Exact Sparse Recovery: Criteria & Algorithms

• Exact Sparse Representation Recovery (ESRR) is a framework defining precise conditions—such as NSP, RIP, and dual certificates—that guarantee unique recovery of sparse signals.
• It employs methodologies like convex programming, greedy pursuit, and algebraic extraction to reconstruct both the support and amplitudes of sparse components with noise robustness.
• ESRR underpins vital applications in compressed sensing, super-resolution, and dictionary learning, offering practical insights for signal demixing and structured recovery.

Exact Sparse Representation Recovery (ESRR) denotes the rigorous and algorithmic conditions under which a high-dimensional object (e.g., a signal, function, measure, or operator) can be reconstructed exactly from a reduced set of measurements, provided that the object has a sparse representation in a given dictionary or basis. In ESRR, both the positions (“support”) and the values (“amplitudes”) of the nonzero or atomic components are recovered uniquely and precisely, without spurious artifacts or overfitting. The theory encompasses discrete, continuous, finite, and infinite-dimensional models, and finds applications across compressed sensing, signal demixing, polynomial interpolation, inverse problems, super-resolution, dictionary learning, and structured representations. Characterizations of ESRR involve geometric or analytic criteria on measurement ensembles (such as the Restricted Isometry Property, Null Space Property, or dual certificates), the convexity or structure of regularizers, and algorithmic strategies (convex programming, greedy pursuit, or algebraic Prony’s methods).

1. Foundational Principles and Mathematical Criteria

ESRR arises from problems of the form: recover a sparse signal $x^*$ (i.e., $\|x^*\|_0 \ll n$), measure, or structured object from linear, nonlinear, or sampled observations $y = \mathcal{A}(x^*) + w$. The essential criterion is that the reconstruction algorithm (such as convex minimization or a greedy scheme) uniquely identifies $x^*$ from $y$ over the model class.

Core Notions

• Null Space Property (NSP): For $A \in \mathbb{R}^{m \times n}$, exact recovery by $\ell_1$-minimization is equivalent to: for all $h \neq 0$ in the nullspace of $A$, $\|h_S\|_1 < \|h_{S^c}\|_1$ for all supports $\|x^*\|_0 \ll n$0 of size $\|x^*\|_0 \ll n$1.
• Restricted Isometry Property (RIP): A matrix $\|x^*\|_0 \ll n$2 obeys RIP of order $\|x^*\|_0 \ll n$3 with constant $\|x^*\|_0 \ll n$4 if $\|x^*\|_0 \ll n$5 for all $\|x^*\|_0 \ll n$6 with $\|x^*\|_0 \ll n$7. Sharp ESRR is guaranteed if $\|x^*\|_0 \ll n$8 for $\|x^*\|_0 \ll n$9 (Cai et al., 2013).
• Dual Certificates: In convex regularized problems with Banach or measure spaces, ESRR is certified by the existence of a dual element (via the Fenchel dual) that “peaks” only on the building blocks of the ground-truth solution (Carioni et al., 2023).

These criteria have algorithmic counterparts in weighted $y = \mathcal{A}(x^*) + w$0, group-structured, or fusion models, each with precise tailored conditions (e.g., WNSP for weighted norms (Zhou et al., 2013), fusion-NSP and block-RIP (0912.4988), or group-atomic width bounds (Rao et al., 2011)).

2. Metric Non-Degenerate Source Condition (MNDSC) and Infinite-Dimensional ESRR

The MNDSC framework (Carioni et al., 2024, Carioni et al., 2023) extends ESRR to optimization in Banach spaces, where sparsity corresponds to a sum of extreme points of the regularizer ball (e.g., Dirac measures, indicator functions).

• Given a regularizer $y = \mathcal{A}(x^*) + w$1, the minimizer $y = \mathcal{A}(x^*) + w$2 is sparse if $y = \mathcal{A}(x^*) + w$3 with $y = \mathcal{A}(x^*) + w$4 extreme.
• The Minimal-Norm Dual Certificate $y = \mathcal{A}(x^*) + w$5, constructed from the Fenchel dual, attains $y = \mathcal{A}(x^*) + w$6 only at true atoms.
• The MNDSC requires (a) localization (dual maximizes precisely at the correct atoms), and (b) strict local concavity of the dual action along curves in a metric space of atoms:

$y = \mathcal{A}(x^*) + w$7

for curves $y = \mathcal{A}(x^*) + w$8 connecting nearby atoms.

Main ESRR Theorem (Carioni et al., 2024): Under MNDSC and a linear independence assumption, the Tikhonov-regularized solution is unique and reconstructs exactly the number and structure of the true atoms, stable to small noise and regularization parameters.

This framework is applicable to TV-regularized deconvolution (BLASSO), BV-regularity, and coupled Wasserstein problems, enabling verifiable ESRR guarantees in settings beyond classical compressed sensing.

3. Algorithmic Strategies and Their ESRR Guarantees

Multiple algorithmic paradigms underpin ESRR, with precise success criteria:

• Convex Programming:
  • $y = \mathcal{A}(x^*) + w$9-minimization and its weighted variants (MIRL1) achieve ESRR under NSP/WNSP/RIP bounds (Zhou et al., 2013). Weighted null-space property identifies exact regions where reweighted $x^*$0 strictly outperforms standard $x^*$1.
  • Group lasso and structured convex atomic norms achieve universal measurement bounds for group-sparse ESRR, improving over standard $x^*$2 scaling (Rao et al., 2011).
• Greedy Pursuit:
  • OMP and its block/generalized variants (OMP-SR, BSR) provide ESRR with coherence or tailored ERC-type conditions (e.g., $x^*$3), and probabilistically achieve $x^*$4 measurements under random-Bernoulli/Gaussian matrices (Yu et al., 2024, Wen et al., 2020).
  • Constrained Matching Pursuit (CMP) characterizes ESRR on convex CP-admissible cones, extending exact recovery to nonnegative and other constrained models under generalized restricted-cone constants (Shen et al., 2019).
• Algebraic/Analytic Extraction:
  • For polynomial and super-resolution models, Prony’s method (Toeplitz or Hankel) achieves ESRR without geometric spacing or minimum separation, and with low sample complexity (2r or r+1 evaluations) (Josz et al., 2017).
  • SDP-based super-resolution (convex TV-minimization with moment constraints) enables ESRR for pointwise measures/Dirac-spike recovery with precise geometric separation and flat extension conditions (Josz et al., 2017).
• Sparse Dictionary Identification:
  • ER-SpUD achieves ESRR for dictionary learning from sample complexity $x^*$5 with Bernoulli-subgaussian sparse codes, uniquely determining both dictionary and coefficients (Spielman et al., 2012).

4. Extensions to Structured and Continuous Models

ESRR is generalized to settings where sparsity is defined over structured, infinite, or group partitions:

• Fusion Frames: Signals sparse in union-of-subspaces are recovered via block $x^*$6-minimization, with fusion-NSP and fusion-coherence/RIP criteria ensuring ESRR (0912.4988).
• Partial Prior Information: Modified mutual coherence bounds extend ESRR in the presence of support-side-information, with exact thresholds for OMP, OLS, and $x^*$7-relaxation (Herzet et al., 2013).
• Continuous Dictionaries: In infinite/parametric atom-indexed models (e.g., spike trains, Gabor expansions), continuous analogs of ERC and kernel admissibility yield ESRR provided the Gramian/cross-correlation matrix fulfills precise invertibility/separation conditions. For 1D completely monotone kernels, no minimal separation is needed; in higher dimension, restricted ERC on grids governs ESRR (Elvira et al., 2019).

5. Applications and Specialized ESRR Scenarios

ESRR theory underpins a broad array of concrete applications:

• Signal Demixing and Group BLASSO: Convex regularization with TV and group structure rigorously recovers component measures or group elements under dual-certificate nondegeneracy; exact atomical support is reconstructed up to small perturbation (Carioni et al., 2024).
• Super-Resolution and Sparse Polynomial Interpolation: TV-minimization and SDP relaxations achieve ESRR in discrete/continuous settings when model-specific entropy/separation criteria are satisfied (e.g., $x^*$8) (Josz et al., 2017).
• Matrix and Operator Identification: ESRR for sparse matrix representations (e.g., time-frequency shift dictionaries) is guaranteed by convex optimization under classical coherence or RIP, even in regimes where standard RIP fails (Pfander et al., 2015).

6. Practical, Computational, and Robustness Aspects

Algorithmic ESRR strategies exhibit distinct statistical and computational characteristics:

• Efficiency: Greedy OMP-SR and its block variants match the ESRR performance of standard OMP but scale to large k/regimes by reducing per-step complexity from $x^*$9 to $y$0 (Yu et al., 2024).
• Noise Robustness: Both SDP-based and LP-based recovery show 1–2% coefficient error in moderate noise regimes; algebraic Prony’s variants remain robust under careful SVD thresholding (Josz et al., 2017).
• Model Overlap and Universality: Group-structured ESRR bounds are universal in group count and size (not overlap or ambient dimension), supporting a wide range of practical structures (Rao et al., 2011).
• Limiting Behavior: In random setting and for decaying or Gaussian-amplitude signals, OMP achieves ESRR with reduced measurements compared to worst-case bounds, controlling error through $y$1 ratios of the signal (Wen et al., 2020).

7. Limitations and Open Challenges

• The verification of MNDSC or curvature-type dual conditions is problem-dependent and may be technically challenging (Carioni et al., 2024).
• ESRR results for highly nonconvex or coupled regularizers (e.g. WDSN, $y$2) require augmented null space properties and specialized RIP bounds (Yu et al., 4 Jun 2026).
• Attaining explicit perturbation rates beyond $y$3 generally demands further regularity or finite-dimensional assumptions (Carioni et al., 2024).
• In some settings, such as continuous dictionaries or high-overlap group-lasso, computationally feasible certificates or scalable algorithms for ESRR remain only partially resolved.

---

Summary Table: ESRR Criteria and Guarantee Types

ESRR Regime	Condition Type	Guarantee
Finite $y$4/BP	NSP, RIP ($y$5)	Unique recovery of all k-sparse solutions
MNDSC/Banach	Dual certificate, strict curvature	Unique atomic decomposition in infinite-d.
OMP/gOMP	Mutual coherence, ERC	Support recovery in k steps probabilistic
Group/fusion-structured	Group-atomic width, block-RIP	Universal Gaussian measurement bound
Continuous dictionary	ERC on parameter grid (kernel)	k-step exact support recovery
Prony/Algebraic	Hankel/Toeplitz rank, SVD	Algebraic extraction with minimal samples

---

This synthesis brings together analytic, algorithmic, and geometric tools across ESRR, delineating the fundamental criteria, practical solvability, and universal features that underlie exact recovery of sparse signals and structured objects in contemporary applied mathematics and signal processing.