Abstract
Finite Rate of Innovation (FRI) theory considers sampling and reconstruction of classes of non-bandlimited signals, such as streams of Diracs. Widely used FRI reconstruction methods including Prony's method and matrix pencil method involve Singular Value Decomposition (SVD). When samples are corrupted with noise, they achieve an optimal performance given by the Cramer-Rao bound yet break down at a certain Signal-to-Noise Ratio (SNR) due to the so-called subspace swap problem. In this paper, we investigate a deep neural network approach for FRI signal reconstruction that directly learns a transformation from signal samples to FRI parameters. Simulations show significant improvement on the breakdown SNR over existing FRI methods.
Introduction
In signal processing, Finite Rate of Innovation (FRI) theory provides a framework for the perfect reconstruction of certain non-bandlimited signals from a finite number of samples. Typical examples include sequences of Dirac deltas or streams of pulses. Traditionally, reconstruction is achieved through algebraic methods such as Prony’s method or the Matrix Pencil method. These techniques rely on Singular Value Decomposition (SVD) to estimate a "signal subspace" from the data.
While mathematically elegant and capable of reaching the Cramer-Rao bound (CRB) in low-noise scenarios, these classical methods exhibit a critical "breakdown SNR." At high noise levels, a "subspace swap" occurs where noise components are incorrectly identified as signal components, leading to a precipitous drop in reconstruction accuracy. This challenge is particularly acute when signal pulses are closely spaced, necessitating extremely high SNR for successful resolution.
Proposed Deep Learning Approach
To address the limitations of subspace-based methods, we propose a Deep Neural Network (DNN) architecture that learns a direct mapping from observed signal samples to the FRI signal parameters (specifically, the locations of the Diracs). By bypassing explicit subspace estimation, the model avoids the subspace swap phenomenon.
Architecture
The proposed network consists of six layers designed to extract features and perform inference:
- 3 Convolutional (Conv) Layers: Using filters of size 3 to extract local features from the sampled data.
- 3 Fully Connected (FC) Layers: To perform the final inference of the Dirac locations.
The network employs Rectified Linear Units (ReLU) as activation functions and is trained using Mean-Squared Error (MSE) loss via the Adam optimizer.
Experimental Results
The performance of the DNN was evaluated using a synthetic dataset of $10^5$ signal samples, each containing $K=2$ Diracs with varying amplitudes and locations. Testing spanned SNR levels from -5 dB to 50 dB.
Robustness to Noise
A key finding is the DNN's significantly lower breakdown threshold compared to classical methods. For a pulse spacing of $\Delta t = 10^{-2}$, the DNN maintains reliable estimation down to 10 dB SNR, whereas classical methods (Cadzow+Prony and Matrix Pencil) fail below 25 dB.
As shown in the heatmaps, the DNN demonstrates a much larger "operating region" where error remains low, particularly in the challenging regime of low SNR and small pulse spacing.
Performance Trade-offs
While the DNN is superior in noisy conditions, classical methods still provide higher precision in very low-noise (high SNR) regimes. This indicates that the DNN functions as a robust estimator that gracefully degrades, whereas classical methods are high-precision estimators that are fragile under noise.
Conclusion and Future Work
This study demonstrates that deep neural networks can effectively reconstruct FRI signals by learning underlying structures without explicit subspace separation. This approach significantly improves robustness to noise and pulse proximity. Future research will explore hybrid models to combine DNN robustness with classical high-SNR precision, as well as scaling the architecture to handle signals with a larger number of Diracs ($K > 2$).
Summary updated