Tikhonov Regularization of Circle-Valued Signals

Laurent Condat

Introduction

Many real-world signals, such as the phase in interferometric radar, wind direction, or pixel hue, are "circle-valued." Unlike standard numerical data, these values are cyclic, where $0^\circ$ and $360^\circ$ represent the same point. Applying standard smoothing techniques like Tikhonov regularization directly to these signals is problematic because they assume a flat, Euclidean space. This often leads to the "jump problem," where the average of $1^\circ$ and $359^\circ$ is incorrectly calculated as $180^\circ$ instead of $0^\circ$. While phase unwrapping is a common workaround, it is notoriously fragile and error-prone. This paper presents a method for regularizing these signals directly on the complex unit circle using a convex optimization framework.

Methodology: Fourier Lifting

The core challenge lies in the non-convex nature of the unit circle. To address this, the author proposes a Fourier Lifting approach that transforms the non-convex optimization into a manageable convex problem:

  1. Probabilistic Modeling: The problem is reframed from finding a single optimal angle to finding an optimal probability distribution of angles at each point. This "lifts" the problem into a higher-dimensional but convex space.
  2. Method of Moments: These distributions are represented using a finite number of their Fourier coefficients (moments) to maintain computational feasibility.
  3. Semidefinite Programming (SDP): The circular constraints are enforced by requiring certain $3 \times 3$ Hermitian matrices to be positive semidefinite. This converts the problem into a well-behaved Semidefinite Program.
  4. Optimization Algorithm: The resulting SDP is solved using the Proximal Method of Multipliers, an iterative algorithm that involves computing eigendecompositions of small matrices at each step.

Experimental Evaluation and Results

The proposed method was evaluated on synthetic 1D signals and 2D images. A critical finding is the tightness of the relaxation: the convex solution frequently corresponds to the exact global optimum of the original non-convex problem, outperforming traditional "unwrap-then-smooth" or naive projection techniques.

1D Interpolation

In tasks where values are missing between fixed endpoints, the method demonstrates superior geometric awareness. As shown in the figure below, the proposed method correctly identifies the "straight line" path in angular space, whereas the baseline method produces a distorted path.

Interpolation comparison between the baseline (blue) and proposed (red) methods.

2D Image Denoising

For 2D phase maps, the Fourier lifting approach produces qualitatively smoother results with less "jagged" level lines compared to baseline techniques or naive circular mean filtering. Quantitatively, the proposed method achieved a lower cost function value ($\approx 2479$) in 2D denoising experiments compared to the baseline ($\approx 2534$) and Gaussian filtering ($\approx 2542$).

2D image denoising results: (a) original, (b) noisy, (c) baseline, (d) proposed method, (e) circular mean filtering.

Conclusion and Future Work

This research provides a robust framework for smoothing signals on graphs where values lie on a circle. By leveraging Fourier lifting and semidefinite programming, the method avoids the pitfalls of phase unwrapping and naive projection. While the current implementation is computationally intensive due to repeated eigendecompositions, future work may focus on optimizing these routines for $3 \times 3$ matrices. Additionally, investigating the limits of relaxation tightness in extremely high-noise scenarios remains an open challenge.

Summary updated

Abstract

It is common to have to process signals or images whose values are cyclic and can be represented as points on the complex circle, like wrapped phases, angles, orientations, or color hues. We consider a Tikhonov-type regularization model to smoothen or interpolate circle-valued signals defined on arbitrary graphs. We propose a convex relaxation of this nonconvex problem as a semidefinite program, and an efficient algorithm to solve it.
BibTeX
@article{condat2021tikhonov,
  author = {Laurent Condat},
  title = {{Tikhonov Regularization of Circle-Valued Signals}},
  journal = {IEEE Trans. Signal Process.},
  year = {2021},
  url = {https://openreview.net/forum?id=PaSk9etJee},
}
Figure