Partial Wave Analysis using Graphics Cards

Partial wave analysis is a key technique in hadron spectroscopy. The use of unbinned likelihood fits on large statistics data samples and ever more complex physics models makes this analysis technique computationally very expensive. Parallel computing techniques, in particular the use of graphics processing units, are a powerful means to speed up analyses; in the contexts of the BES III, Compass and GlueX experiments, parallel analysis frameworks have been created. They provide both fits that are faster by more than two orders of magnitude than legacy code and environments to quickly program and run an analysis. This in turn allows the physicists to focus on the many difficult open problems pertaining to partial wave analysis.

💡 Research Summary

Partial wave analysis (PWA) is a cornerstone technique in hadron spectroscopy, allowing the extraction of resonance parameters by fitting complex amplitude models to unbinned data. Modern experiments such as BES III, COMPASS, GlueX, and the upcoming PANDA produce data sets with millions of events and require increasingly sophisticated models that involve dozens of interfering partial waves. The computational cost of a typical PWA scales as N_data × N_waves² × N_iterations, which quickly reaches billions of floating‑point operations. Traditional CPU‑based implementations, often written in FORTRAN or C++, can take days or weeks to converge, hampering iterative model development and systematic studies.

The paper presents a comprehensive solution based on graphics processing units (GPUs). By exploiting the embarrassingly parallel nature of event‑wise calculations, the authors ported the core of the likelihood evaluation to CUDA and later to the vendor‑neutral OpenCL framework. Each event’s complex amplitudes are computed in single precision on thousands of GPU cores, while the final log‑likelihood sum is accumulated in double precision using a tree‑reduction algorithm to minimize rounding errors. This hybrid precision strategy preserves numerical stability without sacrificing speed.

Three independent software packages are described: a BES III‑specific framework (GPUPWA), the Indiana‑University AMPTOOLS suite, and a ROOT‑integrated package (ROOTPWA). All share a modular design that lets physicists add custom amplitude functions via a simple C++/Python API, interface with established minimizers such as MINUIT or FUMILI, and handle large Monte‑Carlo (MC) samples either by loading them into GPU memory (≈1.5 GB per million events for a 20‑wave model) or by streaming chunks to avoid memory overflow. Benchmarks show speed‑ups of 100–120× over legacy FORTRAN code and 30× over multi‑core CPU OpenMP implementations for realistic analysis configurations.

Beyond raw performance, the authors discuss several technical challenges. Complex parameters must be represented in a real‑valued minimizer; Cartesian and polar representations each introduce pathological correlations or boundary constraints that can impede convergence. GPU memory limits require careful data staging, and the overhead of host‑to‑device transfers can dominate for analyses involving billions of MC events. Numerical precision issues arise in the log‑likelihood accumulation; tree reductions and Kahan‑type compensations are employed to mitigate them.

On the physics side, the paper highlights unresolved problems: enforcing S‑matrix unitarity in multi‑channel models, incorporating detector resolution effects for narrow resonances (e.g., the φ meson), and selecting an adequate waveset when the true set of intermediate states is unknown. Traditional goodness‑of‑fit tests are inadequate for unbinned, high‑dimensional data, making systematic validation difficult.

Future directions include implementing complex‑valued automatic differentiation directly on GPUs to provide exact gradients for minimizers, integrating Bayesian inference tools for robust uncertainty quantification, and scaling the workflow to cloud‑based GPU clusters for massive parameter scans. The authors argue that these advances will transform PWA from a bottleneck into a rapid prototyping environment, enabling the community to explore exotic hadrons, test theoretical predictions, and ultimately usher in a new golden age of hadron spectroscopy.