Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Towards a Theory of Stable Super-Resolution: Model-Based Formulation and Stability Analysis

Zetao Fei, Hai Zhang

TL;DR

This work introduces Model-SR, a model-based framework for stable super-resolution in the frequency domain by leveraging a low-dimensional parameterization of the target signal. The authors establish a two-stage procedure—parameter estimation from low-frequency data and a resolution-enhancing mapping to high-frequency content—underpinned by Lipschitz stability results that depend on parameter separation and the high-frequency cutoff. They develop concrete stability guarantees for point-source, finite-rate-of-innovation (FRI), and general continuous signal models, and show how sparsity (sigma-compatible modeling) can bolster robustness while keeping the nonconvex optimization tractable. Numerical experiments across models validate stable extrapolation up to finite SRF and highlight the practical relevance of model choice and initialization, with extensions to data completion and connections to deep learning outlined for broader applicability.

Abstract

In mathematics, a super-resolution problem can be formulated as acquiring high-frequency data from low-frequency measurements. This extrapolation problem in the frequency domain is well-known to be unstable. We propose a model-based super-resolution framework (Model-SR) for solving the super-resolution problem and analyzing its stability, aiming to narrow the gap between limited theory and the broad empirical success of super-resolution methods. The key rationale is that, to be determined by its low-frequency components, the target signal must possess a low-dimensional structure. Instead of assuming that the signal itself lies on a low-dimensional manifold in the signal space, we assume that it is generated from a model with a low-dimensional parameter space. This shift of perspective allows us to analyze stability directly through the model parameters. Within this framework, we can recover the signal by solving a nonlinear least square problem and achieve super-resolution by extracting its high-frequency components. Theoretically, the resolution-enhancing map is proven to have Lipschitz continuity, with a constant that depends crucially on parameter separation conditions\. This separation condition can be effectively enforced via sparsity modeling, which requires using the minimal number of parameters to represent the measured signal, thereby highlighting the role of sparsity in the stability of super-resolution. Moreover, the Lipschitz constant grows with the high-frequency cutoff, ultimately rendering extrapolation ineffective beyond a certain threshold. We apply the general theory to three concrete models and give the stability estimates for each model. Numerical experiments are conducted to show the super-resolution behavior of the proposed framework. The model-based mathematical framework can be extended to problems with similar structures.

Towards a Theory of Stable Super-Resolution: Model-Based Formulation and Stability Analysis

TL;DR

This work introduces Model-SR, a model-based framework for stable super-resolution in the frequency domain by leveraging a low-dimensional parameterization of the target signal. The authors establish a two-stage procedure—parameter estimation from low-frequency data and a resolution-enhancing mapping to high-frequency content—underpinned by Lipschitz stability results that depend on parameter separation and the high-frequency cutoff. They develop concrete stability guarantees for point-source, finite-rate-of-innovation (FRI), and general continuous signal models, and show how sparsity (sigma-compatible modeling) can bolster robustness while keeping the nonconvex optimization tractable. Numerical experiments across models validate stable extrapolation up to finite SRF and highlight the practical relevance of model choice and initialization, with extensions to data completion and connections to deep learning outlined for broader applicability.

Abstract

In mathematics, a super-resolution problem can be formulated as acquiring high-frequency data from low-frequency measurements. This extrapolation problem in the frequency domain is well-known to be unstable. We propose a model-based super-resolution framework (Model-SR) for solving the super-resolution problem and analyzing its stability, aiming to narrow the gap between limited theory and the broad empirical success of super-resolution methods. The key rationale is that, to be determined by its low-frequency components, the target signal must possess a low-dimensional structure. Instead of assuming that the signal itself lies on a low-dimensional manifold in the signal space, we assume that it is generated from a model with a low-dimensional parameter space. This shift of perspective allows us to analyze stability directly through the model parameters. Within this framework, we can recover the signal by solving a nonlinear least square problem and achieve super-resolution by extracting its high-frequency components. Theoretically, the resolution-enhancing map is proven to have Lipschitz continuity, with a constant that depends crucially on parameter separation conditions\. This separation condition can be effectively enforced via sparsity modeling, which requires using the minimal number of parameters to represent the measured signal, thereby highlighting the role of sparsity in the stability of super-resolution. Moreover, the Lipschitz constant grows with the high-frequency cutoff, ultimately rendering extrapolation ineffective beyond a certain threshold. We apply the general theory to three concrete models and give the stability estimates for each model. Numerical experiments are conducted to show the super-resolution behavior of the proposed framework. The model-based mathematical framework can be extended to problems with similar structures.
Paper Structure (31 sections, 9 theorems, 120 equations, 8 figures)

This paper contains 31 sections, 9 theorems, 120 equations, 8 figures.

Table of Contents

  1. Introduction
  2. Problem outline in the frequency domain
  3. Introduction to main results
  4. Connection with Existing Literature
  5. Organization of the paper
  6. Notations
  7. Model-based Super-resolution Framework
  8. Mathematical Model for Model-SR
  9. Stability Estimate for an Ideal Modeling Pair
  10. Stability for Compatible Modeling Pairs
  11. The Optimization Problem
  12. Point Source Model
  13. Going Beyond Point Source Model
  14. Signals with Finite Rate of Innovation
  15. Towards General Signals
  16. ...and 16 more sections

Key Result

Proposition 2.2

Assume that $U \subset \mathbb R^m$ is a convex compact set. Consider $\mathcal{P}_L \in C^1(\mathbb R^m,\mathcal{H}_L)$ satisfying that Then, for every $\theta,\theta'\in U$, there exists $C_{U}>0$ such that Further, we have

Figures (8)

  • Figure 1: Signal space, low- and high-resolution spaces, and related maps.
  • Figure 2: Model-based super-resolution framework.
  • Figure 3: Two-point source approximation of the signal, followed by resolution enhancement up to $\mathrm{SRF} = 6$. The experiment is conducted under $\mathrm{SNR} = 12.76$ (see the definition of SNR in (\ref{['eq: snr']})).
  • Figure 4: (a) Boxplot of point source position reconstruction error. (b-d) Original and resolution-enhanced signals in the physics domain. The red line represents the ground truth of the point source. The blue line shows the signal profile calculated by iFFT using the original/extrapolated Fourier data. The SNR of the experiment is $20.08$.
  • Figure 5: Reconstruction for closely positioned point sources.
  • ...and 3 more figures

Theorems & Definitions (29)

  • Example 1
  • Example 2
  • Definition 1
  • Definition 2
  • Remark 2.1
  • Example 3
  • Definition 3
  • Proposition 2.2
  • Theorem 2.3
  • proof
  • ...and 19 more