Multi-scale Optimal Transport for Complete Collider Events
Tianji Cai, Nathaniel Craig, Katy Craig, Xinyuan Lin · 2025
Read Multi-scale Optimal Transport for Complete Collider Events by Tianji Cai, Nathaniel Craig, Katy Craig, Xinyuan Lin (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2501-10681.txt.
Introduction
Tianji Cai, Nathaniel Craig, Katy Craig, and Xinyuan Lin build a multi-scale tool for particle smash events in a 2025 paper. A smash sprays jets of particles, and each jet is itself a spray of hits in a sensor. Past tools that weigh how far one spray must shift to match another worked on lone jets only.
Their frame stacks that shift measure: jets as spreads on a curved space, and whole events as spreads of jets. Tests sort events by inner jet build in one task and by jet place links in the next. The nested view widens what the tool can read in smash data.
Catalog record
| CREATOR | Tianji Cai, Nathaniel Craig, Katy Craig, Xinyuan Lin |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 47 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2501-10681
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Multi-scale Optimal Transport for Complete Collider Events Tianji Cai, Nathaniel Craig, Katy Craig, Xinyuan Lin Abstract Building upon the success of optimal transport metrics defined for single collinear jets, we develop a multi-scale framework that models entire collider events as distributions on the manifold of their constituent jets, which are themselves distributions on the ground space of the calorimeter. This hierarchical structure of optimal transport effectively captures relevant physics at different scales. We demonstrate the versatility of our method in two event classification tasks, which respectively emphasize intra-jet substructure and inter-jet spatial correlations. Our results highlight the importance of a nested structure of manifolds in the treatment of full collider events, broadening the applicability of optimal transport methods in collider analyses. I Introduction Distance-based representations of collider data have been shown to provide powerful inputs to modern machine learning methods (see e.g. [ 1 , 2 , 3 , 4 , 5 , 6 , 7 , 8 , 9 ] for broad overviews). Treating collider events as distributions on the calorimeter naturally suggests the use of optimal transport (OT) distances, which quantify the “work” required to rearrange one distribution into another. In recent years, optimal transport methods have opened promising new avenues for analyzing LHC data [ 10 , 11 , 12 , 13 , 14 , 15 , 16 , 17 , 18 , 19 , 20 , 21 , 22 , 23 , 24 ] . 1 1 1 For a broad overview, see [ 25 ] . Other suitable distances between collider events have recently been proposed in [ 26 , 27 ] . Optimal transport has also been employed to great effect within machine learning architectures for collider physics, see e.g. [ 28 , 29 , 30 , 31 , 32 , 33 ] . In the usual application of OT to collider physics, both jets and entire events are considered as transverse momentum distributions in the rapidity-azimuth plane of the calorimeter. Optimal transport distances between these distributions are typically quite effective whenever the underlying task hinges on a single intrinsic scale in the data. However, entire collider events often feature multiple intrinsic scales, such as the scales associated with jet structure and substructure. This raises the natural question: What is the distance between multi-scale collider events? In this paper, we propose a distance between complete collider events that models them as distributions on the manifold of jets. Inspired by previous work, we endow the manifold of jets with a type of 2-Wasserstein (W 2 ) distance, induced by the Euclidean ground distance in the rapidity-azimuth plane of the collider. However, unlike previous work, we introduce additional weights which allow us to adjust the relative importance of jet substructure and the locations of the jets within the full collider event. Then, we endow the manifold of events with another 2-Wasserstein distance, induced by the above ground distance on the manifold of jets. In this way, our approach can be thought of as “W 2 -on-W 2 ”. This hierarchical structure allows us to effectively distinguish between differences between full collider events, at various scales of the data. Our paper is organized as follows. Section II gives an overview of our methodology. First, the usual single-scale optimal transport method is presented, with discussions on suitable representations that account for relevant symmetries acting at the jet and event level. We then introduce the new multi-scale optimal transport framework which incorporates both intra-jet substructure and inter-jet spatial correlations for quantifying the dissimilarity between two complete events. Section III applies the new method to two event classification tasks, i.e., t\bar{t} vs. QCD classification and BSM paired dijet vs. QCD classification, underlining the broad applicability of our method. The paper concludes with section IV , which highlights several promising directions for future work using the multi-scale OT framework. II Methodology II.1 The 2-Wasserstein distance We begin with a brief overview of optimal transport (OT) theory relevant to collider physics, focusing on the p -Wasserstein distance and the linearization of the p=2 case. Consider an event represented by its energy flow, i.e., a discrete distribution \mathcal{E}=\sum_{i}E_{i}\delta_{x_{i}} where the i th constituent particle, assumed to be massless, has an energy E_{i} and is located at position x_{i} in some pre-defined ground space \Omega . Here an ‘‘event’’ could denote either a jet or the calorimeter deposit of an entire collision event. 2 2 2 In the usual application of OT to collider physics, both jets and events are considered as single-scaled transverse momentum distributions in the rapidity-azimuth plane, with no distinction made among possible intrinsic scales in the data. In this subsection, we will refer to both as “events”, but will distinguish them beginning in section II.2 . Given events \mathcal{E} and \mathcal{E}^{\prime} with the same total energy, i.e., \sum_{i}E_{i}=\sum_{j^{\prime}}{E}_{j^{\prime}}^{\prime} , the p -Wasserstein distance between them is defined in terms of the minimization problem \displaystyle\text{W}_{p}(\mathcal{E},\mathcal{E}^{\prime}) \displaystyle=\min_{g_{ij^{\prime}}\in\Gamma(\mathcal{E},\mathcal{E}^{\prime})}\left(\sum_{ij^{\prime}}g_{ij^{\prime}}\big(d_{\text{ground}}(x_{i},{x}_{j^{\prime}}^{\prime})\big)^{p}\right)^{1/p}, \displaystyle\Gamma(\mathcal{E},\mathcal{E}^{\prime}) \displaystyle=\left\{g_{ij^{\prime}}:g_{ij^{\prime}}\geq 0,\sum_{j^{\prime}}g_{ij^{\prime}}=E_{i},\sum_{i}g_{ij^{\prime}}={E}_{j^{\prime}}^{\prime}\right\}, (1) where the transport plan g_{ij^{\prime}}\in\Gamma(\mathcal{E},\mathcal{E}^{\prime}) represents how energy is transferred from particle i in event \mathcal{E} to particle j^{\prime} in event \mathcal{E}^{\prime} . The W p distance measures the difference between two events based on the minimal effort required to transfer all the energy from \mathcal{E} to \mathcal{E}^{\prime} , where the effort of a transport plan g_{ij^{\prime}} is proportional to d_{\text{ground}}(x_{i},x_{j^{\prime}}^{\prime}) , penalizing the distance over which energy is moved in the ground space. The g_{ij^{\prime}} that achieves the minimum in 1 is known as the optimal transport plan . It is clear from the definition that the choice of ground metric encodes the essential physics of the problem at hand. In the context of hadron colliders, a natural ground space is the physical rapidity vs. azimuthal-angle plane of the cylindrical collider, i.e., \Omega=(y,\phi)\subset\mathbb{R}^{2} , with the ground metric being the Euclidean distance d_{\text{ground}}(x_{i},{x}_{j^{\prime}}^{\prime})=\|(y_{i},\phi_{i})-(y_{j^{\prime}}^{\prime},\phi_{j^{\prime}}^{\prime})\| in this y-\phi plane. Furthermore, the “energy” E_{i} is often replaced by the transverse momentum p_{T} of the particle. We refer to 1 defined on this ground space as the single-scale optimal transport problem. Different choices of p penalize the ground distance at different length scales. Two common choices are the W 1 metric, also called the Earth Mover’s distance [ 10 , 11 , 14 , 15 ] , and the 2-Wasserstein (W 2 ) distance [ 12 , 13 , 34 ] . The W 2 distance is unique due to its formal Riemannian structure [ 35 , 36 , 37 ] , which lends itself conveniently to efficient linearization, as discussed below. Therefore, the default choice in our study is p=2 . In general, optimal transport distances are more expensive to compute than classical distances between distributions, such as the l^{2} norm. For events containing n constituent particles, the Bertsekas’ auction algorithm for the W p distance demands \mathcal{O}(n^{3}) operations, while entropic regularization and the Sinkhorn algorithm requires \mathcal{O}(n^{2}\log(n)) [ 38 , 39 , 40 ] . In comparison, after binning two distributions, the l^{2} norm between distributions represented on a grid of size n only requires \mathcal{O}(n) steps. Furthermore, to use OT to compare the mutual differences in a typical collider dataset containing \mathcal{O}(10^{5}) events, one would need to calculate \mathcal{O}(10^{10}) pairwise OT distances, exacerbating the computational demand. One way to mitigate this