these are merged recursively. The distinction is that the local variable-coefficient scattering problems here are solved through the Lippmann–Schwinger equation rather than by directly discretizing the differential operator.
The contribution of this paper is a high-order numerical realization of this field-to-field hierarchical scattering construction. We give the exact continuum merge identities, describe a practical discretization based on high-order local volume solves and boundary traces, and present numerical experiments on a range of variable-media scattering problems. The present implementation deliberately does not exploit the substantial numerical rank structure that appears in many of the merge operators; this provides an immediate avenue for future acceleration.
The remainder of the paper is organized as follows. § 2 defines the local incoming and outgoing fields and the field-to-field scattering operator. § 3 derives the merge operation, first for separated scatterers and then for adjacent boxes. § 4 describes the complete hierarchical continuum algorithm, including the downward recovery of the interior solution. § 5 describes the discretization and implementation. Numerical experiments are presented in § 6, and conclusions are given in § 7.
Local scattering operators
In this section, we give a precise mathematical definition of the local scattering operators that will serve as the basic building blocks of our hierarchical solver. Throughout this section, denotes a label associated with a square or rectangular subdomain , with boundary , cf. Figure 2. These subdomains will later form part of the hierarchical tree.
Figure 2. A box contains the support of a scattering potential . An incoming field induces the scattered field . The fields are represented by their traces and on the artificial boundary , so the local scattering operator maps .
For a single square or rectangular subdomain, we first identify the incoming and outgoing fields, then represent these fields by their boundary traces, and finally show how the resulting boundary-to-boundary scattering operator can be constructed from a local Lippmann–Schwinger solve.
Incoming and outgoing fields
For a box , we define incoming and outgoing fields, following standard practice. Let denote the solution to (3). The outgoing field generated by the sources in is
In other words, is the radiating solution to the -Helmholtz equation generated by the sources inside .
The incoming field for the box is the sum of the externally applied field and the field generated by all sources outside . Thus,
Observe that satisfies the homogeneous Helmholtz equation throughout .
Once the incoming field is known, the Lippmann–Schwinger equation can be solved locally on . Indeed, restricting (3) to and splitting the volume potential into an interior and an exterior part gives
Solving (6) specifies the restriction of to , and hence the outgoing field via (4).
Trace representations
When a scattering operator is to be computed in practice, one must specify how the incoming and outgoing fields are represented. Standard practice in the physics literature is to use truncated expansions in cylindrical harmonics (and, in three dimensions, spherical harmonics). We will instead represent both fields through their boundary traces, by which we simply mean their restrictions to the boundary. For a box , define
Thus, is the incoming trace and is the outgoing trace. (The incoming and outgoing fields considered here are continuous, so the boundary traces are well defined.)
The outgoing trace always uniquely determines : it supplies Dirichlet data for the exterior Helmholtz problem, which is uniquely solvable when coupled with the radiation condition at infinity. The incoming trace supplies Dirichlet data for an interior Helmholtz problem on . For Helmholtz, this problem can in principle be ill-posed. For the present development, make the assumption:
For every box in the hierarchy, the local Dirichlet problem is well posed.
Under this assumption, uniquely determines the incoming field throughout .2
We now define the boundary-to-boundary scattering operator by
The scattering operator encodes the complete response of the material in to any incoming field that is homogeneous Helmholtz in that box.
Constructing a local scattering operator
The scattering operator for a leaf box can be written as a composition of four continuum operators. Let
be the solution operator for the local Dirichlet Helmholtz problem on , so that
be multiplication by the function .
be the locally restricted Lippmann–Schwinger operator
evaluate the outgoing field on the boundary:
Define the local source operator
Then , and the leaf scattering operator is
Formula (10) will be used only on leaf boxes. For larger boxes, the scattering operator will instead be constructed by merging the operators of two smaller boxes.
Remark 1 (Equivalent-source formulation). The field-to-field scattering operator used throughout the main text is not the only possible boundary representation. An alternative is to continue representing the incoming field by its trace , but to represent the outgoing field by an equivalent single-layer source density supported on . This leads to a closely related hierarchical merge in which interactions between children are evaluated directly through single-layer potentials. We use the field-to-field formulation in the main text because it proved simpler to implement and avoids the reconstruction of equivalent source densities. The alternative formulation is recorded in Appendix A.
Merging scattering operators
Suppose that a subdomain is formed as a union of two smaller subdomains and , and that the scattering operators and for the two “children” have already been constructed. We will show how to form the scattering operator for the “parent” using boundary-to-boundary operations only. Once the scattering operator of a box has been built, we never need to revisit its interior during the upward pass.
To introduce the merge in the cleanest possible setting, we first consider two children that are separated in space. We then turn to the case of two adjacent boxes, which is the case required by the hierarchical solver. The second derivation involves additional bookkeeping, but the underlying algebra is the same.
Merging two separated boxes
Consider a scattering potential supported in two disjoint boxes and , and suppose that and have been built, cf. Figure 3. Set

Figure 3. Merging of two scattering operators, as discussed in § 3.1. Let and be two boxes, with the union . If the scattering operators and have been constructed, then the joint scattering operator can be constructed without the need to revisit the interiors of or ; only operators supported on the boundaries are involved. The parent operator maps the incoming trace to the outgoing trace .
Since the boxes are separated, the boundary of the parent is simply the disjoint union of the two child boundaries. (For this introductory derivation, we temporarily allow the parent domain to be disconnected; the hierarchical solver itself will use only adjacent rectangular children.)
Let denote the trace on of an externally applied incoming field. Since , we regard as a two-component trace ordered by . Note that , because in the latter quantities we must also consider the interaction between and . For later convenience (see § 3.2), we introduce
Introduce the block-diagonal child scattering operator
so that
The sibling-generated incoming field is obtained by propagating each outgoing field to the other box. Let denote the operator that takes the outgoing trace on , reconstructs the corresponding radiating exterior Helmholtz field, and evaluates this field on . Define analogously, and set
The total incoming trace on each child is the sum of the externally applied field and the field generated by its sibling. Thus
Combining (12) and (14), we obtain the equilibrium equation for the local incoming fields
or, more compactly,
This motivates the following definition.
Definition 1 (Solution-transfer operator). The solution-transfer operator for the separated pair is
It maps the incoming trace on the parent to the total incoming traces experienced by the two children:
We assume that the local Dirichlet problems and the parent scattering problem are uniquely solvable, so that the inverse in the merge formula (15) exists.
It remains to form the outgoing trace of the composite scatterer. On , the total outgoing field is the sum of the field generated by and the field generated by ; the analogous statement holds on . Consequently,
Combining (12), (16), and (17), we obtain
Thus the parent scattering operator and the information needed to recover the fields on its children are both constructed using boundary-to-boundary operators only.
Remark 2 (Multiple scattering). The solution-transfer operator (15) has a natural physical interpretation in terms of repeated scattering between the children. Observe that
whenever the series converges. The term represents one scattering event followed by propagation to the sibling, represents two successive interactions, and so forth. Thus incorporates all orders of multiple scattering between the children.
Merging two adjacent boxes
Consider next a case where the two boxes and directly touch, as illustrated in Figure 4. This is the situation of immediate relevance to our hierarchical solution process, since it consists of pairwise merging of domains to build scattering operators for larger and larger subdomains. Let
Figure 4. Two adjacent boxes to be merged. The labels 1, 2, 3, 4 refer to boundary pieces. The labels 3 and 4 denote two copies of the same artificial interface, so that the two children may carry distinct trace variables and, after discretization, distinct grids. The support of may cross this interface.
where is the common interface. The parent box has boundary
We use two labels for the common interface because its traces belong to different child boxes, and because the two copies may later be discretized differently.
The adjacent case differs from the separated case in three ways. First, the externally applied incoming trace on the parent boundary does not include its values on the common interface, so these values must be reconstructed. Second, the sibling-evaluation operator contains identity blocks on the common interface; this makes computations cheaper, but requires additional notational machinery. Third, the child outgoing traces include the common interface, which must be eliminated when the outgoing trace of the parent is formed.
Continuing the externally applied incoming field to the child boundaries
Let denote the incoming trace on the parent boundary . The corresponding field satisfies the homogeneous Helmholtz equation throughout , so its trace on the common interface can be reconstructed by solving the interior Dirichlet problem on . We define the incoming continuation operator to map to the trace of this same field on the complete boundaries of the two children. In the boundary ordering , this operator has the form
so that gives the trace of the externally applied incoming field on the complete boundaries of the two children. Since , the third and fourth components of agree at the continuum level; equivalently, and .
Resolving the interaction between the children
We now split the scattering operators of the children according to their exterior and interface boundary pieces:
In the boundary ordering , define
Thus simply applies the scattering operator of each child to the appropriate pair of boundary traces. With the ordering introduced above, the total incoming and outgoing traces on the two children from § 3.1 become
with
The child scattering relations are simply, cf. (12),
Define to be the operator that maps the child outgoing traces to the incoming traces generated on the sibling. On the common interface, no exterior solve is required: the outgoing trace from one child is directly the incoming trace generated on the other child. In the ordering ,
The total incoming trace on each child is the sum of the parent incoming field, continued to the child boundary, and the field generated by its sibling. Thus, cf. (14),
Combining (23) and (25), we find
or
We define the solution-transfer operator for the parent box by
Then
The operator will be used twice: it enters the construction of the parent scattering operator during the upward pass, and it transfers the incoming trace on a parent to the total incoming traces on its children during the downward pass.
Forming the parent outgoing field
The outgoing traces in are tabulated on the complete child boundaries, including the common interface. The outgoing trace of the parent is required only on and must include the fields generated by both children. We therefore introduce the outgoing assembly operator by
In the ordering ,
Combining (23), (27), and (28), we obtain the parent scattering operator
These are the two fundamental merge formulas.
Summary of the merge
The separated and adjacent cases are governed by the same three operations:
continue the externally applied incoming field on the parent to the complete child boundaries;
resolve all multiple scattering between the children;
assemble the child outgoing fields into the outgoing field of the parent.
The first and third operations are represented by and , respectively, while the first two combine to form the solution-transfer operator . If denotes the block operator formed from the two child scattering operators, then the generic merge is
For separated scatterers, is the operator defined in (13), , and . For adjacent boxes, reconstructs the externally applied incoming field on the internal interface, contains identity maps across that interface, and eliminates the interface while accumulating both child contributions on the parent boundary. No approximation has entered the construction in this section.
Remark 3 (Universal operators). The operators , , and are independent of the scattering potential . They depend only on the box geometry and the free-space Helmholtz operator. Likewise, the leaf operators and are universal. All dependence on enters through and in the local Lippmann–Schwinger solve, and hence through , , and . Consequently, the universal operators can be precomputed and reused for different scattering potentials. On the uniform, translation-invariant trees considered here, only one copy is needed for each box size and merge orientation.
Remark 4 (Equivalent formulation and rank structure). The identity
shows that the parent scattering operator can equivalently be written as
This form is useful when considering possible accelerations of the merge. From (22) and (24), we obtain
For boxes that contain a moderate number of wavelengths, numerical experiments indicate that all nonzero blocks in (33) except the and blocks have low numerical rank. The two exceptional blocks describe interactions across the common interface. They are not globally low rank, but admit hierarchically semiseparable (HSS) representations. We do not pursue this structure in the present work. Equation (33) nevertheless points to a particularly simple avenue for accelerating the method in future work. As the number of wavelengths in a box grows, the numerical ranks naturally grow as well.
The hierarchical continuum algorithm
We are now ready to describe the complete continuum algorithm for solving the Lippmann–Schwinger equation (3). The method is hierarchical. It consists of partitioning the computational domain into small leaf boxes, constructing a local scattering operator on every leaf, and then repeatedly merging pairs of child operators in an upward pass through a binary tree. The solution-transfer operators constructed during the upward pass are subsequently used to recover the incoming traces on all leaves in a downward pass.
The hierarchical partition
The computational domain will always be a square or rectangle. Observe that the domain is in this context an artificial boundary box anyway. We first partition into a collection of square, or possibly rectangular, leaf boxes. These should be small enough that the Lippmann–Schwinger equation restricted to a leaf can easily be solved to very high accuracy. We then construct a binary tree by hierarchically merging the boxes, making sure that all boxes on the same level are roughly of the same size; see Figure 5. The boxes are ordered so that if is the parent of , then . The root of the tree, which represents the full domain , has index .

Figure 5. The square domain is split into leaf boxes. These are then gathered into a binary tree of successively larger boxes as described in § 4. One possible enumeration of the boxes in the tree is shown, but note that the only restriction is that if box is the parent of box , then .
Upward pass
Let denote the total number of boxes in the tree. The upward pass constructs the scattering operator for every box and the solution-transfer operator for every nonleaf box. At a leaf, we also retain the local source operator from (9), since it will be needed to recover the volume density after the incoming trace on the leaf has been found.
The root operator maps the incoming trace on to the outgoing trace generated by all sources in . In particular, it determines the complete exterior scattering solution without ever revisiting the interiors of the boxes.
Downward pass and recovery of the solution
Let the incoming trace on the root be
Suppose that the incoming trace has been determined for a nonleaf box with children and . By the definition of the solution-transfer operator,

Figure 6. Upwards (left) and downward (right) pass algorithms for the construction of the global scattering operator and the computation of the interior traces.
For the ordering used in § 3.2, the incoming traces on the two children are then
Applying this operation recursively from the root to the leaves determines the incoming trace on every leaf box.
Once is known on a leaf, its volume density is recovered independently by
and the incoming field for box can be computed by
The total field inside the leaf box is recovered by superposing the incoming field and the field induced by the volume density,
Finally, the local scattered field is given by
In the exterior of , the scattered field can be reconstructed from the root outgoing trace
by solving the radiating exterior Dirichlet problem.
The entire construction in § 2–§ 4 is exact at the continuum level. Approximation enters only when the operators appearing above are discretized, which is the subject of the next section.
Storage efficiency
There are several ways to reduce storage requirements. First, the scattering operators are needed only during the upward pass. Once a parent operator has been constructed, the scattering operators of its children may be discarded, since they are not used in the downward pass. Second, for recovery of the full interior field, the solution-transfer operators must be retained, while the leaf source operators may either be stored or recomputed on demand during the downward pass. Recomputing the ’s reduces memory usage at the expense of additional local solves. Finally, if only the root scattering operator is required, neither the ’s nor the ’s need be retained; in that case all intermediate operators may be discarded as soon as they are no longer needed.
Figure 8. Decomposition of the discretized scattering operator into its constituent operators (see § 5.2).
Discretization and implementation details
This section describes how we turn the continuum algorithm of § 4 into an actual numerical method.
Discretization
All boundary traces such as are represented via collocation. For a leaf box , we place Gauss–Legendre nodes on each side. Typically high is used, say or . For larger boxes, we use the composite Gauss–Legendre grids obtained by retaining the boundary nodes of their children.
To facilitate the interior spectral solve, we also introduce a Chebyshev discretization of . The tensor-product Chebyshev grid provides the interior discretization on which the Lippmann–Schwinger equation is solved. Local discretizations are shown in Figure 7.
Figure 7. Local discretization on a leaf box.
Leaf computations
As introduced in § 2.3, the task is to find the scattering operator that maps the incoming field prescribed on the boundary of a leaf box to the corresponding outgoing scattered field on the same boundary. The leaf computations are executed on tensor-product Chebyshev grids.
Recall that, in the continuous case, maps the incoming field to the outgoing field , where is the boundary trace of , and solves the local Lippmann–Schwinger equation with incoming field . In discretized form, the computation of then involves the following steps:
Interpolate the incoming trace from the Gauss–Legendre nodes to Chebyshev boundary nodes, averaging the two side interpolants at each corner.
Use a local spectral collocation solve for the homogeneous Helmholtz equation to reconstruct the incoming field at all tensor-product Chebyshev nodes.
Solve the local Lippmann–Schwinger equation using a high-order corrected quadrature rule to obtain the (discretized) volume density .
Interpolate the trace of the volume potential generated by to the Gauss–Legendre boundary nodes to obtain the outgoing trace.
Given an incoming field on the Gauss–Legendre boundary nodes, the operator reconstructs the corresponding field on the tensor-product Chebyshev grid. It consists of interpolation from the Gauss–Legendre boundary nodes to the Chebyshev boundary nodes, followed by a Chebyshev spectral collocation solve of the homogeneous Helmholtz equation. Thus,
The induced volume density is then computed from the discretized Lippmann–Schwinger equation,
where is the discretized Lippmann–Schwinger operator, contains the scattering potential scaled by and represents the volume Green’s operator discretized using a high-order corrected quadrature rule. Hence,
Finally, the scattered field generated by is evaluated on the Chebyshev boundary nodes and interpolated to the Gauss–Legendre boundary nodes. Denoting the resulting operator by , we have
Combining the three stages gives
Therefore, the scattering matrix for leaf is
The leaf continuation operators , scattering trace operator and the high-order quadrature matrices are universal when all leaf boxes have the same size. They therefore need to be constructed only once.
Discretization of the merge stage
As already noted in Remark 3, the operators , and from (31) are universal at each level where they exist. At level , there is no continuation operator (leaf nodes have no children), so we can index all discretized continuation operators by , . Similarly (the root node has no siblings) we can index all sibling-evaluation operators by , . For notational convenience we will drop the index for the blocks and operators that constitute , and .
As can be seen from (24) and (29), and can be written as
Figure 9 shows the set-up for and . The construction for is similar. Computationally, the structure of is

Figure 9. Set-up for the discretization of and .
Here maps the boundary data, collocated at piecewise Gauss-Legendre nodes, to interpolated boundary data at a point distribution on the boundary that is suited for a BIE solver. The exterior BIE solver solves for a (discrete) surface density, whose precise form depends on the chosen formulation. Finally, maps the resulting surface density to the values of the Helmholtz solution at the specified exterior target DOFs. The matrix has the same structure, but interior:
In our implementation we use Brakhage–Werner-type indirect BIE formulations for both the exterior and interior problem. Concretely:
with DL and SL discretizations of the boundary double-layer and boundary single-layer operators (here SL is not to be confused with the product of S and L from § 5.2) and is a scalar usually chosen to be . We construct and solve both using the chunkIE package (see [4]).
For a fixed level , and are defined on the same boundary. As such, the matrices DL and SL can be re-used for the construction of and .
Complexity
The complexity of our proposed solver is dominated by the following three terms:
: the construction, at the leaf level, of and .
: the off-line computation, at each level, of , and .
: the inversion, at each level, of .
We set to be the total number of unreduced degrees of freedom. For simplicity we assume that is even.
The leaf scattering computation is dominated by the inversion of , which is of order . The total cost is therefore
Usually, the leaf level matrices are quite small. Even at , they are only of size , making them perfectly suited for highly efficient BLAS level 3 operations.
The cost , by a geometric series argument, is bounded by a constant times the cost of the inversion of at the root, where the size of the system is precisely . Then, trivially,
The cost depends heavily on the chosen BIE solvers. We present the argument for the continuation matrices , here, but the same argument applies to the sibling evaluation operators.
Assume that the system matrices at level are square and of size . With the number of boundary segments of a box at level , we ensure . Naive construction and inversion of the BIE system results in a cost of . The cost of applying interpolation and evaluation by Interp, scales as follows: at level these operators are block diagonal matrices, consisting of blocks. Each of the blocks in the interpolation (resp. evaluation) can be of variable row size (resp. column size), but always has column size (resp. row size) . This means the cost multiplying left and right by these block diagonal matrices is bounded by .
Summing over levels, again using a geometric series argument, finally gives us
The final estimate on the right-hand side obfuscates the dependence of the off-line phase on the chosen BIE solver, and for this reason we present both complexity estimates. While we have presented the asymptotic cost for naive BIE construction and inversion, at large scale both can be accelerated using fast algorithms like the Fast Multipole Method (FMM, see, e.g., [24, 33, 38]), hierarchical matrix techniques (see [5, 23, 30]) and recursive compression (see [25]).
Numerical examples
We present six numerical experiments with various wavenumbers and scattering potentials . The computational domain is the unit square . The numerical computations in this section were carried out using MATLAB. All timings are reported on a workstation with an Intel(R) Xeon(R) Gold 6248R processor (48 cores with 2 threads per core at 3.00 GHz base frequency) and 1.5 TB of RAM. Results are presented in Table 1 and Table 2.
| levels | Number of leaves | Number of points | precomp. | build | eval. | solve |
| 3 | 4 | 1600 | ||||
| 5 | 16 | 6400 | ||||
| 7 | 64 | 25 600 | ||||
| 9 | 256 | 102 400 | ||||
| 11 | 1024 | 409 600 | ||||
| 13 | 4096 | 1 638 400 | ||||
| 15 | 16 384 | 6 553 600 |
Table 1. Computational cost for different numbers of levels, at and .
| Scattering potential | levels | leaves | points | ext. | int. | grid |
| Gaussian bump | 3 | 4 | 1600 | |||
| 5 | 16 | 6400 | ||||
| 7 | 64 | 25 600 | ||||
| 9 | 256 | 102 400 | ||||
| 11 | 1024 | 409 600 | ||||
| Smooth mollifier | 3 | 4 | 1600 | |||
| 5 | 16 | 6400 | ||||
| 7 | 64 | 25 600 | ||||
| 9 | 256 | 102 400 | ||||
| 11 | 1024 | 409 600 | ||||
| Cavity | 3 | 4 | 1600 | |||
| 5 | 16 | 6400 | ||||
| 7 | 64 | 25 600 | ||||
| 9 | 256 | 102 400 | ||||
| 11 | 1024 | 409 600 | ||||
| Lens3 | 3 | 4 | 1600 | |||
| 5 | 16 | 6400 | ||||
| 7 | 64 | 25 600 | ||||
| 9 | 256 | 102 400 | ||||
| 11 | 1024 | 409 600 | ||||
| 13 | 4096 | 1 638 400 | ||||
| 15 | 16 384 | 6 553 600 | ||||
| Photonic crystal4 | 3 | 4 | 1600 | |||
| 5 | 16 | 6400 | ||||
| 7 | 64 | 25 600 | ||||
| 9 | 256 | 102 400 | ||||
| 11 | 1024 | 409 600 | ||||
| 13 | 4096 | 1 638 400 | ||||
| 15 | 16 384 | 6 553 600 | ||||
| Pyramid5 | 3 | 4 | 1600 | |||
| 5 | 16 | 6400 | ||||
| 7 | 64 | 25 600 | ||||
| 9 | 256 | 102 400 | ||||
| 11 | 1024 | 409 600 | ||||
| 13 | 4096 | 1 638 400 | ||||
| 15 | 16 384 | 6 553 600 |
Table 2. Results of the proposed method for six scattering potentials.
We note the following quantities:
The number of levels of the hierarchical tree.
The number of leaves on the finest level.
The total number of degrees of freedom.
Time in seconds to precompute the matrices.
Time in seconds to build scattering matrices .
Time in seconds to evaluate exterior field.
Time in seconds to solve for interior field.
The relative error in the exterior field (total field) evaluated at selected points.
The relative error in the interior field (total field) evaluated at selected off-grid points.
The maximal relative error in the interior field (total field) among all tensor Chebyshev grid points.
All errors are in the -norm and obtained by comparing to a reference solution at and . The total field refers to the sum of the scattered and incoming fields, i.e., , as in § 4.3.
We consider six test problems, where the incident field is given by a plane wave :
Gaussian bump: A Gaussian bump given by with , where is the center of the computational domain (Figure 10(a)). The computational domain is about 4 wavelengths across.

Figure 10. Gaussian bump

caption: Smooth mollifier

caption: Cavity

caption: Lens

caption: Photonic crystal

caption: Pyramid
caption: Scattering potentials .
Smooth mollifier: A smooth mollifier given by for with (Figure 10(b)). The wavenumber is set to , so that the computational domain is about 4 wavelengths across.
Cavity: A cavity potential is given in polar coordinates by (Figure 10(c)). The wavenumber is set to , so that the computational domain is about 8 wavelengths across.
Lens: A vertically-graded lens given by with (Figure 10(d)). The computational domain is about 31.8 wavelengths across.
Photonic crystal: A 20 by 20 square grid of Gaussian bumps with a channel removed with (the first complete band gap) (Figure 10(e)). The computational domain is about 12.2 wavelengths across.
Pyramid: A pyramid potential given by with (Figure 10(f)). The computational domain is about 4 wavelengths across.
Table 1 reports the computational cost of the algorithm for increasing numbers of levels. As the hierarchy is refined from 3 to 15 levels, the precomputation time increases from 13.5 s to 49.3 s, reflecting the additional cost associated with constructing the operators required at each level of the hierarchy. In contrast, the evaluation time remains nearly constant despite the substantial increase in problem size. The build and solve times exhibit more pronounced growth, roughly linear in terms of . These results demonstrate that the computational cost of the hierarchical merge is dominated increasingly by the construction and solution stages as the number of levels increases, whereas the evaluation stage remains relatively insensitive to the overall problem size.
Table 2 summarizes the convergence behavior of the hierarchical merge for six scattering potentials. For the Gaussian bump, smooth mollifier, and cavity test cases, the exterior, interior on-grid and off-grid errors decrease rapidly with refinement, reaching errors between approximately and at the finest resolution. The lens and photonic-crystal examples are more demanding on coarse grids and require substantially finer discretizations before achieving high accuracy. In particular, the lens errors decrease from at three levels to below at fifteen levels, while the photonic-crystal errors exhibit a substantial reduction at eleven levels, and decrease gradually until the finest resolution. Additionally the BIE operators from § 5.3 are refined in this case (i.e., is increased, see § 5.4). The pyramid case, since its -function is only , converges more gradually. Overall, these results demonstrate that the hierarchical merge produces convergent solutions across a range of smooth and highly structured scattering potentials, while the resolution required to attain a given accuracy depends on the complexity of the scattering potential. For the smooth examples, the errors eventually plateau under further refinement. We suspect that this behavior is primarily due to limitations in the accuracy of the boundary integral evaluations used to construct the universal operators, particularly for targets very close to the boundary.

Figure 11. (a) Gaussian bump; (b) Smooth mollifier; (c) Cavity; (d) Lens; (e) Photonic crystal; (f) Pyramid. Total field for different scattering potentials , given the incoming plane wave .
Exterior scattering and multiple right-hand sides
The solver introduced in this manuscript is direct, so it can naturally handle multiple right hand sides. Table 3 reports the timings of simultaneously performing 500 solves, corresponding to 500 angles of incidence of plane waves.
| Precomp. | Build | EEM | direct | from EEM | Memory (GB) | Error | |
| Volume solver | 46.7 | 54.9 | — | 14.92 | — | 10.6 | |
| Exterior scatterer | 45.8 | 53.9 | 1.0 | 0.29 | 1.25 |
Table 3. Cavity scatterer (, , 13 levels): timings and storage for the volume solver and the exterior scatterer, for 500 incidence directions. EEM denotes the exterior evaluation matrix; the error is the difference from a fine solve.
There are also many applications in which one is only interested in the scattered field at prescribed exterior locations. Examples include lens scattering [16, 34, 31], microcavity emission [26], inverse scattering [13, 11, 36, 32, 6] and photonic crystal scattering [27, 35]. Significant effort has gone into designing specialized solvers to compute only the exterior scattered field (see, e.g., [9, 10, 21]). In the present formulation, such a specialized solver is easily obtained by only retaining the root scatterer . As Table 3 shows, this leads to large memory savings and significant speed-up.

Figure 12. Magnitude of the total field for various incoming wave angles . Top: global field. Bottom: Field measured at an exterior receiver array (indicated in black).
Two types of exterior scattering are considered in Table 3, which differ in the way the obtained outgoing field at is evaluated in the exterior of . After the outgoing trace is computed for each incoming field by applying , both use the root level on followed by
each exterior evaluation being performed separately, assuming the exterior evaluation points differ per incoming field (‘direct’), or
using a precomputed exterior evaluation matrix (EEM), mapping scattered fields at to the prescribed, shared, evaluation points (‘from EEM’).
To be precise, assume the exterior points are given in advance. As in § 5.3, evaluates an exterior Helmholtz solution to these points. Then , with the root-level exterior Brakhage–Werner solution operator.
In both cases, the exterior solve is orders of magnitude faster than the volume solve. Solving 500 right-hand sides, even for the volume solve, is not significantly more expensive than solving a single right-hand side (see Table 1).
Conclusions
This paper describes a high-order accurate hierarchical direct solver for the two-dimensional Lippmann–Schwinger equation arising in acoustic scattering from variable media. The essential idea is to construct a reusable field-to-field scattering operator for the inhomogeneous medium, which would avoid a new global volume solve for every incident field. Each leaf subdomain is treated by a local Lippmann–Schwinger solve, while neighboring subdomains are merged recursively through boundary-only operations. As a result, volume solves are only operated on the finest level of the hierarchy, and the remainder of the computation is performed entirely in terms of boundary traces.
The resulting method naturally separates the computation into a reusable precomputation and an application stage. Once the scattering operator for a given medium, geometry, and frequency has been constructed, it can be applied to multiple incident fields without repeating the expensive volume discretization and hierarchical construction. Numerical experiments demonstrate high-order accuracy over a range of variable-media scattering problems, including smooth potentials as well as more challenging set-ups containing cavities, high-frequency problems, strong back-scattering, and other features that can make conventional iterative approaches difficult to use efficiently. The convergence results also illustrate the effect of the regularity of the scattering potential: smooth potentials exhibit rapid convergence, whereas potentials with limited regularity converge more gradually and may eventually be limited by the polynomial order used in the discretization.
Possible future work is to study the potential rank structure of the dense matrices arising throught the hierarchy. Although these matrices are dense, many of them have global low-rank structure, and the remaining matrices often exhibit hierarchical low-rank structure, such as hierarchically semiseparable (HSS) structure. In particular, exploiting low-rank and HSS representations could reduce both the storage requirements and the computational cost of the leaf solves, merging operations, and scattering-operator applications.
Overall, the numerical results demonstrate that the proposed hierarchical formulation provides an accurate and reusable direct solver for two-dimensional Lippmann–Schwinger scattering problems. Future work will focus on systematically exploiting the low-rank and HSS structure of the operators, with the goal of developing a compressed hierarchical implementation that retains the high-order accuracy while substantially reducing memory usage, construction cost, and application time.
Acknowledgements
This work was supported by the Office of Naval Research (FP00011501), by the National Science Foundation (DMS-2608883 and DMS-2313434), and by the Simons Foundation. JB was supported in part by the Peter O’Donnell Jr. Postdoctoral Fellowship at the Oden Institute, and in part by the Swedish Research Council under grant no. 2021-06594 while in residence at Institut Mittag-Leffler in Djursholm, Sweden during the fall semester of 2025.
Equivalent-source formulation
The field-to-field formulation in the main text represents both the incoming and outgoing fields of a box by their Dirichlet traces. There is a closely related alternative in which the incoming field is still represented by its trace , but the outgoing field is represented by an equivalent source density on the boundary. We record the formulation here because its algebra is distinct, as it makes the propagation between boxes a direct layer-potential evaluation, and it may prove useful in future variants of the method. The leaf solve is essentially unchanged; what differs is the representation of the outgoing field and, consequently, the merge.
Throughout this appendix, denotes the equivalent-source scattering operator, rather than the field-to-field scattering operator used in the main text.
Local equivalent-source scattering operator
Let be a box and suppose that the incoming trace is given. As in § 2, first reconstruct the incoming field and solve the locally restricted Lippmann–Schwinger equation. Equivalently, using (9),
The corresponding outgoing field is
Instead of representing by its boundary trace, seek a density on whose single-layer potential reproduces the same exterior field:
Assuming that a unique equivalent density has been selected, define the equivalent-source scattering operator
For the pure single-layer representation, uniqueness can fail at exceptional Helmholtz frequencies. This issue can be avoided by restricting the box sizes, or by replacing the single-layer representation by an appropriate combined-field representation. Since the present appendix is intended only to record the alternative hierarchical formulation, we do not pursue this question further.
Merge of two adjacent boxes
Consider the same geometry and boundary ordering as in § 3.2:
and
The parent incoming trace is continued to the complete child boundaries by the same universal operator used in the field-to-field formulation. Set
Let
denote the equivalent source densities on the two child boundaries, with
Split the child scattering operators according to their exterior and interface boundary pieces,
and form the block operator
Thus maps the total incoming traces on the two child boundaries to the corresponding equivalent source densities.
The field generated by the sources on one child provides part of the incoming field on its sibling. For example,
Here denotes the sibling-generated incoming trace. Each operator is simply a layer-potential evaluation operator: it maps a single-layer density on to the value of the resulting field on . For instance,
Combining the two relations in (51) gives
The merge equilibrium equation is therefore
Using (49), we obtain
It remains to replace the four child densities in by one equivalent density on the parent boundary. Write
where and are supported on and , respectively. Define the equivalent-source assembly operator by
The blocks are chosen so that the density generates the same field outside as the two child single-layer potentials. Equivalently,
Combining (54) and (55) gives the parent scattering operator
Thus the equivalent-source formulation has the same basic hierarchical structure as the field-to-field method: solve a local multiple-scattering equilibrium between the children, and then replace the pair by a single scattering operator for the parent.
Structure of the merge
The product that appears in (54) has a particularly useful block structure. From (50) and (52),
This identity is useful because almost every term in (58) contains at least one factor that is numerically low rank. The principal exceptions are the terms and , which couple the two copies of the common interface. These blocks are not globally low rank, but they have multilevel rank structure. Thus, just as in the field-to-field formulation, the merge contains considerable exploitable structure even though the present appendix does not attempt to use it.
Remark 5 (Comparison with the field-to-field formulation). The two formulations use the same incoming trace and the same local Lippmann–Schwinger solves. Their difference is entirely in the representation of the outgoing field. In the formulation used in the main text, the outgoing state is the boundary trace . This avoids equivalent-source reconstruction and makes the coupling across the common interface particularly simple: two blocks of are identities. In the present formulation, the outgoing state is the single-layer density . Propagation between children is then given directly by the universal single-layer evaluation operator , but the near-field blocks and require singular or nearly singular layer-potential evaluation, and the parent merge requires the additional equivalent-source reconstruction . These considerations are the main reasons that the field-to-field formulation was preferred in the implementation described in this paper.
Remark 6 (Universal operators). The interaction operator and the assembly operator are independent of the scattering potential , as is the incoming continuation operator . They depend only on the geometry and on the free-space Helmholtz operator, and can therefore be precomputed. The separated-boundary blocks of are smoothing and highly compressible. The interface blocks and require near-field evaluation and, at high accuracy, a multilevel representation. These observations provide a possible alternative route to accelerating the method if the equivalent-source formulation is revisited in future work.
References
- [1]J. C. Aguilar and Yu Chen, A high-order, fast algorithm for scattering calculation in two dimensions, Computers & Mathematics with Applications 47 (2004), no. 1, 1–11.
- [2]Sivaram Ambikasaran, Carlos Borges, Lise-Marie Imbert-Gerard, and Leslie Greengard, Fast, adaptive, high-order accurate discretization of the Lippmann–Schwinger equation in two dimensions, SIAM Journal on Scientific Computing 38 (2016), no. 3, A1770–A1787.DOI
- [3]Akash Anand, Ambuj Pandey, B. V. Rathish Kumar, and Jagabandhu Paul, An efficient high-order Nyström scheme for acoustic scattering by inhomogeneous penetrable media with discontinuous material interface, Journal of Computational Physics 311 (2016), 258–274.DOI
- [4]Travis Askham, Manas Rachh, Michael O’Neil, Jeremy Hoskins, Dan Fortunato, and Shidong Jiang, chunkie: a matlab integral equation toolbox, https://github.com/fastalgorithms/chunkie, 2023.
- [5]Mario Bebendorf, Hierarchical matrices, Lecture Notes in Computational Science and Engineering, vol. 63, Springer-Verlag, Berlin, 2008, A means to efficiently solve elliptic boundary value problems. MR 2451321 (2009k:15001)DOI
- [6]Carlos Borges, Adrianna Gillman, and Leslie Greengard, High resolution inverse scattering in two dimensions using recursive linearization, SIAM Journal on Imaging Sciences 10 (2017), no. 2, 641–664.arxiv.org/abs/1608.06871
- [7]Duan Chen, Min Hyung Cho, and Wei Cai, Accurate and efficient Nyström volume integral equation method for electromagnetic scattering of 3-d metamaterials in layered media, SIAM Journal on Scientific Computing 40 (2018), no. 1, B259–B282.DOI
- [8]Yu Chen, A fast, direct algorithm for the Lippmann-Schwinger integral equation in two dimensions, Adv. Comput. Math. 16 (2002), no. 2-3, 175–190, Modeling and computation in optics and electromagnetics. MR 1892248 (2003b:65139)DOI
- [9]Weng Cho Chew, Levent Gurel, Y-M Wang, Gregory Otto, Robert L Wagner, and Qing Huo Liu, A generalized recursive algorithm for wave-scattering solutions in two dimensions, IEEE transactions on microwave theory and techniques 40 (1992), no. 4, 716–723.DOI
- [10]Weng Cho Chew and Cai-Cheng Lu, The use of Huygens’ equivalence principle for solving the volume integral equation of scattering, IEEE Transactions on Antennas and Propagation 41 (1993), no. 7, 897–904.DOI
- [11]Weng Cho Chew and Y. M. Wang, Reconstruction of two-dimensional permittivity distribution using the distorted Born iterative method, IEEE Transactions on Medical Imaging 9 (1990), no. 2, 218–225.pubmed.ncbi.nlm.nih.gov/18222767
- [12]David Colton and Rainer Kress, Integral equation methods in scattering theory, vol. 72, SIAM, 2013.DOI
- [13][——— , Inverse acoustic and electromagnetic scattering theory, 4 ed., Applied Mathematical Sciences, vol. 93, Springer, Cham, 2019.
- [14]Eduardo Corona, Per-Gunnar Martinsson, and Denis Zorin, An o (n) direct solver for integral equations on the plane, Applied and Computational Harmonic Analysis 38 (2015), no. 2, 284–317.arxiv.org/abs/1303.5466
- [15]Ran Duan and Vladimir Rokhlin, High-order quadratures for the solution of scattering problems in two dimensions, J. Comput. Physics 228 (2009), no. 6, 2152–2174.
- [16]Benjamin Fuchs, Sébastien Palud, Laurent Le Coq, Olivier Lafond, Mohamed Himdi, and Sébastien Rondineau, Scattering of spherically and hemispherically stratified lenses fed by any real source, IEEE Transactions on Antennas and Propagation 56 (2008), no. 2, 450–460.DOI
- [17]M. Ganesh, S. C. Hawkins, and R. Hiptmair, Convergence analysis with parameter estimates for a reduced basis acoustic scattering T-matrix method, IMA Journal of Numerical Analysis 32 (2012), no. 4, 1348–1374.DOI
- [18]M. Ganesh and S.C. Hawkins, A fast high order algorithm for multiple scattering from large sound-hard three dimensional configurations, Journal of Computational and Applied Mathematics 362 (2019), 324 – 340.DOI
- [19]M. Ganesh, J.S. Hesthaven, and B. Stamm, A reduced basis method for electromagnetic scattering by multiple particles in three dimensions, Journal of Computational Physics 231 (2012), no. 23, 7756–7779.DOI
- [20]A. Gillman and P. Martinsson, A direct solver with o(n) complexity for variable coefficient elliptic pdes discretized via a high-order composite spectral collocation method, SIAM Journal on Scientific Computing 36 (2014), no. 4, A2023–A2046, arXiv.org report #1307.2665.arxiv.org/abs/1307.2665
- [21]Adrianna Gillman, Alex H. Barnett, and Per-Gunnar Martinsson, A spectrally accurate direct solution technique for frequency-domain scattering problems with variable media, BIT Numerical Mathematics 55 (2015), no. 1, 141–170.arxiv.org/abs/1308.5998
- [22]Abinand Gopal and Per-Gunnar Martinsson, An accelerated, high-order accurate direct solver for the Lippmann–Schwinger equation for acoustic scattering in the plane, Advances in Computational Mathematics 48 (2022), no. 4, 1–31.arxiv.org/abs/2007.12718
- [23]Lars Grasedyck and Wolfgang Hackbusch, Construction and arithmetics of ℋ-matrices, Computing 70 (2003), no. 4, 295–334.DOI
- [24]L. Greengard and V. Rokhlin, A fast algorithm for particle simulations, J. Comput. Phys. 73 (1987), no. 2, 325–348.
- [25]Johan Helsing, Solving Integral Equations on Piecewise Smooth Boundaries Using the RCIP Method: A Tutorial, Abstract and Applied Analysis 2013 (2013), no. SI51, 1 – 20.
- [26]Yushin Kim, Soo-Young Lee, Jung-Wan Ryu, Inbo Kim, Jae-Hyung Han, Heung-Sik Tae, Muhan Choi, and Bumki Min, Designing whispering gallery modes via transformation optics, Nature Photonics 10 (2016), no. 10, 647–652.
- [27]Philip Trøst Kristensen, Peter Lodahl, and Jesper Mørk, Light propagation in finite-sized photonic crystals: multiple scattering using an electric field integral equation, Journal of the Optical Society of America B 27 (2010), no. 2, 228–237.arxiv.org/abs/0908.4240
- [28]Fei Liu and Lexing Ying, Sparsify and sweep: An efficient preconditioner for the Lippmann–Schwinger equation, SIAM Journal on Scientific Computing 40 (2018), no. 2, B379–B404.arxiv.org/abs/1705.09443
- [29]P.G. Martinsson, A direct solver for variable coefficient elliptic pdes discretized via a composite spectral collocation method, Journal of Computational Physics 242 (2013), no. 0, 460 – 479.DOI
- [30]P.G. Martinsson and V. Rokhlin, A fast direct solver for boundary integral equations in two dimensions, J. Comp. Phys. 205 (2005), no. 1, 1–23.
- [31]Hossein Mosallaei and Yahya Rahmat-Samii, Nonuniform Luneburg and two-shell lens antennas: Radiation characteristics and design optimization, IEEE Transactions on Antennas and Propagation 49 (2001), no. 1, 60–69.DOI
- [32]Neslihan Ozmen, Robin Dapp, Michael Zapf, Hartmut Gemmeke, Nicole V Ruiter, and Koen WA Van Dongen, Comparing different ultrasound imaging methods for breast cancer detection, IEEE transactions on ultrasonics, ferroelectrics, and frequency control 62 (2015), no. 4, 637–646.DOI
- [33]V. Rokhlin, Rapid solution of integral equations of scattering theory in two dimensions, J. Comput. Phys. 86 (1990), 414–439.
- [34]John R. Sanford, Scattering by spherically stratified microwave lens antennas, IEEE Transactions on Antennas and Propagation 42 (1994), no. 5, 690–698.DOI
- [35]Mirbek Turduev, E Bor, and Hamza Kurt, Photonic crystal based polarization insensitive flat lens, Journal of Physics D: Applied Physics 50 (2017), no. 27, 275105.DOI
- [36]Zhun Wei and Xudong Chen, Deep-learning schemes for full-wave nonlinear inverse scattering problems, IEEE Transactions on Geoscience and Remote Sensing 57 (2019), no. 4, 1849–1860.DOI
- [37]Lexing Ying, Sparsifying preconditioner for the Lippmann–Schwinger equation, Multiscale Modeling & Simulation 13 (2015), no. 2, 644–660.arxiv.org/abs/1408.4495
- [38]Lexing Ying, George Biros, and Denis Zorin, A kernel-independent adaptive fast multipole algorithm in two and three dimensions, Journal of Computational Physics 196 (2004), no. 2, 591–626.
- [39]Leonardo Zepeda-Nuñez and Hongkai Zhao, Fast alternating bidirectional preconditioner for the 2d high-frequency Lippmann–Schwinger equation, SIAM Journal on Scientific Computing 38 (2016), no. 5, B866–B888.




