Abstract: Retrieving a signal from its triple correlation spectrum, also called bispectrum, arises in a wide range of signal processing problems. Conventional methods do not provide an accurate inversion of bispectrum to the underlying signal. In this paper, we present an approach that uniquely recovers signals with finite spectral support (band-limited signals) from at least $3B$ measurements of its bispectrum function (BF), where $B$ is the signal's bandwidth. Our approach also extends to time-limited signals. We propose a two-step trust region algorithm that minimizes a non-convex objective function. First, we approximate the signal by a spectral algorithm and then refine the attained initialization based on a sequence of gradient iterations. Numerical experiments suggest that our proposed algorithm is able to estimate band-/time-limited signals from its BF for both complete and undersampled observations.
Introduction
In fields such as radar, optics, and seismology, the standard power spectrum (or autocorrelation) of a signal is often insufficient because it discards critical phase information. To capture a complete representation, researchers utilize higher-order statistics, specifically the bispectrum. The bispectrum is a two-dimensional mapping that represents the interactions between different frequency components of a signal.
However, "inverting" the bispectrum—recovering the original signal from this representation—is a notoriously difficult task. It is a non-convex, ill-posed problem complicated by "trivial ambiguities" such as time shifts and global phase rotations. Furthermore, the search space is massive, and existing methods frequently lack theoretical guarantees or fail to scale with signal size.
Methodology
This research introduces a two-stage approach designed to overcome the challenges of bispectrum inversion by exploiting the physical properties of the signals.
1. Continuity-Aware Initialization
The primary hurdle in inversion is that each bispectrum measurement can correspond to multiple mathematical roots. To select the correct candidate, the authors utilize the Paley-Wiener theorem, which implies that the Fourier transform of a signal with finite support must be smooth and continuous ($C^\infty$). By enforcing spectral continuity through finite differences, the proposed initialization algorithm effectively prunes the exponential tree of possible solutions, identifying a high-quality starting point.
2. Trust-Region Refinement
With a robust initialization, the algorithm employs a non-convex trust-region method to minimize the least-squares error between the observed and estimated bispectra. To maintain computational efficiency for high-dimensional signals, a block stochastic gradient descent strategy is used, updating the estimate via small minibatches of the bispectrum data.
Key Results
The proposed framework was validated through synthetic numerical simulations of band-limited signals with Gaussian-shaped power spectra and random phases.
Noise Resilience
The algorithm demonstrates significant robustness to Gaussian noise. As shown in the spectral reconstruction results, the method successfully recovers the intensity and general shape of the underlying power spectrum even at a Signal-to-Noise Ratio (SNR) of 20dB.
Efficiency and Undersampling
A key theoretical contribution is the proof that a signal with bandwidth $B$ requires only $3B$ measurements from the bispectrum for unique recovery. Numerical results confirm that the algorithm maintains low relative error even when a substantial percentage of the bispectrum entries are missing, confirming that the full $N \times N$ matrix is redundant for finite-support signals.
Conclusion and Future Work
This study presents a new theoretical guarantee and a practical algorithm for bispectrum inversion. By leveraging spectral continuity as a constraint, the method bypasses the combinatorial explosion of traditional inversion techniques. While the current approach relies on the integrity of the initial bispectrum columns, future work will explore methods to increase robustness to localized noise and scale the algorithm for multi-dimensional volumetric imaging.
Summary updated