🔬 applied physics

Matching of signal, noise and hardware timescales for filtering and forecasting of correlated noise signals

This paper demonstrates that in physical reservoir computing, the relationship between noise correlation time, reservoir memory, and forecast horizon determines whether correlated noise is filtered or predicted, introducing specific metrics to guide the design of architectures capable of processing stochastic signals across distinct temporal scales.

Original authors: Joshua Donald, Alex Gabbitas, Arthur G. T. Coveney, Sergey Savel'ev, Pavel Borisov

Published 2026-10-08

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Joshua Donald, Alex Gabbitas, Arthur G. T. Coveney, Sergey Savel'ev, Pavel Borisov

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the world of data, time is rarely a straight, clean line. Real-world signals, whether they are the fluctuating price of a digital currency, the shifting patterns of weather, or the rhythm of a human heartbeat, are almost always a mixture of a clear, predictable pattern and a chaotic, random jumble. This random element is often called noise. For decades, scientists and engineers have treated this noise as an enemy to be removed, a static interference that obscures the true message. However, a new perspective suggests that this noise is not just static; it often has its own internal structure and rhythm. When noise changes slowly over time, it carries a hidden pattern that might be just as predictable as the signal itself. The challenge for modern computing is to build systems that can not only filter out the fast, useless jitters but also recognize and forecast the slow, structured randomness that is woven into the data.

Researchers at Loughborough University have taken a significant step toward solving this puzzle by testing how a specific type of physical computer handles these mixed signals. Instead of using traditional software running on silicon chips, they used a physical device made of a thin, porous layer of niobium oxide sandwiched between metal electrodes. This material acts as a "reservoir," a term used to describe a system that naturally holds onto a memory of what happened just a moment ago. When an electrical signal is fed into this device, the material's internal state changes in complex, non-linear ways, creating a rich history of the input that can be read out to make predictions. The researchers wanted to see if this physical hardware could distinguish between noise that moves too fast to be useful and noise that moves slowly enough to be forecasted.

To test this, the team created a series of experiments using both synthetic signals and real-world data. They started with a simple, predictable wave made of two different frequencies and added a layer of random noise to it. They then carefully adjusted how quickly this noise changed, ranging from extremely rapid fluctuations to very slow, drifting variations. They fed these mixed signals into their physical reservoir and asked the system to do two things: first, to strip away the noise and reveal the original clean wave, and second, to predict what the signal would look like in the near future. The results revealed a clear dividing line based on time. When the noise changed very quickly—faster than the physical memory of the device could keep up with—the reservoir acted as a filter, effectively averaging out the chaos and leaving the underlying pattern intact. However, when the noise changed slowly, moving at a pace that matched the device's ability to remember, the system did something different. It stopped treating the noise as interference and instead recognized its slow, rhythmic structure, allowing it to predict the future path of the noisy signal with surprising accuracy.

The researchers then applied this same principle to a much more complex, real-world problem: predicting the volatility of cryptocurrency prices. They analyzed the price fluctuations of Bitcoin, Ethereum, Solana, and Dogecoin, calculating how much the prices varied over different time windows. Just like with their synthetic signals, they found that the system's ability to predict the future depended entirely on the speed of the price changes relative to the hardware's memory. When the market volatility shifted rapidly, the device smoothed it out. But when the market moved with a slower, more persistent rhythm, the physical reservoir could successfully forecast the next steps. This confirmed that the key to success was not just the power of the algorithm, but the matching of timescales between the input data and the physical properties of the computer itself.

To quantify this relationship, the team introduced a new way of measuring the system's performance, which they called a forecasting regime index. This metric essentially compares the speed of the noise to the speed of the machine's memory and the length of the prediction window. They found that when the noise is faster than the machine's memory, the system enters a "filtering regime," where it averages out the rapid changes. When the noise is slower, it enters a "prediction regime," where the machine uses the noise's own structure to make forecasts. This distinction is crucial because it suggests that for physical computers to work effectively with real-world data, engineers cannot simply speed up or slow down the data arbitrarily. They must encode the information in a way that respects the natural memory limits of the hardware.

The study demonstrates that physical reservoir computing is not just a faster way to run software; it is a fundamentally different approach that leverages the natural dynamics of materials to process time. By showing that a simple piece of porous metal can distinguish between useless noise and useful, structured randomness, the researchers have provided a new blueprint for designing future computing systems. These systems could be tailored to specific tasks, filtering out the fast, chaotic noise of a sensor while simultaneously predicting the slow, drifting trends of a financial market. The work suggests that the future of efficient computing lies not in making machines think exactly like humans, but in building hardware that naturally aligns with the timescales of the world it is trying to understand.

Technical Summary: Matching of Signal, Noise, and Hardware Timescales for Filtering and Forecasting of Correlated Noise Signals

Problem Statement
Physical reservoir computing (RC) offers energy-efficient processing of time-dependent data by leveraging the nonlinear dynamics of physical systems. However, a fundamental challenge arises because physical reservoirs possess fixed intrinsic response timescales, whereas real-world signals often comprise a mixture of deterministic components and stochastic (noisy) components spanning multiple timescales. While some machine learning methods address denoising or forecasting, there is a lack of understanding regarding how a physical reservoir specifically handles input signals where the noise correlation time ( $\tau$ ) varies relative to the reservoir's memory and the prediction horizon. Specifically, it remains unclear whether such systems filter out correlated noise or exploit its temporal structure for prediction, and how this behavior depends on the matching of timescales.

Methodology
The authors investigated these dynamics using a nanoporous niobium oxide ( $NbO_x$ ) memristor-based physical reservoir. The experimental platform consisted of two parallel $NbO_x$ layers sandwiched between platinum electrodes, providing four current outputs for a single voltage input. The system was tested using three distinct tasks:

Reconstruction: Filtering colored noise from a deterministic signal to recover the original waveform.
Forecasting (Mixed): Predicting future values of a signal containing both deterministic components and colored noise.
Forecasting (Pure Noise): Predicting future values of the colored noise component alone.

The input signals were synthetic, combining a quasi-periodic deterministic function (two harmonics) with colored noise generated via an Ornstein-Uhlenbeck process. The noise parameters were independently tunable: noise intensity ( $D$ ) and noise correlation time ( $\tau$ ), ranging from $1\tau_0$ to $1000\tau_0$ (where $\tau_0 = 6.3 \times 10^{-7}$ s). Additionally, the system was tested on real-world data: the time-averaged volatility of four cryptocurrencies (Bitcoin, Ethereum, Solana, Dogecoin), where the correlation time was manipulated by varying the averaging window size.

Performance was quantified using Normalized Root Mean Square Error (NRMSE) and Absolute Mean Error (AME). The authors also characterized the reservoir's intrinsic properties, specifically Memory Capacity (MC) and Nonlinearity (NL), across different data window sizes.

Key Contributions and Results
The study establishes that the performance of a physical RC system in handling noisy signals is governed by the relationship between three timescales: the noise correlation time ( $\tau$ ), the reservoir's memory horizon (MH), and the forecast/prediction window size ( $\tau_w$ ).

Two Distinct Forecasting Regimes:
- Reservoir-Based Averaging: When the noise correlation time is significantly shorter than both the MH and the prediction window size (e.g., $\tau < MH$ ), the noise component leaves a memory imprint on the reservoir but is averaged out within the prediction window under the limitations of the reservoir's MH. Consequently, the reservoir output resembles the underlying deterministic signal rather than the noisy input, effectively filtering the noise. In this regime, the system performs better at reconstructing the deterministic signal than predicting the noisy signal itself.
- Algorithm-Based Forecasting: When the noise correlation time is longer than the MH and prediction window (e.g., $\tau > MH$ ), the noise exhibits sufficient temporal structure for the system to learn and predict. Here, the reservoir treats the noise as part of the signal dynamics, allowing for the forecasting of the combined signal or the noise component itself.
Timescale Mismatch and Performance:
- The system exhibited the highest error (worst performance) when the noise correlation time was comparable to the deterministic signal's timescales (around $10\tau_0$ in the synthetic tests), where filtering and prediction become most difficult due to interference.
- Performance improved for very fast noise (filtered out) and very slow noise (predictable).
- In the cryptocurrency volatility tests, the results mirrored the synthetic findings: prediction error decreased as the correlation time increased, confirming that slower-varying stochastic components are more amenable to algorithmic forecasting.
New Metrics:
- Memory Horizon (MH): Defined as the time delay range where the correlation coefficient $r^2 \geq 0.2$ , quantifying the effective short-term memory of the hardware.
- Forecasting Regime Index (FRI): An empirical metric defined as $FRI = \log_{10}[\tau / \sqrt{\tau_m \tau_w}]$ . An FRI < 0 indicates the averaging regime, while FRI > 0 indicates the algorithm-based forecasting regime. This index successfully predicted the transition between filtering and prediction behaviors observed in the experimental data.
Hardware Characterization:
- The reservoir's Memory Capacity peaked at a window size of $15.87\tau_0$ (MC $\approx$ 4.11) but decreased for larger windows as nonlinearity increased.
- The study demonstrated that the optimal window size for a specific task depends on the trade-off between linear memory capacity and the nonlinearity required to process complex signal structures.

Significance and Claims
The paper claims that colored noise should not be viewed merely as a nuisance to be removed, but as a structured component of the learning problem whose predictability depends on timescale matching. The primary contribution is the demonstration that the "computational role" of noise (filtering vs. prediction) is determined by the relative position of the noise correlation time against the hardware's memory horizon and the forecast window.

The authors argue that this finding establishes new design principles for physical RC architectures and data encoding. Specifically, they suggest that input signals should be encoded not just to fit the hardware's voltage and time ranges, but to align the signal's autocorrelation structure and noise spectrum with the reservoir's memory dynamics. This is particularly relevant for high-speed physical reservoirs processing slow real-world signals, where non-uniform time compression may be necessary to map informative temporal structures into the reservoir's effective memory window. The work provides a quantitative framework (via FRI and MH) to guide the development of physical reservoirs capable of simultaneously filtering, analyzing, and predicting stochastic signal components across distinct temporal scales.

More like this