Multipath Matching Pursuit (MMP)

• Multipath Matching Pursuit (MMP) is a class of greedy algorithms that maintains multiple candidate supports for more reliable sparse signal recovery.
• It systematically expands candidate supports using correlation-based selection and pruning strategies, enabling recovery under relaxed conditions even in noisy scenarios.
• MMP variants—including breadth-first, depth-first, and Bayesian extensions—offer practical trade-offs between recovery accuracy and computational complexity for diverse applications.

Multipath Matching Pursuit (MMP) is a class of algorithms in sparse signal recovery that extends the single-path, greedy pursuit framework by maintaining and expanding multiple support hypotheses in parallel. MMP methods systematically explore a small search tree of candidate supports, typically using correlation-based selection and pruning strategies to efficiently recover sparse signals from underdetermined measurements. These algorithms provide a means to mitigate the error-propagation and myopia of traditional approaches like Orthogonal Matching Pursuit (OMP), combining the advantages of tree search with greedy selection for improved recoverability under relaxed conditions.

1. Mathematical Formulation and Core Principles

Given observations $y = \Phi x + n$, where $\Phi\in\mathbb{R}^{m\times N}$ is a known sensing matrix (with $m\ll N$), $x\in\mathbb{R}^N$ is a $k$-sparse signal, and $n$ is additive noise, the goal is to estimate both the support $\mathrm{supp}(x)$ and amplitudes of the nonzero components.

Standard greedy algorithms build up the support iteratively, selecting at each iteration a single new index with maximal correlation to the current residual. The innovation of MMP is to generate and expand multiple promising candidate supports. At each iteration $k$, a pool $S^k$ of partial supports is maintained. For each $T\in S^{k-1}$:

• Compute the residual $\Phi\in\mathbb{R}^{m\times N}$0, where $\Phi\in\mathbb{R}^{m\times N}$1 (Moore–Penrose pseudoinverse).
• Identify the $\Phi\in\mathbb{R}^{m\times N}$2 columns of $\Phi\in\mathbb{R}^{m\times N}$3 most correlated with $\Phi\in\mathbb{R}^{m\times N}$4; for each such column, form a child support $\Phi\in\mathbb{R}^{m\times N}$5.
• Retain all distinct $\Phi\in\mathbb{R}^{m\times N}$6 for the next level.
• After $\Phi\in\mathbb{R}^{m\times N}$7 iterations, select the support in $\Phi\in\mathbb{R}^{m\times N}$8 achieving the minimal final residual (Suhyuk et al., 2013).

This breadth-first process can be paralleled by depth-first variants with a maximum number of explored paths, and is modulated by parameters such as branching factor $\Phi\in\mathbb{R}^{m\times N}$9 and search depth $m\ll N$0.

2. MMP Algorithmic Variants

The MMP framework has evolved into several notable variants, each with specific search, selection, and pruning strategies. Key implementations include:

Breadth-First MMP (MMP-BF)

• At each iteration, expand every surviving support by $m\ll N$1 children, keeping all distinct offspring. Computational cost grows as $m\ll N$2 in the worst case, though practical path overlap often reduces this burden.
• Final selection is based on the child with minimal residual norm (Suhyuk et al., 2013).

Depth-First MMP (MMP-DF)

• Explores the tree one path at a time, backtracking when leaf nodes or residual thresholds are met.
• Allows direct control over the complexity by fixing $m\ll N$3, the maximum number of full candidate paths.

Multi-Branch Matching Pursuit (MBMP)

• Introduces a vector $m\ll N$4 that specifies the number of branches at each tree level.
• Incorporates dictionary and subspace refinement, especially effective in multiple-measurement-vector (MMV) models and rank-aware scenarios (Rossi et al., 2013).

Matching Pursuit with Tree Pruning (TMP)

• Employs an initial pre-selection phase (e.g., via group OMP) to limit the set of indices investigated, reducing tree width from $m\ll N$5 to $m\ll N$6.
• Uses early noncausal support completion and residual-based criteria to prune entire branches whose projected residuals cannot exceed the running optimum (Lee et al., 2014).

Bayesian Multiple Matching Pursuit (BMMP)

• Uses a Bayesian likelihood ratio score, constructed via Davies–Eldar’s rank-aware inner product, to guide index selection.
• Expands each candidate support up to cardinality $m\ll N$7, exploits diversity by constructing $m\ll N$8 different paths, and applies iterative subset prunning and a final sparse Bayesian learning (SBL) or ridge regression step (Kim et al., 2019).

3. Exact and Stable Recovery Guarantees

Theoretical guarantees for MMP depend on the properties of the sensing matrix $m\ll N$9, typically described via the Restricted Isometry Property (RIP), and on the specifics of the algorithmic variant.

• Breadth-First & Depth-First MMP: Under the condition that $x\in\mathbb{R}^N$0 (where $x\in\mathbb{R}^N$1 is the $x\in\mathbb{R}^N$2-RIP constant), MMP exactly recovers any $x\in\mathbb{R}^N$3-sparse $x\in\mathbb{R}^N$4 in the noiseless case (Suhyuk et al., 2013).
• TMP: Achieves exact support recovery if, for example, $x\in\mathbb{R}^N$5 and $x\in\mathbb{R}^N$6 for $x\in\mathbb{R}^N$7, with additional stable recovery bounds in the noisy setting (Lee et al., 2014).
• MBMP: Introduces the MB-coherence condition, whose satisfaction for the chosen branching structure allows recovery with fewer measurements than the classical ERC or Babel/coherence bounds. For instance, with $x\in\mathbb{R}^N$8, moving from single- to triple-branch at level 1 reduces the required measurement product $x\in\mathbb{R}^N$9 for $k$0 recovery from $k$1 to $k$2 (Rossi et al., 2013).

A key distinction is that, while MMP's idealized recovery bounds assume no aggressive pruning, pruning strategies are typically required for practical tractability. The correctness claims above hold provided pruning does not eliminate all correct paths.

4. Empirical Performance and Complexity

Simulation results consistently show that MMP algorithms outperform standard single-path greedy schemes in both noiseless and noisy settings.

• Noiseless Regimes: MMP-BF (with $k$3) achieves significantly higher exact recovery ratios (ERR) than OMP, CoSaMP, and SP, particularly as sparsity increases. For $k$4, MMP-BF's ERR is $k$5 versus $k$6 for OMP (Suhyuk et al., 2013).
• Noisy Regimes: MMP variants achieve MSE close to the oracle least-squares bound, with error decaying as $k$7, while OMP and SP floor earlier (Suhyuk et al., 2013).
• BMMP: Approaches the information-theoretic $k$8 limit ($k$9), outperforming several MAP-based and convex relaxation methods in both accuracy and runtime (Kim et al., 2019).
• TMP: With adaptive pre-selection ($n$0) and pruning, achieves near-oracle recovery for $n$1 up to $n$2 in a $n$3 regime, and MSE within $n$4–$n$5 dB of state of the art at moderate SNR (Lee et al., 2014).

The primary computational cost arises from the exponential growth in the number of candidate paths for breadth-first variants. Depth-first or pruning approaches, use of pre-selection, and limits on maximum node expansions can reduce typical run-times to be comparable with other practical recovery algorithms.

5. Relation to Prior Art and Critical Assessments

Karahanoğlu & Erdogan (Karahanoglu et al., 2015) critically assessed the originality and efficacy of MMP:

• Historical Context: The core approach of MMP—breadth-first (MP:M-L) and depth-first (MP:K) multipath extensions—were established by Cotter and Rao in 2001. The algorithms later named MMP-BF and MMP-DF are rediscoveries of these schemes.
• Comparison to A*OMP: A*OMP, introduced prior to MMP, leverages more sophisticated path cost models and global queue management, achieving higher recovery rates, lower error, and lower computational burden in empirical comparisons.
• Theory-Practice Gap: Recovery guarantees in the MMP literature often ignore the impact of pruning, which is essential in practice. A*OMP offers tighter RIP-based bounds.
• Novelty Limitations: The main contribution of MMP is pedagogical clarity rather than algorithmic advancement; performance is generally eclipsed by sophisticated heuristic-guided searches such as A*OMP (Karahanoglu et al., 2015).

6. Application Domains

MMP and its derivatives are relevant in settings requiring robust sparse support detection in the presence of noise, basis mismatch, or high mutual coherence. Empirically validated application areas include:

• Wireless channel estimation in highly multipath environments (Suhyuk et al., 2013).
• Sparse MRI and imaging tasks, particularly when the dictionary is redundant or ill-conditioned.
• Radar/sonar target detection, notably in MIMO radar, where MBMP leverages signal subspace structure in the MMV model for improved measurement efficiency (Rossi et al., 2013).
• Any context where error propagation from early incorrect selections in OMP or SP critically limits the accuracy or reliability of the support recovery.

7. Algorithmic Innovations: Bayesian and Rank-Aware Extensions

Recent work has emphasized the integration of Bayesian inference and rank-aware metrics into the MMP paradigm. Notable contributions:

• BMMP: Utilizes Davies–Eldar’s rank-aware OMP correlation together with a Bayesian likelihood-ratio test to construct a more powerful index selection criterion. BMMP's iterative, multi-path structure with Bayesian score-based expansion and extended support estimation consistently outperforms state-of-the-art greedy and convex alternatives in both reconstructive accuracy and computational efficiency (Kim et al., 2019).
• MBMP with Subspace Refinement: For MMV and high-rank cases, MBMP exploits orthogonal projections of the measurement matrix and the residuals, replacing standard atom selection with maximization of the projected correlation over multiple snapshots. The rank-awareness is critical for guaranteeing recovery under looser conditions and in higher noise scenarios (Rossi et al., 2013).

A plausible implication is that by leveraging such statistical and structural information, multipath pursuit algorithms can further close the performance gap to ideal oblique or oracle estimators, especially in challenging regimes.

---

References

• "Multipath Matching Pursuit" (Suhyuk et al., 2013)
• "Greedy Sparse Signal Recovery with Tree Pruning" (Lee et al., 2014)
• "Multi-Branch Matching Pursuit with applications to MIMO radar" (Rossi et al., 2013)
• "Bayesian Approach with Extended Support Estimation for Sparse Regression" (Kim et al., 2019)
• "Comments On 'Multipath Matching Pursuit' by Kwon, Wang and Shim" (Karahanoglu et al., 2015)