Encoding and Decoding Temporal Signals with Spiking Bandpass Wavelets

arXiv:2605.09770 · cs.NE, eess.SP, q-bio.NC · Submitted 2026-05-10 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Genomics Radio. Generated commentary on the latest computational biology and genomics papers.

Ines: I'm Ines, and with me are Marcus and Yuki, guest researcher.

Marcus: Today's paper: "Encoding and Decoding Temporal Signals with Spiking Bandpass Wavelets".

Ines: Spike-based encodings are sparse and energy-efficient, but have largely been formulated probabilistically, disconnected from most signal processing literature.

Marcus: First, who's behind it and why it matters.

Paper summary: Ines: So, we're looking at "Encoding and Decoding Temporal Signals with Spiking Bandpass Wavelets," and I want to start by giving you the main idea. The paper tackles how we can take these sparse, energy-efficient spike encodings and fit them into a proper signal processing framework. It claims they can be recast as time-causal wavelet frames that have measurable bandwidths and specific error bounds for reconstruction <ref:2605.09770#pg0>.

Marcus: From my side, what I’m hearing is that the paper is trying to bridge the gap between those probabilistic spike models and established signal processing literature, which has been a real hurdle for applying them widely in actual data science workflows <ref:2605.09770#pg1>. It suggests these wavelets have reconstruction capabilities up to spike quantization and time discretization, which is pretty significant.

Yuki: And from a population genetics standpoint, the fact that they preserve the sparsity and locality of spiking representations seems important because it mirrors how biological signals are often represented in a very localized manner <ref:2605.09770#pg1>. It suggests this mathematical structure might naturally capture certain aspects of temporal patterns found across different biological scales.

Ines: Exactly, Yuki, and that mapping to neuromorphic hardware is what really excites me; it suggests we have a principled way to do Analog-to-Digital Conversion that directly translates to those physical chips <ref:2605.09770#pg0>. It moves away from just having a probabilistic model toward something with reconstruction guarantees.

Marcus: I agree, Ines, and the paper’s focus on providing quantitative bandwidths for these wavelets is what gives us the necessary metrics to compare them against existing continuous wavelet transforms <ref:2605.09770#pg0>. It moves it from being just a conceptual idea to something we can actually test with concrete performance numbers.

Yuki: I wonder how this mathematical framing might connect to broader evolutionary patterns in species, if we consider these temporal representations as a way of organizing information across different temporal scales within a population <ref:2605.09770#pg2>. It hints at a universal way of structuring signals that might be relevant across diverse biological systems.

Ines: That's the big picture, Yuki; it suggests that the mathematical structure itself has inherent properties that relate to how biological information is organized across different time resolutions <ref:2605.09770#pg2>. We’re moving beyond just observing spikes to understanding the underlying signal processing structure they form.

Marcus: And when we look at the methodology, they introduce two specific families of spiking wavelets, the Difference of Truncated Exponentials and the Difference of Time-causal Limit Kernels <ref:2605.09770#pg1>. This is where I start to think about how robust these encoding schemes are when dealing with noisy, real-world genomics data or any complex signal.

Paper summary: Yuki: That distinction between the DoE and DoT families tells me a lot about the complexity they’re trying to capture; one seems more focused on differences between discrete time steps, while the other involves cascading limit kernels <ref:2605.09770#pg1>. This suggests they are targeting different kinds of temporal structures in the underlying signals.

Ines: And when we get to the reconstruction guarantees, they show that for the DoT wavelet, it forms an overcomplete, non-tight frame with bounds A > zero and B = one <ref:2605.09770#pg2>. That’s a solid mathematical property because it means reconstruction is possible even when you have more channels than necessary for the signal itself.

Marcus: An overcomplete frame with a specific bound, like B=one is very useful statistically; it gives us a clear boundary on how much noise we can expect during the reconstruction process <ref:2605.09770#pg2>. The total reconstruction error they bound is linear in the spike threshold theta thr, which is a crucial piece of information for real-time applications.

Yuki: That linearity in the spike threshold theta thr is compelling because it implies that improving the sensitivity of our detection system directly translates to better accuracy in reconstructing the original signal, which feels like a very practical link to biological measurement itself <ref:2605.09770#pg2>.

Ines: It really does suggest that the underlying mechanism—the spike thresholding—is a fundamental part of the reconstruction fidelity, not just an external parameter we tune <ref:2605.09770#pg1>. This ties it back to how neurons process input and generate output spikes in a very direct way.

Marcus: And the experimental validation is where things get interesting; they tested these spiking DoT wavelets on ECG and audio datasets, showing performance at or below the Morlet baseline <ref:2605.09770#pg2>. Achieving zero point zero five eight versus Morlet’s zero point zero six four on MIT-BIH is a tangible result that shows they are competitive without sacrificing the causality requirement <ref:2605.09770#pg2>.

Yuki: Comparing these results to established continuous transforms like the Morlet wavelet gives context to how effective this spiking approach is at capturing temporal features in complex biological data <ref:2605.09770#pg2>. It shows that this discrete, event-driven approach can achieve similar fidelity as traditional methods when applied correctly.

Ines: That comparison to Morlet really grounds the theoretical work in something measurable, showing that the spiking wavelets aren't just mathematically sound but functionally effective for temporal signal analysis <ref:2605.09770#pg2>. This validates the entire premise of recasting spike encoders as a proper signal processing tool.

Marcus: From a data science perspective, I’m paying attention to how they handle the channel overlap characterization using the Gram matrix G jk; increasing that cascade order n sharpens the bandpass response <ref:2605.09770#pg2>. That suggests we have control over how distinct those spectral channels are, which is vital when dealing with overlapping biological signals or batch effects in cohort analysis.

Yuki: Controlling that overlap through the cascade order n seems like a powerful way to tailor the representation to the specific temporal characteristics we’re interested in within a population study <ref:2605.09770#pg2>. It gives us a knob for adjusting our sensitivity to finer temporal details.

Paper summary: Ines: So, it seems like the core message of "Encoding and Decoding Temporal Signals with Spiking Bandpass Wavelets" is that we can build a principled, hardware-friendly framework for spike encoding by treating them as time-causal wavelet frames <ref:2605.09770#pg0>. It’s not just a probabilistic shortcut; it has concrete reconstruction bounds and works on real data.

Marcus: And the implications for hardware are significant because the computation is entirely time-recursive with no need for external state, mapping directly to neuromorphic primitives <ref:2605.09770#pg1>. This makes it a very attractive building block for developing new types of analog-to-digital conversion systems.

Yuki: Considering the broader history of signal processing, this work suggests that the fundamental mathematics underpinning scale-space theory can be successfully adapted to model sparse, event-driven systems like neural firing <ref:2605.09770#pg2>. It broadens the context for how we view temporal information across different scientific domains.

Ines: I think the title itself, "Encoding and Decoding Temporal Signals with Spiking Bandpass Wavelets," captures the essence well because it clearly states what they are doing: moving from probabilistic descriptions to a frame-based, causal representation of time <ref:2605.09770#pg0>.

Marcus: The authors’ work provides a concrete way to quantify reconstruction error in spike-based ADC systems using these wavelets, which is something I need when I'm trying to assess the reliability of any encoding method for large-scale genomic data <ref:2605.09770#pg2>.

Yuki: If we look at the long term, this suggests a new way to analyze temporal organization in biological sequences, perhaps offering a more structured lens than purely statistical approaches <ref:2605.09770#pg2>. It points toward a deeper understanding of how complex temporal patterns emerge from simpler spike events.

Ines: So, to wrap up this discussion on "Encoding and Decoding Temporal Signals with Spiking Bandpass Wavelets," the main thrust is providing a rigorous mathematical foundation for spike encoders by treating them as time-causal wavelet frames with defined reconstruction quality metrics <ref:2605.09770#pg0>.

Marcus: The paper’s contribution lies in showing that these wavelets work empirically on ECG and audio, achieving results comparable to continuous transforms while maintaining the sparsity and causality required for neuromorphic mapping <ref:2605.09770#pg2>.

Yuki: It offers a new mathematical language for discussing temporal signal representation in systems where information is inherently sparse, which could have implications across many fields beyond just neural modeling <ref:2605.09770#pg1>.

Ines: And the practical implication is that this provides a "principled building block for neuromorphic analog-to-digital conversion" by showing how to map these spike representations directly onto hardware primitives <ref:2605.09770#pg1>.

Marcus: Ultimately, this work helps us understand the quantitative limits of reconstructing continuous signals from sparse spike events, which is a necessary step for applying these methods reliably in data science contexts <ref:2605.09770#pg2>.

Conclusion: Ines: So, to wrap up this discussion on "Encoding and Decoding Temporal Signals with Spiking Bandpass Wavelets," the core idea is that they’ve successfully framed spike encodings as time-causal wavelet frames, giving us concrete math for reconstruction error bounds.

Marcus: Exactly, Ines; I think the real takeaway is how they’ve given a rigorous statistical foundation to something that used to feel pretty ad-hoc in neuroscience modeling.

Yuki: From my perspective, it suggests a universal mathematical structure for organizing temporal information across different biological scales, which could help us see deeper into population dynamics.

Ines: And when you look at the authors' work, they’ve really managed to tie this abstract signal processing theory directly to physical constraints of neuromorphic hardware.

Marcus: Right; that direct mapping to primitives is what makes this paper so relevant for anyone working on efficient analog-to-digital conversion systems.

Yuki: It opens up a new way to think about how temporal patterns emerge in complex biological sequences, moving beyond just statistical correlation toward structural representation.

Ines: That’s the big picture; it shifts the focus from just observing spikes to understanding the underlying signal processing architecture they form.

Marcus: And when we consider the implications for genomics data, this provides a quantifiable way to assess how much fidelity we lose when compressing complex temporal signals into sparse spike events.

Yuki: It really frames our understanding of temporal organization in a way that connects microscopic neural activity to macroscopic evolutionary patterns across species.

Ines: So, the authors have essentially provided the blueprint for turning sparse spiking representations into a reliable tool for signal reconstruction in real-world applications.

Marcus: And that blueprint is incredibly useful because it includes explicit error bounds, which is something we desperately need when dealing with noisy cohort data and batch effects.

Yuki: This work suggests a new mathematical language for discussing temporal information in systems where the data is inherently sparse, which could have implications across many fields beyond just neural modeling.

Ines: We're definitely going to keep digging into how these specific wavelet families perform against known continuous transforms next.

Jens E. Pedersen, Tony Lindeberg, Peter Gerstoft

Technical University of Denmark · KTH Royal Institute of Technology

cs.NE, eess.SP, q-bio.NC

Submitted: 2026-05-10

Updated: 2026-10-02

Code: https://github.com/jegp/swavelet

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 90/100

The gist: Spike-based encodings are sparse and energy-efficient, but have largely been formulated probabilistically, disconnected from most signal processing literature.

Key concepts

Scale-space theory
This theory parameterizes a signal across different scales using smoothing kernels. For time-causal operations, it uses the hexp kernel to discretize the scale parameter. This allows researchers to represent a signal as a function of both its amplitude and temporal resolution.
Spiking Wavelet Families (DoE/DoT)
These are two types of wavelets constructed by taking adjacent differences between scale-space representations. The Difference of Truncated Exponentials (DoE) uses kernel differences, while the Difference of Time-causal Limit Kernels (DoT) uses differences between cascaded hexp stages, creating bandpass representations.
Frame Bounds and Reconstruction Error
The analysis establishes mathematical limits on how well the wavelets can reconstruct the original signal. The error is shown to be linear with respect to the spike threshold ($ heta_{thr}$), meaning lower thresholds lead to more accurate reconstructions, which is crucial for practical application.

Terminology

Summary

Spike-based encodings are sparse and energy-efficient, but have largely been formulated probabilistically, disconnected from most signal processing literature. This work recasts spike encoders as time-causal wavelet frames with quantitative bandwidths and reconstruction error bounds, providing a principled framework for spike-based Analog-to-Digital Conversion that maps directly to neuromorphic hardware.

The gist

The proposed wavelets preserve the sparsity and locality of spiking representations, with reconstruction up to spike quantization and time discretization.

Theoretical Foundations of Scale Spaces

Scale-space theory parameterizes a signal over a scale parameter σ into a representation L(t; σ) via convolution with smoothing kernels h(t; σ). For time-causal operations, the scale parameter must be discretized using the truncated exponential kernel, hexp(t, µ), which is based on underlying theoretical results by Schoenberg. Time-causal scale spaces are constructed by cascading K such kernels according to a geometric progression of time constants: µk = c µk−1. The input current implements the time-causal scale-space convolution L(t; µ) = (hexp(·; µ) ∗ f)(t).

Spiking Wavelet Families

The paper introduces two families of spiking time-causal scale covariant wavelets:

  1. Difference of Truncated Exponentials (DoE), based on the kernel ψDoE(t; µk, c) = hexp(t; µk) − hexp(t; µk−1).

  2. Difference of Time-causal Limit Kernels (DoT), based on the kernel ψDoT(t; σk, c) = hΨ(t; σk, c) − hΨ(t; σk−1, c), where hΨ is a cascade of hexp stages.

These wavelets are constructed by taking adjacent differences of scale-space representations: ∆L(t; σk, c) = L(t; σk, c) − L(t; σk−1, c), which yields the bandpass representation ψ(t; σk, c) ∗ f(t).

Frame Bounds and Reconstruction Guarantees

The analysis establishes frame bounds for both wavelets. For the DoT wavelet, it is shown to form an overcomplete, non-tight frame with bounds A > 0 and B 0 and B = 1. The reconstruction follows a dual frame reconstruction scheme using the bandpass decomposition (14). The total reconstruction error is bounded by:

∥f − fe∥∞ ≤ C θthromegaK + X K k=1omegak, where the error is linear in the spike threshold θthr. In the limit case when K → ∞, this converges to ∥f − fe∥∞ ≤ C θthr omega1c(c−1).

Experimental Validation and Performance

Reconstruction experiments on ECG (MIT-BIH) and audio (LibriSpeech) datasets demonstrate that the spiking wavelets perform at or below the Morlet baseline. Specifically, the spiking DoT yielded 0.058 vs Morlet’s 0.060 on MIT-BIH and 0.064 vs LibriSpeech’s 0.12 on LibriSpeech, showing competitive performance while maintaining causality and sparsity. The reconstruction error is also shown to be linear in the spike threshold θthr, as demonstrated by Figure 8.

Hardware Mapping and Implementation

A key advantage is that the computation is entirely time-recursive and no external history or state is required for filter states, enabling real-time processing with bounded memory. The implementation maps directly onto primitives common to neuromorphic platforms because every operation in the spiking wavelets reduces to a leaky-integrator difference with spike thresholding. This positioning makes spiking wavelets a principled building block for neuromorphic analog-to-digital conversion.

Bandwidth and Channel Overlap

The geometric scale spread ensures that the exact expression is independent of individual scales, covering a total bandwidth ratio of c(K−1). The quality factor Q = ωpeak / ∆ω is constant for the DoE wavelet across all channels. For the DoT wavelet, while an exact peak requires numerical evaluation due to phase obstruction, an upper bound on the peak frequency is established as ωDoT peak ≤ 1/µ1. The channel overlap is characterized by the Gram matrix Gjk; increasing cascade order n sharpens the bandpass response and provides a more distinct banded structure.

Stability Constraints

Practical implementation requires stability constraints on the scale ratio c and cascade order n. Numerical leaky-integrator stability imposes a minimum time constant µ at every stage, translating into a minimum time step ∆τ at the finest stage.

Improvements for AI systems

As a fastidious and diligent researcher, I have analyzed this paper, Encoding and Decoding Temporal Signals with Spiking Bandpass Wavelets, which proposes a novel framework for mapping continuous temporal signals onto sparse, event-driven spike trains suitable for neuromorphic hardware.

The core contribution is the development of two time-causal, scale-covariant wavelet families—Difference of Gaussian (DoG), Difference of Time-causal Limit Kernel (DoT), and Difference of Truncated Exponential (DoE)—that provide formal frame bounds and reconstruction error guarantees.

Here are the specific improvements that can be made to AI systems based on this research:

)1. Implementation in Neuromorphic Hardware for Low-Power, Real-Time Processing

The proposed system maps directly to neuromorphic hardware primitives (leaky integrators and spike thresholds). The improved AI system would move beyond current software/GPU implementations for signal processing by deploying the encoding and decoding logic directly onto spiking neural networks (SNNs) or specialized neuromorphic chips.

  • Specific Improvement: Replace traditional Digital Signal Processing (DSP) pipelines (like FFT, standard filtering, or complex wavelet transforms) with the proposed Spiking Wavelet Encoding and Reconstruction algorithm.

  • System Capability: Enables ultra-low latency, event-driven processing where computation only occurs upon a spike. This drastically reduces energy consumption compared to continuous analog or high-frequency digital sampling systems (e.g., for real-time sensor fusion, edge AI in robotics).

)2. Robust Signal Analysis and Feature Extraction from Sparse Data

The paper demonstrates that these spiking wavelets can reconstruct signals like ECG and audio with Normalized RMSE comparable to classical transforms, even under quantization noise (linear in the threshold parameter).

  • Specific Improvement: Develop Spiking Feature Extractors for time-series data. Instead of analyzing raw continuous waveforms, the AI system would use the spike trains as its primary feature representation.

  • System Capability: Improved classification and anomaly detection in noisy, sparse data streams (e.g., monitoring physiological signals like ECG or detecting subtle acoustic events). The scale-covariant nature ensures that the extracted features are robust across different temporal resolutions.

)3. Guaranteed Stability and Error Bounding for Quantized Encoding

The paper derives closed-form reconstruction error bounds (Eq. 32) that are linear in the spike threshold and independent of the number of channels for large K (scale levels).

  • Specific Improvement: Implement a Guaranteed Accuracy Layer in the AI pipeline. This layer would dynamically adjust the required spike threshold or channel count based on a desired tolerance level, using the derived error bounds to ensure acceptable reconstruction quality.

  • System Capability: Critical for safety-critical applications (e.g., medical diagnostics). It allows engineers to guarantee that a specific level of signal fidelity is achieved regardless of minor variations in hardware quantization or input noise.

)4. Adaptive Bandwidth and Frequency Resolution Control

The analysis shows that the quality factor Q scales predictably with the scale ratio 'c' and cascade order 'n', allowing for explicit control over frequency selectivity.

  • Specific Improvement: Create a Dynamic Spectral Filter module within the AI architecture that can adjust its operational parameters (like channel selection or cascade depth) based on the current signal content.

  • System Capability: The system could intelligently switch between high-resolution (dense scale sampling, low 'c') modes for transient events and lower-resolution (coarser scale sampling, high 'c') modes for long-term trend monitoring, optimizing both frequency coverage and computational load in real time.

)5. Neuromorphic Mapping via Spiking Scale Covariance

The system proves that the LIF model implemented by the wavelet is provably scale covariant, meaning temporal scaling of the input signal results in a predictable scaling of the representation.

  • Specific Improvement: Design AI architectures where temporal relationships are inherently learned through this covariance property rather than being explicitly programmed.

  • System Capability: Development of Self-Scaling Temporal Models. The AI would naturally learn how to interpret signals regardless of their sampling rate or temporal stretching, making it highly adaptable to varying data acquisition rates without retraining on new sampling schemes.

Abstract

Spike-based encodings are sparse and energy-efficient. However, previous theory for such spiking representations is largely disconnected from most signal processing literature and does not extend to multi-scale frames. We recast spike-encoders as time-causal overcomplete wavelet frames with closed-form bandwidth tiling and reconstruction error bounds. We show that scale covariance, frame bounds, and reconstruction guarantees hold for any finite-energy input in both encoding and decoding steps. The proposed spiking bandpass wavelets preserve the sparsity and locality of spiking representations: A threshold parameter sets the event rate, and the reconstruction error grows linearly with it, up to spike quantization and time discretization. On ECG and audio datasets, we reach reconstruction accuracy comparable to continuous wavelet transforms, multi-channel integrate-and-fire time encoding machines, and Sigma-Delta modulation. The experiments demonstrate robustness of the proposed spiking bandpass wavelets to noise and spike-time jitter, and that the spike activity shifts across channels as predicted by the covariance property under temporal rescaling. Every operation reduces to a leaky integrate-and-fire difference with thresholding, and both the encoder and the decoder map directly to existing neuromorphic hardware.

Sources

Related papers