Ferkans — Interactive Telecom Tutor

Why Spectral Analysis Matters for Reconstruction

The singular values of $\mathbf{A}$ determine everything about the reconstruction problem: the condition number controls noise amplification, the spectral gap determines whether regularization is needed, and the singular vector structure reveals which scene features are well-resolved and which are invisible to the measurements. Because $\mathbf{A}$ has Kronecker structure, its spectral properties are completely determined by the spectra of the three factor matrices --- and these, in turn, are determined by the physical imaging geometry.

Definition:
Condition Number of the Sensing Matrix

The condition number of $\mathbf{A} \in \mathbb{C}^{M \times N}$ is

$\kappa(\mathbf{A}) = \frac{\sigma_{\max}(\mathbf{A})}{\sigma_{\min}(\mathbf{A})}$

where $\sigma_{\max}$ and $\sigma_{\min}$ are the largest and smallest non-zero singular values. For the least-squares solution of the normal equation

$\hat{\mathbf{c}} = (\mathbf{A}^{H}\mathbf{A})^{-1}\mathbf{A}^{H}\mathbf{y},$

the relative reconstruction error is bounded by

$\frac{\|\hat{\mathbf{c}} - \mathbf{c}\|}{\|\mathbf{c}\|} \leq \kappa(\mathbf{A}) \cdot \frac{\|\mathbf{w}\|}{\|\mathbf{y}\|}.$

For imaging inverse problems, $\kappa(\mathbf{A}) = \infty$ is common when $M < N$ (underdetermined system). Regularization effectively replaces $\sigma_{\min}$ with the regularization parameter $\lambda$ , giving an effective condition number $\kappa_{\text{eff}} = \sigma_{\max}/\lambda$ .

Theorem: SVD of Kronecker-Structured Sensing Matrices

If $\mathbf{A} = \mathbf{A}_{3} \otimes \mathbf{A}_{2} \otimes \mathbf{A}_{1}$ and each factor has SVD $\mathbf{A}_{k} = \mathbf{U}_k \boldsymbol{\Sigma}_k \mathbf{V}_k^H$ , then:

Singular values: The singular values of $\mathbf{A}$ are all triple products

$\sigma_{(i,j,l)}(\mathbf{A}) = \sigma_i(\mathbf{A}_{3}) \cdot \sigma_j(\mathbf{A}_{2}) \cdot \sigma_l(\mathbf{A}_{1})$

for all valid multi-indices $(i, j, l)$ .

Condition number: $\kappa(\mathbf{A}) = \kappa(\mathbf{A}_{3}) \cdot \kappa(\mathbf{A}_{2}) \cdot \kappa(\mathbf{A}_{1}).$

Singular vectors: The left and right singular vectors are Kronecker products of the factor singular vectors: $\mathbf{u}_{(i,j,l)} = \mathbf{u}_i^{(3)} \otimes \mathbf{u}_j^{(2)} \otimes \mathbf{u}_l^{(1)}$ and similarly for the right singular vectors.

The Kronecker structure means the conditioning of the full imaging system is the product of the conditionings in each dimension. If either the angular coverage or the frequency coverage is poor (large $\kappa$ for that factor), the overall problem is even more ill-conditioned than either dimension alone. This motivates joint optimization of both array geometry and frequency allocation.

Proof

Factor the Gram matrix

Using the mixed-product property:

$\mathbf{A}^{H}\mathbf{A} = (\mathbf{A}_{3}^{H}\mathbf{A}_{3}) \otimes (\mathbf{A}_{2}^{H}\mathbf{A}_{2}) \otimes (\mathbf{A}_{1}^{H}\mathbf{A}_{1}).$

Each factor $\mathbf{A}_{k}^{H}\mathbf{A}_{k} = \mathbf{V}_k \boldsymbol{\Sigma}_k^2 \mathbf{V}_k^H$ has eigenvalues $\sigma_i^2(\mathbf{A}_{k})$ .

Apply the eigenvalue property

The eigenvalues of a Kronecker product are all pairwise products of the factor eigenvalues. Therefore, the eigenvalues of $\mathbf{A}^{H}\mathbf{A}$ are $\sigma_i^2(\mathbf{A}_{3}) \cdot \sigma_j^2(\mathbf{A}_{2}) \cdot \sigma_l^2(\mathbf{A}_{1})$ , and the singular values of $\mathbf{A}$ are the square roots.

Condition number

$\kappa(\mathbf{A}) = \frac{\max_{i,j,l} \sigma_i \sigma_j \sigma_l}{\min_{i,j,l} \sigma_i \sigma_j \sigma_l} = \frac{\sigma_{\max}^{(3)}\sigma_{\max}^{(2)}\sigma_{\max}^{(1)}}{\sigma_{\min}^{(3)}\sigma_{\min}^{(2)}\sigma_{\min}^{(1)}} = \kappa(\mathbf{A}_{3}) \kappa(\mathbf{A}_{2}) \kappa(\mathbf{A}_{1}). \quad \blacksquare$ $

Definition:
The Gram Matrix as Point-Spread Function

The Gram matrix $\mathbf{G} = \mathbf{A}^{H}\mathbf{A} \in \mathbb{C}^{N \times N}$ acts as the imaging system's point-spread function (PSF) in discrete form. Applying the matched filter to a point scatterer at voxel $q$ (i.e., $\mathbf{c} = \mathbf{e}_q$ ) gives

$\hat{\mathbf{c}}^{\text{MF}} = \mathbf{A}^{H}(\mathbf{A}\mathbf{e}_q + \mathbf{w}) = \mathbf{G}\mathbf{e}_q + \mathbf{A}^{H}\mathbf{w} = \mathbf{g}_q + \mathbf{A}^{H}\mathbf{w}$

where $\mathbf{g}_q$ is the $q$ -th column of $\mathbf{G}$ . The off-diagonal elements $G_{pq}$ for $p \neq q$ measure the sidelobe leakage from voxel $q$ into voxel $p$ .

The mutual coherence is the worst-case normalized sidelobe:

$\mu(\mathbf{A}) = \max_{p \neq q} \frac{|G_{pq}|}{\sqrt{G_{pp} G_{qq}}}.$

The PSF of a Kronecker-structured sensing matrix is itself a Kronecker product of per-dimension PSFs. The main lobe width is inversely proportional to the k-space coverage extent in each dimension, as established in Ch 06.6.

Example: Conditioning of ULA vs. Distributed MIMO

Compare the condition numbers of the sensing matrix for:

(a) A co-located ULA with $N_t = 4$ , $N_r = 8$ , half-wavelength spacing, $N_f = 16$ subcarriers, imaging a $30^\circ \times 30^\circ$ scene on a $16 \times 16$ grid.

(b) A distributed MIMO with the same total number of elements placed randomly in a $10\lambda \times 10\lambda$ aperture.

Solution

ULA sensing matrix

The ULA creates a filled virtual array of $N_tN_r = 32$ elements. The angular factor $\mathbf{A}_{\text{ang}}$ is a partial DFT matrix (steering vectors sampled at grid angles). For a $30^\circ$ field of view with half-wavelength spacing, the steering vectors are nearly orthogonal, giving $\kappa(\mathbf{A}_{\text{ang}}) \approx 5$ -- $10$ . The frequency factor $\mathbf{A}_{f}$ is also a partial DFT with $\kappa(\mathbf{A}_{f}) \approx 1$ -- $3$ for well-chosen subcarrier spacing. Overall: $\kappa(\mathbf{A}) \approx 5$ -- $30$ .

Distributed MIMO

Random element placement creates a sparse, non-uniform virtual array with large aperture. The PSF has a narrow main lobe (large aperture) but high sidelobes (sparse sampling). The angular factor's Gram matrix has higher off-diagonal coherence due to the non-uniform sampling. Typical condition numbers: $\kappa(\mathbf{A}_{\text{ang}}) \approx 50$ -- $200$ . Overall: $\kappa(\mathbf{A}) \approx 50$ -- $600$ .

Resolution vs. conditioning trade-off

The ULA is better conditioned but has lower resolution (smaller aperture). The distributed array achieves higher resolution but requires stronger regularization due to ill-conditioning. Optimal array design balances resolution and conditioning --- this is a recurring theme in system design (Ch 09).

Condition Number vs. Array Geometry

Explore how the condition number of $\mathbf{A}$ depends on array geometry and frequency allocation. Compare random matrices (well-conditioned), physical ULA matrices (structured), and Kronecker products (condition number is the product of factors). Observe how the singular value distribution changes as you vary the number of antennas and the angular coverage.

Parameters

N_t

(Tx antennas)4

N_r

(Rx antennas)8

N_f

(Frequencies)16

Angular coverage (degrees)60

Point-Spread Function from the Sensing Operator

Visualizes the PSF by computing $\mathbf{A}^{H}\mathbf{A}\mathbf{e}_q$ for a point scatterer at the grid center. The PSF main lobe width is inversely proportional to the k-space coverage extent. Adjust the array parameters to see how the PSF changes shape.

Parameters

N_t

4

N_r

8

N_f

16

Grid size16

Common Mistake: Random Matrix Intuition Fails for Physical Sensing

Mistake:

Applying results from random matrix theory (e.g., Marchenko-Pastur distribution, $\kappa \approx (\sqrt{M/N} + 1)/(\sqrt{M/N} - 1)$ ) to the physical sensing matrix $\mathbf{A}$ .

Correction:

The physical $\mathbf{A}$ is highly structured: its columns are parameterized by continuous angles and frequencies, creating strong correlations absent in random matrices. The singular value distribution of a Kronecker-structured DFT-like matrix is fundamentally different from Marchenko-Pastur. Algorithms that assume random $\mathbf{A}$ (such as AMP --- see Ch 17.1) can diverge catastrophically when applied to the physical operator. Always verify spectral properties numerically for the specific geometry.

Why This Matters: Connection to Massive MIMO Channel Estimation

The conditioning problem we face in RF imaging is structurally identical to the channel estimation problem in massive MIMO (Telecom Ch 16): the pilot observation matrix is a partial DFT (steering vectors at discrete angles), and ill-conditioning arises when the angular coverage is incomplete. In massive MIMO, the condition number of the pilot matrix determines the MSE of the LMMSE channel estimator. The same preconditioning techniques we develop in Section 7.4 apply to both problems.

Quick Check

If the Rx factor has $\kappa(\mathbf{A}_{\text{Rx}}) = 5$ , the Tx factor has $\kappa(\mathbf{A}_{\text{Tx}}) = 3$ , and the frequency factor has $\kappa(\mathbf{A}_{f}) = 2$ , what is $\kappa(\mathbf{A})$ ?

$10$

$30$

$5$ (the maximum)

$\sqrt{38}$

Correction:

30

For Kronecker products, $\kappa(\mathbf{A}) = \kappa(\mathbf{A}_{\text{Rx}}) \cdot \kappa(\mathbf{A}_{\text{Tx}}) \cdot \kappa(\mathbf{A}_{f}) = 5 \times 3 \times 2 = 30$ .

Key Takeaway

The singular values of $\mathbf{A}$ are all triple products of the factor singular values. The condition number is the product of the factor condition numbers, so even mild ill-conditioning in one dimension is amplified multiplicatively. The Gram matrix $\mathbf{A}^{H}\mathbf{A}$ is the discrete PSF, and its off-diagonal elements determine the mutual coherence. Physical sensing matrices have fundamentally different spectral properties from random matrices --- algorithms must be designed accordingly.

Spectral Properties and Condition Number

Why Spectral Analysis Matters for Reconstruction

Definition: Condition Number of the Sensing Matrix

Theorem: SVD of Kronecker-Structured Sensing Matrices

Factor the Gram matrix

Apply the eigenvalue property

Condition number

Definition: The Gram Matrix as Point-Spread Function

Example: Conditioning of ULA vs. Distributed MIMO

ULA sensing matrix

Distributed MIMO

Resolution vs. conditioning trade-off

Condition Number vs. Array Geometry

Parameters

Point-Spread Function from the Sensing Operator

Parameters

Common Mistake: Random Matrix Intuition Fails for Physical Sensing

Why This Matters: Connection to Massive MIMO Channel Estimation

Quick Check

Key Takeaway

Definition:
Condition Number of the Sensing Matrix

Definition:
The Gram Matrix as Point-Spread Function