Ferkans — Interactive Telecom Tutor

RF Diffraction Tomography: From Rays to Waves

The Fourier diffraction theorem (FDT) is the wave-based generalization of the Fourier slice theorem. Instead of straight lines in Fourier space, each view angle fills a semicircular arc (the Ewald sphere). This section derives the FDT under the Born approximation, develops the direct inversion pipeline (gridding + inverse FFT, NUFFT), and introduces iterative alternatives (SIRT, ART, CGLS) for incomplete angular coverage.

Definition:
Born Approximation for Weak Scatterers

Consider an object with contrast function $\chi(\mathbf{r})$ embedded in a homogeneous background with wavenumber $k_0$ . Under the first Born approximation, the scattered field at position $\mathbf{r}$ due to a plane-wave incidence $E_{\text{inc}} = e^{jk_0\hat{\mathbf{k}}_i \cdot \mathbf{r}}$ is:

$E_s(\mathbf{r}) = k_0^2 \int G(\mathbf{r}, \mathbf{r}')\,\chi(\mathbf{r}')\, e^{jk_0\hat{\mathbf{k}}_i \cdot \mathbf{r}'}\,d\mathbf{r}',$

where $G(\mathbf{r}, \mathbf{r}') = \frac{e^{jk_0|\mathbf{r} - \mathbf{r}'|}}{4\pi|\mathbf{r} - \mathbf{r}'|}$ (3D) is the free-space Green's function.

Validity: The Born approximation requires weak scattering: $|E_s| \ll |E_{\text{inc}}|$ throughout the object, which translates to $k_0\|\chi\|_\infty D \ll 1$ where $D$ is the object diameter.

Theorem: The Fourier Diffraction Theorem

(Wolf, 1969; Devaney, 1982). Under the first Born approximation, the far-field scattered data measured along direction $\hat{\mathbf{k}}_s$ for incidence direction $\hat{\mathbf{k}}_i$ at wavenumber $k_0$ is proportional to the object spectrum evaluated at the Ewald vector:

$\tilde{E}_s(k_0, \hat{\mathbf{k}}_s, \hat{\mathbf{k}}_i) \propto \tilde{\chi}\!\left( k_0(\hat{\mathbf{k}}_s - \hat{\mathbf{k}}_i)\right) = \tilde{\chi}(\mathbf{K}),$

where $\mathbf{K} = k_0(\hat{\mathbf{k}}_s - \hat{\mathbf{k}}_i)$ is the Ewald vector.

For fixed $\hat{\mathbf{k}}_i$ and varying $\hat{\mathbf{k}}_s$ in 2D, the locus of $\mathbf{K}$ traces a semicircular arc of radius $k_0$ centered at $-k_0\hat{\mathbf{k}}_i$ -- the Ewald circle (or Ewald sphere in 3D).

Each scattered-field measurement tells us the strength of a particular spatial frequency component of the object. The specific spatial frequency probed is $\mathbf{K} = k_0(\hat{\mathbf{k}}_s - \hat{\mathbf{k}}_i)$ : the difference between outgoing and incoming wave directions, scaled by the wavenumber. Varying the receiver direction sweeps out an arc of probed spatial frequencies.

Proof

Spectral representation of the Green's function

In the far field ( $|\mathbf{r}| \to \infty$ ), the stationary-phase evaluation of $G$ gives: $G(\mathbf{r}, \mathbf{r}') \approx C(r)\, e^{-jk_0\hat{\mathbf{k}}_s \cdot \mathbf{r}'}$ where $\hat{\mathbf{k}}_s = \mathbf{r}/|\mathbf{r}|$ .

Substitution into the Born integral

Substituting into the Born integral:

$E_s \propto \int \chi(\mathbf{r}')\, e^{j(k_0\hat{\mathbf{k}}_i - k_0\hat{\mathbf{k}}_s) \cdot \mathbf{r}'}\,d\mathbf{r}' = \tilde{\chi}(k_0(\hat{\mathbf{k}}_s - \hat{\mathbf{k}}_i)). \quad \blacksquare$

,

Definition:
Rytov Approximation

The Rytov approximation models the total field as $E = E_{\text{inc}}\,e^{\phi_s}$ where $\phi_s$ is a complex phase perturbation. The FDT also holds for the Rytov approximation, with $\phi_s$ replacing $E_s$ :

$\tilde{\phi}_s(k_0, \hat{\mathbf{k}}_s, \hat{\mathbf{k}}_i) \propto \tilde{\chi}(\mathbf{K}).$

Born vs. Rytov: Born is valid when $|E_s| \ll |E_{\text{inc}}|$ (small scattering amplitude); Rytov is valid when $|\nabla \phi_s| \ll k_0$ (small phase gradient), making it more accurate for large smooth objects. For RF imaging of weak scatterers, both give similar results.

Geometry of the Ewald Circle

Incident wave along $\hat{\mathbf{x}}$ . For $\hat{\mathbf{k}}_i = \hat{\mathbf{x}}$ , the Ewald vector is $\mathbf{K} = k_0(\hat{\mathbf{k}}_s - \hat{\mathbf{x}})$ . As $\hat{\mathbf{k}}_s$ varies over the unit circle, $\mathbf{K}$ traces a circle of radius $k_0$ centered at $(-k_0, 0)$ :

$(K_x + k_0)^2 + K_y^2 = k_0^2.$

Comparison with CT. In X-ray CT (ray optics, $k_0 \to \infty$ ), the curvature $1/k_0 \to 0$ and the arc degenerates into a straight line -- recovering the Fourier slice theorem. The FDT reduces to the FST in the high-frequency limit.

Accessible region. For a single frequency and all view angles $\theta_i \in [0, 2\pi)$ , the union of all Ewald circles covers a disk of radius $2k_0$ (since $|\mathbf{K}|_{\max} = 2k_0$ at backscatter). This sets the diffraction limit for single-frequency imaging.

Definition:
k-Space Coverage Map

For a measurement configuration with $N_v$ view angles $\{\theta_v\}$ at a single frequency $k_0$ , the k-space coverage map is the union of all Ewald arcs:

$\mathcal{K} = \bigcup_{v=1}^{N_v} \left\{\mathbf{K} = k_0(\hat{\mathbf{k}}_s - \hat{\mathbf{k}}_i^{(v)}) : \hat{\mathbf{k}}_s \in S^1 \right\}.$

The coverage density $\rho(\mathbf{K})$ counts the number of measurements mapping to a neighborhood of $\mathbf{K}$ in Fourier space. Reconstruction quality depends on both the coverage extent (resolution) and the coverage density (noise robustness).

k-Space Coverage and DT Reconstruction

Left: Ewald circles in the $(K_x, K_y)$ plane for the selected number of Tx view angles. Each arc is color-coded by view angle. The dashed circle shows the maximum accessible spatial frequency $2k_0$ .

Right: Reconstruction of a simple phantom from the Fourier data on these arcs using gridding + inverse FFT.

Increase $N_v$ to fill angular gaps and improve reconstruction quality. Observe how the "rosette" pattern in k-space progressively fills the $2k_0$ disk.

Parameters

Number of view angles

N_v

8

\text{SNR}

(dB)30

Definition:
Density Compensation Function

For non-uniform Fourier sampling on the Ewald arcs, the density compensation function (DCF) assigns weights $w_i$ to each sample $\mathbf{K}_i$ such that the reconstruction

$\hat{\chi}(\mathbf{r}) = \sum_{i=1}^{M} w_i\,\tilde{\chi}(\mathbf{K}_i)\, e^{j\mathbf{K}_i \cdot \mathbf{r}}$

is unbiased. The DCF satisfies the Voronoi condition: $w_i$ is proportional to the area of the Voronoi cell of $\mathbf{K}_i$ in the Fourier plane. Pipe & Menon's iterative algorithm converges in approximately 10 iterations:

$w^{(t+1)} = w^{(t)} / \left[\sum_{j} w_j^{(t)}\,\text{PSF}(\mathbf{K}_i - \mathbf{K}_j)\right].$

Direct Inversion: Gridding + Inverse FFT

Complexity:

O(MW^2 + N\log N)

where

M

= number of samples,

N

= image pixels,

W

= kernel width

Input: Scattered field data

\{E_s(\theta_v, \phi_s, k_0)\}

,

view angles

\{\theta_v\}

, wavenumber

k_0

.

1. Map to k-space: For each measurement, compute

\mathbf{K} = k_0(\hat{\mathbf{k}}_s - \hat{\mathbf{k}}_i^{(v)})

.

2. Compute DCF weights

\{w_i\}

via Pipe-Menon iteration

(10 iterations).

3. Grid onto Cartesian mesh: For each non-uniform sample

(\mathbf{K}_i, w_i \tilde{\chi}(\mathbf{K}_i))

, spread onto

the nearest grid points using a Kaiser-Bessel kernel

(width

W = 6

, oversampling factor

\alpha = 2

).

4. Inverse FFT: Apply 2D IFFT to the gridded data.

5. Deapodize: Divide by the Fourier transform of the

Kaiser-Bessel kernel to correct for the gridding convolution.

Output: Reconstructed contrast function

\hat{\chi}(\mathbf{r})

.

The NUFFT (non-uniform FFT) achieves the same accuracy with guaranteed error bounds: $O(N\log N + M\log(1/\epsilon))$ where $\epsilon$ is the approximation tolerance.

Iterative Alternatives: SIRT, ART, CGLS

Complexity:

O(MQ)

per iteration for SIRT/CGLS;

O(Q)

per sub-iteration for ART

Formulation: Discretize the FDT as a linear system

\mathbf{y} = \mathbf{A}\,\mathbf{c} + \mathbf{w}

, where

\mathbf{A} \in \mathbb{C}^{M \times Q}

encodes the

Ewald-arc sampling pattern.

SIRT (Simultaneous Iterative Reconstruction Technique):

\mathbf{c}^{(t+1)} = \mathbf{c}^{(t)} + \omega\,\mathbf{A}^{H} \mathbf{D}_r^{-1} (\mathbf{y} - \mathbf{A}\,\mathbf{c}^{(t)})

where

\mathbf{D}_r = \text{diag}(\mathbf{A}\,\mathbf{1})

normalizes by the number of rays per pixel.

ART (Algebraic Reconstruction Technique):

Update one measurement at a time (Kaczmarz iteration):

\mathbf{c}^{(t+1)} = \mathbf{c}^{(t)} + \omega\,\frac{y_i - \mathbf{a}_i^H\mathbf{c}^{(t)}}{ \|\mathbf{a}_i\|^2}\,\mathbf{a}_i

CGLS (Conjugate Gradient Least Squares):

Apply CG to the normal equations

\mathbf{A}^{H}\mathbf{A}\,\mathbf{c} = \mathbf{A}^{H}\mathbf{y}

.

Converges in at most

\text{rank}(\mathbf{A})

iterations.

Iterative methods are preferred over direct inversion when angular coverage is incomplete: they can incorporate regularization (e.g., $\ell_1$ or TV penalty) and handle arbitrary sampling geometries. For well-sampled configurations, gridding + IFFT is faster.

Example: Fourier Data from a Single View Angle

2D imaging at $f = 10$ GHz ( $k_0 = 209.4$ rad/m). Incident wave along $\hat{\mathbf{x}}$ . Scattered field measured over $\hat{\mathbf{k}}_s \in [0, 2\pi)$ .

What is the Fourier coverage from this single view, and what resolution does it provide?

Solution

Identify the Ewald circle

A single Ewald circle centered at $(-k_0, 0)$ with radius $k_0$ . This arc passes through the origin (forward scattering, $\hat{\mathbf{k}}_s = \hat{\mathbf{k}}_i$ ) and reaches $(-2k_0, 0)$ at backscatter.

Resolution analysis

Cross-range: the arc's $K_y$ extent spans $[-k_0, k_0]$ , giving $\delta_y \approx \pi/k_0 = \lambda/2$ .

Range: $K_x$ spans $[-2k_0, 0]$ , giving $\delta_x \approx \pi/k_0 = \lambda/2$ .

A single view at a single frequency gives $\lambda/2$ resolution but only along a curved slice -- the image has severe artifacts due to incomplete Fourier coverage.

Example: Gridding vs. CGLS for Incomplete Angular Coverage

A 30 cm phantom is imaged at $f_0 = 10$ GHz with $N_v = 8$ views uniformly spaced over $[0, 2\pi)$ . Compare direct inversion (gridding + IFFT) with CGLS (100 iterations) at $\text{SNR} = 25$ dB.

Solution

Setup

With 8 views, the k-space coverage consists of 8 Ewald arcs forming a rosette pattern with significant angular gaps. The coverage ratio is $\eta \approx 0.35$ .

Gridding + IFFT result

Gridding produces a reconstruction with PSNR $\approx 18$ dB and visible streak artifacts along the uncovered Fourier directions. The DCF partially compensates for non-uniform density but cannot fill the gaps.

CGLS result

CGLS converges to a solution with PSNR $\approx 22$ dB after 50 iterations. The iterative process implicitly applies a spectral filter that smoothly interpolates across the Fourier gaps, reducing streak artifacts at the cost of mild blurring.

When to use each method

Gridding + IFFT is preferred for well-sampled configurations ( $N_v > 30$ ) where speed matters. Iterative methods (CGLS, SIRT) are preferred for limited-view scenarios ( $N_v < 20$ ) where regularization is essential. For sparse scenes, combining CGLS initialization with FISTA refinement (Chapter 13) yields the best results.

Common Mistake: Gridding Artifacts and Kernel Choice

Mistake:

Interpolating non-uniform Fourier data onto a Cartesian grid introduces errors:

Aliasing: A narrow gridding kernel causes energy leakage from outside the image domain.
Smoothing: A wide kernel degrades effective resolution.
Phase errors: Poor interpolation of complex-valued data introduces phase errors that corrupt the reconstruction.

Correction:

Use a Kaiser-Bessel kernel with width $W = 6$ and oversampling factor $\alpha = 2$ (grid is $2\times$ finer than the image). Apply deapodization to compensate for the kernel's spectral roll-off. Validate against direct (non-gridded) reconstruction on a small test case.

Common Mistake: Born Approximation Breaks for Strong Scatterers

Mistake:

The Born approximation (and therefore the FDT) fails when the object is electrically large or has strong dielectric contrast. The scattering parameter $\alpha = k_0 |\chi|_{\max} D$ must satisfy $\alpha \ll 1$ . For a 10 cm object at 10 GHz with $|\chi| = 0.3$ : $\alpha = 209.4 \times 0.3 \times 0.1 = 6.3$ -- the Born approximation is invalid.

Correction:

For strong scatterers, use the distorted Born iterative method (DBIM) or full-waveform inversion, which iteratively linearize around the current estimate. Alternatively, use the Rytov approximation for objects with small phase gradients but large total phase accumulation.

Why This Matters: Diffraction Tomography and MIMO Radar Imaging

In MIMO radar imaging, each Tx-Rx pair probes a specific spatial frequency $\mathbf{K}$ determined by the array geometry. The collection of all Tx-Rx pairs creates a k-space coverage map that is mathematically identical to the multi-view DT coverage. The virtual aperture of a MIMO array (Chapter 8) determines which Ewald arcs are sampled. Increasing the number of Tx/Rx elements is analogous to increasing $N_v$ in DT -- both fill k-space more densely and improve reconstruction quality.

See full treatment in Chapter 8

Quick Check

What happens to the Ewald circle in the Fourier diffraction theorem as the wavenumber $k_0 \to \infty$ (high-frequency limit)?

It shrinks to a point at the origin

Its curvature decreases and it degenerates into a straight line

It becomes a full circle covering all of Fourier space

It rotates by 90 degrees

Correction:

Its curvature decreases and it degenerates into a straight line

The curvature is $1/k_0 \to 0$ , so the arc flattens into a straight radial line, recovering the Fourier slice theorem.

Quick Check

The density compensation function (DCF) in gridding-based reconstruction corrects for:

Additive measurement noise

Non-uniform sampling density in k-space

The Born approximation error

Antenna pattern variations

Correction:

Non-uniform sampling density in k-space

The DCF weights each k-space sample inversely proportional to the local sampling density, preventing bias toward oversampled regions.

Ewald Vector

The spatial frequency vector $\mathbf{K} = k_0(\hat{\mathbf{k}}_s - \hat{\mathbf{k}}_i)$ that links a scattered-field measurement to a point in the object's Fourier space. Named after Paul Ewald, who introduced the geometric construction for X-ray crystallography.

Related: Fourier Diffraction Theorem

Fourier Diffraction Theorem

The theorem stating that, under the Born approximation, the far-field scattered data is proportional to the object spectrum evaluated on an Ewald circle (2D) or Ewald sphere (3D). Generalizes the Fourier slice theorem from ray optics to wave optics.

Related: Ewald Vector

Key Takeaway

The Fourier diffraction theorem is the wave-optics generalization of the Fourier slice theorem: each measurement maps to the object spectrum on an Ewald circle, not a straight line. Direct inversion via gridding + IFFT is fast for well-sampled configurations; iterative methods (SIRT, ART, CGLS) handle incomplete angular coverage. The density compensation function corrects for non-uniform k-space sampling. All methods require the Born or Rytov approximation -- strong scatterers need iterative linearization.

RF Diffraction Tomography

RF Diffraction Tomography: From Rays to Waves

Definition: Born Approximation for Weak Scatterers

Theorem: The Fourier Diffraction Theorem

Spectral representation of the Green's function

Substitution into the Born integral

Definition: Rytov Approximation

Geometry of the Ewald Circle

Definition: k-Space Coverage Map

k-Space Coverage and DT Reconstruction

Parameters

Definition: Density Compensation Function

Direct Inversion: Gridding + Inverse FFT

Iterative Alternatives: SIRT, ART, CGLS

Example: Fourier Data from a Single View Angle

Identify the Ewald circle

Resolution analysis

Example: Gridding vs. CGLS for Incomplete Angular Coverage

Setup

Gridding + IFFT result

CGLS result

When to use each method

Common Mistake: Gridding Artifacts and Kernel Choice

Common Mistake: Born Approximation Breaks for Strong Scatterers

Why This Matters: Diffraction Tomography and MIMO Radar Imaging

Quick Check

Quick Check

Ewald Vector

Fourier Diffraction Theorem

Key Takeaway

Definition:
Born Approximation for Weak Scatterers

Definition:
Rytov Approximation

Definition:
k-Space Coverage Map

Definition:
Density Compensation Function