K-AHS: Adaptive Hierarchical Sensing Method

• K-AHS is an adaptive sensing framework that organizes measurements in a binary tree to efficiently capture sparse or compressible signals in a known transform domain.
• It dynamically selects the K largest measurements at each level, eliminating the need for measurement incoherence or restricted isometry properties found in traditional compressed sensing.
• The method achieves a measurement complexity of O(K log(N/K)) with improved recovery performance in both synthetic and real-world experiments, offering a practical alternative to convex optimization approaches.

K-Adaptive Hierarchical Sensing (K-AHS) is an adaptive sampling framework designed for efficiently acquiring signals that are sparse or compressible within a known transform domain. The fundamental innovation of K-AHS is its online adaptation: the selection of each measurement vector is dynamically conditioned on previous measurements, leveraging a hierarchical structure to locate and directly measure the significant coefficients of the signal in the sparse domain. K-AHS achieves a measurement complexity comparable to traditional compressed sensing (CS), but operates without the need for measurement incoherence or restricted isometry, and typically eliminates the requirement for an inverse optimization step at reconstruction (Schütze et al., 2018).

1. Algorithmic Structure and Sensing Tree Construction

K-AHS operates on signals $x\in \mathbb{R}^N$ assumed to be sparse or compressible in a predetermined orthonormal basis $\Psi\in\mathbb{R}^{N\times N}$, producing transform coefficients $a = \Psi x$ with $K\ll N$ dominant entries. The algorithm organizes a set of linear sensing vectors into a perfect binary tree of height $L_{\max} = \lceil \log_2 N \rceil$. Each node $(\ell, n)$ at tree level $\ell$ ($0 \leq \ell \leq L_{\max}$) defines a vector $\Phi_{\ell, n} \in \mathbb{R}^N$ recursively:

• Leaves ($\ell=0$): $\Psi\in\mathbb{R}^{N\times N}$0, where $\Psi\in\mathbb{R}^{N\times N}$1.
• Internal nodes ($\Psi\in\mathbb{R}^{N\times N}$2): $\Psi\in\mathbb{R}^{N\times N}$3.

A measurement at node $\Psi\in\mathbb{R}^{N\times N}$4 is $\Psi\in\mathbb{R}^{N\times N}$5, representing the sum of a contiguous block of $\Psi\in\mathbb{R}^{N\times N}$6 coefficients.

The initial sampling commences at level $\Psi\in\mathbb{R}^{N\times N}$7, guaranteeing the level contains $\Psi\in\mathbb{R}^{N\times N}$8 nodes. At each level, the algorithm selects the $\Psi\in\mathbb{R}^{N\times N}$9 nodes with the largest magnitude $a = \Psi x$0 and recursively expands them to their respective children until the leaves are reached. Ultimately, $a = \Psi x$1 coefficients are directly sampled at the leaf level and are used to reconstruct an estimate $a = \Psi x$2, with all non-measured coefficients set to zero. The final signal estimate is formed by inverse transforming: $a = \Psi x$3.

Pseudocode Summary:

$\ell=0$0

2. Measurement Complexity

The total number of measurements $a = \Psi x$4 performed by K-AHS is given by

$a = \Psi x$5

where $a = \Psi x$6 measurements are acquired at the initial coarse sampling, and $a = \Psi x$7 correspond to expansions at each subsequent level. With $a = \Psi x$8 and $a = \Psi x$9, this yields

$K\ll N$0

(up to lower order terms), matching the asymptotic $K\ll N$1 measurement complexity of nonadaptive $K\ll N$2-based compressed sensing schemes.

3. Theoretical Guarantees

Sampling Complexity

For any $K\ll N$3, K-AHS ensures

$K\ll N$4

by construction.

Sufficient Condition for Recovery of Largest Coefficients

Let $K\ll N$5 denote indices of the $K\ll N$6 largest-magnitude coefficients. For the sums at level $K\ll N$7,

$K\ll N$8

where $K\ll N$9 are sorted by magnitude. If $L_{\max} = \lceil \log_2 N \rceil$0, then each member of $L_{\max} = \lceil \log_2 N \rceil$1 is contained in a “winning” sum at level $L_{\max} = \lceil \log_2 N \rceil$2 and propagated to the leaf level.

Specific model insights:

• Exactly $L_{\max} = \lceil \log_2 N \rceil$3-sparse: $L_{\max} = \lceil \log_2 N \rceil$4, implying perfect recovery with high probability.
• Exponential decay $L_{\max} = \lceil \log_2 N \rceil$5: For $L_{\max} = \lceil \log_2 N \rceil$6, $L_{\max} = \lceil \log_2 N \rceil$7 ensures selection of the top $L_{\max} = \lceil \log_2 N \rceil$8 coefficients.
• Power-law decay $L_{\max} = \lceil \log_2 N \rceil$9: Recovery of $(\ell, n)$0 is guaranteed for $(\ell, n)$1 such that $(\ell, n)$2.

This sufficient (but not necessary) condition provides formal guarantees for accurate identification of dominant coefficients in diverse decay scenarios.

4. Adaptivity, Transform Domain, and Relaxed Requirements

K-AHS determines each set of measurements adaptively: at each level, only the children of subregions corresponding to the largest previously measured values are sampled. The measurement strategy is inherently data-driven rather than predetermined.

A critical distinction from traditional CS is that K-AHS imposes no requirement for incoherence between sampling vectors and the sparsifying basis, nor for satisfaction of the Restricted Isometry Property (RIP). The only criterion is that the basis $(\ell, n)$3 renders the signal sparse or compressible. Final reconstruction is direct, without the need for $(\ell, n)$4-minimization or convex optimization.

For natural signals exhibiting heavy-tailed coefficient distributions, the probability of significant destructive cancellation within the sensing tree is minimal, so dominant terms are unlikely to be eliminated during the adaptive descent.

5. Empirical Results and Performance

Synthetic Signals

• $(\ell, n)$5-Sparse ($(\ell, n)$6): For $(\ell, n)$7, perfect ($(\ell, n)$8) detection of the $(\ell, n)$9 largest coefficients when $\ell$0; for $\ell$1, detection rates degrade gracefully (approx.\ $\ell$2 for $\ell$3).
• Exponential decay ($\ell$4): K-AHS accurately identifies top $\ell$5 terms for $\ell$6 in line with theory.
• Power law ($\ell$7): Near-perfect detection of the largest coefficient for all tested $\ell$8, with rapidly increasing captured signal energy as $\ell$9 increases.

Real-World Images

Image experiments (e.g., $0 \leq \ell \leq L_{\max}$0 grayscale images) employed both orthogonal Haar and biorthogonal CDF97 wavelets as transforms. Performance metrics compared K-AHS to baseline CS (using noiselet measurements and $0 \leq \ell \leq L_{\max}$1 recovery via the NESTA solver):

• Rate-Distortion: For $0 \leq \ell \leq L_{\max}$2, K-AHS exceeded $0 \leq \ell \leq L_{\max}$3-CS by $0 \leq \ell \leq L_{\max}$4--$0 \leq \ell \leq L_{\max}$5 dB in PSNR across both wavelet domains.
• Example ($0 \leq \ell \leq L_{\max}$6 on “Cameraman”):
  • Haar: CS: $0 \leq \ell \leq L_{\max}$7 dB, K-AHS: $0 \leq \ell \leq L_{\max}$8 dB.
  • CDF97: CS: $0 \leq \ell \leq L_{\max}$9 dB, K-AHS: $\Phi_{\ell, n} \in \mathbb{R}^N$0 dB.
• Visual Observations: Reconstructions via K-AHS generally presented reduced high-frequency artifacts and better restoration of smooth regions, with occasional coarser blocks and mild ringing.
• Spatial Sensing Maps: Aggregating the sum of absolute entries of selected “winner” vectors at each level, K-AHS initially emphasizes large-scale regions, focusing on finer structural details (edges, corners, texture) as is recursively descends.

6. Limitations and Future Research Directions

K-AHS incurs a trade-off between memory and computation: full storage of all $\Phi_{\ell, n} \in \mathbb{R}^N$1 tree vectors demands $\Phi_{\ell, n} \in \mathbb{R}^N$2 memory. In practice, this can be mitigated by computing $\Phi_{\ell, n} \in \mathbb{R}^N$3 on the fly via sparse indicator transforms, with cost $\Phi_{\ell, n} \in \mathbb{R}^N$4 per vector if $\Phi_{\ell, n} \in \mathbb{R}^N$5 is a fast transform (e.g., wavelets, DCT).

The sufficient recovery condition does not preclude loss of significant coefficients in edge cases where large values cancel due to sign anti-alignment. Theoretical analysis covers only sufficient, not necessary, recovery guarantees.

Performance is also contingent on the suitability of the chosen basis $\Phi_{\ell, n} \in \mathbb{R}^N$6—effectiveness depends on the degree to which $\Phi_{\ell, n} \in \mathbb{R}^N$7 sparsifies the signal. Future research directions outlined in (Schütze et al., 2018) include:

• Learning optimal node weights in tree sum combinations (i.e., $\Phi_{\ell, n} \in \mathbb{R}^N$8) to reduce cancellation.
• Extending the methodology to undercomplete ($\Phi_{\ell, n} \in \mathbb{R}^N$9) or redundant dictionaries.
• Developing hybrid schemes that integrate K-AHS adaptivity with convex reconstruction of residuals.

K-AHS thus offers a rigorous, algorithmically simple alternative to traditional compressed sensing, achieving similar or improved measurement efficiency and practical performance, particularly in settings where signal structure is known a priori and large-scale convex reconstruction is undesired or impractical (Schütze et al., 2018).