Restricted Isometry Constant (RIC) in Compressed Sensing

• RIC is a measure that quantifies the near-isometric behavior of a sensing matrix on sparse vectors in compressed sensing.
• It determines phase transitions by setting thresholds for recovery algorithms, impacting both convex and nonconvex optimization methods.
• Its computation is NP-hard, leading to reliance on probabilistic bounds and relaxations for practical estimation in high dimensions.

The restricted isometry constant (RIC) is a central quantitative metric in compressed sensing, sparse signal recovery, and high-dimensional data analysis. It measures to what extent a matrix acts nearly isometrically when restricted to sparse vectors, providing necessary and sufficient conditions for the success of various recovery algorithms, both convex and nonconvex. RIC thresholds govern phase transitions, inform algorithm selection, and underlie sample complexity analyses in modern compressed sensing theory. The RIC framework has shaped both theoretical understanding and algorithmic design, but is also marked by significant challenges in its computation, estimation, and sharpness for diverse matrix classes and recovery methods.

1. Formal Definition and Basic Properties

Given a matrix $A\in\mathbb{R}^{m\times n}$ and integer $s\geq 1$, the $s$-th order restricted isometry constant (RIC) $\delta_s(A)$ is defined as the smallest $\delta\geq0$ such that, for all $s$-sparse vectors $x$,

$$(1-\delta)\|x\|_2^2 \leq \|Ax\|_2^2 \leq (1+\delta)\|x\|_2^2$$

Equivalently,

$$1-\delta \leq \frac{\|Ax\|_2^2}{\|x\|_2^2} \leq 1+\delta \qquad \forall x: \|x\|_0 \leq s$$

RICs are monotone in order: if $s_1\leq s_2$ then $s\geq 1$0. They can also be expressed in eigenvalue language: for each subset $s\geq 1$1, $s\geq 1$2, let $s\geq 1$3 denote the associated submatrix. Then,

$s\geq 1$4

The minimal $s\geq 1$5 for which $s\geq 1$6 is said to have the restricted isometry property (RIP) of order $s\geq 1$7 and level $s\geq 1$8.

2. RIC in Sparse Recovery: Thresholds and the Null-Space Property

RICs constitute a critical metric guaranteeing exact and stable recovery of $s\geq 1$9-sparse vectors. For recovery via (potentially nonconvex) $s$0-minimization ($s$1): $s$2 Necessary and sufficient conditions are formulated via the null-space property (NSP): $s$3 RIC lower bounds are obtained by relating the NSP to the RIC via sharp inequalities between $s$4 and $s$5 norms. For $s$6 or $s$7 (including both convex and certain archetypal nonconvex relaxations), the exact-recovery threshold is: $s$8 This threshold is sharp for $s$9-minimization and, as recently established, for $\delta_s(A)$0-minimization as well, showing an equivalence in recovery capability between these relaxations under the same RIC bound (Zhou et al., 2013).

For more general $\delta_s(A)$1, nontrivial improvements are obtained on $\delta_s(A)$2. Explicit examples:

• For $\delta_s(A)$3, if $\delta_s(A)$4 then recovery holds for all $\delta_s(A)$5.
• For $\delta_s(A)$6, if $\delta_s(A)$7 then recovery holds for all $\delta_s(A)$8 (Zhou et al., 2013). Furthermore, in the minimal-sparsity regime, for $\delta_s(A)$9, RIC bounds specialize to $\delta\geq0$0 (even $\delta\geq0$1) and $\delta\geq0$2 (odd $\delta\geq0$3), pinpointing the sharpest-known thresholds.

3. RIC Bounds for Recovery Algorithms

The RIC framework enables the comparison of various recovery algorithms:

Algorithm	Sufficient RIC Threshold ($\delta\geq0$4)	Reference
$\delta\geq0$5-minimization	$\delta\geq0$6	(Zhou et al., 2013)
OMP	$\delta\geq0$7	(Mo et al., 2012, Mo, 2015)
gOMP	$\delta\geq0$8	(Chen et al., 2016)
SP	$\delta\geq0$9	(Song et al., 2013)
CoSaMP	$s$0	(Song et al., 2013)
$s$1-min ($s$2)	see explicit bounds in (Zhou et al., 2013, Hsia et al., 2013, Song et al., 2013)

These thresholds often match, or are provably sharp for, the breakdown points of the corresponding algorithms. For instance, OMP’s bound is sharp: a matrix can have $s$3 and yet OMP will fail on some $s$4-sparse vector. The same tightness applies to gOMP and certain block-sparse reductions (Mo, 2015, Chen et al., 2016).

4. Computational Complexity and Practical Estimation

Determining whether a given matrix $s$5 satisfies the RIP for specified $s$6 and $s$7, or computing $s$8 exactly, is strongly intractable. Both the exact computation and decision versions of the RIC are NP-hard, and even coNP-complete for some formulations (Tillmann et al., 2012). No polynomial-time (or even pseudo-polynomial time) approximation within arbitrarily small factors exists unless P=NP. This computational barrier arises via direct reduction from the spark problem, which is also NP-complete: determining whether $s$9 has a dependent subset of $x$0 columns.

As a result, practical and theoretical work on RIC estimation focuses on:

• Efficient upper and lower bounds, often via semidefinite relaxations, convex programming, or probabilistic methods.
• Probabilistic guarantees for random matrix ensembles, such as Gaussian or subgaussian matrices, where high-probability uniform RIP is attainable with $x$1 rows.

Due to NP-hardness, almost all practical recovery and analysis methods rely on proxies or replace worst-case RICs with high-probability bounds for structured or random $x$2.

5. RIC Bounds for Random and Structured Matrices

The typical mechanism for bounding RICs in random or structured ensembles is via concentration inequalities for the extreme singular values (or equivalently, eigenvalues) of submatrices.

Notable classes and scaling laws:

• Gaussian/Random Matrices: With probability at least $x$3, $x$4 satisfies $x$5 when $x$6. Sharper large-deviation and replica-based bounds further optimize $x$7 and asymptotic tightness (Bah et al., 2012, Stojnic, 2013, Sakata et al., 2015, James et al., 2014).
• Partial Random Circulant/Gabor: For such structured matrices, $x$8 rows suffice to ensure $x$9 with overwhelming probability (Krahmer et al., 2012).
• Khatri-Rao/Kronecker Products: The RIC of a Khatri-Rao product $$(1-\delta)\|x\|_2^2 \leq \|Ax\|_2^2 \leq (1+\delta)\|x\|_2^2$$0 is at most $$(1-\delta)\|x\|_2^2 \leq \|Ax\|_2^2 \leq (1+\delta)\|x\|_2^2$$1, making such product matrices strictly stronger isometries. For Kronecker products, $$(1-\delta)\|x\|_2^2 \leq \|Ax\|_2^2 \leq (1+\delta)\|x\|_2^2$$2 (Khanna et al., 2017, He et al., 2024).

Extreme-value theory and random matrix concentration provide not only the scaling laws but the precise finite-$$(1-\delta)\|x\|_2^2 \leq \|Ax\|_2^2 \leq (1+\delta)\|x\|_2^2$$3 performance, including sharp constants and the distributions of left/right RICs. Notably, in the Gaussian ensemble, left and right RICs asymptotically follow Weibull (min) and Gumbel (max) limit laws, respectively (James et al., 2014).

6. Role in Compressed Sensing Phase Transitions and Algorithmic Design

The RIC determines the algorithmic boundary (“phase transition”) for which classes of sparse recovery succeed uniformly. For example, in compressed sensing:

• Sharp RICs imply that for $$(1-\delta)\|x\|_2^2 \leq \|Ax\|_2^2 \leq (1+\delta)\|x\|_2^2$$4, $$(1-\delta)\|x\|_2^2 \leq \|Ax\|_2^2 \leq (1+\delta)\|x\|_2^2$$5-minimization exactly recovers all $$(1-\delta)\|x\|_2^2 \leq \|Ax\|_2^2 \leq (1+\delta)\|x\|_2^2$$6-sparse vectors.
• For OMP, exact (and only exact) recovery for every $$(1-\delta)\|x\|_2^2 \leq \|Ax\|_2^2 \leq (1+\delta)\|x\|_2^2$$7-sparse signal in $$(1-\delta)\|x\|_2^2 \leq \|Ax\|_2^2 \leq (1+\delta)\|x\|_2^2$$8 steps is achievable when $$(1-\delta)\|x\|_2^2 \leq \|Ax\|_2^2 \leq (1+\delta)\|x\|_2^2$$9 (Mo et al., 2012, Mo et al., 2011, Mo, 2015).
• Extensions to block-sparsity, joint sparsity, and group-structured dictionaries similarly depend on structured RICs.

In practical design, knowledge of attainable RIC levels in candidate matrix ensembles guides the selection of $$1-\delta \leq \frac{\|Ax\|_2^2}{\|x\|_2^2} \leq 1+\delta \qquad \forall x: \|x\|_0 \leq s$$0 (row count) required for reliable recovery, and dictates when higher computational-complexity algorithms (e.g., convex programming) are justified versus fast greedy algorithms (e.g., OMP, gOMP, SP, CoSaMP).

7. Open Problems and Future Directions

Despite rigorous advances, several open questions remain:

• Sharpness for General Classes: While exact sharp bounds are available for some algorithms and ensembles, closing gaps between sufficiency and necessity for $$1-\delta \leq \frac{\|Ax\|_2^2}{\|x\|_2^2} \leq 1+\delta \qquad \forall x: \|x\|_0 \leq s$$1 (with $$1-\delta \leq \frac{\|Ax\|_2^2}{\|x\|_2^2} \leq 1+\delta \qquad \forall x: \|x\|_0 \leq s$$2 or for nonconvex $$1-\delta \leq \frac{\|Ax\|_2^2}{\|x\|_2^2} \leq 1+\delta \qquad \forall x: \|x\|_0 \leq s$$3) remains a prominent goal.
• RICs in Highly Structured or Deterministic Matrices: Achieving RIP with near-optimal $$1-\delta \leq \frac{\|Ax\|_2^2}{\|x\|_2^2} \leq 1+\delta \qquad \forall x: \|x\|_0 \leq s$$4 for deterministic constructions or those with significant algebraic structure is significantly less understood than for random matrices.
• Beyond RIC: Alternative Metrics: Given NP-hardness and extremality, complementary notions (mutual coherence, statistical RIP, etc.) are actively explored for practical verifiability.

Recent techniques leveraging advances in statistical mechanics (replica symmetric and replica symmetric breaking analysis), concentration of measure, chaos processes, and probabilistic order-statistics provide ongoing progress and increasingly tight upper and lower bounds for RICs in both random and structured settings (Sakata et al., 2015, Stojnic, 2013, James et al., 2014).

---

References: (Zhou et al., 2013, Tillmann et al., 2012, Mo et al., 2012, Mo, 2015, Chen et al., 2016, Mo et al., 2011, Hsia et al., 2013, Song et al., 2013, Song et al., 2013, Bah et al., 2012, James et al., 2014, Khanna et al., 2017, He et al., 2024, Stojnic, 2013, Sakata et al., 2015, Dallaporta et al., 2016, Krahmer et al., 2012, Bah et al., 2010, Elzanaty et al., 2018).