Beyond the Grid: Continuous Dictionary Pursuit for Interpretable Signal Decomposition

Nick Janssen, Melanie Schaller, Bodo RosenhahnarXiv 2026 · arXiv

In the world of signal processing, we often face a frustrating trade-off: we can have the high accuracy of "black-box" deep learning models, or the transparency of classical "glass-box" methods that frequently struggle with real-world signal complexities. A new paper from researchers at Leibniz University Hannover, titled "Beyond the Grid: Continuous Dictionary Pursuit for Interpretable Signal Decomposition," proposes a way to have both.

TL;DR

Continuous Dictionary Pursuit (CDP) is a "glass-box" framework that decomposes signals into a sparse set of physically meaningful, closed-form functions. By optimizing parameters over a continuous manifold rather than a discrete grid, it eliminates spectral leakage and achieves a superior sparsity-accuracy trade-off compared to both classical and neural baselines.

What Problem Does This Paper Address?

Traditional signal decomposition methods like Orthogonal Matching Pursuit (OMP) rely on a fixed, discrete "dictionary" of atoms (like sine waves at specific frequencies). This leads to the grid mismatch problem: if a signal's true frequency falls between two grid points, the algorithm must use multiple atoms to approximate that single feature. This "spectral leakage" inflates model complexity and destroys interpretability.

While deep learning models avoid these grid issues and achieve high accuracy, they are notoriously difficult to interpret. In safety-critical fields like healthcare or autonomous systems, understanding why a model produced a specific output is often a regulatory requirement. The authors aim to bridge this gap by creating a model that is as accurate as it is readable.

Methods

The core of the paper is Continuous Dictionary Pursuit (CDP). Unlike prior methods that might only optimize for sine waves, CDP uses a heterogeneous library of differentiable, parametric functions (atoms).

Figure 2: Parametric Dictionary Library: Example parameterizations of atom types: (a) SineAtom, (b) LinearAtom, (c) ExponentialAtom, (d) GaussianAtom, (e) SquareAtom, (f) SawtoothAtom, (g) StepAtom, (h) ModulatedAtom (for example SineAtom · StepAtom).

Key Innovations:

  • Continuous Optimization: Instead of picking from a list, CDP uses gradient-based optimization (L-BFGS) to learn the exact parameters (frequency, amplitude, phase, etc.) of closed-form functions directly from the data.
  • Spectral Sharpness Annealing: To handle discontinuous waveforms like square or sawtooth waves—which are usually non-differentiable—the authors use smooth approximations that "sharpen" over time during training.
  • Greedy Iterative Strategy: The algorithm identifies one component at a time, subtracts it from the signal, and then searches for the next component in the residual.
  • Global Refinement: Periodically, the framework re-optimizes all previously selected atoms together to correct for any suboptimal choices made in earlier greedy steps.
Figure 1: Continuous Dictionary Pursuit: The iterative greedy procedure initializes atoms from data- driven priors, identifies a candidate via truncated optimization, fully optimizes it with L-BFGS [30], and adds it to the additive model until the atom budget is exhausted or the residual error stagnates.

Datasets

The authors evaluated CDP on both synthetic and real-world data:

  • Synthetic Benchmarks: A custom synthesizer generated signals with known ground truths, including harmonics, trends, and non-stationary amplitude shifts.
  • Real-World LTSF Benchmarks: 13 diverse time-series datasets from the Long-Term Time Series Forecasting (LTSF) suite, including Electricity, ETT (Transformer temperature), Exchange, PEMS (Traffic), Solar, Traffic, and Weather.

Key Insights and Ideas

  • "Aha" Moment on Sparsity: The researchers found that by moving "off the grid," they didn't just get better accuracy—they got much simpler models. Because one continuous atom can perfectly match one physical feature, CDP requires orders of magnitude fewer parameters than deep learning models to achieve comparable reconstruction fidelity.
  • Curriculum Learning: One of the most creative applications in the paper is using CDP atoms as "curriculum targets." Instead of training a complex neural model on a noisy raw signal, they first train it to predict the simple, interpretable atoms extracted by CDP, gradually increasing the complexity.
  • Physical Mapping: Because the atoms are closed-form functions, their parameters map directly to physical units (e.g., a 24-hour cycle in traffic data), providing immediate domain insight without post-hoc analysis.

Key Results

  • Parameter Recovery: On synthetic datasets, CDP recovered generative parameters to within 1% for stationary signals. It successfully resolved off-grid frequencies where standard OMP failed due to spectral leakage.
Figure 3: Parameter recovery on synthetic datasets: CDP reconstructs the true mathematical components from noisy targets (σ = 0.1): (a) linear trend, (b) sine, (c) square wave, (d) off-grid frequency superposition, (e) damped sine via a modulated atom, (f) non-stationary amplitude shift. Reconstruction and prediction MSE are given per panel.
  • Pareto Efficiency: In the "sparse regime" (10 to 100 parameters), CDP achieved the best accuracy-sparsity trade-off among all nine baselines, including N-BEATS and other continuous pursuit methods.
Figure 4: Pareto frontier of parameter count versus training reconstruction MSE: CDP and CDP (Modulated) achieve the optimal trade-off in the sparse regime across all baselines. The lower convex envelope excludes Analysis K-SVD, whose near-zero training residual reflects in-sample denoising rather than a parametric decomposition (subsection A.8).
  • Forecasting Boost: Using CDP atoms for curriculum learning reduced out-of-sample forecasting error for state-space models (S4D) by 25% to 29% on several traffic datasets.

Limitations and Future Work

While powerful, CDP is not a silver bullet. The authors acknowledge several limitations:

  • Non-Convexity: The greedy optimization is susceptible to local minima, though their data-driven initialization (using FFT) helps mitigate this.
  • Library Dependence: The model can only decompose signals into the types of functions present in its library. If a signal contains a structure completely foreign to the library, the fit will degrade.
  • Computational Cost: Fitting the model is more expensive than simple grid-based methods (taking seconds to minutes rather than milliseconds), though inference on the resulting closed-form model is nearly instantaneous.
  • Univariate Focus: Currently, the method is designed for univariate signals. Extending this to multivariate data where atoms are shared across channels is a stated goal for future work.

Summary updated

Abstract

Deep learning models achieve high accuracy in signal processing but remain difficult to interpret. Classical matching pursuit and sparse coding offer better transparency but suffer from spectral leakage and grid mismatch when signal frequencies do not align with discrete dictionary atoms. We introduce Continuous Dictionary Pursuit (CDP), a framework for interpretable signal decomposition that learns the parameters of closed-form functions directly from data. CDP optimizes over a continuous manifold of differentiable atoms and, unlike prior continuous pursuit methods that operate on a single analytic family, searches a heterogeneous library spanning periodic, trend, transient, and discontinuous structures. The algorithm employs a greedy iterative strategy using data-driven priors and gradient-based optimization to isolate individual signal components, with discontinuous waveforms made differentiable through spectral sharpness annealing. To ensure a sparse atom set for reconstruction, the framework utilizes a stopping criterion based on a predefined atom budget and residual error convergence. Evaluations on synthetic and real-world datasets show that CDP removes the grid discretization error for atom families contained in its library, recovers generative parameters to within $1\%$ on stationary synthetic signals expressible in that library, and achieves the best sparsity-accuracy trade-off among classical, continuous-pursuit, and neural baselines in the sparse regime. Beyond decomposition, we demonstrate that the extracted atoms can serve as structured curriculum targets for training sequence models, improving forecasting accuracy over standard end-to-end training on most evaluated datasets. The method offers a glass-box alternative for signal decomposition by providing high reconstruction fidelity alongside direct physical interpretability.
BibTeX
@misc{janssen2026beyond,
  author = {Nick Janssen and Melanie Schaller and Bodo Rosenhahn},
  title = {{Beyond the Grid: Continuous Dictionary Pursuit for Interpretable Signal Decomposition}},
  year = {2026},
  eprint = {2611.09163},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.09163},
}
Figure