Binary Wavelet Codes for Fast Signal Reconstruction

• Binary wavelet codes are schemes that reconstruct signals from binary measurements using Walsh–Hadamard and compactly supported wavelet bases.
• They leverage dyadic structures to decompose dense matrix operations into FWHTs, diagonal scaling, and boundary corrections, achieving O(N log N) complexity.
• These methods are crucial for large-scale inverse problems like compressive sensing and medical imaging, enabling faster iterative reconstructions with reduced memory usage.

Binary wavelet codes refer to schemes and computational algorithms for efficiently reconstructing a signal from samples in a binary measurement basis—most notably, the Walsh–Hadamard basis—using wavelet functions for reconstruction. These codes and fast operator implementations exploit the special structure of Walsh–Hadamard transforms and compactly supported wavelets to enable efficient, accurate recovery of wavelet coefficients from binary data, which is pivotal for generalized sampling, compressive sensing (CS), and related inverse problems, particularly in high-dimensional settings such as medical imaging (Antun, 2021).

1. Foundations: Binary Sampling and Wavelet Bases

Binary wavelet coding arises in scenarios where signals are acquired by measuring their projections onto binary-valued function systems, typically the sequency-ordered Walsh functions $\{w_n(x)\}$. On $\mathbb{L}^2([0,1]^d)$, the $d$-dimensional sampling functions are tensor-products of one-dimensional Walsh functions. Measurements are then given by

$$y_{n_1,\dotsc,n_d} = \int_{[0,1]^d} f(x) \prod_{j=1}^d w_{n_j}(x_j)\,dx$$

for a function $f$. The discrete Walsh–Hadamard transform (FWHT), which acts on vectors by combining them with Walsh functions evaluated at dyadic grid points, enables these measurements to be computed in $\mathcal{O}(N\log N)$ time for grid size $N=2^L$ per axis, circumventing storage of the full $N\times N$ Hadamard matrix.

For reconstruction, compactly supported orthonormal wavelet bases are used. Each wavelet basis is generated from a scaling function $\phi$ and wavelet function $\psi$ with $\mathbb{L}^2([0,1]^d)$0 vanishing moments, and the basis at resolution level $\mathbb{L}^2([0,1]^d)$1 for $\mathbb{L}^2([0,1]^d)$2 uses translated and scaled copies:

$\mathbb{L}^2([0,1]^d)$3

where “rep” denotes handling of boundary conditions (periodic or vanishing-moments-preserving).

2. The Change-of-Basis Matrix and Fast Operator Structure

The crux of binary wavelet coding is the change-of-basis matrix $\mathbb{L}^2([0,1]^d)$4 between the Walsh–Hadamard sampling basis and the wavelet basis:

$\mathbb{L}^2([0,1]^d)$5

Computing matrix-vector products with $\mathbb{L}^2([0,1]^d)$6, with appropriate projections $\mathbb{L}^2([0,1]^d)$7, is central to all recovery algorithms but is naively an $\mathbb{L}^2([0,1]^d)$8 operation with prohibitive memory requirements for large-scale problems.

A pivotal result based on the dyadic-shift property of Walsh functions rewrites the interior (non-boundary) entries using a sparse sum involving the Walsh transforms of shifted scaling functions:

$\mathbb{L}^2([0,1]^d)$9

where $d$0 is the continuous Walsh transform. This characterization enables the decomposition of the change-of-basis multiplication into fast Walsh–Hadamard transforms, diagonal scaling, and minor boundary corrections.

3. Efficient Matrix–Vector Multiplication: Algorithm Design

For $d$1 and $d$2 (so $d$3 or $d$4 in $d$5D), the multiplication $d$6 for wavelet coefficient vector $d$7 is split into contributions from left boundary ($d$8), right boundary ($d$9), and middle columns ($$y_{n_1,\dotsc,n_d} = \int_{[0,1]^d} f(x) \prod_{j=1}^d w_{n_j}(x_j)\,dx$$0).

Middle Columns

For middle indices, the sum is expressed as:

$$y_{n_1,\dotsc,n_d} = \int_{[0,1]^d} f(x) \prod_{j=1}^d w_{n_j}(x_j)\,dx$$1

with

• $$y_{n_1,\dotsc,n_d} = \int_{[0,1]^d} f(x) \prod_{j=1}^d w_{n_j}(x_j)\,dx$$2 applies a length-$$y_{n_1,\dotsc,n_d} = \int_{[0,1]^d} f(x) \prod_{j=1}^d w_{n_j}(x_j)\,dx$$3 FWHT to a zero-padded and shifted $$y_{n_1,\dotsc,n_d} = \int_{[0,1]^d} f(x) \prod_{j=1}^d w_{n_j}(x_j)\,dx$$4,
• $$y_{n_1,\dotsc,n_d} = \int_{[0,1]^d} f(x) \prod_{j=1}^d w_{n_j}(x_j)\,dx$$5 is a diagonal scaling with Walsh-transform-derived weights.

Each $$y_{n_1,\dotsc,n_d} = \int_{[0,1]^d} f(x) \prod_{j=1}^d w_{n_j}(x_j)\,dx$$6 is computed via one FWHT, and $$y_{n_1,\dotsc,n_d} = \int_{[0,1]^d} f(x) \prod_{j=1}^d w_{n_j}(x_j)\,dx$$7 uses only $$y_{n_1,\dotsc,n_d} = \int_{[0,1]^d} f(x) \prod_{j=1}^d w_{n_j}(x_j)\,dx$$8 precomputed weights per $$y_{n_1,\dotsc,n_d} = \int_{[0,1]^d} f(x) \prod_{j=1}^d w_{n_j}(x_j)\,dx$$9. The computational cost is $f$0.

Boundary Columns

For boundary indices, the inner products reduce to short sums ($f$1 terms). The boundary contributions can be assembled directly in $f$2.

Complete Algorithm (1D) Structure

Step	Operation	Complexity
Middle terms	$f$3 FWHTs + diagonal scales	$f$4
Boundary terms	Direct, short inner products	$f$5
Adjoint multiply	Same, with FWHTs and sub-selection	$f$6

The algorithm generalizes to $f$7D by applying the 1D operator to columns, then rows, with total cost $f$8.

4. The Dyadic Regime, Stability, and Storage Analysis

Stable recovery necessitates a sampling–reconstruction rate $f$9 for integer $\mathcal{O}(N\log N)$0 (typically $\mathcal{O}(N\log N)$1 or $\mathcal{O}(N\log N)$2), ensuring the smallest singular value of $\mathcal{O}(N\log N)$3 remains bounded away from zero.

Storage demands are dominated by the precomputed diagonal-weight arrays $\mathcal{O}(N\log N)$4, which depend only on the finite set $\mathcal{O}(N\log N)$5 and the small set of shifts $\mathcal{O}(N\log N)$6. Hence, storage is $\mathcal{O}(N\log N)$7, independent of $\mathcal{O}(N\log N)$8 and $\mathcal{O}(N\log N)$9.

5. Integration into Signal Reconstruction Frameworks

Fast Walsh–wavelet change-of-basis multiplication is crucial for iterative algorithms underlying generalized sampling, infinite-dimensional compressive sensing, and parameterised-background data-weak (PBDW) reconstruction. These methods typically involve repeated matrix–vector and adjoint–vector operations with $N=2^L$0. By replacing direct $N=2^L$1 multiplication with the $N=2^L$2 algorithm, significant reductions in both runtime and memory requirements are realized.

In CS, wavelet coefficients are often the object of $N=2^L$3-minimization or similar pursuits; pre- or post-application of the discrete wavelet transform (DWT) carries an added $N=2^L$4 cost, negligible relative to the main operator.

Implementation must address:

• Accurate treatment of boundary wavelet inner products (selection of periodic vs. vanishing-moment-preserving formulation).
• The moderate computational overhead of up to $N=2^L$5 FWHTs per multiplication (with $N=2^L$6 up to $N=2^L$7 practical).
• The scheme is inherently restricted to dyadic sizes; for non-dyadic or irregular settings, generalizations such as interpolation or alternative shift lemma formulations are required.

6. Limitations and Precision Considerations

When the sampling grid or reconstruction size is non-dyadic, direct application of the presented fast transform is not possible—alternative approaches must be used. Finite-precision and numerical stability are determined by the subspace angle between the sampling and reconstruction spaces. Sufficiently large $N=2^L$8 ensures the stability of the operator, specifically that the smallest singular value of $N=2^L$9 does not approach zero.

A plausible implication is that boundaries, grid-size restrictions, and the choice of $N\times N$0 are critical for guaranteeing effective, robust binary wavelet coding in practical scenarios.

7. Summary and Practical Significance

By exploiting the dyadic structure of Walsh–Hadamard functions and the locality of compactly supported wavelets, binary wavelet coding reorganizes dense matrix multiplications into a small number of FWHTs, diagonal scaling, and direct boundary corrections. This structure facilitates $N\times N$1 forward and adjoint operators, enabling high-efficiency wavelet recovery from binary samples on very large problems in one and two dimensions. These advancements directly impact medical imaging and related fields, where rapid and accurate signal reconstruction from binary data is essential (Antun, 2021).