Source-linked AI summary
Frame bounds and bandwidth tiling for time-causal bandpass wavelets
Jens E. Pedersen, Tony Lindeberg, Peter Gerstoft
TL;DR
Time-causal wavelet filterbanks are suited to streaming, but their frame guarantees and conditioning have not been characterized. The paper derives closed-form frame and spectral analyses for DoG, DoT, and DoE wavelets and validates them numerically against recursive filters. DoT has B/A = 2.38 at c = 2, while DoE trades conditioning for a single integrator per channel.
Problem
Time-causal wavelet filterbanks for streaming lack characterized frame bounds, conditioning, and quality-factor analyses, properties relevant to stability under noise and quantization.
Source: §1 · p. 1, §5.1 · p. 4Method
Results
B/A = 2.38 at c = 2 for DoT, while DoE has B/A = 11.63 and uses a single integrator per channel.
Source: §6 · p. 4Takeaways & Limitations
The spectral characterization and frame geometries provide rules for choosing scale ratio c from a target band and channel count Kmax.
Source: §6 · p. 4
Abstract
from arXiv · showhide
Wavelets and frames that provide strong guarantees for signal representations are a cornerstone in signal processing theory, but have long been restricted to non-causal settings. Recent work extended wavelet analysis to time-causal systems that operate on streaming signals, with exact reconstruction and time-recursive implementations. However, the frame guarantees that make such representations reliable under noise and quantization are uncharacterized. We derive closed-form frame bounds and conditioning on band-limited domains, together with peak frequencies, constant-Q bandwidth tiling, and spectral decay rates for time-causal bandpass wavelets. Our work provides design rules for the scale ratio and channel count, which we validate numerically against discrete implementations of the causal wavelets in terms of recursive filters.
1. INTRODUCTION
Time-causal wavelet filterbanks support real-time streaming, but their frame bounds, conditioning, and quality-factor analyses remain uncharacterized. This work introduces DoE and analyzes frame properties, bandwidth tiling, and recursive-filter implementations for causal wavelets.
1. INTRODUCTION
- Time-causal bandpass filterbanks lack characterized frame bounds, conditioning, and quality-factor analyses despite their suitability for real-time applications.These properties are relevant to representation reliability under noise and quantization.Source: §1 · p. 1
- The paper introduces Difference-of-truncated-Exponentials (DoE) and derives closed-form frame bounds and conditioning for DoG, DoT, and DoE.It also analyzes bandwidth and constant-Q tiling, then numerically validates the closed forms against recursive-filter implementations.Source: §1 · p. 1
- The DoE is presented as a simplified case of the DoT wavelet.Source: §1 · p. 1
2. SCALE SPACES, WAVELETS, AND FRAMES
Scale spaces represent signals across smoothing scales, while wavelets form zero-mean scale-and-shift components. Frame bounds characterize coefficient capture and reconstruction stability.
2. SCALE SPACES, WAVELETS, AND FRAMES
- Scale-space theory represents a signal over scale by convolution with smoothing kernels, with scale covariance relating temporal rescaling to scale rescaling.For non-causal domains the Gaussian is the unique continuous-semigroup variation-diminishing kernel; causality admits the one-sided truncated exponential.Source: §2 · p. 1
- Geometrically spaced truncated-exponential stages build a time-causal scale space that is scale covariant for temporal scaling factors s = c^j.The scale ratio c is greater than one.Source: §2 · p. 1
- Wavelet coefficients represent scale-and-shift components, and admissibility entails a zero-mean wavelet.Source: §2 · p. 1
- A frame has lower and upper bounds A and B; its energy-capture function gives these bounds, and B/A governs conditioning and stability.The dual frame reconstructs signals exactly.Source: §2 · p. 1
3. SCALE-COVARIANT BANDPASS REPRESENTATIONS
Adjacent differences of scale-space representations produce scale-covariant bandpass channels that remove constant components. The section constructs DoG, DoT, and DoE wavelets and describes their reconstruction and causal behavior.
3. SCALE-COVARIANT BANDPASS REPRESENTATIONS
- Adjacent differences of scale-space representations create discrete, scale-covariant bandpass channels that remove constant components.Summing the bandpass channels with a coarsest-scale lowpass residual reconstructs the signal exactly.Source: §3 · p. 2
- The DoG is based on Gaussian differences, while DoT uses differences of time-causal limit kernels.The time-causal limit kernel restores variation diminution and scale covariance by cascading infinitely many truncated-exponential stages.Source: §3 · p. 2
- Truncating the DoT construction to N = 1 yields a single truncated-exponential term and the DoE wavelet.DoT and DoE are admissible because their smoothing kernels each integrate to one, so their difference integrates to zero.Source: §3 · p. 2
- Unlike the non-causal DoG, the DoT and DoE necessarily lag behind because their responses are causal.Source: §3 · p. 2
4. SPECTRAL CHARACTERIZATION
The section characterizes frame bounds, peak frequencies, −3 dB bandwidths, and spectral decay from squared magnitudes, and shows how geometric scale spacing yields constant-Q tiling across channels.
4. SPECTRAL CHARACTERIZATION
- Geometric scale spacing shifts each channel’s peak and −3 dB edges together, giving constant-Q bandwidths and a peak-frequency span of c^(K−1).The characterization reads bounds, peak frequency, bandwidth, and high-frequency decay from a squared magnitude.Source: §4 · p. 2
4.1. Difference-of-Gaussians
- For DoG, on [0, Ω], A = SDoG(Ω); on bounded bands its frame is overcomplete and non-tight, and B/A = 8.00 at c = 2 independently of K, with a closed-form peak.Source: §4.1 · p. 3
- DoG’s −3 dB edges are found numerically because they solve a difference of two Gaussians.The channel is a difference of Gaussians.Source: §4.1 · p. 3
4.2. Difference-of-time-causal-limit-kernels
- For the DoT’s N→∞ limit kernel at c = 2, on ω ∈ [0, 1/σ0), A = 0.419 and B/A = 2.38; both bounds are independent of K.Source: §4.2 · p. 3, Table 1 · p. 3
- The DoT N-fold cascade decays as O(ω^-N), or 6N dB/octave.Its peak is not available in simple closed form; a unique bounded solution determines its location.Source: §4.2 · p. 3
4.3. Difference-of-truncated-exponentials
- B/A → 11.66 as K → ∞ for DoE; at c = 2 and K = 12, it is a non-tight frame that does not converge exactly.Its lowpass term vanishes only as K → ∞.Source: §4.3 · p. 3, Table 1 · p. 3
4.3. Difference-of-truncated-Exponentials
- DoE peaks at the geometric mean of adjacent time constants, and its fractional bandwidth depends only on c.A single DoE channel decays as O(ω^-1), or 6 dB/octave.Source: §4.3 · p. 3
5. FRAME GEOMETRIES
The section relates frame quality to channel overlap and scale spacing, showing how scale ratio and cascade order trade frequency coverage against conditioning. It also gives constraints for choosing channel spacing and count in practical implementations.
5.1. Channel overlap and Gram matrix structure
- Tightness constrains aggregate energy capture but does not require orthogonal channels, so off-diagonal overlap may remain even in a tight frame.The normalized Gram matrix is the identity precisely when channels are orthogonal.Source: §5.1 · p. 4
- The Gram matrix captures interactions among channels, while its condition number measures frame conditioning; large values mean reconstruction is more sensitive to noise and quantization.The normalized Gram matrix’s off-diagonal entries reveal spectral overlap, and rapid off-diagonal decay corresponds to approximately independent information across distant channels.Source: §5.1 · p. 4
5.2. Choosing the distance between scales
- Increasing cascade order N sharpens the bandpass response and reduces adjacent-channel overlap, while increasing scale distance c reduces overlap at the cost of denser coverage.The caption compares energy capture and conditioning across N, from DoE at N = 1 toward the DoT limit.Source: §5.2 · p. 4, Figure 2 · p. 4
5.2.1. Closed form for c from the frequency range
- For a fixed target frequency range, geometric variance spacing and f ≈ 1/(2π√τ) determine c; increasing K drives c toward 1 and worsens conditioning through greater channel overlap.Equation (29) relates the channel count, frequency range, and scale spacing.Source: §5.2.1 · p. 4
5.2.2. Stability constraints on c
- Stability constraints can require reducing DoT cascade order N, while a floor c ≥ 1.05 keeps channels meaningfully distinct; realized conditioning interpolates monotonically from DoE toward the DoT limit as N increases.The stability requirement uses αfloor = 0.01; the resulting minimum c sets a maximum useful channel count for a given frequency range.Source: §5.2.2 · p. 4
6. CONCLUSION
The paper characterizes DoE, DoT, and non-causal DoG wavelets through frame bounds, conditioning, peak frequencies, bandwidth tiling, and spectral decay. DoT is best-conditioned, while DoE uses one integrator per channel in exchange for weaker conditioning; the analysis also supplies scale-ratio and channel-count design rules.
6. CONCLUSION
- DoT is best-conditioned (B/A = 2.38), while DoE trades conditioning (B/A = 11.63) for a single integrator per channel.DoT requires N cascaded filters per channel.Source: §6 · p. 4
- The paper characterizes DoE, DoT, and non-causal DoG in terms of frame bounds, conditioning, peak frequencies, constant-Q bandwidth tiling, and spectral decay.Source: §6 · p. 4
- The spectral and frame analyses provide rules for selecting scale ratio c from a target band and channel count Kmax.Source: §6 · p. 4