Multi-Branch Matching Pursuit

• Multi-Branch Matching Pursuit is a structured greedy algorithm for sparse recovery that explores multiple candidate supports via a prescribed tree-search strategy.
• The algorithm improves recovery guarantees by balancing branching factors and measurement resources using the MB-coherence condition.
• MBMP applies to both SMV and MMV problems, incorporating rank-aware techniques to efficiently recover sparse signals even under high dictionary coherence.

Multi-Branch Matching Pursuit (MBMP) is a structured greedy algorithmic framework for sparse recovery over overcomplete dictionaries, characterized by simultaneous exploration of multiple candidate support sets through a prescribed tree-search strategy. By balancing the number of candidate solution branches against available measurements, MBMP achieves significantly improved recovery guarantees over classical single-path greedy methods, especially in settings where dictionary coherence imposes stringent theoretical limits. The MBMP paradigm is applicable to both single- and multiple-measurement vector (SMV/MMV) problems and incorporates rank-aware techniques when multiple signals share joint support. The theoretical justification for MBMP is supplied by the MB-coherence condition, which yields strictly weaker requirements on the sensing matrix or dictionary than prior results based on cumulative coherence or the Neumann Exact Recovery Condition (ERC) (Rossi et al., 2013).

1. Formal Definition and Algorithmic Structure

MBMP is defined for the sparse recovery problem: Given $Y\in\mathbb{C}^{m\times l}$ and dictionary $A\in\mathbb{C}^{m\times n}$, find the K-sparse support $C$ minimizing the Frobenius norm of the projected residual,

$\min_{C: |C|\le K} \|\Pi_{A_C^\perp}Y\|_F,$

where $\Pi_{A_C^\perp}=I-A_CA_C^\dagger$ is the orthogonal projector onto the complement of the column space of $A_C$.

MBMP proceeds by constructing a tree with depth $K$; at each level $i$, each node (partial support $C$) spawns $d_i$ child nodes, each corresponding to an augmented support. The selection of children is governed by maximizing a rank-aware subspace correlation with the current residual. The number of children $A\in\mathbb{C}^{m\times n}$0 (the branch vector) is a key parameter that modulates the balance between computational burden and recovery guarantees.

At each expansion, newly added indices are selected greedy-wise among those not already present in $A\in\mathbb{C}^{m\times n}$1, according to the largest projections of the current residual subspace onto refined dictionary atoms. The process continues until the tree reaches depth $A\in\mathbb{C}^{m\times n}$2, after which the leaf node with the smallest residual norm is returned as the recovered support (Rossi et al., 2013).

2. Theoretical Guarantees: MB-Coherence and Recovery Conditions

The theoretical power of MBMP is founded on the MB-coherence condition. For a given partial support $A\in\mathbb{C}^{m\times n}$3 and remaining indices, define

• Dictionary mutual coherence: $A\in\mathbb{C}^{m\times n}$4
• Out-in-ratio (OIR): The ratio of maximal subspace projection for atoms outside the true support to that within it.

The MB-ERC (Multi-Branch Exact Recovery Condition) for a true support $A\in\mathbb{C}^{m\times n}$5 and partial support $A\in\mathbb{C}^{m\times n}$6, parametrized by $A\in\mathbb{C}^{m\times n}$7, states:

$A\in\mathbb{C}^{m\times n}$8

with $A\in\mathbb{C}^{m\times n}$9. The uniform MB-coherence bound at the root (noiseless, OIR=0) simplifies to:

$C$0

Hence, increasing the branching factor $C$1 allows MBMP to succeed with strictly more coherent dictionaries than permitted by single-path (TMP, $C$2) methods.

MBMP theoretical results rigorously prove that, if the MB-coherence condition is satisfied for all partial supports throughout the search, exact support recovery is guaranteed, even in the MMV (multi-snapshot) setting, so long as $C$3 (Rossi et al., 2013).

3. Greedy, Rank-Aware, and Tree-Search Principles

MBMP synthesizes three algorithmic principles:

• Greedy Expansion: Sequentially constructs candidate supports by maximizing a local criterion analogous to that in OMP or ORMP.
• Rank Awareness: In MMV, leverages the orthogonalized measurement subspace at each partial support to select candidate atoms, enhancing signal discrimination when multiple measurements share sparse structure.
• Multi-Branch Tree Search: Systematically explores multiple hypotheses at each step, with the branch vector $C$4 specifying the trade-off between hardware (e.g., number of measurements) and computational resources.

This framework generalizes earlier schemes—TMP/ORMP are recovered for $C$5—and strictly improves recovery capacity as $C$6 increases, as quantified by MB-coherence (Rossi et al., 2013).

4. Computational Complexity and Trade-offs

The computational complexity of MBMP is

$C$7

significantly higher than the $C$8 of single-path TMP, due to the exponential growth in the number of nodes traversed with increased branching.

A crucial benefit is that MBMP admits a flexible exchange: with higher branching (even modest $C$9), the required number of measurements $\min_{C: |C|\le K} \|\Pi_{A_C^\perp}Y\|_F,$0 for successful recovery can be reduced, at the price of exponentially (or combinatorially) increased runtime and memory. Empirical studies in MIMO radar demonstrate that increasing $\min_{C: |C|\le K} \|\Pi_{A_C^\perp}Y\|_F,$1 from 1 to 2 can decrease hardware requirements for the same recovery probability, sometimes by 5–10% (Rossi et al., 2013).

In high-SNR regimes, MBMP approaches Oracle-like recovery error rates, especially as branching is increased. Pruning is strictly enforced by not allowing more than $\min_{C: |C|\le K} \|\Pi_{A_C^\perp}Y\|_F,$2 child expansions per parent node, ensuring computational tractability for moderate $\min_{C: |C|\le K} \|\Pi_{A_C^\perp}Y\|_F,$3 and $\min_{C: |C|\le K} \|\Pi_{A_C^\perp}Y\|_F,$4.

5. Comparison: MBMP vs. Classical TMP and Other Greedy Algorithms

MBMP (with $\min_{C: |C|\le K} \|\Pi_{A_C^\perp}Y\|_F,$5) reduces to the tree-pruned OMP/ORMP or TMP (Lee et al., 2014), which explore a single greedy path. In this regime, recovery is controlled by cumulative coherence or Neumann ERC, and measurement requirements are comparatively stringent ($\min_{C: |C|\le K} \|\Pi_{A_C^\perp}Y\|_F,$6, $\min_{C: |C|\le K} \|\Pi_{A_C^\perp}Y\|_F,$7 for random matrices).

With $\min_{C: |C|\le K} \|\Pi_{A_C^\perp}Y\|_F,$8, MBMP admits dictionaries where $\min_{C: |C|\le K} \|\Pi_{A_C^\perp}Y\|_F,$9, strictly subsuming TMP's guarantee. TMP and related single-path methods trade off assurance for lower computation; MBMP provides a controllable spectrum interpolating between the two extremes (Rossi et al., 2013).

Empirical studies confirm that MBMP delivers substantial performance gains, especially for moderate or large support sizes $\Pi_{A_C^\perp}=I-A_CA_C^\dagger$0, with exact recovery ratios and MSEs significantly exceeding single-path greedies, approaching those of exhaustive search or Oracle estimators when branching is sufficient (Lee et al., 2014).

6. Domain-Specific Implementations and Practical Guidelines

In practical deployments (e.g., MIMO radar, compressive imaging), MBMP's branch vector should be chosen to satisfy

$\Pi_{A_C^\perp}=I-A_CA_C^\dagger$1

where $\Pi_{A_C^\perp}=I-A_CA_C^\dagger$2 is the target dictionary's mutual coherence. Moderately increasing $\Pi_{A_C^\perp}=I-A_CA_C^\dagger$3 can yield substantial reductions in measurement requirements, justifying the computational overhead for hardware-constrained scenarios (Rossi et al., 2013).

The algorithm is most effective when tree depth is matched to sparsity $\Pi_{A_C^\perp}=I-A_CA_C^\dagger$4, and pruning schedules are set flat (constant $\Pi_{A_C^\perp}=I-A_CA_C^\dagger$5) except near the tree leaves, where $\Pi_{A_C^\perp}=I-A_CA_C^\dagger$6 can be reduced as rank-awareness facilitates unambiguous support completion.

In MMV settings with pronounced joint-sparsity, MBMP's rank-aware selection substantially improves the recovery threshold versus standard (rank-blind) algorithms.

7. Impact and Applications

MBMP's principal significance lies in its ability to interpolate between classical greedy and intractable combinatorial search regimes, enabling exact or near-exact sparse recovery from fewer measurements across a broad range of overcomplete dictionaries. Its flexible hardware–software trade-off is especially valuable in settings where data acquisition is expensive or limited, e.g., array processing, imaging, and compressive sensing. The development and theoretical validation of the MB-coherence framework constitute a substantial expansion of the allowable dictionary design space for sparse representation with rigorous guarantees (Rossi et al., 2013).