Research map / Architectures and efficiency / RNNs and normalization
Universality of Real Minimal Complexity Reservoir
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proves that real-valued Simple Cycle Reservoirs (cycle permutation coupling with binary ±1 input weights) are universal approximators of time-invariant fading-memory filters, using orthogonal dilation and real canonical forms, and introduces a bipartite-matching method to reduce SCR size.
Problem
Simple cycle reservoirs (SCRs) are highly constrained, hardware-friendly reservoir computing models, but their theoretical universality was only known over the complex domain; the real-domain case was open.
Approach
Replaces the complex unitary dilation and eigendecomposition used in prior complex-domain proofs with orthogonal dilation and real Schur canonical forms, so that cyclic permutation coupling with ±1 input weights approximates any real linear reservoir; a maximum-matching procedure on roots of unity reduces the required number of units.
Key results
- State mean square differences between the original linear reservoir and its cyclic dilation decay toward zero exponentially as the dilation dimension N grows, on ETTm2 and ECL.
- Maximum-matching reduction yields cyclic dilation sizes n_C of roughly 300-400 units for initial reservoirs of size 80-160, far below the theoretical upper bound.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Increase the dilation dimension | Cyclic dilation of a linear reservoir · ETTm2 and ECL | state MSE | Helped |
| Use maximum matching to choose the cyclic dilation dimension | Cyclic dilation of an orthogonal reservoir · Initial reservoir sizes 80-160 | reservoir dimension | Helped |
Limitations
- Experiments use small original reservoir (n=5) and a linear readout, so they only validate structural approximation rather than task performance.
- Theoretical dimension bound for the cyclic dilation is large; the matching-based reduction is empirical and has no tight guarantee.
Experimental setup
Datasets: ETTm2 ECL
Benchmarks: ETTm2 ECL
Metrics: mean square difference of reservoir states prediction loss
Builds on: Echo State Networks simple cycle reservoirs Egervary dilation theorem complex-domain SCR universality proof
Closest papers by meaning
- Reservoir Computing meets Recurrent Kernels and Structured Transforms (NeurIPS 2020 · similarity 0.90)
- Frequency Domain Reservoir Computing (NeurIPS 2026 · similarity 0.87)
- ParalESN: Enabling parallel information processing in Reservoir Computing (ICML 2026 · similarity 0.87)
- Spectral-Aware Reservoir Computing for Fast and Accurate Time Series Classification (ICML 2025 · similarity 0.87)
- Accelerating Linear Recurrent Neural Networks for the Edge with Unstructured Sparsity (ICML 2025 · similarity 0.85)
- Recurrent neural networks: vanishing and exploding gradients are not the end of the story (NeurIPS 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.