Support Recovery Capabilities

• Support recovery capabilities are theoretical and algorithmic criteria for precisely identifying the nonzero indices in sparse signals, essential for applications like compressed sensing and high-dimensional statistics.
• These methods are evaluated based on sample complexity, error tolerance, and robustness to noise, with models ranging from classical linear measurements to one-bit and federated settings.
• Recovery algorithms span convex relaxations, greedy approaches, and deep learning techniques, each offering specific trade-offs in accuracy, computational efficiency, and structural adaptability.

Support recovery capabilities refer to the theoretical and algorithmic criteria under which the support (i.e., the set of nonzero indices) of a sparse signal or parameter vector can be exactly or approximately identified from observations. This concept is foundational in sparse regression, compressed sensing, high-dimensional statistics, signal deconvolution, mixture models, federated inference, and a variety of structured signal recovery tasks. Support recovery is analyzed in terms of sample complexity, algorithmic procedures, error tolerance, robustness to noise, and structural priors, with guarantees that are often stated precisely as sufficient and/or necessary conditions on signal, measurement, and problem parameters.

1. Problem Formulations Across Models

Support recovery problems are defined with respect to a measurement model, a sparsity constraint, and a decoding or estimation objective. Consider several representative settings:

• Linear model (classical/compressed sensing):

$y = A x + z$

where $x\in\mathbb{R}^n$ is $k$-sparse, $A\in\mathbb{R}^{m\times n}$, and $z$ is (possibly zero) noise. The goal is to determine $\operatorname{supp}(x)$ (Jin et al., 2010).

• One-bit compressed sensing:

$y = \operatorname{sign}(\Phi x)$

with extreme quantization, restricting the available information to the signs of projections only (Mazumdar et al., 2021, Matsumoto et al., 2022).

• Noiseless/noisy deconvolution and grid-free inference:

$y = \Phi m_0 + w$

where $m_0$ is a sparse measure (Diracs) and $\Phi$ a convolution operator (Duval et al., 2013).

• Mixture models with sparse parameters:

Sampling from mixtures $x\in\mathbb{R}^n$0, each $x\in\mathbb{R}^n$1 sparse; recover supports $x\in\mathbb{R}^n$2 (Mazumdar et al., 2022).

• Federated/distributed regression:

Multiple clients with local regression models; joint support recovery at a server using communication-efficient aggregation (Barik et al., 2020).

• Structured and hierarchical sparsity:

Models with block, hierarchical, or group-sparse supports; recovery of compound support structure (Lu et al., 9 Nov 2025).

The inference task is instantiated either as exact support recovery (identifying the full support with no errors), approximate recovery (allowing a bounded fraction of errors), or superset/extended-support recovery (selecting a superset containing the true support, possibly with limited false positives).

2. Fundamental Information-Theoretic Limits

Support recovery feasibility is often characterized by sharp necessary and sufficient conditions on the sample complexity, in terms of sparsity $x\in\mathbb{R}^n$3, ambient dimension $x\in\mathbb{R}^n$4, noise parameters, and signal coefficients.

• Linear Model: For $x\in\mathbb{R}^n$5-sparse signals and measurement noise of variance $x\in\mathbb{R}^n$6, the number of measurements needed for exact recovery with vanishing error (assuming optimal decoding) satisfies:

$x\in\mathbb{R}^n$7

where $x\in\mathbb{R}^n$8 is a function of the nonzero amplitudes and SNR (Jin et al., 2010). For sublinear sparsity ($x\in\mathbb{R}^n$9), $k$0 is necessary and sufficient.

• One-Bit Compressed Sensing (Universal): For exact universal support recovery (recovering support for all $k$1-sparse signals using the same measurement matrix), $k$2 measurements are required and sufficient (Mazumdar et al., 2021). Allowing a controlled fraction of false positives reduces the rate to $k$3 (Mazumdar et al., 2021), and further to $k$4 given known signal dynamic range (Mazumdar et al., 2021, Matsumoto et al., 2022).
• Mixture Models: For recovering the support set of each sparse parameter in an $k$5-component mixture of canonical distributions, the sample complexity is polylogarithmic in ambient dimension and polynomial in $k$6, e.g., $k$7 for binary mixtures (Mazumdar et al., 2022).
• Federated Regression: With mutually independent features, per-client sample size $k$8 and number of clients $k$9 suffice for exact support recovery, matching central sample complexity (Barik et al., 2020).
• Noise and Incoherence: The minimal sample size also depends on measurement matrix incoherence and the signal-to-noise ratio (SNR) (0908.0744). Lower bounds are typically dictated by Fano's inequality and upper bounds by Chernoff analysis or multiuser capacity analogies.

3. Recovery Algorithms and Structural Conditions

Support recovery can be achieved by a diverse set of algorithms, each supported by precise theoretical sufficiency guarantees under different structural or statistical assumptions.

Convex Relaxations and Certificates

• $A\in\mathbb{R}^{m\times n}$0 Regularization (Basis Pursuit, LASSO):

Guarantees depend on Restricted Isometry Property (RIP), irrepresentable conditions, or dual certificate conditions (Kang et al., 2012, Degraux et al., 2016).

• Total Variation (TV) and Dual Certificates:

In continuous deconvolution, exact support recovery is characterized by existence of a dual certificate saturating at the true spike locations and no more (Duval et al., 2013); the nondegenerate source condition provides necessary and sufficient criteria.

Greedy and Combinatorial Algorithms

• Orthogonal Matching Pursuit (OMP), Hierarchical OMP, Block-OMP:

Recovery depends on coherence, block coherence, and power ratios, with explicit sufficient conditions even for structured or hierarchical supports (Lu et al., 9 Nov 2025, Shen et al., 2019, Saidi et al., 2021).

• Constrained Matching Pursuit (CMP):

Guarantees for models with convex constraint sets, exploiting coordinate-projection admissibility and constraint-dependent RIP/ROC constants (Shen et al., 2019).

Novel and Enhanced Procedures

• LiRE (Error-correction module):

Given a baseline algorithm's support estimate, LiRE applies combinatorial least-squares selection to correct up to a sublinear number of errors, provided minimal RIP assumptions (Mehrabi et al., 2021). Demonstrated to reduce measurements required for perfect support over OMP, BP, and CoSaMP.

• RAWLS and RLS (Random/Averaged Least Squares):

Aggregating and averaging many small-scale LS fits, followed by greedy support extraction, empirically outperform classical methods in exact support recovery, especially in high-noise or high-sparsity regimes (Lindenbaum et al., 2020, Lindenbaum et al., 2021).

• Support Exploration Algorithm (SEA) based on STE:

An STE-inspired method applying k-sparse least-squares projections to a dense proxy iteratively. RIP-based conditions, although stringent, enable theoretical guarantees. Empirically excellent, especially in high-coherence regimes where standard RIP fails (Mohamed et al., 2023).

Belief Propagation and Bayesian Hypothesis Test

• BHT-BP:

Bayesian hypothesis tests driven by belief propagation messages provide robust support recovery in moderate SNR regimes, outperforming OMP and Lasso in some settings, particularly for highly sparse or very noisy scenarios (Kang et al., 2012).

Deep Learning/Auto-encoder Approaches

• Model-driven and NN-based MMV Decoders:

Deep architectures that unroll MMV algorithms (AMP, coordinate descent, etc.) and combine them with trainable correction layers exhibit improved jointly sparse support recovery over classical approaches in MIMO/massive-IoT applications (Cui et al., 2020).

4. Error Tolerance and Robustness

Significant work quantifies the error-tolerance and robustness characteristics of support recovery algorithms:

• Approximate Recovery: OMP achieves arbitrarily low error rates in support identification provided the SNR is high enough relative to the minimum-to-average ratio of the signal, independent of the amplitude maximum (Lu et al., 2020).
• Extended/Approximate Support: Relaxed recovery objectives (approximate or superset recovery) allow much lower measurement complexity, especially in one-bit CS, at the expense of allowing a controlled number/ratio of false positives or negatives (Matsumoto et al., 2022, Mazumdar et al., 2021).
• Noisy Deconvolution: Under TV regularization and the NDSC, spike locations are recovered with $A\in\mathbb{R}^{m\times n}$1 accuracy and are robust to bounded or small Gaussian noise (Duval et al., 2013).
• Federated Settings: Median aggregation of support vectors is robust to up to 50% adversarial clients or client dropouts, with negligible loss in exact recovery probability (Barik et al., 2020).

5. Structural and Hierarchical Support Models

Support recovery has been extended and analyzed for complex structural sparsity:

• Hierarchical Block Sparsity (HiBOMP-P): Sufficient conditions are given in terms of hierarchical block-MIP and sub-coherence, showing that even with partial/incorrect prior support information, reliable recovery is possible and often strictly improved compared to non-hierarchical or non-PSI algorithms (Lu et al., 9 Nov 2025).
• Group/joint MMV Recovery: In MMV models, identifiability and sample complexity depend on the number of jointly sparse vectors and the rank or SNR configuration, with precise detection thresholds (0908.0744, Cui et al., 2020).
• Mixture Models: Recovery of supports for mixture components is possible with only logarithmic sample complexity in the ambient dimension, using combinatorial moment-based or spectral algorithms (Mazumdar et al., 2022).

6. Comparative Analysis and Trade-offs

A broad comparative perspective emerges:

Method/Setting	Exact Sufficient m (polylog factors suppressed)	Universality	Error Tolerance	Structural Adaptation
Universal 1bCS (exact)	$A\in\mathbb{R}^{m\times n}$2	Yes	No	None
Universal 1bCS (superset, approx)	$A\in\mathbb{R}^{m\times n}$3 or $A\in\mathbb{R}^{m\times n}$4	Yes	Yes ($A\in\mathbb{R}^{m\times n}$5)	Dynamic range, rationals, signs
Dense CS, i.i.d. Gaussian	$A\in\mathbb{R}^{m\times n}$6	No	Yes	Some (RIP), group possible
MMV (jointly sparse)	$A\in\mathbb{R}^{m\times n}$7	Yes	No/Approx	Rank/SNR separation, block structures
OMP (RIP)	$A\in\mathbb{R}^{m\times n}$8, error rate $A\in\mathbb{R}^{m\times n}$9 as SNR$z$0	Yes (RIP)	Yes	Block/hierarchical (extended)
TV/LASSO (spike deconv)	depends on NDSC & TV-certificates	Yes	Yes	Infinite ambient, grid/discrete
Federated regression	$z$1 (independent)	Yes	No	Robust to failures/poisoning
Hierarchical Block-Sparse/PSI	Structure-dependent, explicit coherence bounds	Yes	No	Multi-mode, subblock, PSI-guided

Additional observations:

• Algorithms that tolerate a controlled fraction of false positives/negatives efficiently break quadratic barriers in measurement complexity (Mazumdar et al., 2021, Matsumoto et al., 2022).
• The existence of appropriate dual certificates or low block-coherence often strictly improves the phase transitions relative to generic mutual coherence or RIP bounds (Duval et al., 2013, Saidi et al., 2021).
• Empirical studies consistently show that even minor algorithmic enhancements—e.g., LiRE post-processing, averaging, or hierarchical selection—can yield substantial improvements in sample complexity or recovery probability compared to classical greedy or convex approaches (Mehrabi et al., 2021, Lu et al., 9 Nov 2025, Lindenbaum et al., 2021, Lindenbaum et al., 2020).

7. Limitations and Future Directions

Current limitations include:

• RIP sufficiency: The sharpest theoretical guarantees depend on RIP, which may not hold for highly coherent or structured matrices; many practical settings violate these properties (Mohamed et al., 2023).
• Tightness for intermediate regimes: Gaps remain between necessary and sufficient sample complexity in certain sparsity/dimension regimes, with closing these recognized as an open problem (Jin et al., 2010).
• Extensions to non-linear and nonconvex problems: Support recovery theory for nonconvex, deep, or structured neural models is not yet sharply characterized, although heuristics (e.g., STE/SEA) exhibit promising empirical behavior (Mohamed et al., 2023).
• Finite sample and high-noise regimes: Many asymptotic results may be pessimistic for moderate sample sizes or high noise unless further refinements are introduced (Lu et al., 2020, Lindenbaum et al., 2021).

Continued research is proceeding on adaptivity to prior or side information (PSI), exploitation of signal structure (block/group/hierarchy), improved robustness to adversarial or federated constraints, and the integration of statistical, combinatorial, and deep learning tools. The evolving understanding of support recovery capabilities thus underpins progress in high-dimensional statistical inference, signal processing, and modern data-driven applications.