abelkan: Routh-Hurwitz Certified Mittag-Leffler Kernels, Anchor-Residual Spectral Decoders, and O(L) Power-Law Attention for Non-Markovian Extrapolation
Year: 2026
Type: preprint
Abstract: I present a unified technical report and experimental benchmark suite for abelkan (v0.4.0–v0.6.0, github.com/Lolycut/abelkan), a PyTorch library for non-Markovian, irregular-time, and long-context sequence modeling. First, prove via Routh-Hurwitz minors that the [3/4] Padé denominator Q4(z) of MittagLefflerPade34 is strictly Hurwitz-stable across all 501 grid points α ∈ [0.40, 0.90] (max Re(z) = −1.3230 < 0), certifying zero poles on {Re(z) ≥ 0}. Second, across 14 architectures × 8 multi-domain tasks × 3 seeds, abelkan-v5-Universal ranks #1 overall (Mean R² = +0.373; +0.664 on the 7 non-battery tasks), outperforming ODE-RNN (+0.354), iTransformer+RevIN (+0.193), FNO-1D+Detrend (+0.133), PI-KAN-PointNet (−3.499), and DeepOKAN (−17.754). Third, we introduce an O(L) Sum-of-Exponentials (SOE-Abel) decomposition that matches dense O(L²) AbelAttention to 0.114% kernel L2 error (M = 8), achieves < 1.4 × 10⁻¹⁵ equivalence with O(log L) Blelloch associative and chunkwise scans, and eliminates the bfloat16 long-range tail collapse (116.7% → 0.24%, 492× gain) via bf16-io + fp32-acc. Finally, on a laptop NVIDIA RTX 3050, a 15.58M-parameter SOE-Abel mini-LLM trained strictly on L_train = 128 achieves 99.22% (1×), 98.44% (4×), 85.94% (8×), and 41.41% (16×) on MQAR extrapolation, while SOE-Abel Chunk64 scales cleanly to L = 32,768 (2.65 GB VRAM) where dense O(L²) attention hits OOM.
Author: Kakhno, Ilya
Institution: Belarusian State University
Language: English
Subfield: Signal Processing
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7226014959
