In traditional signal processing, we often assume that the world is discrete. To find the location of a star in a telescope image or a molecule in a microscope, we divide the space into a grid of pixels and ask: "Which pixel contains the source?"
However, physical reality is continuous. If a point source lies between two grid cells, a discrete model faces a "basis mismatch." No matter how fine we make the grid, the source will never perfectly align, leading to artifacts and limited precision. A series of papers, led by researchers like Emmanuel J. Candès and Carlos Fernandez-Granda, has developed a mathematical framework to solve this problem "off-the-grid." By using convex optimization in the space of continuous measures, we can recover point sources with theoretically infinite precision from low-frequency, blurry data.
The Problem with Grids
When we observe a signal—whether it's a radar echo or a blurred image—we are usually looking at a low-pass filtered version of reality. High-frequency details are lost. In a discrete setting, we try to recover these details by solving an $\ell_1$-minimization problem (like Basis Pursuit) over a predefined grid.
As Clarice Poon explains in The geometry of off-the-grid compressed sensing, this discretization is fundamentally flawed. If the true signal consists of "spikes" (Diracs) at arbitrary locations, forcing them onto a grid creates a model error that cannot be eliminated by simply increasing the grid resolution.
Total Variation of Measures
The breakthrough idea, formalized in Towards a Mathematical Theory of Super-Resolution, is to replace the discrete $\ell_1$ norm with its continuous-domain analog: the Total Variation (TV) norm of a measure.
In this context, "Total Variation" does not refer to the integral of the gradient (common in image denoising). Instead, for a collection of point sources, the TV norm is simply the sum of the absolute values of their amplitudes. By minimizing this norm subject to the constraint that the recovered signal matches our low-frequency observations, we promote sparsity in continuous space.
This optimization doesn't require a grid. It searches the infinite-dimensional space of measures to find the simplest (most sparse) set of spikes that explains the data. Carlos Fernandez-Granda demonstrated that this can be solved efficiently using semi-definite programming, effectively "super-resolving" the signal.
The Separation Condition
Why does this work? Intuition suggests that if two spikes are too close together, they will look like a single blurred blob, making them impossible to distinguish.
Emmanuel J. Candès and Carlos Fernandez-Granda proved that exact recovery is guaranteed as long as the spikes are sufficiently separated. Specifically, if the distance between any two spikes is greater than a certain threshold (roughly $2/f_c$, where $f_c$ is the cutoff frequency of the measurements), the TV-norm minimization will perfectly identify their locations and amplitudes.
This guarantee has been extended to various geometries beyond the real line:
- Two Dimensions: Joseph P. McDonald and colleagues established a sampling theorem for 2D deconvolution, showing how to perfectly reverse Gaussian blur in A Sampling Theorem for Deconvolution in Two Dimensions.
- Spherical Geometry: Tamir Bendory and Shai Dekel proved that point sources on a sphere can be recovered from low-degree spherical harmonics in Exact recovery of Dirac ensembles from the projection onto spaces of spherical harmonics.
Robustness to Noise: The Beurling LASSO
In the real world, measurements are noisy. To handle this, the framework evolves into the Beurling LASSO (BLASSO). Instead of forcing an exact match with the data, the BLASSO minimizes a combination of the TV norm and a data-fidelity term (the squared error).
Jean‐Marc Azäis, Fabrice Gamboa, and Yohann de Castro showed in Spike detection from inaccurate samplings that the BLASSO is robust. Even with noise, the recovered spikes will be very close to the true locations. Carlos Fernandez-Granda further proved in Support detection in super-resolution that the localization error for a specific spike depends primarily on its own amplitude and the noise level, rather than being corrupted by other nearby spikes.
Beyond Theory: Radar and Pulses
While much of this work is foundational mathematics, it has direct implications for sensing technologies. For instance, Shai Dekel and Tamir Bendory applied these principles to the recovery of "streams of pulses," which is critical for radar and ultrasound (Robust Recovery of Stream of Pulses using Convex Optimization).
More recently, the concept of treating point-cloud extraction as an inverse deconvolution problem has moved into deep learning. Tomer Borreda and Shengyu Huang used these insights in RadarGen to generate high-fidelity automotive radar data, demonstrating that the "off-the-grid" perspective remains the gold standard for spatial accuracy.