Bio-inspired learning algorithm for time series using Loewner equation
Yusuke Kosaka Shibasaki 1 2 · 2025
Read Bio-inspired learning algorithm for time series using Loewner equation by Yusuke Kosaka Shibasaki 1 2 (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-2506-12372.txt.
Introduction
Yusuke Kosaka Shibasaki builds learning methods for time series from the Loewner equation. That equation is a tool from complex analysis and statistical physics. Links between statistical physics and machine learning are much discussed. But learning methods drawn from biological systems are still in growth. The Loewner equation encodes a growing curve into a driving signal. The author uses that unique encoding.
Two simple methods are proposed for one-dimensional time series. The first uses Gaussian process regression. It uses the normality of the Loewner driving force got from a curve built from the series. The second uses a fluctuation-dissipation relation from Loewner theory. It gauges how sensitive non-linear dynamics are under small shifts. Both methods were tested on neuronal dynamics from a leaky integrate-and-fire model.
Catalog record
| CREATOR | Yusuke Kosaka Shibasaki 1 2 |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 42 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2506-12372
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Bio-inspired learning algorithm for time series using Loewner equation Yusuke Kosaka Shibasaki 1 2 Abstract Though the relationship between the theoretical statistical physics and machine learning techniques has been a well-discussed topic, the studies on the mechanism of learning inspired by the biological system are still developing. In this study, we investigate the application methods of Loewner equation to the learning algorithm particularly focusing on its statistical-mechanical aspects. We suggest two simple methods of learning of one-dimensional time series based on the unique encoding property of the discrete Loewner evolution. The first one is the Gaussian process (GP) regression using the normality of the distribution of Loewner driving force corresponding to the curve composed from the time series. The second one is the fluctuation-dissipation relation (FDR) for the time series, which is derived from the Loewner theory, measuring the sensitivity of the nonlinear dynamics under the small perturbation. These methods were numerically tested dealing with the neuronal dynamics generated by the leaky integrate-and-fire model. In addition, we discuss the similarity between the mapping mechanism of the present method and the structure of biological information processing from a point of view of self-organization system theory. Publication type: Paper keywords Loewner equation; gaussian process regression; fluctuation-dissipation; neuronal dynamics; biological learning 1 Introduction Whereas the relationship between the technique of the machine learning in artificial intelligence (AI) and statistical physics is an essential issue in the current information-based science and humanities, one of the approaches that fills this gap is treating it as the problem of the complex systems [1-4]. Because the learning algorithm itself has been developed derived from the studies to model the functions of the brain (See, e.g., [5,6]), it is closely related to the nonlinear dynamics of the neural system and the question of how we treat its non-equilibrium (but robust) state as a dynamical system. As is well-known, the neural network method significantly contributes to the machine learning technique in AI enabling the computation of the nonlinear and high-dimensional problems [5-8], and this fact is also studied in statistical physics context. In the several studies, the methods in the research of non-equilibrium statistical physics (e.g., stochastic thermodynamics) have been applied to learning algorithm [9-11], where the mechanism of learning is interpreted in the physical sense. Such attempts are important not only for the development of the machine learning algorithm but also for our understanding of the physical meaning of the AI. As research of the general mechanisms of nonlinear (and thus non-equilibrium) phenomena in the real-world systems, the studies of Loewner equation [12-19] provide a worth discussing issue to consider the above problem. The previous studies on the physical application of the chordal Loewner equation [12] suggested that the unique encoding property of stochastic Loewner evolution (SLE) and discrete Loewner evolution [13,14], involving the chaotic dynamical system [16]. As is well-known, the chordal Loewner equation consists of two components, that is, the complex-valued curve \gamma_{\left[0,s\right]} in the upper half-plane \mathbb{H} and real-valued 1D function U_{s} . For the practical application, the increments \eta_{s}\left(n\right) of the Loewner driving function U_{s} converted from the discretized points on the curve \gamma_{\left[0,s\right]} obtained via the discrete Loewner evolution is useful for the analyses of the 2D lattice systems [18] and 1D time series [19]. We refer to this dynamical variable \eta_{s}\left(n\right) as the Loewner driving force, which has a one-to-one correspondence to the curve \gamma_{\left[0,s\right]} that we analyze. For the stationarity and mixing property of the Loewner driving force [16-18], the entropy of \eta_{s}\left(n\right) plays a central role in its formalism and the interpretation of complexity in the nonlinear dynamics [18,19]. We refer to this entropy as the Loewner entropy S_{Loew} , and it is expected to be an important quantity for the information analyses based on the Loewner equation. Because this Loewner equation-based information processing is enabled by the unique characteristic of conformal mapping system, the implementation of this theoretical framework to the machine learning method has a novelty and will provide some benefits to the understanding of the biological learning. For the above motivation, in this paper, we discuss the application method of discrete Loewner evolution to the learning algorithm. Particularly, we demonstrate that the relations obtained from the statistical-physical applications of Loewner equation (e.g., fluctuation-dissipation relation (FDR) and statistics of the Loewner driving force and entropy [18,19]) work as those help to answer this fundamental problem between machine learning and physics. As one of the simplest methods of learning, we briefly review the notions of “regression” in the machine learning [7,20]. This paper consists of the following parts. In Sec. 2.1., we present the basics of the Loewner equation, driving force, and entropy. In Sec. 2.2., as an example of the most familiar methods of leaning, the Gaussian process (GP) regression is briefly described. In Sec. 3.1., following the previous section, a possible form of the GP regression using discrete Loewner evolution is demonstrated. In Sec. 3.2., the FDR using the Loewner entropy is discussed in the context of machine learning. Subsequently, in Sec. 4., the numerical simulation of the present method is performed using time series obtained by the leaky integrate-and-fire model. Finally, in Sec. 5 and 6, the discussions and conclusion of the present study are remarked, respectively. 2 Model 2.1 Loewner Equation, Driving Force, and Entropy Let us consider the curve \gamma_{\left[0,s\right]} starting from the origin O of the upper half-plane \mathbb{H} . The family of the conformal map g_{s}\left(z\right)=z+2s/z+O\left(z^{-2}\right) (as z\rightarrow\infty ) that transforms the region \mathbb{H}\setminus\gamma_{\left[0,s\right]} to \mathbb{H} is given by the following (chordal) Loewner evolution [12,14-19]: \frac{\partial g_{s}\left(z\right)}{\partial s}=\frac{2}{g_{s}\left(z\right)-U_{s}},\ \ \ \ \ \ \ \ g_{0}\left(z\right)=z\in\mathbb{H}. (1) Here, U_{s} is a real-valued 1D function called the driving function. We note that Eq. (1) becomes the standard stochastic Loewner evolution (SLE) if we choose the driving function as U_{s}=\sqrt{\kappa}B_{s} , where B_{s} is the standard Brownian motion, i.e., Wiener process and \kappa is the constant parameter that determines the phases and fractal dimension d_{f} of the curve. The relation between the coordinate of the curve and driving function is expressed as: \lim_{z\rightarrow\gamma_{s}}{g_{s}\left(z\right)}=U_{s}. (2) The above relation ensures the one-to-one correspondence relation of the curve and driving function. For the practical calculation of simulating the curve \gamma_{\left[0,s\right]} from the driving function U_{s} , some discretizing methods of the Loewner equation (e.g., Refs. [13,14]). In this study, we employ the zipper-algorithm using the vertical slit map [14] briefly described as follows. Let us denote the curve \gamma_{\left[0,s\right]} as that consists of the sequence of discretized points on \mathbb{H} , i.e., \gamma_{\left[0,s\right]}=\{z_{0}(=0),z_{1},\ldots,z_{n},\ldots,z_{N}\} . Using the map h_{n}\left(z\right):=\sqrt{(z-\Delta U_{s_{n}})^{2}+4\Delta s_{n}} where \Delta U_{s_{n}}:=U_{s_{n}}-U_{s_{n-1}} and \Delta s_{n}:=s_{n}-s_{n-1} . The sequence of the increments of the driving function \{\Delta U_{s_{n}}\} and \{\Delta s_{n}\} corresponding to an arbitrary curve \gamma_{\left[0,s\right]} is obtained from the following recursive rule [14,16-18]: \begin{gathered}w_{1}=z_{1},\ w_{2}=h_{1}\left(z_{2}\right),\ldots,\ w_{n}=h_{n-1}\circ\ h_{n-2}\circ\cdots\circ\ h_{1}\left(z_{n}\right),\ \ldots,\\ \ w_{N}=\ h_{N-1}\circ h_{N-2}\circ\cdots\circ h_{1}\left(z_{N}\right)\end{gathered} (3) where \{\Delta U_{s_{n}}\} and \{\Delta s_{n}\} are obtained as [14,16-18]: w_{n}=\Delta U_{s_{n}}+2i\sqrt{\Delta s_{n}},\ \ \ \ \ i=\sqrt{-1}. (4) For the analyses of the geometrical properties of the \gamma_{\left[0,s\right]} , we define the following dynamical variable: \eta_{s}\left(n\right):=\frac{\Delta U_{s_{n}}}{\sqrt{\Delta s_{n}}},\ \ \ \ \ \Delta s_{n}\rightarrow 0^{+}, (5) which we call the Loewner driving force. For an arbitrary curve \gamma_{[0,s]} ,