Adaptive Spectral Block (ASB) in ML & THz

• Adaptive Spectral Block (ASB) is a method for dynamic spectral partitioning that improves deep learning and THz communication by applying adaptive thresholding and filtering.
• In deep learning, ASB leverages FFT-based transformation, learnable thresholds, and dual spectral filters to integrate global context with local noise mitigation.
• In THz spectrum allocation, ASB optimizes variable sub-band widths through convex optimization, boosting throughput under frequency-dependent loss conditions.

The term “Adaptive Spectral Block” (ASB) denotes two distinct but rigorous concepts in contemporary research: one in deep time series representation learning (Eldele et al., 2024) and another in terahertz (THz) communication spectrum allocation (Shafie et al., 2021). In the context of TSLANet, the ASB is a network module leveraging learnable frequency-domain thresholding and filtering to enhance robustness and long-range pattern extraction in time series. In THz communications, ASB refers to flexible, optimization-driven assignment of variable-width frequency sub-bands to maximize system throughput under severe frequency-dependent loss. Both concepts employ adaptive partitioning or selection in the spectral domain, but with distinct mathematical instantiations and engineering objectives.

1. Mathematical Foundation of the Adaptive Spectral Block in Deep Learning

In TSLANet, the Adaptive Spectral Block is constructed around the Discrete Fourier Transform (DFT) of input embeddings. For a multichannel patch embedding $\bm{S}_{\mathrm{PE}} \in \mathbb{R}^{C \times p'}$, a 1D FFT is applied along the length dimension to obtain the spectrum: $$\bm{F} = \mathcal{F}[\bm{S}_{\mathrm{PE}}] \in \mathbb{C}^{C \times L'}.$$ The power spectrum $\bm{P} = |\bm{F}|^2$ is computed per frequency. An adaptive, learnable threshold $\theta$ is used to generate a binary mask $\bm{M} = (\bm{P} > \theta)$, yielding a filtered spectrum $\bm{F}_{\mathrm{filtered}} = \bm{F} \odot \bm{M}$. Two spectral filters, $\bm{W}_G$ (global) and $\bm{W}_L$ (local), are learned and multiplicatively applied: $$\bm{F}_G = \bm{W}_G \odot \bm{F}, \quad \bm{F}_L = \bm{W}_L \odot \bm{F}_{\mathrm{filtered}}.$$ The outputs combine as $\bm{F}_{\mathrm{integrated}} = \bm{F}_G + \bm{F}_L$, with the enhanced time-domain representation recovered by inverse FFT: $$\bm{S}' = \mathcal{F}^{-1}[\bm{F}_{\mathrm{integrated}}].$$ This structure allows the ASB to integrate full-spectrum (long-range) and adaptively masked (local, noise-suppressed) spectral features (Eldele et al., 2024).

2. Data Flow and Algorithmic Structure of the ASB

The ASB operates according to the following sequential procedure:

• Input: patch embedding $\bm{S}_{\mathrm{PE}}$
• Spectrum extraction: FFT yields $\bm{F}$
• Power analysis: compute $\bm{P} = |\bm{F}|^2$
• Adaptive masking: $\bm{M} = (\bm{P} > \theta)$; $\bm{F}_{\mathrm{filtered}} = \bm{F} \odot \bm{M}$
• Global and local filtering: Apply $\bm{W}_G$, $\bm{W}_L$ to respective spectral paths
• Fusion: sum filtered outputs
• Output: inverse FFT to time domain

Pseudocode for a PyTorch-compatible implementation is provided in the main text (Eldele et al., 2024):

def adaptive_spectral_block(x_in, θ, W_G, W_L): X_fft = fft(x_in) P = (X_fft.abs()**2).sum(dim=1, keepdim=True) M = (P > θ).float() X_filtered = X_fft * M F_G = X_fft * W_G F_L = X_filtered * W_L F_int = F_G + F_L x_out = ifft(F_int).real return x_out

The threshold $\theta$ is trainable via backpropagation and may be parameterized as either a direct scalar or a quantile of the power distribution. This adaptive mechanism distinguishes signal-dominated frequencies from noise, effecting dynamic spectral denoising.

3. Adaptive Spectral Block in Terahertz Spectrum Allocation

ASB in spectrum allocation, as studied for THz communication systems, refers to Adaptive Sub-band Bandwidth. Here, the available spectrum $B_{\mathrm{tot}}$ is partitioned into $S$ sub-bands $\{B_s\}$ of variable rather than equal width, with

$$\sum_{s=1}^S B_s + (S-1)B_g = B_{\mathrm{tot}}, \quad 0 \leq B_s \leq B_{\max}$$

where $B_g$ is the guard band between sub-bands (Shafie et al., 2021). Bandwidth allocation is formulated as a mixed-integer optimization problem, maximizing the minimum user rate subject to power, SNR, and assignment constraints. In this context, ASB allows the spectrum manager to assign narrower bands where frequency-dependent absorption loss $K(f)$ is high and wider bands where loss is low, improving aggregate user throughput in power- and blockage-limited regimes. The mathematical program integrates assignment variables $x_{i,j,s}$, per-link power $P_{i,j,s}$, and bandwidths $B_s$, leveraging convexity via auxiliary variables and penalty terms for tractable solution via successive convex approximation.

4. Empirical and Analytical Performance Evaluation

In time series applications, ablation experiments indicate that removing the ASB results in a marked accuracy decline (FordA: 93.1%→87.3%) and increased forecasting mean squared error (ETTh1: 0.413→0.421). The local branch of the ASB is especially important for noise mitigation; exclusion of this component further degrades performance under significant noise. Noise-robustness analyses show that TSLANet’s ASB maintains stable accuracy as Gaussian noise increases, significantly outperforming transformer architectures without ASB, especially on noisy data (Eldele et al., 2024).

For spectrum allocation, enabling ASB (adaptive sub-bands) in THz wireless networks yields a 13–33% higher aggregate throughput over equal sub-band widths, with maximal gain realized under low-power constraints. Gains are further amplified by multi-connectivity, up to approximately 26% outperformance when the number of simultaneous access point associations grows. These results are explained by the ability of adaptive partitioning to equalize SNR across spectrally non-uniform loss landscapes (Shafie et al., 2021).

ASB Context	Core Mechanism	Empirical Gain
TSLANet (deep learning)	FFT, adaptive threshold, dual spectral filters	93.1% vs 87.3% accuracy; robust to noise
THz spectrum (comm)	Variable sub-band width, convex optimization	13–33% increased throughput

5. Architectural Integration and Computational Complexity

In TSLANet, each network layer stacks the ASB followed by an Interactive Convolution Block (ICB), establishing a sequence of spectral and spatial–temporal operations. The ASB operates in $\mathcal{O}(N \log N)$ time per layer due to the FFT/IFFT-based filtering, a significant advantage over $\mathcal{O}(N^2)$ self-attention in transformers for long sequences. In THz systems, the SCA-based optimization problem governing ASB executes on timescales sufficient for adaptive spectrum management (hundreds of milliseconds), accommodating moderately dynamic environments without prohibitive complexity.

6. Design Principles, Practical Considerations, and Applications

The spectral ASB in TSLANet is underpinned by learnable thresholds and filters, with all parameters (including $\theta$) updated end-to-end using standard backpropagation. In spectrum allocation, $B_s$ values are confined within operational bands where absorption models are monotonic; the design imposes lower and upper limits to balance hardware constraints, spectral efficiency, and interference suppression via guard bands. In both cases, the ASB paradigm allows for selective, data-driven partitioning of the spectrum—frequency components in deep models, and communication bands in wireless.

Key application domains include:

• Multivariate time series classification, forecasting, and anomaly detection under nonstationary noise (deep learning ASB)
• Multi-connectivity, power-limited, and indoor THz wireless communication environments (THz allocation ASB)

Design requires fitting environmental parameters (e.g., HITRAN-driven absorption curves), choosing appropriate penalty weights for assignment relaxation, and ensuring the real-time adaptivity of the decision engine (Shafie et al., 2021).

7. Significance and Interdisciplinary Implications

The ASB’s frequency-selective mechanism—learning or optimizing spectrum allocation—demonstrates strong performance improvements in both deep representation learning and physical-layer resource management. In neural architectures, it formalizes a method for reducing the impact of spurious high-frequency components while maintaining global contextual awareness. In THz communications, it represents a principled approach to overcoming severe frequency- and distance-selective propagation effects that are unique to molecular absorption phenomena.

A plausible implication is that adaptive spectral partitioning may become a generic principle across domains where high-dimensional signals or resources are subject to nonuniform or nonstationary spectral characteristics, especially in environments requiring resilience to noise or dynamically varying loss.

References:

TSLANet/Adaptive Spectral Block in deep representation learning (Eldele et al., 2024); Adaptive Sub-band Bandwidth in THz spectrum allocation (Shafie et al., 2021).