Research / Scattering amplitudes
Research entry

What can a neural fit tell us about two pions at threshold?

Smith and colleagues fit pion scattering data with neural networks constrained by dispersion relations. Their public mesh starts 0.86 MeV above the two-pion threshold. Its first two energies give a way to check the reported scattering lengths in two isospin channels.

First published Scientific update Sources checked

Two numbers at zero relative momentum

As two pions approach zero relative momentum, their lowest-angular-momentum scattering amplitudes approach constants. There are two allowed S-wave isospin channels, I = 0 and I = 2. Their scattering lengths also supply the two subtraction constants in the Roy dispersion equations, which link low-energy partial waves through integrals over other energies and channels.

Table IV gives a positive scattering length of 0.222 for I = 0 and a negative one of −0.041 for I = 2, in inverse pion masses. The central 68% intervals across the fitted replicas are [0.208, 0.236] and [−0.047, −0.036]. The fit includes a pionium measurement constraining the absolute difference between the two lengths.

What the networks learned

Smith and colleagues trained neural networks to represent seven pion-pion partial waves as functions of energy. The waves share part of the network and have separate outputs. The authors encoded threshold behavior and unitarity bounds in the output maps and added a penalty for violations of the Roy equations in the three dominant low-energy waves. They used changes in fits under experimental reweighting to identify mutually compatible data sets.

Training on resampled measurements produced an ensemble of possible amplitudes. The authors retained 10,000 models for the released mesh. Each replica carries multiple waves together, so a reader can use the same replica across channels when evaluating a derived quantity. The threshold numbers are predictions of this constrained fit to data.

From a phase shift to the scattering length

For both S waves, the release gives the phase shift δ in degrees, the inelasticity η and the real and imaginary parts of the partial wave. With center-of-mass energy E and pion mass 139.57 MeV, Eq. (6) relates these quantities. The limiting value of the dimensionless real S wave gives the numerical scattering length in units of inverse pion mass.

We checked the real part directly at 280, 280.5 and 281 MeV for both S waves of three replicas. Converting the phases to radians reproduced all 18 stored values within 2 × 10⁻¹⁵ in absolute value. Treating degrees as radians made all 18 comparisons fail the 10⁻¹² tolerance.

Re⁡t0I(E2)=η0Isin⁡(2δ0I)2ρ(E),ρ(E)=1−4mπ2E2\operatorname{Re}t_0^I(E^2)=\frac{\eta_0^I\sin(2\delta_0^I)}{2\rho(E)},\qquad \rho(E)=\sqrt{1-\frac{4m_\pi^2}{E^2}}
Above threshold, the phase shift in the sine is in radians. The numerical scattering length in inverse pion masses is the limit of this real part as E approaches 2mπ from above. [1]
Worked calculation · Checking the pion scattering lengths at threshold →

Approaching threshold from the released mesh

The first mesh point is at 280 MeV, just above threshold at 279.14 MeV. We drew a straight line in squared relative momentum, q² = E²/4 − mπ², through the deposited median real amplitudes at 280 and 280.5 MeV and evaluated it at q² = 0. The intercepts were 0.221886 for I = 0 and −0.040534 for I = 2. Both lie within the published intervals.

Extending the second point to 281 MeV gave 0.221885 and −0.040478. The latter change is only about 0.000055, yet it changes the I = 2 value rounded to three decimal places. The result supports agreement at the scale of the reported ensemble spread. Appendix F describes the authors’ four-point Richardson extrapolation for the network threshold values. Our two-point check uses pointwise summary curves, which cannot recover the distribution of individual replica limits.

Limits of this result

  • The check uses two or three released mesh points. Threshold extrapolation of the median or either pointwise band edge need not equal the corresponding percentile of replica scattering lengths.
  • The quoted intervals describe the authors’ selected production ensemble. This check covers neither retraining nor the experimental selection and Roy-equation residuals.

Checks performed

Near-threshold mesh extrapolation and replica normalization check

Compared two S-wave intercepts from the deposited median curves against Table IV and varied the extrapolation span. Reconstructed 18 real amplitudes from phases and inelasticities on three identified replicas. The calculation used published numerical outputs.

Sources & citations

  1. S-matrix informed neural networks for amplitude analysisWyatt A. Smith, Arkaitz Rodas, Marius D. Thomas, César Fernández-Ramírez, Giorgio Foti, Lin Qiu, Adam P. Szczepaniak and Alessandro Pilloni · 2026-08-24

    arXiv:2608.23750v1. Eqs. (5)–(6), partial-wave convention. Eqs. (10)–(11), Roy equations. Eqs. (19)–(20), threshold definition and pionium input. §§IV–V, representation, training and selection. §VI C, Table IV and Fig. 9, scattering lengths. Appendix F, initialization and numerical settings.

  2. Production SINN ensemble for pion scatteringWyatt A. Smith · 2026-08-19

    Zenodo 22016314, version SINN, CC BY 4.0. The archive binds Git revision 91b1ece3543aba0d6d783c6142c388e71437d903. Under final_v7-lt4p0, metadata.json documents the mass and units, manifest.csv identifies selected replicas, and the median and ci68 Git LFS objects contain the pointwise summaries.

Index terms pion scattering · pion-pion · ππ · S-matrix informed neural networks · SINN · hadron physics · Roy equations · scattering length · threshold extrapolation · isospin · arXiv:2608.23750