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: yes, but not in the way you might expect. The *structural* challenge of multi-step decays is qualitatively different, but modern machine learning can handle it with the right representation.

Why Multi-Step Decays Are Harder

In a two-body or prompt decay, you observe final particles and work backward to a single parent. The problem is relatively clean: given a muon and a kaon, is their invariant mass consistent with a B meson?

Multi-step decays are messier. A Λ_c⁺ hadron decays into a proton and a K⁰_s; the K⁰_s then decays further into two pions. You now have a *tree structure* of decays, and you must simultaneously:

1. Identify which pions came from the K⁰_s (not from combinatorial background).

2. Confirm that pair forms a valid K⁰_s.

3. Match that K⁰_s to the proton to form the Λ_c⁺.

The combinatorial background grows explosively—you're sifting through many possible ways to pair particles [AG-2025.04-1624].

How Feature Extraction Differs

Traditional approaches might flatten decay trees into simple variables (invariant masses, angles). But recent work shows that explicitly encoding the hierarchical structure extracts far more information.

The PASCL method, for instance, uses a "lowest common ancestor generations" (LCAG) matrix to represent the decay tree topology as a compact graph structure [AG-2024.02-1078]. This tells the neural network not just the properties of individual particles, but how they relate to each other across decay steps. The model then applies graph-based attention mechanisms to learn which inter-particle relations matter most.

This is qualitatively different from two-body decays, where you'd just use the four-momenta and masses directly [AG-2024.01-684].

Practical Payoff

The payoff is real: PASCL significantly outperforms baseline models on physics-based datasets by leveraging the full tree structure, whereas a simple MLP on flattened features would miss the crucial information that certain particles are siblings in the decay chain [AG-2024.02-1078].

For charm-hadron reconstruction via three-body decays, XGBoost and deep neural networks achieve high efficiency by learning to weight kinematic features that distinguish signal from the enormous combinatorial noise—but the gain comes from feature engineering informed by decay topology, not from the algorithms alone [AG-2025.04-1624].

Bottom line: Multi-step decays do require *different thinking* about feature representation—you need to encode the decay tree structure itself—but they don't require fundamentally new learning algorithms, just smarter feature design and graph-aware architectures.