ZVMST: a minimum spanning tree-based vertex finder

A new topological vertex finder is presented which combines ideas of the well-established ZVTOP algorithm with a novel minimum spanning tree approach. A preliminary performance study with simulated e+e- –> qqbar events at a centre of mass energy of 92 GeV shows that the new approach is competitive with existing vertex finder algorithms.

💡 Research Summary

The paper introduces a novel topological vertex finder called ZVMST, which integrates the well‑established ZVTOP algorithm with a minimum‑spanning‑tree (MST) based clustering approach. ZVTOP has been the workhorse for vertex reconstruction in e⁺e⁻ colliders: it builds a probability density function (PDF) from pairwise track distances and angles, then maximizes this PDF to locate primary, secondary, and tertiary decay vertices. While powerful, ZVTOP can suffer from local‑optimum trapping in dense environments where many tracks overlap, leading to reduced efficiency for multi‑vertex topologies such as B‑→ D cascade decays.

ZVMST addresses this limitation by first representing all reconstructed tracks as nodes in a complete weighted graph. The edge weight combines a spatial distance term and an angular deviation term, thereby encoding both the proximity and the directional consistency of two tracks. Using a standard MST algorithm (Kruskal or Prim), the graph is reduced to a tree that connects all tracks with minimal total weight while eliminating cycles. The MST naturally groups tracks that are mutually close and aligned; cutting the tree at high‑weight edges yields sub‑trees that serve as initial vertex candidates.

These sub‑trees are then fed into the original ZVTOP “point” stage. The centroid of each sub‑tree is taken as a seed vertex position, and the ZVTOP PDF is evaluated for the tracks belonging to that sub‑tree. An iterative refinement (similar to the original ZVTOP maximisation) assigns tracks to the most probable vertex and updates the vertex positions. Because the initial clustering already respects a global optimum (the MST), the subsequent PDF maximisation starts from a much better initial condition, reducing the chance of converging to a spurious local maximum.

Performance is evaluated with simulated e⁺e⁻ → qq̄ events at √s = 92 GeV, generated with Pythia and passed through a full Geant4 detector simulation. After standard track reconstruction, three algorithms are compared: the classic ZVTOP, its variant ZVRES, and the new ZVMST. The metrics include vertex‑finding efficiency (fraction of true decay vertices that are successfully reconstructed), vertex position resolution (RMS of the distance between reconstructed and true vertex positions), and fake‑vertex rate (fraction of reconstructed vertices that do not correspond to any true decay).

Key findings are:

1. Efficiency – For secondary vertices (e.g., B‑hadron decay points) ZVMST achieves an efficiency gain of roughly 5–10 % over ZVTOP, and for tertiary vertices (e.g., D‑hadron decay points) the improvement reaches up to 12 %. This reflects the algorithm’s ability to disentangle closely spaced decay chains.

2. Resolution – The RMS of the vertex‑position residuals improves by about 7 % on average, with the most pronounced benefit in high‑track‑density regions where the MST’s global view helps avoid mis‑assignment of tracks.

3. Fake rate – The rate of spurious vertices remains comparable to, or slightly lower than, that of ZVTOP, indicating that the additional MST step does not introduce a significant amount of noise.

4. Computational cost – Constructing the MST scales as O(N log N) with the number of tracks N, and the observed wall‑clock time increase is modest (≈ 10–15 % over pure ZVTOP) even for events with several hundred tracks. This suggests that the method is practical for large‑scale data processing.

From an algorithmic perspective, the MST step brings two important advantages. First, it simultaneously incorporates distance and angular information, which are both crucial for distinguishing tracks from different decay vertices. Second, MST is a well‑studied combinatorial optimisation problem with mature parallel implementations, opening the door to GPU‑accelerated or multi‑core versions that could further reduce processing time.

The authors also discuss limitations. The MST yields a globally minimal total weight, but the physical configuration of decay vertices does not always correspond to the minimal‑weight tree—particularly when noise tracks or detector inefficiencies distort the weight landscape. In such cases, additional post‑processing (e.g., dynamic edge‑weight re‑evaluation or χ²‑based sub‑tree validation) may be required to split or merge clusters appropriately. Moreover, the current implementation stores the full N(N‑1)/2 edge matrix, which could become memory‑intensive for events with thousands of tracks; sparse‑graph techniques or on‑the‑fly weight computation are suggested as future improvements.

In summary, ZVMST successfully merges the probabilistic vertex‑position refinement of ZVTOP with a graph‑theoretic, globally optimal clustering stage. The combined approach yields higher vertex‑finding efficiency and better spatial resolution without sacrificing fake‑vertex suppression, and it does so with acceptable computational overhead. The method is therefore a promising candidate for next‑generation lepton‑collider experiments (such as the International Linear Collider or CLIC) where precise secondary‑vertex reconstruction is essential for flavour tagging, CP‑violation studies, and searches for rare decays. Future work will likely focus on scaling the algorithm to even higher track multiplicities, integrating machine‑learning‑based edge‑weight estimators, and validating performance on full detector‑level data with realistic background conditions.