---
title: Universality of Linear Recurrences Followed by Non-linear Projections: Finite-Width Guarantees and Benefits of Complex Eigenvalues
url: https://www.ml-quant.com/papers/arxiv/2307.11888/
site: ML-Quant (https://www.ml-quant.com)
updated: 2026-09-26
license: Summaries CC BY 4.0; links go to the original sources
index: https://www.ml-quant.com/llms.txt
identifier: arXiv:2307.11888
source_url: https://arxiv.org/abs/2307.11888
featured: 2024-03-13
citations: 41
topic: Econometrics & Forecasting
---


# Universality of Linear Recurrences Followed by Non-linear Projections: Finite-Width Guarantees and Benefits of Complex Eigenvalues

The study reveals that deep neural networks using linear complex-valued RNNs and MLPs can accurately approximate regular causal sequence-to-sequence maps, with complex eigenvalues near unit disk aiding in information storage.

- Source: https://arxiv.org/abs/2307.11888
- Identifier: arXiv:2307.11888
- Released: 2023-07-21
- First featured: Quant Letter No. 40 (2024-03-13): https://www.ml-quant.com/issues/2024-03-13/
- Citations (Semantic Scholar): 41
- Published in: not yet
- Topic: Econometrics & Forecasting

## Related

- [Distributional Refinement Network: Distributional Forecasting via Deep Learning](https://www.ml-quant.com/papers/arxiv/2406.00998/): The article introduces the Distributional Refinement Network (DRN), a model that enhances predictive performance and interpretability in actuarial modelling by merging a baseline model with a flexible neural network.
- [Machine learning framework for forecasting sales of new products with short life cycles](https://www.ml-quant.com/papers/repec/eee-intfor-v-39-y-2023-i-4-p-1874-1894/): For predicting sales of short-lived products, simple ARIMAX is more effective than deep neural networks, but DNNs perform well when Gaussian white noise is added, a study finds.
- [Forecasting Probability Distributions of Financial Returns with Deep Neural Networks](https://www.ml-quant.com/papers/arxiv/2508.18921/): The research shows that deep neural networks can accurately forecast financial return distributions and are competitive with traditional models for risk assessment and portfolio management.
- [TKAN: Temporal Kolmogorov-Arnold Networks](https://www.ml-quant.com/papers/ssrn/4825654/): The article presents Temporal Kolomogorov-Arnold Networks (TKANs), a new neural network design that merges the benefits of Recurrent Neural Networks and Long Short-Term Memory for improved multistep time series forecasting.
- [A Dynamic Regime-Switching Model Using Gated Recurrent Straight-Through Units](https://www.ml-quant.com/papers/ssrn/4810879/): The Gated Recurrent Straightthrough Unit (GRSTU), a new deep learning model, outperforms statistical jump models in identifying regime changes in the S&P500 index, especially with smaller datasets.
- [Diffusive Gibbs Sampling](https://www.ml-quant.com/papers/arxiv/2402.03008/): The article introduces Diffusive Gibbs Sampling (DiGS), a new method for sampling from multi-modal distributions, which performs better in tasks like Bayesian neural networks and molecular dynamics.