Characterizing Nonlinear Dynamics via Smooth Prototype Equivalences

Roy Friedman, Noa Moriel, Matthew Ricci, Guy Pelc, Yair Weiss, Mor Nitzan · 2025

Read Characterizing Nonlinear Dynamics via Smooth Prototype Equivalences by Roy Friedman, Noa Moriel, Matthew Ricci, Guy Pelc, Yair Weiss, Mor Nitzan (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2503-10336.txt.

Introduction

Roy Friedman and co-authors present smooth prototype equivalences, a method to read long-term dynamics from sparse data in 2025. Scientists often get only few, noisy glimpses of a moving system. Matching those glimpses to known prototype behaviors can reveal where the system settles.

Invertible neural networks warp prototype space onto data space with smooth deformations. The warp locates invariant sets like fixed points and limit cycles without any equations. Regime labels come from how far warped data sit from each prototype. Tests beat rival methods on oscillating systems. Gene drivers in a synthetic cell circuit and cell cycle paths in real single-cell data also emerged.

Catalog record

CREATORRoy Friedman, Noa Moriel, Matthew Ricci, Guy Pelc, Yair Weiss, Mor Nitzan
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE55 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2503-10336

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Characterizing Nonlinear Dynamics via Smooth Prototype Equivalences Roy Friedman, Noa Moriel, Matthew Ricci, Guy Pelc, Yair Weiss, Mor Nitzan Abstract Characterizing dynamical systems given limited measurements is a common challenge throughout the physical and biological sciences. However, this task is challenging, especially due to transient variability in systems with equivalent long-term dynamics. We address this by introducing smooth prototype equivalences (SPE), a framework that fits a diffeomorphism using normalizing flows to distinct prototypes — simplified dynamical systems that define equivalence classes of behavior. SPE enables classification by comparing the deformation loss of the observed sparse, high-dimensional measurements to the prototype dynamics. Furthermore, our approach enables estimation of the invariant sets of the observed dynamics through the learned mapping from prototype space to data space. Our method outperforms existing techniques in the classification of oscillatory systems and can efficiently identify invariant structures like limit cycles and fixed points in an equation-free manner, even when only a small, noisy subset of the phase space is observed. Finally, we show how our method can be used for the detection of biological processes like the cell cycle trajectory from high-dimensional single-cell gene expression data. 1 Introduction Predicting the long-term behavior of dynamical systems from sparse data is a challenging problem in data science and computational physics ( Ghadami & Epureanu 2022 ) . The problem of identifying these behaviors, referred to as invariant sets , arises throughout the physical and life sciences. For example, in the early detection of dangerous self-excited oscillations in aeronautical design ( Weisshaar 2012 ; Dimitriadis 2017 ) and in the classification of cell cycle dynamics in computational biology ( Chervov & Zinovyev 2022 ; Moriel et al. 2023 ) . Even if governing equations are known, there are few guarantees on the existence of long-term equilibrium or oscillatory behaviors ( Hilbert 1902 ; Smale 2000 ) , and this situation is exacerbated in realistic cases where equations are unknown and behavior must be estimated from sparse, noisy and/or high-dimensional data. Figure 1: Schematic for the process of fitting dynamical prototypes using smooth prototype equivalences (SPE). The observed dynamics are compared to each of the chosen prototypes, which involves training a diffeomorphism H_{\theta}(x) , parameterized by a neural network, for each prototype g(y) . After training each of the diffeomorphisms, the observed dynamics are classified to the prototype with the smallest equivalence loss. Once the diffeomorphism H_{\theta}(x) is trained, we have access to a mapping between the data space and the prototype space, which allows us to estimate the long-term behavior, i.e. invariant set, of the dynamics underlying the observed data. We construe this problem as a form of prototype learning, where long-term structures are identified and localized by comparison to archetypal class exemplars ( Bonilla & Robles-Kelly 2012 ) . Prototype methods have their conceptual origin in cognitive science ( Rosch 1973 ) and have seen widespread use across machine learning applications, especially in service of few-shot learning ( Chen et al. 2023 ) . Here, we propose smooth prototype equivalences (SPE 1 1 1 Code available in https://github.com/nitzanlab/prototype-equivalences . ), a method that fits a diffeomorphism from the given data to so-called normal forms ( Nayfeh 2008 ) , a natural choice for prototypes in the realm of dynamical systems. Normal forms are systems, \dot{x}=g(x) , whose governing equations take a mathematically concise form and which describe canonical dynamical behaviors. These normal forms are then the basis of equivalence classes \{f_{H}\}_{H\in\mathcal{H}} of topologically identical dynamics via: f_{H}(x)=\left(\frac{\partial H(x)}{\partial x}\right)^{-1}g(H(x)), (1) for a family of diffeomorphisms \mathcal{H} . In this case, we say that f_{H}(\cdot) is smoothly equivalent to the prototype g(\cdot) . In particular, the action of H preserves the existence of invariant sets under the dynamics in both systems (e.g. fixed points, limit cycles, etc.). This means that the problem of characterizing and localizing the invariant set of a dynamical system given its measurements is reduced to that of identifying the prototype that is equivalent to it and mapping that prototype’s invariant structure to the data domain. Below, we demonstrate how this can be done by efficiently optimizing a diffeomorphism to such a prototype. When the equivalent prototype of the system is unknown, we classify the behavior according to the most similar prototype out of several candidates, resulting in a robust framework for long-term dynamical behavior which outperforms existing baselines. Leveraging the fact that normalizing flows ( Papamakarios et al. 2021 ) can approximate diffeomorphisms, H\in\mathcal{H} , we optimize normalizing flows which transform samples, (x,\dot{x}) , to a predefined prototype system. Our use of prototypes of pre-defined, simple dynamical structure means that we can easily use our optimized normalizing flows to locate invariant sets (i.e. limit cycles, fixed points) in complex samples with unknown governing equations. The minimal residual loss across parallel optimizations to candidate prototypes is used to make the classification decision, extending existing work on data-driven dynamical equivalence ( Chen et al. 2024 ; Skufca & Bollt 2008 ) to the classification domain. Importantly, we show how using normal forms as hardwired knowledge can greatly increase the robustness of classification to the sparse measurement and few-sample setting. Unlike earlier studies, we show how SPE can identify low-dimensional dynamics embedded in higher dimensional data, leading to a form of normalizing-flow-based model reduction. We focus on data domains described by prototypes representing oscillatory dynamics. This is a key test case for both theoretical and practical reasons. On the theoretical front, sustained, isolated oscillations (i.e. limit cycles) are only possible in nonlinear systems and the problem of detecting and localizing these cycles is notoriously challenging (it is one of David Hilbert’s famous unsolved mathematical challenges ( Hilbert 1902 ) ) even in systems with known, simple forms. On the practical front, the emergence of oscillations is of widespread importance in real-world nonlinear systems, particularly in engineering and biological applications. We showcase SPE on oscillations in gene expression, a critical dynamical phenomenon for various biological functions ( Novák & Tyson 2008 ) whose dis-regulation has bearing on the proper function of cells and whole tissues, such as in the case of the runaway cell division in cancer ( Suski et al. 2021 ; Ghafouri-Fard et al. 2022 ; Cavalu et al. 2024 ) . Further, the sparsity of single-cell RNA-sequencing gene expression measurements as well as their high-dimensionality (providing information for the activity levels of tens-of-thousands of genes in thousands of cells simultaneously) makes this domain an important showcase for our framework ( Bergen et al. 2021 ; Zheng et al. 2023 ) . Our concrete contributions include the following: • Efficient identification and localization of invariant structures from dynamical data . By our novel synthesis of prototype methods with data-driven dynamical systems, we strongly regularize the challenging problem of dynamical classification, specifically in the paradigmatic case of oscillatory dynamics. We first demonstrate how SPE can efficiently identify invariant structures in a synthetic data set of two-dimensional oscillatory dynamical systems drawn from across the physical and life sciences. Further, the optimized normalizing flows used in prototype matching can be used to identify invariant structures in the sample data, despite the absence of closed-form governing equations. These results are then extended to higher-dimensional systems, situating SPE as a form of

Read the complete text