Mutual Incoherence Property (MIP) in Compressed Sensing

• Mutual Incoherence Property (MIP) is defined as the maximum normalized correlation between distinct columns of a sensing matrix, vital for sparse signal recovery.
• Low MIP values provide clear, computable recovery conditions that guarantee exact and stable recovery for sparse, block-sparse, and tensor-structured signals.
• Extensions like block, hierarchical, and tensor MIP enable enhanced recovery algorithms, improving performance under noisy measurements and complex structured sparsity.

The mutual incoherence property (MIP) is a central concept in sparse signal recovery, particularly within the context of compressed sensing and greedy selection algorithms. MIP quantifies the worst-case normalized correlation between columns—or blocks of columns—in a sensing matrix or dictionary. Small MIP values are directly linked to sufficient conditions for the exact and stable recovery of sparse, block-sparse, and even hierarchically block-sparse signals, under both noiseless and noisy measurements. Analogous block, hierarchical, and tensor generalizations of MIP have become central in modern frameworks for structured sparsity.

1. Formal Definition of the Mutual Incoherence Property

Let $A \in \mathbb{R}^{m \times n}$ be a measurement matrix with columns $A_1, \ldots, A_n$ normalized to unit $\ell_2$-norm. The mutual incoherence is defined as

$$\mu(A) = \max_{i \neq j} |\langle A_i, A_j \rangle|.$$

This measures the largest absolute inner product between any two distinct columns, quantifying their degree of similarity. For block-sparse frameworks, standard MIP generalizes to the block-coherence

$$\mu_B(A) = \max_{i \neq j} \frac{\rho(A[i]^T A[j])}{d},$$

where $A[i] \in \mathbb{R}^{m \times d}$ denotes the $i$th block and $\rho(\cdot)$ is the spectral norm, and the intra-block coherence

$$\nu(A) = \max_i \max_{p \neq q} |\langle A[i]_p, A[i]_q \rangle|,$$

where $A[i]_p$ is the $A_1, \ldots, A_n$0th column of block $A_1, \ldots, A_n$1 (Lu et al., 2022, Lu et al., 9 Nov 2025). Hierarchical and tensor extensions introduce further coherence measures over groupings of columns or Kronecker-structured atoms (Lu et al., 2024).

2. Fundamental Recovery Guarantees via MIP

Mutual incoherence yields explicit, easily computable sufficient conditions for uniform sparse recovery by greedy algorithms, convex optimization, and their structured variants.

2.1 Classical OMP Recovery Condition

For $A_1, \ldots, A_n$2-sparse signals $A_1, \ldots, A_n$3 measured as $A_1, \ldots, A_n$4, the classic result is: $A_1, \ldots, A_n$5 (Wang et al., 2011, Li et al., 2018). This threshold is tight: for $A_1, \ldots, A_n$6, exact recovery is not guaranteed for all $A_1, \ldots, A_n$7-sparse supports. The proof is inductive, showing by explicit projection estimates that at each OMP iteration, the true support element achieves strictly larger correlation with the residual than any incorrect atom, provided the stated $A_1, \ldots, A_n$8-bound holds.

2.2 OLS, BOLS, and Block-Structured Extensions

For orthogonal least squares-type algorithms, the analogous noiseless condition (for small $A_1, \ldots, A_n$9) is

$\ell_2$0

and for block OLS (blocks of size $\ell_2$1), the critical threshold is

$\ell_2$2

substantially relaxing the OMP/BOMP constraints (Lu et al., 2022).

Noisy-case guarantees require the minimal nonzero coefficient magnitude to exceed a scaled noise level, with scaling factor depending on $\ell_2$3 (Zhang et al., 2021).

3. Block, Hierarchical, and Tensor Generalizations

3.1 Block and Hierarchical Coherence

For block-sparse signals, MIP is formulated on the aggregation of entire blocks: $\ell_2$4 with sub-coherence $\ell_2$5 capturing inner-block coupling. Hierarchical block-sparse settings require consideration of groupings at various block lengths: $\ell_2$6 and sub-coherence at block-length $\ell_2$7, $\ell_2$8, as in (Lu et al., 9 Nov 2025).

3.2 Tensor MIP

For $\ell_2$9-mode tensor measurements, with measurement set $$\mu(A) = \max_{i \neq j} |\langle A_i, A_j \rangle|.$$0, the mutual block coherence is

$$\mu(A) = \max_{i \neq j} |\langle A_i, A_j \rangle|.$$1

and the mutual sub-coherence $$\mu(A) = \max_{i \neq j} |\langle A_i, A_j \rangle|.$$2 is defined analogously for columns rather than blocks (Lu et al., 2024). These generalize the scalar and block MIP by aggregating n-way structure and cross-coherence.

4. Tight Recovery and Stability Bounds Based on MIP

4.1 Exact Recovery

Most results state that for models with block or tensor structure, versions of the following hold:

• If block- or tensor-MIP is below a computable threshold (involving block size, block- or tensor-coherence, and possibly intra-block parameters), greedy algorithms (e.g., BOMP, BOLS, T-GBOMP) provably recover all supports up to a prescribed sparsity level.

For example, in the tensor model (Lu et al., 2024): $$\mu(A) = \max_{i \neq j} |\langle A_i, A_j \rangle|.$$3 with $$\mu(A) = \max_{i \neq j} |\langle A_i, A_j \rangle|.$$4 and $$\mu(A) = \max_{i \neq j} |\langle A_i, A_j \rangle|.$$5 encoding the supports of the true tensor and selected elements, respectively.

4.2 Stable Recovery Under Noise

For noisy data $$\mu(A) = \max_{i \neq j} |\langle A_i, A_j \rangle|.$$6, let the noise satisfy $$\mu(A) = \max_{i \neq j} |\langle A_i, A_j \rangle|.$$7 and coherence $$\mu(A) = \max_{i \neq j} |\langle A_i, A_j \rangle|.$$8. The Lasso solution $$\mu(A) = \max_{i \neq j} |\langle A_i, A_j \rangle|.$$9 obeys

$$\mu_B(A) = \max_{i \neq j} \frac{\rho(A[i]^T A[j])}{d},$$0

(Li et al., 2018). Lower bounds: minimax risk scales as $$\mu_B(A) = \max_{i \neq j} \frac{\rho(A[i]^T A[j])}{d},$$1, i.e., small $$\mu_B(A) = \max_{i \neq j} \frac{\rho(A[i]^T A[j])}{d},$$2 directly reduces achievable error rates.

In the multiple-measurement-vector setting (e.g., SOMP) (Zhang et al., 2021), when $$\mu_B(A) = \max_{i \neq j} \frac{\rho(A[i]^T A[j])}{d},$$3 and noise spectral norm $$\mu_B(A) = \max_{i \neq j} \frac{\rho(A[i]^T A[j])}{d},$$4,

$$\mu_B(A) = \max_{i \neq j} \frac{\rho(A[i]^T A[j])}{d},$$5

guarantees correct support recovery. For random (e.g., Gaussian) noise, the recovery probability is lower-bounded in terms of the Tracy–Widom law.

5. Algorithmic Implications and Measurement Design

Low MIP is achieved with random constructions (e.g., subgaussian, random partial Fourier), or via explicit design such as Grassmannian packing, frame theory, and Gram–Schmidt orthogonalization. For multi-mode or tensorized frameworks, reducing cross-terms in all dictionary factors lowers the aggregate tensor MIP $$\mu_B(A) = \max_{i \neq j} \frac{\rho(A[i]^T A[j])}{d},$$6; in block/hierarchical setups, minimizing $$\mu_B(A) = \max_{i \neq j} \frac{\rho(A[i]^T A[j])}{d},$$7 and intra-block $$\mu_B(A) = \max_{i \neq j} \frac{\rho(A[i]^T A[j])}{d},$$8 directly enhances the recoverable sparsity range (Lu et al., 2024, Lu et al., 9 Nov 2025).

Block, tensor, and hierarchical generalizations enable the extension of MIP-based guarantees to structured sparsity domains—including the exploitation of prior support information, which can further relax recovery conditions even when such information is not perfectly aligned with the true support (Lu et al., 9 Nov 2025).

6. Comparison to Other Recovery Criteria and Scaling Laws

MIP-based conditions are more tractable than Restricted Isometry Property (RIP) checks, which are often NP-hard to verify. While RIP-based guarantees can be optimal in the sense of the number of measurements $$\mu_B(A) = \max_{i \neq j} \frac{\rho(A[i]^T A[j])}{d},$$9, for practical algorithm analysis (e.g., greedy selection, $A[i] \in \mathbb{R}^{m \times d}$0 minimization) coherence-based MIP criteria offer explicit, deterministic, and, under small $A[i] \in \mathbb{R}^{m \times d}$1, nearly tight thresholds (Li et al., 2018, Lu et al., 2022).

Table: Summary of MIP Recovery Thresholds for Selected Algorithms

Algorithm/Model	Main MIP Threshold	Structural Extension
OMP (scalar sparse)	$A[i] \in \mathbb{R}^{m \times d}$2	(Wang et al., 2011, Li et al., 2018)
OLS, MOLS	$A[i] \in \mathbb{R}^{m \times d}$3	(Lu et al., 2022)
BOMP, BOLS	$A[i] \in \mathbb{R}^{m \times d}$4	$A[i] \in \mathbb{R}^{m \times d}$5 model-dependent
Tensor (T-GBOMP)	$A[i] \in \mathbb{R}^{m \times d}$6	(Lu et al., 2024)
Lasso	$A[i] \in \mathbb{R}^{m \times d}$7	Stable $A[i] \in \mathbb{R}^{m \times d}$8 error

7. Advanced MIP Concepts: Hierarchical and Prior Information

Hierarchical MIP introduces coherence measurements over variable block sizes and aggregated, possibly non-contiguous, groupings of columns. In these settings, the worst-case hierarchical block-coherence $A[i] \in \mathbb{R}^{m \times d}$9 and sub-coherence $i$0 provide recovery thresholds that adapt to arbitrary grouping structures (Lu et al., 9 Nov 2025). Incorporation of prior support information modifies the critical bounds: even zero-overlap prior sets, through augmentation of the candidate support, can improve recovery thresholds—a result not available in classical (non-hierarchical) frameworks.

Recovery conditions under hierarchical MIP can be written explicitly. For example, with perfect hierarchical-block orthogonality, exact recovery in mode $i$1 requires: $i$2 (Lu et al., 9 Nov 2025). For noisy measurements and Lasso-type approaches, the inclusion of MIP in oracle inequalities links practical error rates to what would be achieved by an ideal “oracle” estimator.

---

The mutual incoherence property and its block, tensor, and hierarchical generalizations provide a unified, sharp, and computationally practical framework for analyzing sparse and structured-sparse signal recovery. These concepts underlie the design of sampling matrices, the analysis of recovery algorithms, and the study of trade-offs between sparsity, noise robustness, and data dimensionality.