Inverse Born Series: Nonlinear Inversion

• IBS is a nonlinear inversion method that expands the forward scattering map into multilinear operators and recursively constructs inverse operators to recover unknown coefficients.
• IBS applies to diverse models such as Helmholtz, Schrödinger, Dirac, and Kerr, demonstrating versatility in handling weak-scattering problems and nonlinearities.
• IBS convergence relies on the smallness of measured data and regularization of the linear operator, with recursive, reduced, and restarted formulations balancing accuracy and computational cost.

Searching arXiv for recent and foundational papers on the inverse Born series to ground the article in cited literature. Inverse Born Series (IBS) denotes a class of nonlinear inversion procedures that recover an unknown coefficient, potential, or susceptibility from measured scattering data by formally inverting the forward Born expansion term by term. In the contemporary literature, IBS appears both as an abstract Banach-space construction for nonlinear forward maps and as a concrete reconstruction mechanism for Helmholtz, Schrödinger, diffusion, Dirac, Kerr, and second-harmonic-generation models. Its basic architecture is consistent across these settings: one expands the forward data map into multilinear operators $K_m$, chooses a bounded inverse or regularized pseudoinverse of the linearized operator $K_1$, and recursively constructs inverse operators $\mathcal K_m$ so that the resulting power series recovers the unknown when the relevant convergence hypotheses hold (Hoskins et al., 2022, Lee et al., 2021).

1. Abstract operator formulation

In its most general form, IBS is defined for a nonlinear forward map between Banach spaces. Hoskins and Schotland write the forward map as

$$F(\eta)=\Phi(\eta):=\sum_{m=1}^{\infty}K_m(\eta,\eta,\dots,\eta),$$

where $K_m:X^m\to Y$ is an $m$-linear forward operator, and seek an inverse map

$$I(\phi)=\sum_{m=1}^{\infty}\mathcal K_m(\phi,\phi,\dots,\phi)$$

with $\phi=\Phi(\eta)$ (Hoskins et al., 2022).

The first inverse coefficient is obtained from the linearized forward operator: $\mathcal K_1(\phi)=K_1^+(\phi),$ where $K_1^+$ is a bounded inverse or a regularized pseudoinverse of $K_1$0. Higher-order inverse terms are then determined recursively so that substituting the forward series into the inverse series reproduces the identity on the unknown. In the recursive form used in several problem classes,

$K_1$1

which makes clear that each inverse coefficient is an algebraic combination of $K_1$2 and the forward operators $K_1$3 (McNeill, 2024, Defilippis et al., 2022).

This formulation is local rather than global. The inverse series is not introduced as a universal inverse valid for arbitrary data; instead, it is a structured local expansion whose convergence depends on the size of the measured data, the norms of the multilinear forward operators, and the behavior of the chosen pseudoinverse $K_1$4 (Hoskins et al., 2022, Bardsley et al., 2013).

2. Forward scattering models from which IBS is derived

IBS inherits its structure from the forward scattering model. In multiple light scattering, Lee et al. work from the Lippmann–Schwinger equation

$K_1$5

with scattering potential $K_1$6, and in discrete voxel form

$K_1$7

This is the setting in which the usual weak-scattering Born expansion diverges once the spectral radius exceeds unity, motivating the modified Born construction used for inversion of multiple scattering (Lee et al., 2021).

For semilinear Helmholtz problems with Kerr nonlinearities, the forward field satisfies

$K_1$8

or, in the purely nonlinear case,

$K_1$9

and the forward Born operators are obtained by repeated application of the Neumann Green’s function in a Lippmann–Schwinger representation (Defilippis et al., 2022, Defilippis et al., 2024).

The same structural pattern extends to more specialized systems. For second harmonic generation, the unknown coefficient $\mathcal K_m$0 couples the two field components $\mathcal K_m$1 and $\mathcal K_m$2, and the forward Born operators satisfy explicit multilinear recursions built from the kernels $\mathcal K_m$3 and $\mathcal K_m$4 (McNeill, 2024). For two-coefficient Helmholtz inverse scattering, Cakoni, Meng, and Zhou introduce a forward series for the pair $\mathcal K_m$5 from far-field data at two distinct frequencies $\mathcal K_m$6, with $\mathcal K_m$7 combining the two-frequency information through a matrix $\mathcal K_m$8 and a Fourier-type operator $\mathcal K_m$9 (Cakoni et al., 15 Mar 2025). For Dirac equations arising in waveguide arrays, Schotland and Yu derive Born and inverse Born expansions for both a chiral hyperbolic model and an anti-chiral elliptic model, with the forward series written in a unified Banach-space framework $$F(\eta)=\Phi(\eta):=\sum_{m=1}^{\infty}K_m(\eta,\eta,\dots,\eta),$$0, $$F(\eta)=\Phi(\eta):=\sum_{m=1}^{\infty}K_m(\eta,\eta,\dots,\eta),$$1 or $$F(\eta)=\Phi(\eta):=\sum_{m=1}^{\infty}K_m(\eta,\eta,\dots,\eta),$$2 depending on the model (Schotland et al., 2 May 2026).

A central point across these examples is that IBS is not tied to one PDE class. The same multilinear inversion principle is reused once the forward map admits a convergent Born expansion and the leading operator $$F(\eta)=\Phi(\eta):=\sum_{m=1}^{\infty}K_m(\eta,\eta,\dots,\eta),$$3 can be stably inverted or regularized (Hoskins et al., 2022, Bardsley et al., 2013).

3. Convergence theory and error control

The convergence theory of IBS is typically formulated through multilinear bounds on the forward operators. A representative hypothesis assumes constants $$F(\eta)=\Phi(\eta):=\sum_{m=1}^{\infty}K_m(\eta,\eta,\dots,\eta),$$4 such that

$$F(\eta)=\Phi(\eta):=\sum_{m=1}^{\infty}K_m(\eta,\eta,\dots,\eta),$$5

With $$F(\eta)=\Phi(\eta):=\sum_{m=1}^{\infty}K_m(\eta,\eta,\dots,\eta),$$6, several problem-specific analyses obtain the radius

$$F(\eta)=\Phi(\eta):=\sum_{m=1}^{\infty}K_m(\eta,\eta,\dots,\eta),$$7

and prove convergence of the inverse Born series when $$F(\eta)=\Phi(\eta):=\sum_{m=1}^{\infty}K_m(\eta,\eta,\dots,\eta),$$8 (McNeill, 2024, Schotland et al., 2 May 2026, Schotland et al., 22 May 2026, Cakoni et al., 15 Mar 2025).

Hoskins and Schotland analyze the same problem in a Banach-space setting using geometric function theory. Their results separate the inverse problem into a holomorphic perturbation of the identity and establish convergence and approximation error under qualitatively weaker conditions than previously known. An important feature of that analysis is the explicit separation between truncation error, which comes from cutting off the inverse series after finitely many terms, and projection error, which comes from replacing an exact inverse of $$F(\eta)=\Phi(\eta):=\sum_{m=1}^{\infty}K_m(\eta,\eta,\dots,\eta),$$9 by a regularized pseudoinverse (Hoskins et al., 2022).

For semilinear Helmholtz models, the convergence statements are sharpened in two directions. First, the Kerr and second-harmonic-generation papers derive boundedness estimates for the forward Born operators and then invoke fixed point theory to obtain explicit smallness conditions for convergence of the forward Born series itself (Defilippis et al., 2022, McNeill, 2024). Second, “Nonlinearity helps the convergence of the inverse Born series” shows that if the coefficient of the linear term is known, an arbitrarily strong Kerr nonlinearity can be reconstructed, for sufficiently small data, and that similar convergence results hold for general polynomial nonlinearities (Defilippis et al., 2024). That result is notable because it rules out a simplistic identification of “stronger nonlinearity” with “worse IBS convergence” in every formulation.

Error estimates are likewise model dependent but structurally similar. When both the forward and inverse series converge, the error of the $K_m:X^m\to Y$0-term truncation is bounded by a geometric tail plus a term involving the defect of the pseudoinverse, typically through factors such as $K_m:X^m\to Y$1 or $K_m:X^m\to Y$2 (Hoskins et al., 2022, Schotland et al., 2 May 2026, Bardsley et al., 2013).

4. Recursive, restarted, and reduced forms

The most direct implementation of IBS computes $K_m:X^m\to Y$3 recursively from all lower-order forward and inverse coefficients. This is formally exact within the convergence regime, but its combinatorial structure is costly: the number of compositions appearing at order $K_m:X^m\to Y$4 grows rapidly (Ishida et al., 29 Nov 2025, Bardsley et al., 2013).

One response is the restarted inverse Born series. In the Banach-space formulation of “Restarted inverse Born series for the Schrödinger problem with discrete internal measurements,” the inverse series is re-expanded around the current iterate rather than evaluated once around a fixed reference. The resulting family of methods, denoted RIBS($K_m:X^m\to Y$5), contains the Gauss–Newton method at $K_m:X^m\to Y$6 and the Chebyshev–Halley method at $K_m:X^m\to Y$7 (Bardsley et al., 2013). This places IBS within a broader hierarchy of local nonlinear solvers rather than isolating it as a purely formal power-series device.

A second response is reduction of the recursive structure. In the Dirac inverse-scattering work of Schotland and Yu, the reduced inverse Born series (RIBS) retains only the “diagonal” second-order interactions: $K_m:X^m\to Y$8 and for $K_m:X^m\to Y$9,

$m$0

The paper reports that this reduced series often achieves almost the same reconstruction accuracy as the full IBS while lowering storage and cost (Schotland et al., 2 May 2026).

A related development appears in “Iterative inversion schemes for the Born series and the reduced inverse Born series,” which revisits a Newton-type iterative scheme

$m$1

and derives a fast variant that uses only $m$2 and $m$3: $m$4 The paper states that the standard recursive implementation grows exponentially when nonlinear terms are taken into account, whereas the fast variant has per-step cost equal to one application of $m$5 and one $m$6; it further shows the relation between this fast scheme and the reduced inverse Born series (Ishida et al., 29 Nov 2025).

The optical multiple-scattering work of Lee et al. adopts a different practical compromise. Although an inverse series for the unknown scattering potential can be written formally, the reconstruction is carried out by minimizing $m$7 and computing gradients by back-propagation through the same convergent modified Born propagator used in the forward model, followed by proximal-gradient descent with total-variation regularization (FISTA) and optional nonnegativity (Lee et al., 2021).

5. Applications and reported numerical behavior

In optical diffraction tomography, Lee et al. apply a modified-Born inverse solver to three-dimensional refractive-index reconstruction of optically thick specimens. Their reported examples include arrays of $m$8-$m$9m beads and red blood cells axially aligned by optical tweezers, a live CAR-T cell attacking a target, green algae, and a $$I(\phi)=\sum_{m=1}^{\infty}\mathcal K_m(\phi,\phi,\dots,\phi)$$0-$$I(\phi)=\sum_{m=1}^{\infty}\mathcal K_m(\phi,\phi,\dots,\phi)$$1m human pancreas tissue slab. The paper states that Rytov-only reconstructions gave relative errors $$I(\phi)=\sum_{m=1}^{\infty}\mathcal K_m(\phi,\phi,\dots,\phi)$$2, while the modified-Born inverse gave $$I(\phi)=\sum_{m=1}^{\infty}\mathcal K_m(\phi,\phi,\dots,\phi)$$3; that the mean-squared error in RI was an order of magnitude lower than Rytov-based methods; that the CAR-T-cell tomographic video was acquired at $$I(\phi)=\sum_{m=1}^{\infty}\mathcal K_m(\phi,\phi,\dots,\phi)$$4s intervals over $$I(\phi)=\sum_{m=1}^{\infty}\mathcal K_m(\phi,\phi,\dots,\phi)$$5min; and that a $$I(\phi)=\sum_{m=1}^{\infty}\mathcal K_m(\phi,\phi,\dots,\phi)$$6-$$I(\phi)=\sum_{m=1}^{\infty}\mathcal K_m(\phi,\phi,\dots,\phi)$$7m sphere forward solve took $$I(\phi)=\sum_{m=1}^{\infty}\mathcal K_m(\phi,\phi,\dots,\phi)$$8s with FDTD versus $$I(\phi)=\sum_{m=1}^{\infty}\mathcal K_m(\phi,\phi,\dots,\phi)$$9s with the modified Born series, with each inversion iteration requiring about $\phi=\Phi(\eta)$0s and reconstructions converging in $\phi=\Phi(\eta)$1s on a single GPU (Lee et al., 2021).

For Kerr-type nonlinear scattering, the numerical results are more explicitly tied to the theoretical convergence radii. The 2022 Kerr paper reports one-dimensional and two-dimensional finite-element experiments in FEniCS, including simultaneous reconstruction of linear and nonlinear susceptibilities, and notes that reconstruction degrades above the Born-series radius (Defilippis et al., 2022). The 2024 nonlinear-only Kerr paper studies a two-dimensional unit disk with 16 source locations, 32 detectors, and two frequencies $\phi=\Phi(\eta)$2, reporting that for very high-contrast target $\phi=\Phi(\eta)$3 and a small source amplitude $\phi=\Phi(\eta)$4, already the first term $\phi=\Phi(\eta)$5 captures nearly all of the structure, whereas for moderate contrast and $\phi=\Phi(\eta)$6, 5–10 terms of the IBS noticeably improve on the linearized reconstruction and recover sharp interfaces (Defilippis et al., 2024).

For two function-valued coefficients in the Helmholtz equation, Cakoni, Meng, and Zhou construct a regularized inverse Born approximation and series from multi-frequency far-field data at two distinct frequencies. Their preliminary numerical tests on $\phi=\Phi(\eta)$7, with true contrasts $\phi=\Phi(\eta)$8 given by combinations of Gaussians of peak amplitude up to $\phi=\Phi(\eta)$9 and $\mathcal K_1(\phi)=K_1^+(\phi),$0 noise added to far-field data, report relative $\mathcal K_1(\phi)=K_1^+(\phi),$1-errors falling from $\mathcal K_1(\phi)=K_1^+(\phi),$2 at $\mathcal K_1(\phi)=K_1^+(\phi),$3 down to $\mathcal K_1(\phi)=K_1^+(\phi),$4 after 3–4 IBS terms (Cakoni et al., 15 Mar 2025).

Schotland and Yu validate both IBS and RIBS for inverse scattering in Dirac equations arising in waveguide arrays. In the chiral model, low contrast $\mathcal K_1(\phi)=K_1^+(\phi),$5 is reported to require one or two terms, medium contrast $\mathcal K_1(\phi)=K_1^+(\phi),$6 requires $\mathcal K_1(\phi)=K_1^+(\phi),$7 terms, and high contrast $\mathcal K_1(\phi)=K_1^+(\phi),$8 leads to breakdown. In the anti-chiral model, low contrast $\mathcal K_1(\phi)=K_1^+(\phi),$9 gives approximately $K_1^+$0 relative $K_1^+$1-error at $K_1^+$2, medium contrast $K_1^+$3 gives approximately $K_1^+$4 at $K_1^+$5, and high contrast $K_1^+$6 shows no convergence beyond $K_1^+$7; the paper also concludes that the anti-chiral model typically gives better reconstructions than the chiral one due to richer boundary data (Schotland et al., 2 May 2026).

The phaseless-data setting further broadens the application range. “Reconstruction methods for inverse scattering problems with phaseless data” extends IBS to total-field intensity, far-field total-field intensity, and far-field scattered-field intensity. In that paper, phaseless total-field IBS converges only for low contrast, a Fourier-based reciprocity method is used to reconstruct the linear term in the far field, and a polarization identity is used to recover phase information for scattered-field data before applying IBS; the reported experiments indicate reliable low-contrast reconstructions and failure for higher contrasts, consistent with the stated convergence radius (Schotland et al., 22 May 2026).

6. Limitations, distinctions, and neighboring notions

IBS is intrinsically local. The modern literature repeatedly ties its success to smallness of the unknown, smallness of the measured data, or smallness of the first reconstructed iterate $K_1^+$8. The fast iterative and reduced-series formulations do not remove that limitation; they change the computational pathway through the same local regime. In particular, the 2025 fast-iteration paper states that the method requires smallness of $K_1^+$9 or $K_1$00 so that the Born series converges, that regularization of $K_1$01 is critical to avoid amplifying noise, and that the method reconstructs only components in the subspace spanned by large singular values of $K_1$02 (Ishida et al., 29 Nov 2025).

A second limitation is that richer data and stronger forward solvers do not automatically imply larger convergence radii. The phaseless-data study reports that all proposed phaseless IBS-based methods reconstruct low-contrast potentials reliably, whereas high contrast is beyond the convergence radius, and the optical multiple-scattering study replaces the standard weak-scattering Born expansion by a modified Born propagator precisely because the classical series diverges beyond the weak-scattering regime (Schotland et al., 22 May 2026, Lee et al., 2021).

The literature also contains a distinct object called the inverted Born series. In the pionless effective-field-theory analysis of $K_1$03-wave quartet $K_1$04 scattering, the author expands the inverse on-shell $K_1$05-matrix around the unitary limit and shows that, unlike the ordinary Born series, the inverted Born series converges rapidly for cutoffs $K_1$06 MeV. This construction is an expansion of $K_1$07, not the Banach-space multilinear inverse map built from $K_1$08 and $K_1$09 that defines IBS in the inverse-scattering literature (Ando, 2013). The distinction matters because the two series serve different analytical purposes even though the names are similar.

At the same time, it would be inaccurate to reduce IBS to a purely weak-nonlinearity method in every context. The 2024 Kerr analysis proves that an arbitrarily strong Kerr nonlinearity can be reconstructed, for sufficiently small data, when the coefficient of the linear term is known, and extends the same conclusion to general polynomial nonlinearities (Defilippis et al., 2024). A plausible implication is that the decisive quantity for IBS is not “nonlinearity” in isolation but the interaction between model parameterization, the chosen linear inverse $K_1$10, and the amplitude regime in which the data are generated.