⚛️ high-energy theory

Limits of the inverse scattering problem

This paper investigates the minimum speed limit required for particles to enable the reconstruction of a potential via an inverse scattering problem, utilizing both theoretical examples and a machine learning pipeline to generalize the traditional Radon transform approach.

Original authors: Matvei Fedin, Kirill Gubarev, Andrey Morozov

Published 2026-08-14

📖 4 min read🧠 Deep dive

CC BY 4.0

Original authors: Matvei Fedin, Kirill Gubarev, Andrey Morozov

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a detective trying to figure out what a mysterious, invisible room looks like, but you are forbidden from stepping inside. Instead, you have a fleet of tiny, super-fast drones. You launch them from the walls, and they zip through the room. If the room is empty, the drones fly in straight lines and hit the opposite wall exactly where you'd expect. But if there are invisible hills, valleys, or magnetic whirlpools inside, the drones get pushed off course. They might speed up, slow down, or curve around obstacles. By watching exactly where they exit and how fast they are going when they leave, you can try to map out the invisible landscape inside. This is the core idea of "inverse scattering," a technique used in everything from medical X-rays to peering inside the Earth's crust. Usually, scientists assume these particles move so fast that they barely notice the bumps in the road, flying in straight lines like laser beams. This makes the math easy, like drawing a straight line on a piece of paper. But what happens if the particles aren't super-fast? What if they are slow enough that the invisible hills actually bend their paths, making the math messy and the straight-line assumption useless? This is the puzzle scientists have been wrestling with: how slow can a particle be before we lose the ability to see the room clearly?

This paper, titled "Limits of the inverse scattering problem," dives headfirst into that messy middle ground. The authors, a team of physicists and data scientists, ask a simple but tricky question: How low can the speed of these particles drop before we can no longer reconstruct the shape of the force field inside the medium? They explore this using two main tools: old-school theoretical math to understand the rules of the game, and a modern "Machine Learning" pipeline—a type of artificial intelligence trained on millions of simulated scenarios—to act as a super-smart detective.

The team starts by confirming what we already know: if particles move infinitely fast, the problem is solved. It's just like a standard X-ray scan where the path is a straight line. But as they slow the particles down, the paths start to curve, and the simple math breaks down. They found that there is a critical "tipping point." If the particle's energy (its speed) is high enough to easily overcome the strongest "hill" or "valley" in the potential field, the reconstruction works well. However, once the particle's energy drops to a level comparable to the difference in potential energy between the edge of the room and the deepest part of the interior, the problem becomes incredibly difficult.

To test this, the authors created a digital playground. They simulated various invisible landscapes, from simple smooth hills to complex, bumpy terrains. They then trained a neural network (their "Force-Field Prediction Model") to look at the exit data of the particles and guess the shape of the invisible field. The results were revealing. The AI could reconstruct the field with high precision when the particles were fast. But as the speed dropped, the error in the reconstruction began to climb sharply. The paper suggests that the minimum speed required isn't just a random number; it is directly tied to the energy difference between the boundary and the interior of the domain. Specifically, the kinetic energy of the particle needs to be greater than the maximum difference in potential energy inside the room.

The authors also discovered that different reconstruction methods fail in different ways. The traditional math-based method (Radon transformation) tends to lose the "height" of the hills, flattening them out while keeping their general shape. The AI method, on the other hand, tends to keep the hills' peaks and valleys but starts adding "noise" and weird geometric distortions as the particles get too slow. This suggests that there isn't just one way to fail; the nature of the error changes depending on how you try to solve it.

Ultimately, the paper doesn't claim to have solved the problem for all possible speeds. Instead, it provides a clear map of where the solution works and where it breaks. It suggests that for the potential to be restorable, the particle's energy must exceed the potential energy difference between the inside and the outside. Below this threshold, the scattering map becomes too nonlinear and chaotic for current methods to untangle reliably. The study serves as a warning and a guide: if you want to see the invisible world with slow particles, you need to make sure they have enough "oomph" to climb the highest hill in the room, or else your map will be full of holes and distortions.

=== SUMMARY ===

Technical Summary: Limits of the Inverse Scattering Problem

1. Problem Statement

The paper addresses the inverse scattering problem for Newtonian particles moving through a medium with an unknown potential field $U(x)$ . Unlike classical tomography (e.g., X-ray or Radon transforms) where particles travel in straight lines at infinite speed, this study considers particles with finite initial velocity $v$ . The goal is to reconstruct the force field $F(x) = -\nabla U(x)$ based on boundary measurements: the initial launch coordinates and velocity, and the final exit coordinates, velocity, and traversal time.

The central question is determining the lower limit of particle speed (or conversely, the upper limit of potential strength) below which the potential cannot be reliably restored. The authors posit that when the initial kinetic energy becomes comparable to the characteristic potential variation between the domain boundary and its interior, the scattering map becomes strongly nonlinear, rendering standard linear inversion techniques (like the Radon transform) insufficient.

2. Methodology

Theoretical Framework

The authors develop a systematic perturbative expansion of the scattering data in inverse powers of the velocity ( $v^{-1}$ ).

Infinite-Speed Limit: As $v \to \infty$ , the trajectory approaches a straight chord, and the momentum change corresponds to the vector Radon transform of the force field, $J_\infty(L) = \int_L F(x) ds$ . This allows exact reconstruction via the inverse Radon transform.
Finite-Speed Corrections: For finite $v$ , the trajectory deviates from a straight line. The authors derive a series expansion for the measured vector ray integral $J_v(L)$ :
$J_v(L) = J_\infty(L) + \frac{\delta U}{mv^2} Q[F](L) + \left(\frac{\delta U}{mv^2}\right)^2 P[F](L) + O(v^{-6})$
where $\delta U$ is a characteristic potential energy spread, and $Q[F]$ and $P[F]$ are functionals of the force field and its derivatives.
Iterative Reconstruction: Based on this expansion, the authors propose two methods to recover $F$ $F$ from finite-speed data:
1. Single-speed iteration: Starting from a basic Radon reconstruction, iteratively subtract correction terms derived from the expansion.
2. Multi-speed extrapolation: Using measurements at multiple velocities to algebraically eliminate correction terms and extrapolate to the $v \to \infty$ limit (analogous to renormalization group flow).

Analytical Examples

Two exactly solvable cases are analyzed to verify the perturbative framework:

Harmonic Potential ( $U \propto r^2$ ): The authors derive an exact relation between the finite-speed integral $J_v$ and the infinite-speed limit $J_\infty$ . They demonstrate that the first-order correction vanishes ( $Q[F]=0$ ) and the series converges only when the kinetic energy exceeds a critical threshold ( $E > \Delta U$ ).
Constant Force ( $F = \text{const}$ ): An exact closed-form solution is derived, and the series expansion is verified to match the general perturbative formulas.

Machine Learning Approach

To address the general non-linear regime where analytical inversion is difficult, the authors employ a deep learning pipeline:

Forward Model: A CUDA-based solver integrates Newton's equations of motion for particles on a unit disk using the Runge-Kutta method.
Dataset: Synthetic datasets are generated using diverse potential field families (Gaussian RBF, Matérn, truncated Fourier series, Perlin noise, etc.) to ensure robustness against overfitting to specific topologies.
Architecture: A "Force-Field Prediction Model" (FFPM) based on a Johnson-style generator (encoder $\to$ residual bottleneck $\to$ decoder) is trained. The input is a 6-channel tensor comprising boundary measurements (final velocity direction, magnitude change, residence time) and a coarse Radon reconstruction. The output is the reconstructed force field.
Training: The model uses a combination of $L_1$ reconstruction loss, a two-scale GAN loss, and a contrastive loss term (PatchNCE) to align local features with the Radon prior.

3. Key Results

Theoretical Limits

The analysis of the harmonic and constant force examples reveals a critical transition point. The reconstruction quality degrades significantly when the dimensionless parameter $\epsilon = \delta U / (mv^2)$ approaches $1/2$ . This corresponds to the condition where the initial kinetic energy $E_{kin}$ is comparable to the potential energy difference between the boundary and the interior:
$E_{kin} \approx |U_{in} - U_{out}|$
Below this energy threshold, the scattering map loses injectivity (e.g., multiple injection angles map to the same exit point), and the perturbative series diverges.

Numerical Findings

Error Scaling: The reconstruction error (measured by relative Mean Absolute Error) remains low for high energy ratios ( $E_{kin}/E_{pot} \gg 1$ ) but increases sharply when the ratio drops into the range $[1, 3]$ .
Generalization: The FFPM model, trained on a specific distribution of random fields, successfully generalizes to a centered paraboloid potential (an out-of-distribution test case). The degradation in performance for the paraboloid occurs at the same energy threshold as for the training distribution, suggesting the limit is physical rather than a failure of model generalization.
Failure Modes:
- Radon Approximation: In strong fields, the Radon reconstruction preserves the coarse shape (symmetry) but loses amplitude, effectively flattening the potential.
- FFPM: The neural network retains the location of potential extrema (wells) but introduces noise, anisotropic deformations, and symmetry breaking in strong fields.
Input Informativeness: Correlation analysis shows that the particle's residence time (traversal time) is the most informative channel for estimating the potential amplitude, far more so than the final velocity direction or magnitude.

4. Significance and Claims

The paper claims to establish a quantitative "applicability limit" for inverse scattering problems involving finite-velocity particles. The primary contribution is the identification of the energy ratio $E_{kin}/E_{pot}$ as the governing parameter for reconstructability.

Physical Insight: The work demonstrates that the transition from a solvable linear regime (Radon transform) to an unsolvable or highly non-linear regime is governed by the accessibility of the potential landscape. If the particle's energy is insufficient to overcome local potential barriers relative to the boundary, information about the interior is lost.
Methodological Contribution: The paper provides a systematic perturbative expansion for finite-speed corrections and validates a machine learning pipeline that can operate near these theoretical limits, outperforming linear inversion in the non-linear regime.
Modesty: The authors explicitly state that for general potentials, the proposed non-linear correction formula (Eq. 6.1) is a conjecture and not an exact identity, serving instead as a valuable first approximation. They acknowledge that the current study uses synthetic data and that future work is needed to address noise, missing trajectories, and non-conservative force fields.

The paper concludes that while the inverse problem is theoretically solvable for finite speeds, the practical limit of reconstruction is strictly bounded by the ratio of kinetic to potential energy scales.