Abstract
This paper considers the problem of estimating chirp parameters from a noisy mixture of chirps. While a rich body of work exists in this area, challenges remain when extending these techniques to chirps of higher order polynomials. We formulate this as a non-convex optimization problem and propose a modified Langevin Monte Carlo (LMC) sampler that exploits the average curvature of the objective function to reliably find the minimizer. Results show that our Curvature-guided LMC (CG-LMC) algorithm is robust and succeeds even in low SNR regimes, making it viable for practical applications.
Introduction
Chirp signals—signals whose frequency changes over time—are fundamental to a wide array of technologies and scientific fields, including radar, sonar, and biomedical imaging. In real-world scenarios, these signals often appear as "mixtures" of multiple overlapping chirps with complex, higher-order polynomial phases and amplitudes.
Estimating the parameters of such mixtures is a significant challenge for standard optimization. The mathematical landscape of the error function is notoriously non-convex, characterized by a "jagged" surface filled with thousands of local minima and expansive flat regions where gradients vanish. Traditional methods frequently become trapped in these local dips or fail to navigate the flat zones, particularly when the signal is obscured by high levels of noise.
Methodology
We frame the parameter estimation as a non-convex optimization problem and solve it using a modified Langevin Monte Carlo (LMC) sampler. Our approach, Curvature-guided LMC (CG-LMC), introduces several key innovations to navigate the difficult optimization landscape:
- LMC and Stochasticity: LMC acts as a particle moving through the landscape, following the gradient toward the minimum while utilizing stochastic "thermal noise" to escape shallow local minima.
- Gaussian Smoothing: To mitigate the jaggedness of the objective function, we apply Gaussian smoothing. This effectively "blurs" the landscape, making the general direction of the global minimum more apparent to the optimizer.
- Curvature Guidance: The primary innovation is the adaptive tuning of the smoothing parameter ($\sigma$) using the trace of the Hessian. The Hessian trace serves as a measure of local curvature and acts as a compass for the optimizer:
- In flat regions (far from the solution), $\sigma$ remains large to maintain a consistent "tilt" toward the goal.
- As the optimizer detects increasing curvature (approaching the solution), it automatically reduces $\sigma$ to sharpen the landscape and settle into the exact global minimum.
- Multi-scale Initialization: The process begins using shorter segments of the signal, which naturally possess broader and more accessible minima, before transitioning to the full-length signal for high-precision estimation.
Experimental Setup
The CG-LMC algorithm was evaluated using synthetic mixtures of two chirps ($N_c = 2$) with 4th-order polynomial phases and 3rd-order polynomial amplitudes. We tested the algorithm at various noise levels, including a highly noisy 3 dB SNR and a moderate 12 dB SNR. We compared our results against standard LMC and Noise-Annealed LMC (NA-LMC), the latter of which uses a pre-defined, non-adaptive smoothing schedule.
Key Results and Discussion
Our experiments yielded several critical insights into the performance of CG-LMC:
- Superior Accuracy in Low SNR: At 3 dB SNR, CG-LMC produced parameter estimates with significantly lower Mean Absolute Error (MAE) and lower standard deviation than both standard LMC and NA-LMC.
- Robustness to Initialization: One of the most significant advantages of CG-LMC is its reliability. In multiple trials starting from "far" initializations, CG-LMC successfully converged to the global minimum every time. In contrast, NA-LMC often failed to converge when the starting point was not sufficiently close to the solution.
- Generalization: The algorithm demonstrated consistent performance across five different chirp mixtures, suggesting that the curvature-guided approach is robust to varying signal characteristics and not overfit to a specific configuration.
Conclusion and Future Work
CG-LMC provides a robust framework for estimating parameters in complex, multi-chirp environments. By using local curvature to adaptively tune the optimization process, it overcomes the traditional barriers of non-convexity and noise.
While the computational cost of calculating the Hessian trace is higher than simple gradient descent, the gains in reliability are substantial. Future research will focus on scaling the algorithm for larger numbers of overlapping chirps and applying the method to real-world data, such as the separation of aortic and pulmonary components in electronic heart sound recordings. Additionally, we intend to compare these empirical results against the theoretical Cramér-Rao bound to further quantify the algorithm's efficiency.
Summary updated