Do multi-step decay chains require fundamentally different feature extraction than two-body or prompt decays?

# Multi-step decay chains and feature extraction

The short answer: not fundamentally different, but more demanding.

The key distinction is *structural complexity*, not a qualitatively new problem. A two-body decay gives you two final particles; a multi-step chain (like Λ_c⁺ → pK⁰_s, K⁰_s → π⁺π⁻) gives you a hierarchical tree of intermediate and final states [AG-2025.04-1624]. This tree structure—which particles decayed into which—is invisible to the detector; you only see the leaves (the final particles like protons and pions). Your job remains the same: pick the right combination of detected particles from the combinatorial mess and assign them to the correct decay chain.

What changes in practice:

Structural encoding: Multi-step decays benefit from explicitly modeling the tree topology. A recent approach uses a matrix representation called lowest common ancestor generations (LCAG) that encodes which particles are siblings in the decay tree, letting graph neural networks reason about the hierarchical relationships [AG-2024.02-1078]. This is more powerful than treating all particles as an unordered set, but it's an *enhancement*, not a new paradigm.

Feature scope: You still extract kinematic features—momentum, angles, invariant masses—but in multi-step chains you can compute invariant masses at intermediate stages (e.g., the K⁰_s mass from the two pions), which adds discriminating power [AG-2025.04-1624]. This is a richer feature space, not a fundamentally different one.

Background suppression: The combinatorial background grows worse. In a three-body decay like Λ_c⁺ → pK⁰_s (itself a two-body decay of K⁰_s), standard machine learning classifiers—XGBoost, deep neural networks—handle it well by learning subtle correlations between features [AG-2025.04-1624]. The same algorithms work for complex multi-step decays.

The takeaway: multi-step chains are *harder* because you have more particles to track and more combinatorics to untangle, but the learning problem is continuous, not categorical. The same supervised learning toolkit applies; you just need to feed the network the right representation of the tree.