Selective Averaging with Application to Phase Reduction

Gerd Schmidt, Jeff Moehlis, Frank AllgöwerIEICE 2014 · DOI

TL;DR

This paper introduces "selective averaging," a novel mathematical technique that extends standard averaging methods for dynamical systems. It allows for averaging specific terms in a system's equations while leaving others untouched, which is crucial for analyzing complex systems like coupled oscillators with external inputs. This method provides a more rigorous justification for simplified models used in fields like neuroscience.

What Problem Does This Paper Address?

Analyzing complex dynamical systems, especially those with many interacting components (like coupled oscillators), is a significant challenge. Traditional methods like averaging simplify these systems by smoothing out fast oscillations, but they often average all terms. This can lead to oversimplification, losing important information. For instance, in systems of coupled oscillators that also receive external inputs (e.g., neural networks), it's often desirable to average the coupling terms (which depend on phase differences) but not the external input terms, as these can have distinct effects. Existing averaging techniques don't readily support this selective treatment. This paper addresses this gap by developing a mathematical framework for "selective averaging," enabling a more nuanced and accurate simplification of such systems.

Methods

The core contribution is the development of the "Selective Averaging Theorem." This theorem extends existing averaging techniques by allowing for the averaging of specific components of a vector field while leaving others untouched. The authors achieve this by:

  • Defining Local Averages: They introduce the concept of a "local average" of a function, which is the average over a short time interval.
  • Lemmas for Approximation: Several lemmas are presented to bound the difference between the original system's solution and a simplified system where only certain terms are averaged. These lemmas leverage techniques like Gronwall's Lemma to establish bounds on the error.
  • The Selective Averaging Theorem: This theorem formally states that if a system's vector field is decomposed into two parts (one to be averaged, one not), the solution of the original system can be approximated by the solution of a system where only the designated part is averaged. The error is shown to be small, particularly on a rescaled "slow" time scale.
  • Application to Phase Reduction: The authors demonstrate the utility of their method by applying it to the phase reduction of coupled oscillators with external input. This involves transforming the system into phase variables and then selectively averaging the coupling terms while keeping the external input term as is.

Datasets

Not discussed in the paper.

Key Insights and Ideas

  • Selective Averaging is Powerful: The ability to selectively average terms is not just a mathematical curiosity; it's essential for creating accurate simplified models of complex systems where different components have different dynamics.
  • Justification for Simplified Models: The paper provides a rigorous mathematical basis for models that are often used heuristically in fields like neuroscience (e.g., neural control problems), where coupling terms are averaged but external inputs are not.
  • Phase Reduction with Inputs: The application to phase reduction shows how this technique can disentangle the effects of internal coupling from external forcing in oscillatory systems, leading to a clearer understanding of their collective behavior.
  • Error Bounds: The lemmas provide concrete bounds on the approximation error, giving confidence in the validity of the selective averaging approach.

Key Results

The paper presents two main theorems:

  1. Selective Averaging Theorem (Theorem 2.4): For a system $x = \epsilon f^0(x,t) + \epsilon f^1(x,t)$, if $f^0$ is T-periodic and $f^1$ is not averaged, the solution $x(t)$ is close to the solution $y(t)$ of $y = \epsilon \bar{f}^0(y) + \epsilon f^1(y,t)$ (where $\bar{f}^0$ is the average of $f^0$), with the difference $||x(t) - y(t)|| = O(\epsilon)$ on the time scale $1/\epsilon$.
  2. General Selective Averaging Theorem (Theorem 4.3): This generalizes the concept to systems where $f^0$ is a KBM (Krylov-Bogoliubov-Mitropolsky) vector field with order function $\delta_0(\epsilon)$. The solution $x(t)$ of $x = \epsilon f^0(x,t) + \epsilon f^1(x,t)$ is shown to be close to the solution $z(t)$ of $z = \epsilon f^0(z) + \epsilon f^1(z,t)$, with $||x(t) - z(t)|| = O(\sqrt{\delta_0(\epsilon)})$.

The application to phase reduction demonstrates that the coupling terms can be averaged to depend only on phase differences, while the external input term remains, leading to a simplified phase model.

Limitations and Future Work

  • Assumptions on $f^0$ and $f^1$: The theorems rely on specific assumptions about the vector field components, such as continuity, Lipschitz continuity, and periodicity for $f^0$. The applicability to systems not meeting these criteria might require further investigation.
  • KBM Vector Fields: The general theorem's reliance on KBM vector fields for $f^0$ is a specific condition. Extending the selective averaging to non-KBM $f^0$ would be a valuable future direction.
  • Specific Application: While the paper demonstrates the method on coupled oscillators with input, exploring its application to other complex systems (e.g., biological networks, climate models) could reveal further insights.
  • Quantitative Error Analysis: While error bounds are provided, further quantitative analysis for specific systems could be beneficial for practical applications.

Summary updated

Abstract

For a class of vector fields, we show that one can selectively average terms which are of the same order in a small parameter, giving an extension of standard averaging results. Such selective averaging is illustrated for the phase reduction of a system of oscillators with both coupling and external input, for which the coupling can be averaged to give a term which only depends on phase differences, while the external input term is not averaged.
BibTeX
@inproceedings{schmidt2014selective,
  author = {Gerd Schmidt and Jeff Moehlis and Frank Allgöwer},
  title = {{Selective Averaging with Application to Phase Reduction}},
  booktitle = {IEICE},
  year = {2014},
  doi = {10.15248/proc.2.491},
}
Figure