Yang-Mills glueball masses from spectral reconstruction

The Big Picture

No one has ever seen a glueball. These hypothetical particles, made entirely of gluons (the force carriers that bind quarks together inside protons and neutrons), have been predicted by the Standard Model for decades, but they’ve never been definitively spotted in a detector.

The reason gluons should form bound states is straightforward: they carry their own version of the charge governing the strong nuclear force, which physicists call “color charge.” Unlike photons, which don’t interact with each other, gluons do. So in principle they can clump together into a glueball, no quarks needed.

The problem is practical, not theoretical. Glueballs are expected to share their quantum numbers with ordinary mesons, particles made from quarks and gluons. Every experimental glueball candidate so far has a plausible conventional explanation.

To break the impasse, theorists need mass predictions sharp enough to guide experimentalists. A recent collaboration tackled this by combining quantum field theory techniques with a machine learning method from Bayesian statistics. They computed the masses of the two lightest glueballs: the scalar type at 1870 MeV and the pseudo-scalar type at 2700 MeV. (An MeV, or megaelectronvolt, is the standard unit of mass in particle physics; a proton weighs about 938 MeV.)

Their approach sidesteps some longstanding obstacles by extracting glueball properties directly from interaction vertices rather than solving a separate bound-state equation.

Key Insight: The team treated glueball masses as spectral peaks in the four-gluon interaction vertex and used Gaussian process regression to reconstruct real-time dynamics from imaginary-time calculations, extracting precise glueball masses without solving a bound-state equation directly.

How It Works

The equations governing the strong nuclear force are most tractable in Euclidean space, where time is replaced by imaginary time. This turns wild oscillatory integrals into well-behaved ones. But physical particle masses live in Minkowski space, the real spacetime of the universe. Getting from one to the other requires inverting a mathematical transform that is famously ill-conditioned: small errors in the input can blow up into large errors in the output.

The team’s workflow handles this in three stages:

1. Compute Euclidean data with the functional renormalization group (fRG). The fRG is a non-perturbative method that works even when the interaction strength is too large for standard approximation schemes. It integrates out quantum fluctuations scale by scale, building up the full interacting theory from a tractable starting point. Here, the team applied it to the four-gluon vertex (the interaction amplitude describing four gluons meeting at a point), focusing on channels with the right symmetry properties for each glueball type.

2. Identify the right tensor structures. Different glueballs are distinguished by their tensor structure, the geometric pattern of how gluon fields combine at the interaction point. For the scalar glueball ($J^{PC} = 0^{++}$, zero spin and symmetric under certain reflections), the natural arrangement of the four-gluon vertex suffices. For the pseudo-scalar ($J^{PC} = 0^{-+}$, which flips sign under those same reflections), the team uses a structure involving the antisymmetric tensor $\varepsilon_{\mu\nu\rho\sigma}$. Keeping these two channels from bleeding into each other is essential; mixing them would make spectral reconstruction exponentially harder.

3. Reconstruct the spectral function with Gaussian process regression (GPR). GPR is a Bayesian technique that treats the unknown spectral function (which encodes all the frequencies at which a quantum field can vibrate) as a random field with a prior enforcing smoothness. Peaks in the spectral function correspond to real particles.

Given the Euclidean data, GPR updates that prior to a posterior, yielding both a best estimate and rigorous uncertainty bands. It handles the ill-posed inversion without claiming to know more than the data support.

Where the spectral function peaks, a resonance lives, and a resonance is a particle. The scalar glueball shows up as a clean peak in the scalar channel; the pseudo-scalar appears in the pseudo-scalar channel.

The payoff of this approach is self-consistency. Rather than solving a separate bound-state equation (the usual Bethe-Salpeter approach, which models two particles attracting each other into a composite state), the team extracts glueball masses directly from the same four-gluon correlation functions used throughout the fRG calculation. The glueballs are already there inside the vertex; you just need the right tool to see them.

The final results: $m_{sc} = 1870 \pm 75$ MeV for the scalar glueball and $m_{ps} = 2700 \pm 120$ MeV for the pseudo-scalar. Both agree well with independent lattice QCD calculations and with functional bound-state methods, a nontrivial cross-check given how different the computational approaches are.

Why It Matters

Glueball spectroscopy tests our understanding of the strong nuclear force at its most basic level. If gluons can form stable bound states on their own, that’s a qualitatively new kind of matter. The GlueX experiment at Jefferson Lab and BESIII in Beijing are actively searching, and sharper mass predictions from theory improve the odds of a discovery.

This work also makes a case for Gaussian process regression as a practical tool in quantum field theory. Extracting spectral functions with uncertainties from Euclidean correlators has been a persistent bottleneck in finite-temperature QCD, heavy-ion physics, and transport calculations. The same framework applies wherever real-time physics must be coaxed out of imaginary-time data: quark-gluon plasma properties, exotic hadron spectroscopy, and beyond.

Bottom Line: Combining functional renormalization group calculations with Gaussian process regression yields precise, uncertainty-quantified glueball mass predictions that align with lattice results, while offering a reusable method for extracting real-time physics from Euclidean field theory data.

---