Logarithmic Filter Grouping Techniques

• Logarithmic filter grouping is a technique that arranges filters, poles, or channels along a logarithmic scale using exponential spacing for efficient and perceptually aligned designs.
• It is applied in fractional-order filters, auditory filterbanks, and CNNs, offering precise frequency control and reducing parameters by up to 50% in neural networks.
• Its mathematical foundation leverages logarithmic spacing to enable adaptive designs with minimized edge effects and constant-Q perceptual analysis.

Logarithmic filter grouping is a family of techniques in signal processing and machine learning in which filters, poles, zeros, or convolutional channels are assigned or arranged according to exponential (often base-2) spacing along a logarithmic axis, typically frequency or channel index. This approach is grounded on the observation that both natural signals and learned representations in systems such as convolutional neural networks (CNNs) or auditory filterbanks exhibit strongly nonuniform—often approximately logarithmic—distributions of spectral or spatial characteristics. Logarithmic grouping enables compact, efficient, and perceptually relevant designs for applications ranging from fractional-order filtering to deep learning and auditory analysis.

1. Mathematical Principles of Logarithmic Grouping

The central mathematical foundation of logarithmic filter grouping is the logarithmic spacing of key filter parameters—whether filter center frequencies, convolutional group widths, or the placement of poles and zeros in analog/digital filters. For instance, equispacing over a logarithmic axis for filter center frequencies or pole locations implies that

$$\xi_k = \xi_0 \, r^{\,k}, \quad k=0,1,\dots,N-1, \quad r = \left(\frac{\xi_\mathrm{max}}{\xi_\mathrm{min}}\right)^{\frac{1}{N-1}}$$

where $\xi$ generically denotes frequency, channel index, or any underlying variable being grouped (Lin, 2017, Devakumar et al., 2024).

In the context of fractional-order filtering, closed-form formulae for exponentially-spaced pole-zero pairs yield transfer functions with arbitrary log–magnitude slopes, directly controlling the logarithmic rolloff (Smith et al., 2016). For CNN filter groupings, base-2 or higher logarithmic decay in channel group sizes matches the observed nonuniformity of filter specialization across different layers (Lee et al., 2017).

2. Logarithmic Grouping in Analog and Digital Filters

Logarithmic filter grouping manifests classically in linear systems as arrangements of poles and zeros along exponential grids in the $s$-domain:

• Define a first pole $p_0<0$ and a constant ratio $r>1$.
• Generate $N$ poles and interleaved zeros:

$$p_k = p_0\,r^k, \qquad z_k = p_k\,r^{-\alpha}, \qquad \alpha \in (-1, 1)$$

where $\alpha$ controls the fractional log-magnitude slope (Smith et al., 2016).

This construction produces filters that approximate variable-order differentiators/integrators, Chebyshev-optimal in the log–frequency domain as the array size increases. Edge effects near the frequency band boundaries are minimized by extending the array with $K$ extra pole-zero pairs at each end, reducing ripple inside the desired band. The spectral slope parameter $\alpha$ can be modulated in real time by shifting the zeros, an operation that preserves stability and minimizes computational cost. For filter banks, distinct log-spaced groups can be assigned to adjacent sub-bands, supporting efficient and fully parametric multiband filtering (Smith et al., 2016).

3. Logarithmic Grouping in Convolutional Neural Networks

In deep learning, particularly CNNs, logarithmic filter grouping organizes convolutional filters into groups whose widths decay logarithmically, typically following a base-2 rule. Given a layer of $c$ channels and $K$ groups, group sizes are assigned as:

$$g_k = \left\lfloor \frac{c}{2^k} \right\rfloor, \quad k=1,\dots,K-1;\quad g_K = \left\lfloor \frac{c}{2^{K-1}} \right\rfloor;\quad \sum_{k=1}^K g_k = c$$

Grouped convolutions are then performed per group with matching channel dimensions. This structuring closely mirrors the empirically observed spectrum of filter specializations (e.g., the first layer of AlexNet shows a 53:28:15 split, indicative of logarithmic decay) (Lee et al., 2017).

Logarithmic grouping in shallow CNNs improves parameter efficiency and preserves accuracy relative to uniform grouping. Benchmarks on Multi-PIE (facial expression) and CIFAR-10 (object classification) show that, for comparable accuracy, logarithmic grouping reduces parameters by approximately 20–50% compared to uniform grouping. The principle extends naturally to grouped architectures where depthwise separability or multi-branch designs are used, and is expected to hold for other nonlinear or power-law scaling bases (Lee et al., 2017).

4. Logarithmic Spacing in Auditory and Gabor Filter Banks

Auditory and Gabor filterbanks use logarithmic spacing to ensure consistent resolution and coverage across the frequency spectrum. Given a range $[f_\mathrm{min}, f_\mathrm{max}]$ and $N$ subbands, logarithmic centers are defined by

$$f_n = f_{\min} \left(\frac{f_\mathrm{max}}{f_\mathrm{min}}\right)^{\frac{n-1}{N-1}},\quad n=1,\dots,N$$

This yields equispacing on a logarithmic scale, which is required for constant-Q analysis and perceptual relevance in audio processing (Lin, 2017).

Filter banks constructed with logarithmic grouping also allow for constant frequency coverage (coverage ratio $\eta_C^{(n)}$) and, in wavelet designs, enable Gabor- or log-Gabor-like transforms with constant relative bandwidths. In multidimensional settings, as in (Devakumar et al., 2024), Gaussians are placed along log-frequency axes and their inverse Fourier transforms yield steerable, scale-invariant Gabor-like kernels.

5. Algorithms and Implementation Considerations

Implementation of logarithmic grouping consistently relies on parameterization by desired bandwidth, slope, or coverage. Parameter selection workflows include:

• For analog/digital filters: choose $N$, overshoot $K$, band endpoints, solve for $p_0$ and $r$, and assign zeros per desired slope $\alpha$ (Smith et al., 2016).
• For auditory filterbanks: set $f_\mathrm{min}$, $f_\mathrm{max}$, $N$, compute geometric ratio $R$, and assign centers as $f_n = f_\mathrm{min}R^{n-1}$ (Lin, 2017).
• For CNNs: specify $c$, $K$, assign group widths $g_k$ logarithmically, and partition convolutions accordingly (Lee et al., 2017).

Bandwidth overlap, stability, and edge effects require careful control of the array or group extension regions. Real-time modulation of slope or group boundaries can be achieved by shifting zeros (filters) or reassigning channels, enabling adaptive or fully parametric designs.

6. Applications and Performance Characteristics

Logarithmic filter grouping is employed in at least four principal contexts:

Domain	Logarithmic Grouping Role	Cited Example
Fractional order filters	Exponential arrangement of poles/zeros, real-time slope tuning	(Smith et al., 2016)
CNNs	Channel groups decay log-wise for grouped convolution, reduced params	(Lee et al., 2017)
Auditory filterbanks	Log-spaced center frequencies, constant coverage/Q	(Lin, 2017)
Gabor/log-Gabor banks	Gaussians on log-frequency, scale invariance, steerable orientations	(Devakumar et al., 2024)

Performance benefits include Chebyshev-like equal-ripple log–magnitude approximation, adaptive resolution matching perceptual or statistical features of real-world data, and efficient parameterization (often requiring only two degrees of freedom). In neural networks, log grouping enables up to 50% parameter reduction with accuracy degradation typically ≤1% (Lee et al., 2017).

7. Generalizations, Limitations, and Future Research

Generalizations include extension to nonlinear, power-law, or exponential spacing bases other than log-base-2—the effectiveness of alternate scaling schemes remains a subject of investigation (Lee et al., 2017). The creation of adaptive or learnable group boundaries, integration into deeper architectures, and synergy with recent advances in attention-based and depthwise-separable convolutional architectures are promising directions.

Limitations of current approaches noted in the literature include restriction of demonstrated efficacy to shallow networks, hand-designed rather than learned group sizes, and possible suboptimality of fixed base-2 assignment for specific tasks. In perceptual modeling, parameters such as ERB constants must be re-estimated for individuals/events where strict fidelity is required (Lin, 2017). A plausible implication is that learnable or data-adaptive logarithmic groupings may offer further improvements in both signal processing and machine learning domains.

Logarithmic filter grouping uniquely connects principles from signal analysis, perceptual modeling, and machine learning, delivering efficient, parametric constructions that combine constant relative resolution, adaptive coverage, and computational scalability (Smith et al., 2016, Lee et al., 2017, Lin, 2017, Devakumar et al., 2024).