Ferkans — Interactive Telecom Tutor

When First-Order Approximations Fail

When $\kappa_{0} a |\chi_0|$ is not small, the Born and Rytov approximations introduce significant model error. This section presents methods that iteratively refine the internal field and contrast estimate, progressively accounting for multiple scattering. These methods are more computationally expensive but extend imaging to higher contrast and larger objects.

Definition:
The Born Iterative Method (BIM)

The BIM uses the free-space Green's function at every iteration but updates the total field:

Compute $u^{(k)}$ using the current $\chi^{(k)}$ via forward solve of the Lippmann-Schwinger equation.
Form the Born-linearized system with $G_0$ (free-space) and $u^{(k)}$ (current total field).
Solve for $\chi^{(k+1)}$ .

BIM is cheaper per iteration (no Green's function recomputation) but converges more slowly than DBIM for high-contrast objects.

Definition:
The Distorted Born Iterative Method (DBIM)

The DBIM alternates between two steps:

Forward solve: Given the current contrast estimate $\chi^{(k)}$ , compute the total field $u^{(k)}$ by solving the Lippmann-Schwinger equation.
Linearized update: The scattered field perturbation satisfies:

$\delta u^{\text{sca}} = \kappa_{0}^{2} \int_D G^{(k)}(\mathbf{r}, \mathbf{r}')\, \delta\chi(\mathbf{r}')\,u^{(k)}(\mathbf{r}')\,d\mathbf{r}',$

where $G^{(k)}$ is the Green's function of the current background (including $\chi^{(k)}$ ), not free space.

The update $\delta\chi$ is found by solving the linearized system, and $\chi^{(k+1)} = \chi^{(k)} + \delta\chi$ .

Distorted Born Iterative Method (DBIM)

Complexity:

O(N^2)

per forward solve +

O(N^2)

for Green's function update +

O(MN)

for linearized inverse

Input: Measured data

\mathbf{y}

, incident fields

\{u_p^{\text{inc}}\}

, regularization

\lambda

Output: Contrast estimate

\hat{\chi}

1. Initialize:

\chi^{(0)} \leftarrow \chi_{\text{Born}}

(Born reconstruction)

2. for

k = 0, 1, 2, \ldots

until convergence:

a. Forward solve: Compute

u^{(k)}

from

(\mathbf{I} - \mathbf{G}\text{diag}(\chi^{(k)}))\mathbf{u}^{(k)} = \mathbf{u}^{\text{inc}}

b. Green's function update: Compute

G^{(k)}

for background

\chi^{(k)}

c. Linearize: Form Jacobian

\mathbf{J}^{(k)}

with entries

J_{m,n} = \kappa_{0}^{2} G^{(k)}(\mathbf{r}_{m}, \mathbf{p}_{n}) u^{(k)}(\mathbf{p}_{n})

d. Update:

\delta\chi = (\mathbf{J}^{(k)H}\mathbf{J}^{(k)} + \lambda\mathbf{I})^{-1}\mathbf{J}^{(k)H}(\mathbf{y} - F(\chi^{(k)}))

e.

\chi^{(k+1)} \leftarrow \chi^{(k)} + \delta\chi

3. return

\chi^{(k+1)}

The DBIM is equivalent to Gauss-Newton optimization when no regularization penalty gradient is included.

Definition:
Contrast Source Inversion (CSI)

The CSI method introduces the contrast source $w = \chi u$ as the primary unknown. The cost functional jointly minimizes data and domain residuals:

$\min_{w, \chi} \; \alpha\|\mathcal{G}_S w - \mathbf{y}\|^2 + (1 - \alpha)\|w - \chi(u^{\text{inc}} + \mathcal{G}w)\|^2.$

CSI avoids forward solves altogether — updates to $w$ and $\chi$ are alternating projections. This makes it significantly faster than DBIM for large problems, though convergence guarantees are weaker.

Computational Cost of Iterative Scattering Methods

Method	Cost per iteration	Convergence	Contrast range
Born (1 step)	$O(MN)$	—	$\kappa_{0} a\|\chi\| \ll 1$
BIM	$O(N^2)$ forward + $O(MN)$ inverse	Slow	Moderate
DBIM	$O(N^2)$ forward + $O(N^2)$ Green's + $O(MN)$	Fast	High
Gauss-Newton	$O(N^2)$ forward + $O(MN^2)$ Jacobian	Quadratic (local)	High
CSI	$O(MN)$ per alternation	Linear	Moderate-high

Definition:
Method of Moments for Full-Wave Solutions

The Method of Moments (MoM) discretizes the Lippmann-Schwinger equation directly. Expanding $\chi u$ in basis functions $\{b_n(\mathbf{r})\}_{n=1}^N$ , testing with $\{t_m(\mathbf{r})\}_{m=1}^N$ (Galerkin or point-matching) yields:

$(\mathbf{I} - \mathbf{G}\,\text{diag}(\chi))\,\mathbf{u} = \mathbf{u}^{\text{inc}},$

where $[\mathbf{G}]_{mn} = \kappa_{0}^{2} \int\int t_m(\mathbf{r})\, G(\mathbf{r}, \mathbf{r}')\,b_n(\mathbf{r}')\,d\mathbf{r}\,d\mathbf{r}'$ .

MoM provides the exact (within discretization) forward solution needed for BIM/DBIM iterations.

⚠️Engineering Note

Computational Feasibility of Iterative Methods

For 3D imaging with $N > 10^6$ voxels, even a single MoM forward solve requires $O(N^2)$ storage and $O(N^3)$ operations for direct methods. Practical approaches include:

FFT acceleration: For uniform grids, the Green's function convolution becomes a Fourier-domain multiplication: $O(N \log N)$ .
Fast multipole method (FMM): Reduces matrix-vector products to $O(N \log N)$ for arbitrary geometries.
Conjugate gradient solvers: Iterative solution of the MoM system, requiring only matrix-vector products.

Even with these accelerations, DBIM for 3D problems remains computationally challenging, motivating the Born-linear approaches and learned reconstructions in later chapters.

Practical Constraints

•
3D MoM requires $O(N^2)$ memory for the dense interaction matrix
•
Each DBIM iteration requires one forward solve plus one linearized inverse solve
•
FFT acceleration requires uniform grid discretization

Example: BIM vs DBIM Convergence

A 2D circular inhomogeneity with $\chi_0 = 0.5$ and $\kappa_{0} a = 2$ is imaged with 16 transmitters and 16 receivers.

Compare BIM and DBIM convergence by tracking the relative reconstruction error $\|\chi^{(k)} - \chi_{\text{true}}\| / \|\chi_{\text{true}}\|$ over iterations.

Solution

Expected behavior

BIM: Converges slowly because it uses the free-space Green's function, which does not account for the current contrast estimate. Typical: 15--20 iterations to reduce error below 10%.
DBIM: Converges faster because the distorted Green's function better captures the scattering environment. Typical: 5--8 iterations for the same error threshold.
Born (iteration 0): With $\kappa_{0} a |\chi_0| = 1.0$ , the Born estimate has $\sim 30$ % error.

Historical Note: The Evolution of Iterative Inverse Scattering

1989--1997

The Born iterative method was introduced by Wang and Chew in 1989, followed by the distorted Born iterative method (DBIM) by Chew and Wang in 1990. These methods brought nonlinear inverse scattering from a purely theoretical pursuit to a practical computational tool. The contrast source inversion method by van den Berg and Kleinman (1997) provided a crucial speedup by eliminating forward solves, making large-scale 3D inversions feasible for the first time.

,

Common Mistake: Initialization Matters for Iterative Methods

Mistake:

Starting iterative methods (BIM, DBIM) from zero initial contrast ( $\chi^{(0)} = 0$ ) instead of the Born or backpropagation estimate. With a poor initial guess, iterative methods may converge to a local minimum or diverge entirely.

Correction:

Always initialize with the Born approximation reconstruction or the backpropagation image $\hat{\ntn{refl_vec}}^{\text{BP}} = \mathbf{A}^{H} \ntn{img}$ . This provides a good starting point that is already close to the global minimum for weakly scattering objects.

Key Takeaway

BIM and DBIM iteratively refine Born-like linearizations, updating the total internal field at each step. DBIM also updates the background Green's function, providing faster convergence at higher computational cost. The Method of Moments provides exact forward solutions. Contrast Source Inversion avoids forward solves by jointly optimizing the contrast source and contrast. All iterative methods require good initialization — the Born or Rytov approximation is the natural starting point.

Beyond First-Order: Iterative Methods