Field Matching: an Electrostatic Paradigm to Generate and Transfer Data

Alexander Kolesov; Alexander Korotin; Manukhov Stepan; Vladimir V. Palyulin

REVIEW 4 major objections 3 minor 2 cited by

Field Matching: an Electrostatic Paradigm to Generate and Transfer Data

T0 review · 4 major / 3 minor · reviewed 2026-08-09 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read This paper claims that riding electrostatic field lines between two charged plates provably maps one data distribution onto another, so a single learned field handles both noise-to-image generation and image-to-image translation.

desk verdict A genuinely new electrostatic formulation for distribution transfer whose implemented algorithm is not the map the theorem covers; worth a serious referee but only with the gap addressed. read the letter →

arxiv 2502.02367 v3 pith:2XFCTDJ5 submitted 2025-02-04 cs.LG cs.AIcs.CV

Alexander Kolesov , Manukhov Stepan , Vladimir V. Palyulin , Alexander Korotin This is my paper · ORCID

classification cs.LGcs.AIcs.CV

keywords electrostaticfieldmatchinggenerativemodelingdistributiontransferelectriclinesPoissonflowmodelsneuralODEunpairedimage-to-imagetranslation

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Electrostatic Field Matching (EFM) treats two data distributions as positive and negative charge densities on two parallel plates of a $(D+1)$-dimensional capacitor and learns the resulting electric field with a neural network. To transfer a sample, one integrates along the field line from its starting plate until the line reaches the other plate. The paper's central claim is Theorem 3.1: because the two charges cancel and occupy bounded regions, field lines almost surely run from the positive plate to the negative plate, so moving along lines sends samples distributed as $P$ to samples distributed as $Q$. This makes one framework serve both noise-to-data generation and data-to-data translation, demonstrated on toy, colored-MNIST, and CIFAR-10 experiments. If the claim is right, EFM is a space-conditional transport rule that competes with time-conditional flow-matching methods.

What carries the argument

The central object is the $D$-dimensional electrostatic field $E(\tilde{x}) = \frac{1}{S_D}\int \frac{\tilde{x} - \tilde{x}'}{\|\tilde{x} - \tilde{x}'\|^{D+1}} q(\tilde{x}') d\tilde{x}'$ produced by the two plate charges, whose integral curves are the field lines $\frac{d\tilde{x}}{dt} = E(\tilde{x})$. Three classical facts carry the argument: Gauss's theorem (flux through a closed surface equals enclosed charge), the circulation theorem (no closed field-line loops), and the consequence drawn in Lemma A.6 that for a compactly supported, zero-total-charge system, field lines almost surely run from positive to negative charge. The transport itself is the stochastic map $T$ of equations (15)-(18), which routes each point along forward- or backward-oriented field lines with probabilities proportional to the flux and stops at the opposite plate with probability $\nu$ when a line crosses it. The proof of Theorem 3.1 is a chain of flux-conservation equalities along stream tubes (bundles of neighboring field lines) showing that every region of the target plate receives exactly the sample count its flux prescribes. The practical engine is the neural network $f_\theta$ fitted to the normalized field $E/\|E\|$, with the $z$-coordinate ODE of equation (23) used for inference.

What would settle it

Take two well-separated Gaussians as $P$ and $Q$, compute the exact field, and compare the full stochastic map $T$ (which continues past $z = L$ with probability $\nu$) against the practical rule that always stops at the first crossing; the Wasserstein distance between mapped samples and $Q$ should be zero for the full map and measurably positive for the truncated one if the theorem's continuation step is load-bearing. A second check: in the noise-to-image setting, count trajectories that escape a large bounding box before reaching the target plate — Lemma A.6 forces that fraction to zero for compact sources, so a non-negligible fraction would show the premise is violated.

Watch

Extended reading notes

Core claim

On its own terms, the paper's discovery is that the physical capacitor provides a complete generative transport: assign charge density $q_+ = P$ on the hyperplane $z = 0$ and $q_- = -Q$ on $z = L$, and the superposition field $E(\tilde{x}) = E_+(\tilde{x}) + E_-(\tilde{x})$ of this neutral system has field lines that, by Lemma A.6, almost surely start on positive charge and end on negative charge. The paper defines a stochastic transport map $T$ that follows these lines, choosing between forward- and backward-oriented trajectories with probabilities proportional to the electric flux, and proves in Theorem 3.1 that if $x_+ \sim P$ then $T(x_+) \sim Q$ almost surely. The proof works by flux conservation: the fraction of sample trajectories arriving at any region of the target plate equals the electric flux through that region, which Gauss's theorem identifies with the target density. In practice the field is learned by fitting a neural network to the normalized ground-truth field at points interpolated between the plates, and inference is an ODE integration in the $z$-coordinate, stopped when the trajectory reaches the opposite plate.

Load-bearing premise

The guarantee rests on the fact that in a zero-total-charge system of bounded charges, every field line runs from the positive plate to the negative plate almost surely; that fact requires both distributions to be compactly supported, and the paper's noise-to-image experiments use Gaussian noise, which is not compactly supported — a violation the paper does not flag.

Editorial extensions

If this is right

A single learned electrostatic field serves both unconditional generation (noise to data) and unpaired translation (data to data); only the plate contents change.
The transport is stochastic but exact in distribution: the empirical measure of mapped samples converges almost surely to $Q$.
Inference is a geometric ODE: the integration stops when the trajectory reaches the opposite plate at $z = L$, with no learned time horizon.
The inter-plate distance $L$ is a key hyperparameter: small $L$ keeps field lines nearly straight, large $L$ makes them curved and harder to fit.
The guarantee covers the full support of $Q$ only when both forward- and backward-oriented field lines are used, which is why the authors note their experiments rely on forward lines alone.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Inference: because the practical implementation always stops at the first crossing of $z = L$ and skips the continuation probability $\nu$, the guarantee of Theorem 3.1 should degrade exactly where the mean shift between $P$ and $Q$ is large; a histogram comparison of mapped samples against $Q$ would reveal where.
Inference: the noise-to-image experiments use Gaussian white noise, which is not compactly supported, so Lemma A.6's premise is formally violated there; a testable extension is to measure how much mapped mass lands outside the data manifold for heavier-tailed sources.
Inference: the field-line map suggests a transport-cost or displacement interpretation the paper does not pursue, since the exact field determines a deterministic flux-proportional routing between the two densities.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

The two-plate electrostatic formulation is genuinely new. PFGM only handles noise-to-data with a hemisphere at infinity; putting source and target distributions on two charged plates and proving that field-line transport pushes P to Q is a real step beyond that. Theorem 3.1 is the right kind of contribution: a clean construction, a standard flux-conservation proof, and it doesn't depend on any fitted parameters or prior results from the authors. They also ship code and are unusually honest about limitations. That is all to the good.

The soft spot is not the physics; it is the distance between the theorem and the algorithm. Theorem 3.1 covers the stochastic map T defined in (16)-(17), with first-crossing continuation and backward-oriented lines. Algorithm 2 stops every trajectory at the first crossing of z=L and drops the backward component entirely. The authors admit this in Section 5 and Appendix C.2, saying the practice 'may lead to incorrect learning of the target data distribution.' That means the theoretical guarantee does not apply to the actual samplers used in the experiments. No error bound or corrected variant is supplied, so the empirical results are not backed by the theorem. The compact-support assumption is a second, independent gap: the noise-to-data experiments use Gaussian noise, which is not compactly supported, so Lemma A.6 and Theorem 3.1 do not apply there either. The paper does not flag this. These are real issues, but they are not hidden or fatal in the sense of the theory being wrong; they are open problems in bridging a clean theoretical construction to a practical sampler.

The experiments are proof-of-concept, not evidence of SOTA. The image results are qualitative and the baselines are a mixed bag. For a paper whose selling point is the theory, that is acceptable at this stage, but it means the practical claims should be read cautiously.

Who is this for? Anyone working on flow matching, PFGM, or distribution transfer will want to see this construction. It deserves a serious referee, not a desk rejection. My recommendation: send it to review, but tell the authors the central condition is to either modify the algorithm to match the theorem (e.g., implement the stopping probability and include backward lines where needed) or provide a quantitative bound on the error introduced by stopping early. A few FID or distribution-distance numbers on the toy and MNIST tasks would also help separate approximation error from algorithmic bias.

Referee Report

4 major / 3 minor

Summary. The paper proposes Electrostatic Field Matching (EFM), a generative and distribution-transfer method inspired by the electrostatics of a capacitor. Source and target distributions are placed as positive and negative charge densities on two parallel hyperplanes in (D+1)-dimensional space, the electrostatic field is approximated by a neural network, and samples are transported along field lines. The central theoretical claim, Theorem 3.1, states that a stochastic map T combining forward-oriented and backward-oriented field lines transfers P to Q almost surely. The paper reports proof-of-concept experiments on a 2D toy problem, colored MNIST translation, and MNIST/CIFAR-10 generation, and the authors acknowledge in Section 5 and Appendices C.2-C.3 that the practical algorithm stops trajectories at the first crossing of z=L and uses only forward-oriented lines.

Significance. The idea of using electrostatic flux conservation to construct an exact transport map is novel in the data-to-data setting and the derivation from Gauss's law is a genuine, non-circular theoretical contribution. The paper provides a self-contained proof, a detailed appendix, and a public code link, which are strengths. If the implemented algorithm matched the theorem, this would be a valuable addition to the generative-modeling toolbox. However, the significance is currently tempered by a clear mismatch between the theorem's stochastic map and the deployed deterministic first-crossing sampler, as well as by the violation of the theorem's compact-support assumption in the noise-to-image experiments.

major comments (4)

[Section 3.3, Algorithm 2, Eqs. (16)-(17)] Theorem 3.1 is proven for the stochastic map T defined by Eqs. (16)-(17), which includes continuing past z=L with probability 1-ν(x-_F) and the backward-oriented component TB selected with probability 1-µ(x+). Algorithm 2, however, stops every trajectory at the first crossing of z=L and uses only forward-oriented lines. The text in Section 3.3 admits this ('in practice we simply stop'), and Appendix C.2 states that this practice 'may lead to incorrect learning of the target data distribution.' Therefore Theorem 3.1 does not establish that the output of Algorithm 2 is distributed as Q, and no error bound is provided. This is the central load-bearing gap: either the sampling procedure must implement the continuation and backward choices, or the correctness claim must be replaced with a quantified approximate statement.
[Section 4.3, Theorem 3.1 and Lemma A.6] Theorem 3.1 and Lemma A.6 assume compactly supported distributions P and Q. In the noise-to-image experiments of Section 4.3, the source P is white Gaussian noise, which has unbounded support, so the theorem's assumptions are violated. The paper does not flag this violation. Since Lemma A.6 is what guarantees that field lines terminate on the target distribution almost surely, the theoretical justification does not cover these experiments. The authors should either use a bounded prior (e.g., a truncated Gaussian) or supply a separate argument extending the result to non-compact sources.
[Section 3.3, Eqs. (22)-(23)] Algorithm 2 learns the normalized field fθ via Eq. (22) and integrates the approximate ODE (23), replacing both the exact electric field E and the stochastic map T. Normalization can distort the dynamics near points where Ez is small, and the stopping rule 'stop when z=L' is executed with the learned, normalized field. No bound or quantitative statement connects the approximation error in fθ to the deviation of the output distribution from Q. Given that the theorem is exact only for the ideal stochastic map, the manuscript needs at least an explicit statement of what the experiments are intended to demonstrate about the approximate algorithm.
[Appendix B, proof of Theorem B.5, step 11] The final normalization step of the proof writes dΦ'_F/Φ0 + dΦ'_B/Φ0 = Q dS, but Φ0 is never defined. Earlier steps establish dn'_F/n → dΦ'_F and dn'_B/n → dΦ'_B, and Gauss's theorem gives the total flux through the target plate as the total charge, which is 1, but the normalization constant and the sign/direction conventions in the chain of equalities (53) should be stated explicitly. This is a gap in the written proof, though it appears fixable.

minor comments (3)

[Section 3.3, Eq. (20) and Algorithm 1] The interpolation formula in Eq. (20) reads ex = (t/L)ex- + (1 - t/L)ex+ + eε, but Algorithm 1 writes ex = (t/L)ex+ + (1 - t/L)ex- + eε. The coefficients are swapped; the two should be made consistent.
[Throughout] There are several typographical errors: 'betwenn' (Table 1), 'neccesity' (Section 4.1), 'dimentional' (Section 2.2), 'Hovewer' (Appendix C.1), and inconsistent cross-reference labels 'M3.1', 'M4.1', and 'M5.3' instead of section numbers. These should be corrected.
[Appendix C.2, last sentence] The sentence 'Note that such a motion should indeed end on the distribution Q (Lemma A.6), and also that further movement is impossible (since E+_z(L) < 0, and between plates z ∈ (0,L) we always have Ez > 0, in particular E−_z(L) > 0)' is hard to parse; the sign conventions for the left and right limits E±_z(L) should be defined explicitly in this appendix.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the EFM transport theorem is derived from Gauss's law and flux conservation, with no fitted parameter or load-bearing self-citation.

full rationale

The derivation chain is self-contained. The paper sets q_+ = P and q_- = -Q (Sec. 3.2), defines the field E by Coulomb superposition (Eq. 14), and defines the stochastic maps T_F and T_B by field-line integration with stopping and forward-selection probabilities nu and mu given explicitly in Eqs. (15) and (18) as electric-flux ratios. Theorem 3.1 is proved in Appendix B using consequences of Gauss's law (Lemmas B.1-B.4) and flux conservation along field tubes (Lemmas A.3 and A.6). No free parameter is fitted to target data and then renamed a prediction: nu and mu are derived field quantities, not learned parameters, and the proof does not assume T(P)=Q. The only self-citation (Gushchin et al. 2024 for the colored-MNIST 3-to-2 task setup) is not load-bearing for the theorem or the method. The paper explicitly acknowledges that the implemented Algorithm 2 stops at the first crossing of z=L and uses only forward-oriented lines (Sec. 3.3; Sec. 5; App. C.2, C.3), so the exact Theorem 3.1 statement does not cover the implemented sampler; this is a correctness or robustness gap, not circularity. Similarly, the Gaussian white-noise prior in Sec. 4.3 violates the compact-support assumption of Lemma A.6, an unstated assumption violation rather than a definitional reduction. Overall, no load-bearing step of the derivation reduces to its own input by construction.

Assumptions & free parameters 3 free parameters · 5 assumptions · 1 invented entities

The central theorem rests on standard electrostatics plus compact-support assumptions. The hand-tuned hyperparameters L and sigma affect practical performance but not the theoretical guarantee. No new physical entity is proposed beyond the capacitor construction used to define the field.

free parameters (3)

inter-plate distance L = L=6, 10, 30, 500, 5000 by experiment
Hand-tuned per task; the paper shows quality degrades as L grows and gives no selection rule (Appendix C.1).
training volume noise scale sigma = 0.001 to 0.05 across experiments
Hyperparameter controlling spread of training points in (21); chosen empirically.
epsilon offset near plates = small positive value, not specified in main text
The forward map starts at z=epsilon and the left/right field limits are evaluated at +/- epsilon; the exact value is not reported.

assumptions (5)

standard math Gauss's theorem and circulation theorem in D-dimensional electrostatics (Landau and Lifshitz)
Foundations for Lemmas A.3 to A.7 and for the flux arguments in the proof of Theorem B.5.
domain assumption Distributions P and Q have compact support (Theorem 3.1)
Used in Lemma A.6 to guarantee all field lines terminate on the plates; violated by the Gaussian white noise experiments in Section 4.3.
domain assumption The learned normalized field f_theta approximates the true normalized field well enough along ODE trajectories
Training objective (22) is an unnormalized MSE with no error control; inference (23) uses this approximation without a bound.
ad hoc to paper Stopping probabilities nu and mu in (15) and (18) are exactly computable
The theorem's map T depends on these probabilities, but the implementation says accurate estimation is non-trivial and therefore stops at the first crossing.
domain assumption Training volume (20) with Gaussian noise adequately covers the region where field lines travel
No theoretical coverage guarantee; Appendix C.4 shows alternatives behave similarly on one toy example only.

invented entities (1)

D+1 dimensional capacitor with distributions as plate charges
purpose: Defines the electric field whose flow transports samples from P to Q
A mathematical construction for the algorithm, not an independently measurable physical entity.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Field Matching: an Electrostatic Paradigm to Generate and Transfer Data." pith.science (2026). https://pith.science/paper/2XFCTDJ5

@misc{pith2026250202367,
  author       = {Pith},
  title        = {Pith review of: Field Matching: an Electrostatic Paradigm to Generate and Transfer Data},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2XFCTDJ5}},
  note         = {Machine review of arXiv:2502.02367}
}

read the original 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

Figures

Figures reproduced from arXiv: 2502.02367 by the authors.

**Figure 1.** Our Electrostatic field matching (EFM) method. Two data distributions P(x +) and Q(x −), x ± ∈ R D are placed in the space R D+1 in the planes z = 0 and z = L, respectively. The distribution P(x +) is assigned a positive charge, and the distribution Q(x −) – a negative charge. These charges create an electric field E(ex), where ex = (x, z) ∈ R D+1. The lines of the field begin at positive charges and end at negativ… view at source ↗

**Figure 2.** An illustration of the Gauss’s theorem. where dl is the length element of the closed loop ℓ. This length element is a vector whose length is determined by the infinitesimal segment dl, see [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** An illustration of the electric field circulation theorem. It follows from the circulation theorem that there are no field lines which form closed loops, see Lemma A.7. 2.1.2. D-DIMENSIONAL ELECTROSTATICS The generalization of electrostatic equations for higher dimensions appears in discussions related to the influence of extra dimensions on physics (Ehrenfest, 1917; Gurevich & Mostepanenko, 1971; Caruso et al., 20… view at source ↗

Figures from the paper (19 more)

**Figure 4.** Figure 4: PFGM concept. The original data have a distribution P0(ex), which is assigned a positive charge that produces an electric field E(ex). Near the plate (Fig. 4a), the field lines can have a complex structure, while away from the plate (Fig. 4b) the charge looks like a po…

**Figure 5.** Figure 5: An illustration of the forward map TF . In the first case, the movement ends. In the second case, it is necessary to stop the movement at the point ex − F with probability ν(x − F ) and continue the movement into the region z > L with probability 1 − ν(x − F ), where …

**Figure 6.** Figure 6: Illustrative 2D Gaussian→Swiss Roll experiment: input and target distributions P(x +) and Q(x −) together with the result of the distribution transfer learned with our EFM method for distances L = 6 and L = 30 between the capacitor plates. P(x +) and the Swiss Roll dis…

**Figure 8.** Figure 8: Electric field line structure for the Gaussian→Swiss Roll experiment with L = 6 and L = 30. It can be seen that at large distances, the field lines are more curved than at small distances. defined correspondence, see (Zhu et al., 2017, [PITH_FULL_IMAGE:figures/full_fi…

**Figure 7.** Figure 7 [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]

**Figure 10.** Figure 10: Image-to-Image translation. Pictures from the initial distribution, the result of applying our EFM method as well as the Flow Matching method are presented respectively. (a) White noise samples from P(x +), which are placed on the left plate z = 0. (b) Samples from ou…

**Figure 11.** Figure 11: Noise-to-Image generation. Pictures from the initial distribution (Fig. 11a), the result of our EFM method (Fig. 11b) as well as the PFGM method (Fig. 11c) are presented. (a) White noise samples from P(x +) which are placed on the left plate z = 0. (b) Samples from ou…

**Figure 12.** Figure 12: Noise-to-Image generation. Pictures from the initial distribution (Fig. 12a), the result of our EFM method (Fig. 12b) as well as the PFGM method (Fig. 12c) are presented. accurately translate the shape and color of the initial digits 3. In Appendix D, we demonstrate t…

**Figure 13.** Figure 13: The electric field flux through an arbitrary stream tube. Lemma A.3 (Conservation of the field flux.). The electric field flux is conserved along a stream surface if there are no charges inside that surface. Proof. Consider an arbitrary stream tube ( [PITH_FULL_IMAGE…

**Figure 14.** Figure 14: The field flux of a point charge q0 through an arbitrary surface Σ seen at solid angle Ω. Φ = q0Ω SD−1 . (27) Proof. Divide Σ into small surface elements dS. The total flux is the integral over the entire surface, Φ = R Σ dΦ. By definition of flux (Definition A.1), an…

**Figure 15.** Figure 15: Closed loop of an electric field line. This situation is impossible due to the circulation theorem. Lemma A.7 (Absence of closed electric field loops). Electric field lines cannot form closed loops (as shown in [PITH_FULL_IMAGE:figures/full_fig_p013_15.png]

**Figure 16.** Figure 16: The considered area. The flux through this surface consists of three summands: dΦ +, the flux in the positive direction of the z-axis; dΦ −, the flux in the negative direction of the z-axis; and dΦlat, the flux through the lateral surface: dΦfull = dΦ + + dΦ − + dΦlat…

**Figure 17.** Figure 17: An element of volume dS′ is selected on the distribution Q(·). dn′ is the number of points x − i falling into this volume. Some points came to the distribution Q(·) from the front (denote them by dn′ F ), and some came from the back (dn′ B). 5. During the motion again…

**Figure 18.** Figure 18: Intersection points. Their corresponding area elements are dS′ = dS0 → dS1 → · · · → dSN → dSN+1 = dS. (46) Point counts in these areas are dn′ = dn0 → dn1 → · · · → dnN → dnN+1 = dn, (47) where dnk (k = 0, ..., N + 1) is number of points from sample {xi} n i=1 or fro…

**Figure 19.** Figure 19: The influence of the interplate distance L on the performance of the Noise-to-Image generation of CIFAR-10 dataset. Pictures EFM method with L = 50 (Fig. 19a), L = 500 (Fig. 19b) as well as L = 5000 (Fig. 19c) are presented. C.2. The problem of lines going beyond z = …

**Figure 20.** Figure 20: A toy experiment Gaussian→ 2Gaussians. The two Gaussians are significantly separated from each other, so a large number of lines flying out of the region z = L is observed. Further integration along the field lines still leads to the target distribution. using backwar…

**Figure 21.** Figure 21: An illustration of forward-oriented and backward-oriented lines. When source (P) and target (Q) distributions exhibit significant mean shifts (µP ̸= µQ), the number of forward-oriented lines extending beyond the z = L boundary can be quite large. In such a situation, …

**Figure 22.** Figure 22: showcases the equivalent performance for both approaches in Gaussian-to-Swiss-roll transport. (a) Training volume for our EFM method that is defined by interpolation from (20). (b) Training volume for our EFM method that is defined by uniform cube mesh between plates.…

**Figure 23.** Figure 23: Image-to-Image translation experiment (Colored MNIST dataset, 3 → 2). The results of alternative translation methods: Cycle-GAN (Zhu et al., 2017), DDIB (Su et al., 2023) and DSBM (De Bortoli et al., 2024) 21 [PITH_FULL_IMAGE:figures/full_fig_p021_23.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

Overclocking Electrostatic Generative Models
cs.LG 2025-09 conditional novelty 6.0 of 10

IPFM trains a few-step generator to reproduce the teacher PFGM++ electrostatic field, matches or beats teacher FID on CIFAR-10 and FFHQ, and recovers SiD as D goes to infinity.
Data Generation without Function Estimation
cs.LG 2025-07 conditional novelty 6.0 of 10

Deterministic particle gradient descent, run forward to uniformity and backward from a new point, generates samples from a data distribution without estimating a score function.

Reference graph

Works this paper leans on

17 extracted references · 16 canonical work pages · cited by 2 Pith papers

[1]

The full flux is a sum of fluxes through the ends of the tube and through its lateral surface: Φf ull= Φ1 + Φ2 + Φlat

We have to prove that Φ′ 1 = Φ2. The full flux is a sum of fluxes through the ends of the tube and through its lateral surface: Φf ull= Φ1 + Φ2 + Φlat. (25) By the definition of a stream tube, the flux through the lateral surface must be zero: Φlat = 0, because the field E and the normal to the lateral surface are orthogonal. Thus, Φf ull= Φ1 + Φ2 = Φ2 − Φ′

work page
[2]

Then some number dn′ ≤ dn of lines inside current tube (corresponding to sample {x+ i }n i=1 ∼ P) must terminate, while the others should continue moving

E+ z (ex−) and E− z (ex−) have the same sign. Then some number dn′ ≤ dn of lines inside current tube (corresponding to sample {x+ i }n i=1 ∼ P) must terminate, while the others should continue moving. This portion dn′ can be found from the line termination property in Q. Lemma B.4 (Line Termination). If E+ z (ex−) and E− z (ex−) have the same sign upon cr...

work page
[3]

American Journal of Physics, 91(2):153–158, 2023. Chen, R. T. and Lipman, Y . Flow matching on general geometries. In The Twelfth International Conference on Learning Representations, 2024. De Bortoli, V ., Korshunova, I., Mnih, A., and Doucet, A. Schrodinger bridge flow for unpaired data translation. Advances in Neural Information Processing Systems, 37:...

arXiv 2023
[4]

(26) It follows that Φ′ 1 = Φ2

From Gauss’s theorem (6), we derive: Φf ull= x ∂M E · dS = Z M q(x)dx = 0. (26) It follows that Φ′ 1 = Φ2. Corollary A.4 (Impossibility of line termination in empty space). An electric field line cannot terminate in empty space. Proof. Otherwise, the field flux inside the current tube surrounding the termination point will not be conserved. Intuitively, n...

work page 1971
[5]

Then the field line motion terminates in this case

E+ z (ex−) and E− z (ex−) have opposite signs. Then the field line motion terminates in this case. 14 Field Matching: an Electrostatic Paradigm to Generate and Transfer Data

work page
[6]

Let dn′ denote the number of points x− i within this region

Consider a D-dimensional area element dS′ on Q(·). Let dn′ denote the number of points x− i within this region. We define the empirical distribution: ˆQn(x−)dS′ = dn′ n . (42) Our goal is to prove: ˆQn(·) almost surely − − − − − − − → n→∞ Q(·). (43)

work page
[7]

17), forward-oriented (dn′ F ) and backward-oriented (dn′ B) trajectories: dn′ = dn′ F + dn′ B

The dn′ points arrive at Q(·) through two pathways (Fig. 17), forward-oriented (dn′ F ) and backward-oriented (dn′ B) trajectories: dn′ = dn′ F + dn′ B. (44)

work page
[8]

Since P(·) and Q(·) have compact supports and the total charge related to the system of the two plates is zero, by Lemma A.6, these tubes must start from P(·) almost surely

Consider the current tube terminating at dS′ corresponding to forward-oriented arrivals (dn′ F ). Since P(·) and Q(·) have compact supports and the total charge related to the system of the two plates is zero, by Lemma A.6, these tubes must start from P(·) almost surely. 15 Field Matching: an Electrostatic Paradigm to Generate and Transfer Data Figure 17....

work page

Show all 17 references

[9]

During the motion against field lines from Q to P, multiple crossings of z = 0 and/or z = L may occur (N = 0, 1 or 2, see Fig. 18). Denote the intersection points: x− = x0 → x1 → · · · →xN → xN +1 = x+. (45) Figure 18. Intersection points. Their corresponding area elements are...
[10]

Only the charged planes ( z = 0 or z = L) can alter the count: • At zk = 0: Line count increases by dnk

The dnk are not arbitrary but related by flux conservation. Only the charged planes ( z = 0 or z = L) can alter the count: • At zk = 0: Line count increases by dnk. • At zk = L: Line count decreases by dnk. 16 Field Matching: an Electrostatic Paradigm to Generate and Transfer ...
[11]

− − − − → n→∞ dΦN +1 ≡ dΦ, (50)

Due to the first Lemma B.2 on flow: dnN +1 n ≡ dn n a.s. − − − − → n→∞ dΦN +1 ≡ dΦ, (50)
[12]

− − − − → n→∞ dΦafter,k − dΦbefore,k

Due to the second Lemma B.3 on the flow, and because of the line termination Lemma B.4: (−1)fk · dnk n a.s. − − − − → n→∞ dΦafter,k − dΦbefore,k. (51)
[13]

According to the law of conservation of flux along the tube (Lemma A.3) from xi to xi−1: dΦafter,k = dΦbefore,k−1. (52)
[14]

+ (−1)fN dnN n + dn n ⇒ dn′ F n = (−1)f1 dn1 n +

Whence we obtain a chain of equalities: 0 = N +1X k=0 (−1)fk dnk = − dn′ F n + (−1)f1 dn1 n + ... + (−1)fN dnN n + dn n ⇒ dn′ F n = (−1)f1 dn1 n + ... + (−1)fN dnN n + dn n a.s. − − − − → n→∞ a.s. − − − − → n→∞ −dΦafter,1 − dΦbefore,1 + ... + dΦafter,N − dΦbefore,N + dΦN +1 = ...
[15]

− − − − → n→∞ dΦ′ F Φ0 + dΦ′ B Φ0 = QdS

Then, by virtue of the Gauss’s theorem (6), we finally have ˆQndS′ = dn′ n = dn′ F n + dn′ B n a.s. − − − − → n→∞ dΦ′ F Φ0 + dΦ′ B Φ0 = QdS. (56) This completes the proof. 17 Field Matching: an Electrostatic Paradigm to Generate and Transfer Data C. Extended discussion of limi...
[16]

The field line motion terminates in this case

E+ z (L) and E− z (L) have opposite signs. The field line motion terminates in this case
[17]

Then, the movement should end at this point with probability ν(x−), and with probability 1 − ν(x−), the movement should continue (recall (15))

E+ z (L) and E− z (L) have the same sign. Then, the movement should end at this point with probability ν(x−), and with probability 1 − ν(x−), the movement should continue (recall (15)). Overall, one needs a stochastic choice for terminating and/or continuing the movement. To m...

2015

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.

[1] [1]

The full flux is a sum of fluxes through the ends of the tube and through its lateral surface: Φf ull= Φ1 + Φ2 + Φlat

We have to prove that Φ′ 1 = Φ2. The full flux is a sum of fluxes through the ends of the tube and through its lateral surface: Φf ull= Φ1 + Φ2 + Φlat. (25) By the definition of a stream tube, the flux through the lateral surface must be zero: Φlat = 0, because the field E and the normal to the lateral surface are orthogonal. Thus, Φf ull= Φ1 + Φ2 = Φ2 − Φ′

work page

[2] [2]

Then some number dn′ ≤ dn of lines inside current tube (corresponding to sample {x+ i }n i=1 ∼ P) must terminate, while the others should continue moving

E+ z (ex−) and E− z (ex−) have the same sign. Then some number dn′ ≤ dn of lines inside current tube (corresponding to sample {x+ i }n i=1 ∼ P) must terminate, while the others should continue moving. This portion dn′ can be found from the line termination property in Q. Lemma B.4 (Line Termination). If E+ z (ex−) and E− z (ex−) have the same sign upon cr...

work page

[3] [3]

American Journal of Physics, 91(2):153–158, 2023. Chen, R. T. and Lipman, Y . Flow matching on general geometries. In The Twelfth International Conference on Learning Representations, 2024. De Bortoli, V ., Korshunova, I., Mnih, A., and Doucet, A. Schrodinger bridge flow for unpaired data translation. Advances in Neural Information Processing Systems, 37:...

arXiv 2023

[4] [4]

(26) It follows that Φ′ 1 = Φ2

From Gauss’s theorem (6), we derive: Φf ull= x ∂M E · dS = Z M q(x)dx = 0. (26) It follows that Φ′ 1 = Φ2. Corollary A.4 (Impossibility of line termination in empty space). An electric field line cannot terminate in empty space. Proof. Otherwise, the field flux inside the current tube surrounding the termination point will not be conserved. Intuitively, n...

work page 1971

[5] [5]

Then the field line motion terminates in this case

E+ z (ex−) and E− z (ex−) have opposite signs. Then the field line motion terminates in this case. 14 Field Matching: an Electrostatic Paradigm to Generate and Transfer Data

work page

[6] [6]

Let dn′ denote the number of points x− i within this region

Consider a D-dimensional area element dS′ on Q(·). Let dn′ denote the number of points x− i within this region. We define the empirical distribution: ˆQn(x−)dS′ = dn′ n . (42) Our goal is to prove: ˆQn(·) almost surely − − − − − − − → n→∞ Q(·). (43)

work page

[7] [7]

17), forward-oriented (dn′ F ) and backward-oriented (dn′ B) trajectories: dn′ = dn′ F + dn′ B

The dn′ points arrive at Q(·) through two pathways (Fig. 17), forward-oriented (dn′ F ) and backward-oriented (dn′ B) trajectories: dn′ = dn′ F + dn′ B. (44)

work page

[8] [8]

Since P(·) and Q(·) have compact supports and the total charge related to the system of the two plates is zero, by Lemma A.6, these tubes must start from P(·) almost surely

Consider the current tube terminating at dS′ corresponding to forward-oriented arrivals (dn′ F ). Since P(·) and Q(·) have compact supports and the total charge related to the system of the two plates is zero, by Lemma A.6, these tubes must start from P(·) almost surely. 15 Field Matching: an Electrostatic Paradigm to Generate and Transfer Data Figure 17....

work page