The Crystal Ball function is a probability-density function used in particle physics when a measured peak is approximately Gaussian on one side but has a power-law tail on the other. It is commonly used to model detector response and energy-loss effects such as bremsstrahlung. It is not an array-prediction or “crystal ball problem” algorithm.
At a practical level, the function has four shape and location parameters: the peak location μ, width σ, tail-transition location α, and tail exponent n. A fit uses the Gaussian core to represent the central detector resolution and the tail to represent asymmetric non-Gaussian loss or reconstruction effects. ROOT provides a maintained Crystal Ball PDF example and implementation reference.
Mathematical definition
Define the standardized variable:
t = (x − μ) / σ
For a conventional low-side tail, the unnormalized Crystal Ball function is
f(x) = N × exp(−t² / 2) for t > −|α|
f(x) = N × A × (B − t)^(−n) for t ≤ −|α|
where
A = (n / |α|)^n × exp(−α² / 2)
B = n / |α| − |α|
and N is a normalization constant. The definitions of A and B make the Gaussian and power-law pieces meet continuously at the transition. Software packages may choose a high-side tail instead, or expose a sign convention that reverses the side; check the implementation before comparing fitted parameter values.
What each parameter means
μ(mean): the center of the Gaussian peak.σ(sigma): the Gaussian width; it controls the central resolution.α(alpha): how far from the center the tail begins, in units ofσ.n: the steepness of the power-law tail. Larger values make the tail fall more quickly.N: the amplitude or normalization, depending on whether the fit is normalized as a probability density.
Why particle-physics analyses use it
A pure Gaussian assumes symmetric measurement fluctuations. Real reconstructed masses and energies often have a longer tail on one side because particles can radiate energy, detector material can cause loss, or reconstruction can be incomplete. The Crystal Ball function keeps a well-behaved Gaussian peak while giving the fit enough flexibility to describe that asymmetric tail.
A typical workflow is to fit a signal-region observable with a Crystal Ball signal shape plus a separately motivated background model. The signal mean and width can then estimate the peak position and experimental resolution, while the tail parameters quantify the non-Gaussian response. They should not be interpreted as universal detector constants without a study of the dataset, selections, and fit model.
Minimal fitting checklist
- Confirm whether the library uses a low-side or high-side tail and document its sign convention.
- Choose initial
μandσvalues from the observed peak rather than leaving all parameters unconstrained. - Use physically defensible parameter bounds; an unconstrained tail can absorb background structure.
- Plot residuals and inspect the fit near the transition point, not only the headline goodness-of-fit statistic.
- Report the exact implementation and parameter convention alongside results so another analyst can reproduce the model.
Further reading
The Crystal Ball function is useful because it is simple enough to fit routinely while still representing a common real-world departure from symmetric Gaussian resolution.