Chebyshev State-Space Compression: Provably Stable Near-Minimax Continuous Memory in Sub-Kilobyte SRAM for Real-Time Edge Microcontrollers

Deploying deep recurrent neural networks on microcontrollers is fundamentally bottlenecked by non-linear gating overhead and substantial runtime workspace arenas (16-32 KB in TensorFlow Lite Micro). We present Chebyshev-Compressor, a continuous-time state-space architecture based on a Decoupled State-Space Readout (DSSR) formulation. By projecting continuous sliding-window temporal transport onto shifted Chebyshev polynomials of the first kind, the architecture achieves near-optimal uniform minimax signal approximation with logarithmic Lebesgue constant growth, avoiding the Runge divergence of equispaced interpolations. Inflow boundary conditions are enforced weakly via a Chebyshev-Gauss-Radau (CGR) Simultaneous Approximation Term (SAT). We prove an exact discrete Summation-By-Parts (SBP) Lyapunov dissipation identity via an arrowhead Schur complement condition, guaranteeing strictly Hurwitz continuous-time eigenvalues and unconditional discrete-time unit-disk stability under double-precision Bilinear (Tustin) transformation. Downstream cross-channel non-linearities are decoupled into a static Readout MLP optimized via single-pass precomputed state distillation. The compiled C99 engine requires only 608 bytes of algorithmic active SRAM (448 bytes persistent state and 160 bytes stack scratch; 744 bytes peak compiler stack frame on ARM Cortex-M4F) with strict IEEE-754 bit-masking quarantine immune to aggressive -ffast-math compiler elision. On NASA C-MAPSS FD001 and a synthetic 3-axis vibration benchmark (AUROC 0.988), our engine executes in 7.4 microseconds on x86-64, 3.8-24.6 microseconds on an ESP32 (240 MHz), and 12.4-68.5 microseconds on an STM32F401RE (84 MHz), maintaining strict numerical parity between PyTorch and C99 outputs within 0.0001. Source code and reproducibility scripts: https://github.com/Genesis410/chebyshev-compressor

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-21
DOI
https://doi.org/10.5281/zenodo.24408631
Primary Topic
Advanced Neural Network Applications
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Chebyshev State-Space Compression: Provably Stable Near-Minimax Continuous Memory in Sub-Kilobyte SRAM for Real-Time Edge Microcontrollers

Ihsan Sukandar
1 links
Zenodo (CERN European Organization for Nuclear Research)
Advanced Neural Network Applications
preprint

Chebyshev State-Space Compression: Provably Stable Near-Minimax Continuous Memory in Sub-Kilobyte SRAM for Real-Time Edge Microcontrollers

Ihsan Sukandar
preprint en

Abstract

Deploying deep recurrent neural networks on microcontrollers is fundamentally bottlenecked by non-linear gating overhead and substantial runtime workspace arenas (16-32 KB in TensorFlow Lite Micro). We present Chebyshev-Compressor, a continuous-time state-space architecture based on a Decoupled State-Space Readout (DSSR) formulation. By projecting continuous sliding-window temporal transport onto shifted Chebyshev polynomials of the first kind, the architecture achieves near-optimal uniform minimax signal approximation with logarithmic Lebesgue constant growth, avoiding the Runge divergence of equispaced interpolations. Inflow boundary conditions are enforced weakly via a Chebyshev-Gauss-Radau (CGR) Simultaneous Approximation Term (SAT). We prove an exact discrete Summation-By-Parts (SBP) Lyapunov dissipation identity via an arrowhead Schur complement condition, guaranteeing strictly Hurwitz continuous-time eigenvalues and unconditional discrete-time unit-disk stability under double-precision Bilinear (Tustin) transformation. Downstream cross-channel non-linearities are decoupled into a static Readout MLP optimized via single-pass precomputed state distillation. The compiled C99 engine requires only 608 bytes of algorithmic active SRAM (448 bytes persistent state and 160 bytes stack scratch; 744 bytes peak compiler stack frame on ARM Cortex-M4F) with strict IEEE-754 bit-masking quarantine immune to aggressive -ffast-math compiler elision. On NASA C-MAPSS FD001 and a synthetic 3-axis vibration benchmark (AUROC 0.988), our engine executes in 7.4 microseconds on x86-64, 3.8-24.6 microseconds on an ESP32 (240 MHz), and 12.4-68.5 microseconds on an STM32F401RE (84 MHz), maintaining strict numerical parity between PyTorch and C99 outputs within 0.0001. Source code and reproducibility scripts: https://github.com/Genesis410/chebyshev-compressor

Zenodo (CERN European Organization for Nuclear Research)
Advanced Neural Network Applications
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon
Chebyshev State-Space Compression: Provably Stable Near-Minimax Continuous Memory in Sub-Kilobyte SRAM for Real-Time Edge Microcontrollers — Ihsan Sukandar · Zenodo (CERN European Organization for Nuclear Research) (2026) | TGRS Research Map | TGRS