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, multi-step decay chains introduce distinct challenges that require different architectural choices than prompt or two-body decays — but the distinction is more about *structural complexity* than fundamental physics.

Why the difference matters

In a prompt or two-body decay, you observe a simple event: a particle decays directly into its final products, which hit your detector. Feature extraction is relatively straightforward — you measure momenta, angles, invariant masses, and feed them into standard classifiers.

With multi-step chains (like $\Lambda_c^+ \to pK_s^0$ where $K_s^0 \to \pi^+\pi^-$), the challenge is topological: you must infer a *hierarchical tree structure* from particles you can only see at the very end [AG-2024.02-1078]. This isn't just about measuring properties; it's about reconstructing *which particles came from which intermediate decay*.

How practitioners adapt

For multi-step chains, the field has moved toward graph-based and tree-aware architectures:

PASCL introduces a compact "lowest common ancestor generations" (LCAG) matrix representation to encode the full decay tree structure, then applies supervised graph contrastive learning to learn inter-particle relations across the entire chain [AG-2024.02-1078]. This is fundamentally different from asking "do these two particles form a two-body resonance?"

For charm hadron reconstruction via three-body decay ($\Lambda_c^+ \to pK_s^0, K_s^0 \to \pi^+\pi^-$), traditional feature extraction struggles with combinatorial background — not just separating signal from noise, but identifying *which* pion pair actually came from the $K_s^0$. Machine learning (XGBoost, DNNs) handles this by learning the kinematic correlations across all steps at once [AG-2025.04-1624].

When simpler methods still work

Interestingly, highly specialized architectures (Graph Attention Networks, Particle Transformers) don't always dramatically outperform simpler approaches on all decay problems [AG-2024.08-1184]. For tasks where the decay structure is well-constrained (e.g., heavy-flavor semi-leptonic decays with clear kinematic signatures), basic features like transverse momentum, pseudorapidity, and impact parameter can achieve ~98% accuracy even in standard boosted-tree classifiers [AG-2025.08-1021].

Bottom line

Multi-step chains *do* require different thinking — you need to represent *hierarchies*, not just pairs of properties. But "different" means choosing graph or tree-aware models; it doesn't mean the underlying physics is categorically new. Two-body and prompt decays are point problems; multi-step chains are *structural* problems.