Solving the Forced Soliton Equation Numerically, Then Searching the Solutions for Exact Reductions

A soliton is a localized lump of field that holds its shape as it moves. In quantum field theory one wants to know how such a lump responds when it is struck, and the answer is packaged in "form factors," numbers that say how likely the soliton is to absorb a momentum kick $P$ and remain a soliton. Melnikov, Papageorgakis and Royston showed that, in the semiclassical limit, computing these form factors at any momentum transfer reduces to solving a single equation. They call it the forced soliton equation: a time-dependent, second-order, wave-like integro-differential equation that describes a kink being accelerated by an external source (Forced Soliton Equation and Semiclassical Soliton Form Factors, Accelerating solitons). The authors state that no explicit time-dependent solutions are known. They therefore work perturbatively at small $P$, where they recover the relativistic energy $H=\sqrt{P^2+M^2}$ with $M$ the one-loop-corrected mass. The physically interesting regime, a kick comparable to the soliton's own mass ($P\sim M$), has not been computed.

This note proposes to close that gap in two stages that differ in how they are judged. The first stage is a careful conventional numerical solution of the forced soliton equation across $P/M$ from 0 to order 1. It yields the first non-perturbative semiclassical form factors in this model. The second stage treats those solutions as data and runs an automated search for hidden symmetries. The specific hypothesis is that, for a constant force, the solution becomes stationary in some accelerated (boost-type) frame, which would reduce the equation to a solvable ordinary differential equation. The point is to use numerics to find analytic structure, rather than to use machine learning to replace analysis. Either outcome is informative. A confirmed reduction gives the closed form the original authors lacked. A clean failure, with quantified residuals, says that the analytic route is closed and the numerical form factors are the final answer.

Core Claim

The forced soliton equation is a well-posed hyperbolic problem; the original papers point to the Hughes–Kato–Marsden theory of quasilinear hyperbolic systems. A spectral method-of-lines discretization should therefore solve it to high accuracy across $P\sim M$, provided the nonlocal integral term is handled carefully. Once accurate solutions exist across a family of forcing strengths, symmetry discovery from trajectory data becomes possible. Earlier work could not attempt this, because the method needs numerical trajectories and none existed. We claim that searching for invariances of these computed solutions will either recover an exact reduction or rule out the natural candidate reductions to a stated tolerance. Both are results the perturbative analysis cannot give.

Relation to Prior Work

The motivating work is Melnikov et al., who derived the equation from a saddle point of a phase-space path integral that keeps all time derivatives. They identified the absence of explicit solutions as the obstacle and explicitly suggested numerical exploration. We take that suggestion literally and add an analytic follow-up they did not attempt.

Two lines of work supply methods. For the solver, the Gaussian-process approach of Solving and learning nonlinear PDEs with Gaussian processes treats a nonlinear equation as constraints on a Gaussian prior and linearizes by Gauss–Newton. It reaches relative errors of $10^{-7}$ on smooth elliptic and Burgers problems in 2 to 10 iterations. We use it as an independent cross-check, not as the primary solver, because its dense kernel factorization costs $O(N^3)$ in the number of collocation points $N$, and it handles moving localized fronts and radiation tails poorly without hand-tuned regularization.

For the symmetry search, Discovering Symmetries of ODEs by Symbolic Regression writes the symmetry condition as a loss evaluated on numerical trajectories and searches expression graphs for the symmetry generators. It succeeded on 10 of 10 nonlinear ODEs where Mathematica's automatic mode succeeded on none. Its limits are compact generator spaces and the need for accurate trajectories. We extend it from ordinary differential equations to a (1+1)-dimensional field equation. We keep the search tractable by restricting candidate generators to the physically motivated family: translations, boosts, scalings, and their combinations with field-dependent shifts. We do not attempt open-ended expression search.

Physics-informed neural networks (PINNs) for solitons constrain the design rather than supply it. These networks are trained to make the equation's residual small at sample points. The data-driven localized wave solutions of the derivative nonlinear Schrödinger equation by using improved PINN approach reports relative $L_2$ errors of $2$ to $9\times10^{-2}$ on known solutions, with accuracy degrading as the time window grows. N-double poles solutions for nonlocal Hirota equation with nonzero boundary conditions using Riemann–Hilbert method and PINN algorithm reaches $2.69\times10^{-4}$ but needs about 1262 seconds and relies on exact solutions to supply boundary data. Both studies treat isospectral, unforced equations with known answers. Neither suggests that PINNs would beat a spectral solver on a long-time, nonlocal, forced problem, so we do not build on them.

Research Program

The first component is the solver. We discretize space with a Fourier or Chebyshev basis on a domain large enough, or with absorbing layers, to let radiation leave. We evaluate the integral term by quadrature on the same grid and advance in time with a high-order implicit–explicit integrator. The integral is the main source of cost and error, so we implement it two ways, direct quadrature and a fast-transform evaluation where its kernel permits, and require the two to agree. Correctness is established in three ways. Errors must fall at the expected spectral rate under grid refinement. The equation's residual must be small on solutions interpolated to a finer grid. The extracted energy–momentum relation must reproduce $H=\sqrt{P^2+M^2}$ at small $P$, the only analytic anchor in the file.

From the solutions we extract the semiclassical form factor as a function of $P/M$ by the same prescription the original papers use at small $P$. We report the full curve together with its deviation from naive extrapolation of the low-$P$ result.

The second component is the symmetry search. As a calibration task, we first run it on an equation with exact answers: the modified KdV equation with self-consistent sources in non-uniform media of A modified KdV equation with self-consistent sources in non-uniform media and soliton dynamics, whose sources accelerate and reshape solitons in a controlled, integrable way. The non-isospectral symmetry algebras derived by hand in New symmetries for the Ablowitz–Ladik hierarchies show that time-dependent spectral parameters still leave a closed algebra of symmetries. That result gives us ground truth for what a correct search should find in forced integrable settings.

We then apply the search to the forced soliton solutions for constant and slowly varying forces. If a generator survives with residual at solver precision across forcing strengths, we perform the reduction by hand, aided by the computer-algebra package of The MAPLE package TDDS for computing Thomas decompositions of systems of nonlinear PDEs to split the reduced system into simpler cases. We then attempt closed-form integration. We also test whether the reduced ordinary differential equation can be linearized, using the checks of Algorithmic Verification of Linearizability for Ordinary Differential Equations. This matters because reductions often leave intractable equations behind.

Baselines

The do-nothing reference is the existing small-momentum perturbative expansion of Melnikov et al., extended naively to large $P$, together with the adiabatic collective-coordinate picture of a rigid kink whose position follows Newton's law. The file contains no reported form-factor value at $P\sim M$ to compare against. Its only quantitative anchor is the limit $H=\sqrt{P^2+M^2}$; no reported result was found on Lacuna for the forced soliton equation at finite $P/M$. Our solver beats this reference if the form factor at $P\sim M$ departs from the extrapolation by more than the solver's error bar, and if refinement shows the departure is converged.

The strongest competing solver is the Gaussian-process method of Chen et al., run with the released Jax code Gaussian-Process-Slover-for-High-Freq-PDE. Its strength is its reported $10^{-7}$ relative error in 2 to 10 Gauss–Newton iterations on smooth problems; it is the strongest because it is the only learning-based solver in the file with accuracy near that of spectral methods. For fairness it receives the same space–time domain, the same forcing profiles, and the same wall-clock budget as the spectral solver, with kernel choice and regularization tuned on the small-$P$ case only. We beat it if, at matched compute, our error at $P\sim M$ is lower and stays bounded over the long time windows needed for the solution to settle.

An improved PINN in the style of Pu et al. runs under the same budget. It is a sanity reference, given its reported soliton-case relative $L_2$ error of $5.75\times10^{-2}$ at 1418 seconds.

For the symmetry search, the baseline is direct computer algebra: Mathematica's SYM and Maple Thomas decompositions applied to the forced soliton equation itself, with no data. The ablation that isolates our key ingredient runs the same restricted generator search with the trajectory loss replaced by symbolic checking of the equation alone. If the data-free version finds the same generators, the numerical data added nothing. On the calibration task, the search counts as working only if it recovers the known symmetries of the mKdV equation with sources.

Why not just…?

Why not just use computer algebra to find the symmetries directly? Kahlmeyer et al. found that Mathematica's automatic symmetry routine failed on all 10 of their nonlinear ODE benchmarks and failed in 30% of cases even with human guidance. The forced soliton equation is harder still, being nonlocal and second-order in time. We nonetheless run computer algebra as a baseline rather than assume it fails, and the data-free ablation shows directly whether trajectory data is what makes the search succeed.

Why not just solve it with a PINN or Gaussian process instead of a classical solver? For a forward problem with no unknown parameters, the evidence in the file favours classical methods. PINN errors on soliton problems sit at $10^{-2}$ to $10^{-4}$ and need tens of minutes per run. Pu et al. also document error growth over long time windows, which is exactly what an accelerating, radiating kink requires. The Gaussian-process solver is the real threat on smooth problems, which is why it is our headline competitor under matched compute rather than a straw man.

Why not just extend perturbation theory in $P/M$? Melnikov et al. derived the full equation precisely because the small-momentum expansion is restricted to $P\ll M$ and cannot represent the field distortion and radiation of strong acceleration. Our form-factor curve tests this quantitatively by showing where, and by how much, the perturbative series departs from the converged numerics.

Why not just stop after the numerics? This is a fair objection. The numerical form factors are the main physics deliverable and stand on their own. The symmetry stage is a bounded, cheap addition, aimed at a specific hypothesis about accelerated-frame stationarity. If it fails, we report the residual of each candidate generator as a negative result rather than inflating the claim.

Decisive Checks

The solver is trusted only if three conditions all hold: it converges at the spectral rate, it reproduces $H=\sqrt{P^2+M^2}$ at small $P$, and the quadrature and fast-transform versions of the integral term agree. The symmetry pipeline is trusted only if it recovers the known non-isospectral symmetries on the mKdV-with-sources task. The central result is decisive in both directions. Either a boost-type generator holds at solver precision across forcing strengths, and the closed form derived from it matches the numerics to that precision, or every candidate in the restricted family is excluded with residuals well above solver error. In the second case the numerical form-factor curve at $P\sim M$ is the answer.