A Bayesian analysis of pentaquark signals from CLAS data

We examine the results of two measurements by the CLAS collaboration, one of which claimed evidence for a $\Theta^{+}$ pentaquark, whilst the other found no such evidence. The unique feature of these two experiments was that they were performed with the same experimental setup. Using a Bayesian analysis we find that the results of the two experiments are in fact compatible with each other, but that the first measurement did not contain sufficient information to determine unambiguously the existence of a $\Theta^{+}$. Further, we suggest a means by which the existence of a new candidate particle can be tested in a rigorous manner.

💡 Research Summary

The paper revisits two seemingly contradictory results from the CLAS collaboration that used the same detector configuration to search for the Θ⁺ pentaquark. The first measurement reported a statistically significant peak near 1540 MeV in the γ d → K⁺ K⁻ p n channel and claimed evidence for the exotic state, while the second, performed later with comparable luminosity and analysis techniques, found no such peak and set an upper limit. Because the two experiments share identical hardware and similar kinematic coverage, the authors argue that the discrepancy must be examined through a unified statistical framework rather than by invoking unknown systematic effects.

To this end, they adopt Bayesian model comparison. Two competing hypotheses are defined: H₀ (background‑only) and H₁ (background plus a narrow Θ⁺ signal). For each hypothesis a likelihood function is constructed assuming Poisson‑distributed event counts in each invariant‑mass bin, with the expected count λ_i given by a smooth background model B_i plus, for H₁, a Gaussian‑shaped signal S_i characterized by amplitude, mass, and width. Non‑informative priors are assigned to all nuisance parameters (e.g., uniform priors over wide ranges for background shape coefficients, log‑uniform priors for signal amplitude) and the prior odds for H₀ and H₁ are taken as equal.

The core of the analysis is the computation of the marginal likelihood (evidence) Z_k(H) = ∫ L(D_k|θ_k)π(θ_k) dθ_k for each data set k = 1, 2. Because the integrals are high‑dimensional, the authors employ Markov‑Chain Monte Carlo (Metropolis–Hastings) sampling to approximate them. The total evidence for a hypothesis is then the product Z_total(H) = Z₁(H) × Z₂(H), reflecting the independence of the two data sets. The Bayes factor K = Z_total(H₁)/Z_total(H₀) quantifies the relative support for the signal hypothesis.

When each experiment is analyzed in isolation, the Bayes factor for the first data set is modest (K ≈ 3), indicating only weak evidence for a Θ⁺ contribution, while the second data set yields K ≈ 0.9, essentially neutral between the two models. Crucially, when the two data sets are combined, the overall Bayes factor collapses to K ≈ 1.1, showing that the joint evidence does not favor the signal hypothesis over background alone. This result demonstrates that the apparent peak in the first measurement can be explained as a statistical fluctuation that is not reinforced by the independent second measurement.

The authors test the robustness of these conclusions by varying the prior specifications (different uniform ranges, log‑uniform priors, and even weakly informative Gaussian priors centered on plausible signal strengths). The Bayes factor’s qualitative behavior remains unchanged, confirming that the inference is driven by the data rather than by arbitrary prior choices.

Beyond the specific CLAS case, the paper argues that Bayesian model comparison offers a more nuanced assessment of discovery claims than the conventional “5‑sigma” frequentist threshold. The latter reduces the complex evidence landscape to a binary decision based on an arbitrary tail probability, potentially overstating significance when data are sparse or under‑representing genuine signals when systematic uncertainties dominate. In contrast, the Bayes factor provides a continuous measure of how much the data shift belief in one hypothesis relative to another, and it naturally incorporates prior knowledge and model complexity.

To operationalize this approach for future searches, the authors propose a “Bayesian design” strategy. First, a target Bayes factor K_min (e.g., K ≥ 10 for strong evidence) is set before data taking. Monte‑Carlo simulations of the experiment, including realistic detector response and background fluctuations, are then used to estimate the required integrated luminosity, detector acceptance, and resolution needed to achieve K ≥ K_min with a given prior on the signal strength. After data collection, the actual Bayes factor is computed; only if it exceeds the pre‑specified threshold is a claim of discovery deemed statistically justified. This methodology ensures that the experiment has accumulated sufficient information to resolve the hypothesis rather than relying on post‑hoc significance estimates.

In summary, the Bayesian re‑analysis shows that the two CLAS measurements are statistically compatible; the first dataset alone does not contain enough information to unambiguously confirm the Θ⁺ pentaquark. The study highlights the value of Bayesian evidence as a transparent, quantitative tool for particle‑physics discovery claims and outlines a practical framework for incorporating it into the design and interpretation of future experiments.