Ferkans — Interactive Telecom Tutor

What We Do Not Yet Understand

Despite the rapid progress in learned RF imaging, fundamental theoretical questions remain open. This section surveys the theoretical frontiers where new results would have the highest impact on algorithm design and system understanding. These are not incremental refinements but foundational questions whose answers could reshape the field.

Definition:
Sample Complexity of Learned RF Imaging

Question: how many RF measurements are needed to train a neural network that achieves $\epsilon$ -optimal reconstruction?

For classical compressed sensing, the sample complexity is $m = \mathcal{O}(s \log(n/s))$ where $s$ is the sparsity and $n$ the signal dimension. For learned methods, the sample complexity depends on:

The network architecture (capacity, inductive bias);
The training data distribution (scene complexity);
The forward model (measurement diversity).

Tight bounds are known only for simple architectures (linear networks, shallow ReLU) and idealised forward models. For practical architectures (deep unrolling, NeRF), the sample complexity is an open problem.

Definition:
Resolution Limits with Learned Priors

Classical resolution is limited by measurement bandwidth $W$ and aperture $D$ :

$\Delta r = \frac{c}{2W}, \qquad \Delta x = \frac{\lambda}{2 \sin\theta_{\max}} \approx \frac{\lambda R}{D},$

where $R$ is the range and $D$ the aperture. Learned methods appear to achieve super-resolution by exploiting priors. Open theoretical questions:

Can learned methods truly exceed the Rayleigh limit? Or are they interpolating within the noise margin?
How does the super-resolution factor depend on SNR? At low SNR, the prior dominates and "super-resolution" may hallucinate targets. Where is the boundary?
Information-theoretic limits: what is the maximum resolution achievable by any method, given measurements $\mathbf{y}$ and a prior distribution $p(\boldsymbol{\gamma})$ ?

Theorem: Imaging Fisher Information and Resolution

For the linear model $\mathbf{y} = \mathbf{A}\boldsymbol{\gamma} + \mathbf{w}$ with $\mathbf{w} \sim \mathcal{CN}(\mathbf{0}, \sigma^2 \mathbf{I})$ , the Fisher information matrix for the scene $\boldsymbol{\gamma}$ is:

$\mathcal{I}(\boldsymbol{\gamma}) = \frac{1}{\sigma^2}\mathbf{A}^H\mathbf{A}.$

The Cramer-Rao bound on the estimation error of the $n$ -th voxel is $\mathrm{var}(\hat{\gamma}_n) \geq [\mathcal{I}^{-1}]_{nn}$ . The effective imaging resolution at position $n$ is determined by the local PSF width, which equals the width of the $n$ -th column of $(\mathbf{A}^H\mathbf{A})^{-1}\mathbf{A}^H$ .

The Fisher information tells us how much information each measurement carries about each voxel. Voxels at the edge of the field of view have lower Fisher information (fewer measurements "see" them), hence worse resolution. The CRB provides a fundamental lower bound that no unbiased estimator can beat.

Proof

Fisher information derivation

The log-likelihood is $\ell(\boldsymbol{\gamma}) = -\frac{1}{\sigma^2}\|\mathbf{y} - \mathbf{A}\boldsymbol{\gamma}\|^2 + \text{const}$ . The Fisher information is $\mathcal{I} = -\mathbb{E}[\nabla^2_{\boldsymbol{\gamma}} \ell] = \frac{1}{\sigma^2}\mathbf{A}^H\mathbf{A}$ .

Resolution connection

The CRB matrix $\mathcal{I}^{-1} = \sigma^2 (\mathbf{A}^H\mathbf{A})^{-1}$ has diagonal entries that grow as the singular values of $\mathbf{A}$ decrease. Components of $\boldsymbol{\gamma}$ corresponding to small singular values are poorly resolved -- this is the spectral view of the resolution limit. $\square$

$\text{Information-Theoretic Resolution Bound}$

Shows the Cramer-Rao lower bound on voxel estimation error as a function of aperture and bandwidth. Demonstrates the fundamental trade-off between measurement resources and achievable resolution.

Parameters

W

(MHz)200

D

(m)1

\text{SNR}

(dB)20

f_0

(GHz)28

Definition:
Imaging Resolution and Channel Capacity

A deep connection links imaging resolution to channel capacity. Consider an ISAC system: the sensing measurements $\mathbf{y}$ can be viewed as a communication channel from the scene $\boldsymbol{\gamma}$ (source) through the sensing operator $\mathbf{A}$ (channel) to the measurements (receiver):

$C_{\mathrm{imaging}} = \max_{p(\boldsymbol{\gamma})} I(\boldsymbol{\gamma}; \mathbf{y}) = \log\det\!\left(\mathbf{I} + \frac{\text{SNR}}{N} \mathbf{A}^H\mathbf{A}\right),$

where $I(\cdot;\cdot)$ is mutual information and the maximum is over Gaussian scene distributions. This "imaging capacity" determines the maximum number of resolvable scene degrees of freedom. The connection to MIMO channel capacity (Book ITA, Book MIMO) is exact: the sensing matrix $\mathbf{A}$ plays the role of the channel matrix $\mathbf{H}$ .

Example: Degrees of Freedom from Imaging Capacity

A MIMO radar with $N_t = 16$ transmit and $N_r = 16$ receive antennas at 28 GHz images a 2D scene. The sensing matrix $\mathbf{A} \in \mathbb{C}^{256 \times N}$ has singular values $\sigma_k$ decaying as $\sigma_k \propto k^{-1}$ . At $\text{SNR} = 20$ dB, compute the effective number of resolvable degrees of freedom.

Solution

Imaging capacity

$C = \sum_{k=1}^{256} \log_2(1 + \text{SNR} \cdot \sigma_k^2 / N)$ . With $\sigma_k = C_0 / k$ and SNR $= 100$ : $C = \sum_k \log_2(1 + 100 C_0^2 / (Nk^2))$ .

Effective DoF

The effective degrees of freedom is the number of terms where $\text{SNR} \cdot \sigma_k^2 / N > 1$ , i.e., $k < \sqrt{100 C_0^2 / N}$ . For $C_0^2/N \approx 1$ : $k_{\max} = 10$ . Thus $\sim 10$ degrees of freedom are resolvable, corresponding to $\sim 10$ independent scene features that can be distinguished.

Interpretation

The decay rate of the singular values determines the effective rank of the imaging problem. Faster decay (more ill-posed) $\to$ fewer resolvable DoF. Learned priors can recover details beyond the $k_{\max}$ threshold, but only if the prior correctly constrains the reconstruction. $\square$

Definition:
Generalisation Theory for Imaging Networks

Question: when does a learned imaging method generalise to scenes not seen during training?

Standard ML generalisation bounds (VC dimension, Rademacher complexity) are too loose for practical networks ( $\sim 10^6$ parameters). Physics-informed bounds are tighter:

Networks that respect the forward model (unrolled algorithms) have lower effective complexity than generic networks.
The forward model constrains the output to the measurement-consistent set, reducing the hypothesis space.
Stability bounds (Lipschitz continuity of the reconstruction map) provide robustness guarantees.

Connecting these physics-informed bounds to practical network architectures remains an important open problem.

Promising Theoretical Directions

Neural tangent kernel (NTK) analysis of unrolled networks: characterise the implicit regularisation of deep unrolling at initialisation and during training.
Information-theoretic imaging bounds: extend the rate-distortion framework to characterise fundamental limits of imaging quality given measurement constraints.
Uncertainty quantification: rigorous confidence sets for neural network reconstructions. Conformal prediction provides distribution-free coverage guarantees.
Optimal measurement design: given a learned reconstruction method, what is the optimal measurement matrix $\mathbf{A}$ ? This connects to experimental design and active learning.
Algorithmic stability: does a small perturbation in the measurements cause a small change in the reconstruction? Essential for safety-critical applications (autonomous driving, medical imaging).

Historical Note: The Super-Resolution Debate

1992-present

The question of whether computational methods can exceed the classical diffraction limit has a long history. In 1992, Donoho showed that sparsity-based methods can resolve features below the Rayleigh limit -- but only at sufficient SNR. Candes and Fernandez-Granda (2014) proved that atomic norm minimisation achieves exact super-resolution of point sources separated by at least $\sim 2/f_c$ , where $f_c$ is the cutoff frequency. For RF imaging, the debate continues: learned methods routinely claim $2\times$ -- $4\times$ super-resolution, but distinguishing genuine resolution improvement from hallucination remains an open challenge.

Common Mistake: Super-Resolution vs Hallucination

Mistake:

Claiming super-resolution without testing whether the network is genuinely resolving sub-Rayleigh features or hallucinating targets based on the training distribution.

Correction:

Test super-resolution claims with: (1) a resolution chart (two targets at varying separations, plot detection probability vs separation); (2) noise sensitivity (genuine super-resolution degrades with SNR; hallucination does not); (3) out-of-distribution test (use scene types absent from training); (4) uncertainty maps (high uncertainty near the resolution limit indicates prior-dominated reconstruction).

Imaging Capacity

The mutual information $I(\boldsymbol{\gamma}; \mathbf{y})$ between the scene and measurements, maximised over scene distributions. Determines the maximum number of resolvable scene degrees of freedom. Analogous to channel capacity in communications.

Related: Foundation Model

Key Takeaway

The theoretical frontiers of RF imaging include sample complexity of learned methods, resolution limits with priors, the imaging capacity framework connecting resolution to channel capacity, and generalisation theory for physics-informed networks. The connection between imaging resolution and MIMO capacity is exact and deeply informative: the singular values of the sensing matrix determine both the number of resolvable scene features and the information-theoretic measurement limit.

Theoretical Frontiers