Field Matching: an Electrostatic Paradigm to Generate and Transfer Data

Alexander Kolesov, Stepan I. Manukhov, Vladimir V. Palyulin, Alexander Korotin · 2025

Read Field Matching: an Electrostatic Paradigm to Generate and Transfer Data by Alexander Kolesov, Stepan I. Manukhov, Vladimir V. Palyulin, Alexander Korotin (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-2502-02367.txt.

Introduction

Alexander Kolesov and three co-authors make and move data by use of charge on plates. Their method sets a source spread on one plate of a capacitor. A target spread sits on the other plate, with charge of the flip sign. A neural net learns the field of force between the plates. Points then ride the field lines from one plate to the next.

Points that land, the proof shows, match the target spread. That one trick serves two jobs. It can make new data, and it can move data from one form to next. Toy tasks and image tasks show the method at work. The physics does real work here, not just a nice image. The field's shape moves the points where other plans lean on learned maps or noise steps. Their code is free for all to use.

Catalog record

CREATORAlexander Kolesov, Stepan I. Manukhov, Vladimir V. Palyulin, Alexander Korotin
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE66 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2502-02367

Browse the full library · About the Tower of Babel

Full text

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

Field Matching: an Electrostatic Paradigm to Generate and Transfer Data Alexander Kolesov, Stepan I. Manukhov, Vladimir V. Palyulin, Alexander Korotin Abstract We propose Electrostatic Field Matching (EFM), a novel method that is suitable for both generative modeling and distribution transfer tasks. Our approach is inspired by the physics of an electrical capacitor. We place source and target distributions on the capacitor plates and assign them positive and negative charges, respectively. Then we learn the electrostatic field of the capacitor using a neural network approximator. To map the distributions to each other, we start at one plate of the capacitor and move the samples along the learned electrostatic field lines until they reach the other plate. We theoretically justify that this approach provably yields the distribution transfer. In practice, we demonstrate the performance of our EFM in toy and image data experiments. Our code is available at https://github.com/justkolesov/FieldMatching . † † affiliationnotice: Equal contribution 1 Introduction The basic task of generative modeling is to learn a transformation between two distributions accessible by i.i.d. samples. The typical scenarios considered are noise-to-data ( Goodfellow et al., 2014 ) and data-to-data ( Zhu et al., 2017 ) . These are usually referred to as the unconditional data generation and data translation, respectively. Physics is often at the heart of the principles of generative modeling. One of the first attempts to link generative models and physics was made in Energy-Based models ( LeCun & Huang, 2005 , EBM) . They parameterize data distributions using the Gibbs-Boltzmann distribution density and generate data through simulation of Langevin dynamics ( Du & Mordatch, 2019 ; Song & Kingma, 2021 ) . Diffusion Models ( Sohl-Dickstein et al., 2015 ; Ho et al., 2020 , DM) is a popular class of generative models which is inspired by the nonequilibrium thermodynamics . The diffusion models consist of forward and backward stochastic processes ( Song et al., 2021 ) . The forward process corrupts the data via injection of Gaussian noise; the backward process reverses the forward process and recovers the data. Poisson Flow Generative Models ( Xu et al., 2022 ; Xu et al., 2023 , PFGM) use ideas from the electrostatic theory for the data generation process, recovering an electric field between a hyperplane of the data and a hemisphere of large radius. Both DM and PFGM use physical principles to corrupt data, simplifying the data distribution to a tractable one. As a result, they are only used directly for noise-to-data tasks. Figure 1 : Our Electrostatic field matching (EFM) method. Two data distributions \mathbb{P}(\textbf{x}^{+}) and \mathbb{Q}(\textbf{x}^{-}),\;\textbf{x}^{\pm}\in\mathbb{R}^{D} are placed in the space \mathbb{R}^{D+1} in the planes z=0 and z=L , respectively. The distribution \mathbb{P}(\textbf{x}^{+}) is assigned a positive charge, and the distribution \mathbb{Q}(\textbf{x}^{-}) – a negative charge. These charges create an electric field {\textbf{E}}(\widetilde{\textbf{x}}) , where \widetilde{\textbf{x}}=(\textbf{x},z)\in\mathbb{R}^{D+1} . The lines of the field begin at positive charges and end at negative charges. Movement along the electric field lines provably (see our Theorem 3.1 ) transforms the distribution \mathbb{P}(\textbf{x}^{+}) into the distribution \mathbb{Q}(\textbf{x}^{-}) . Recently, modifications of DM have appeared that can learn diffusion in a data-to-data scenario. Diffusion Bridge Matching ( Shi et al., 2024 ; Albergo & Vanden-Eijnden, 2023 ; Peluchetti, 2023 , BM) is an SDE-based method that recovers the continuous-time Markovian process between data distributions. Flow matching ( Lipman et al., 2023 ; Liu et al., 2023 ; Klein et al., 2024 ; Chen & Lipman, 2024 ; Xie et al., 2024 , FM) is the limiting case of BM that learns ODE-based transformation between distributions. However, there is no method based on electrostatic theory that can be applied to data-to-data translation tasks. Contributions. We propose and theoretically justify a new paradigm for generative modeling called Electrostatic Field Matching (EFM). It is based on the electrostatic theory and suitable for both noise-to-data and data-to-data generative scenarios. We provide proof-of-concept experiments on low- and high-dimensional generative modeling tasks. 2 Background and Related Works 2.1 Basic physics To understand the physics behind the electrostatic field matching method, let us recall some basic background from standard Maxwell’s 3D-electrostatics and then generalize it to the case of D dimensions. Information on Maxwell’s electrostatics can be found in any electricity textbook, for instance ( Landau & Lifshitz, 1971 , Chapter 5) . 2.1.1 Maxwell’s electrostatics 1 1 1 All formulas are written in the Heaviside–Lorentz system of units, where Planck’s constant \hbar=1 , the speed of light c=1 , and the electric constant, which stands as a multiplier in Coulomb’s law (see ( 1 )), is k=1/(4\pi) . This system of units is convenient for our purposes because formulas look particularly simple in this system of units (the Gauss’s theorem and the circulation theorem). The field of a point charge. Let a point charge q\in\mathbb{R} be located at a point \textbf{x}^{\prime}\in\mathbb{R}^{3} . At a point \textbf{x}\in\mathbb{R}^{3} it creates an electric field 3 3 3 The meaning of electric field is as follows. If a charge q_{0} is placed in an electric field, then the force acting on q_{0} equals to \textbf{F}=q_{0}\textbf{E} . Using ( 1 ), we obtain Coulomb’s law of interaction of point charges: \textbf{F}=k\frac{qq_{0}}{||\textbf{x}-\textbf{x}^{\prime}||^{3}}(\textbf{x}-\textbf{x}^{\prime}) , where k=\frac{1}{4\pi} . \textbf{E}(\textbf{x})\in\mathbb{R}^{3} equal to: \textbf{E}(\textbf{x})=\frac{q}{4\pi}\frac{\textbf{x}-\textbf{x}^{\prime}}{||\textbf{x}-\textbf{x}^{\prime}||^{3}}. (1) The superposition principle. If point charges q_{1},q_{2},...,q_{N} are located at points \textbf{x}_{1},\textbf{x}_{2},...,\textbf{x}_{N} , they create independent fields \textbf{E}_{1}(\textbf{x}),\textbf{E}_{2}(\textbf{x}),...,\textbf{E}_{N}(\textbf{x}) at a given point \textbf{x}\in\mathbb{R}^{3} . All these charges together produce the following field: \textbf{E}(\textbf{x})=\sum_{n=1}^{N}\textbf{E}_{n}(\textbf{x})=\sum_{n=1}^{N}\frac{q_{n}}{4\pi}\frac{(\textbf{x}-\textbf{x}_{n})}{||\textbf{x}-\textbf{x}_{n}||^{3}}, (2) In the general case, we are dealing with a continuously distributed charge q(\textbf{x}) . Then the superposition principle can be written as: \textbf{E}(\textbf{x})=\int\frac{1}{4\pi}\frac{(\textbf{x}-\textbf{x}^{\prime})}{||\textbf{x}-\textbf{x}^{\prime}||^{3}}q(\textbf{x}^{\prime})d\textbf{x}^{\prime}. (3) The charge distribution q(\textbf{x}) can have values greater than zero (positive charge) or less than zero (negative charge). An electric field strength line is a curve \textbf{x}(t)\in\mathbb{R}^{3},\;t\in[a,b]\subset\mathbb{R} whose tangent to each point is parallel to the electric field at that point. In other words: \frac{d\textbf{x}(t)}{dt}=\textbf{E}(\textbf{x}(t)),\text{where }t\in[a,b]\subset\mathbb{R}. (4) Electric field lines are the key concept in our work. The basic properties of these lines follow from Gauss’s theorem and the circulation theorem formulated below. The rigorous formulations and proofs of the field line properties are given in Appendix A (for a more general D -dimentional case). The electric field flux . Consider an element of area dS . It is a vector whose length is equal to the considered area dS and whose direction is orthogonal to this area. For closed surfaces, the direction is always selected outward. The electric field flux E through this element is d\Phi=\textbf{E}\cdot\textbf{dS} , i.e., the inner product between vectors E and dS . The field flux through a finite surface \Sigma is given by \Phi=\int_{\Sigma}d\Phi=\iint_{\Sigma}\mathbf{E}\cdot\mathbf{dS}. (5) It indicates the density of field lines passing through \Sigma . Gauss’s theorem ( Landau & Lifshitz, 1971 , \wasyparagraph 31) . For any closed two-dimensional surface \partial M , which bounds the set M\subset\mathbb{R}^{3} (see Fig. 2 ), the electric field flux is equal to the total charge enclosed by this surface: \iint_{\partial M}\textbf{E}\cdot\textbf{dS}=\int_{M}q(\textbf{x})d\textbf{x}. (6) Figure 2 : An illustration of the Gauss’s theorem. Intuitively, the Gauss’s theorem states that the number of lines passing through any closed surface is determined only by the charge inside that surface (and is proportional to it). In particular, it follows from Gauss’s theorem that the electric field line must

Read the complete text