WorksPhysics and Astronomy
Collider event distances from the phase space sphere
Abstract
Can the distribution of collider events in phase space be used to distinguish the processes that produced them? We present a distance-based classification method that uses the global topology of Lorentz-invariant phase space and retains reconstructed masses. To tackle particularly challenging classification tasks that may benefit from higher event resolution, we develop a polynomial-time approximation to compute the distances. The resulting classifiers achieve strong discrimination across several simulated benchmarks, including fully hadronic four-top production against .
Introduction
Collider observables are functions on phase space that probe the dynamics of the underlying field theory, making phase space itself an essential substrate for understanding the Standard Model and searching for physics beyond it. A textbook understanding of phase space focuses almost solely on its volume measure, as that is all that is needed to calculate cross sections and decay rates in quantum field theory [1, 2, 3, 4]. In this paper, we take a step back and use the global topology of relativistic phase space as a starting point for constructing event distances. Treating phase space geometrically enables a sensible definition of distances between entire collider events. Progress towards the understanding and application of phase space has recently been made in Refs. [5, 6]. We find significantly improved performance on classification tasks relative to previous applications of phase space distances [7].
Lorentz-invariant phase space () is the domain of integration for the matrix element in a relativistic Quantum Field Theory. Cross sections and other observables are expressed as integrals over phase space,
Much of the attention naturally goes to the matrix element, and correctly so, as it encodes all the dynamics of the theory. Phase space enforces the kinematic constraints of spacetime and the particle mass shells, and it isn’t precisely ignored in the broader literature: it is central in event generation and Monte Carlo integration, unitarity, EFT bases, amplitude bases and many other areas [5, 8, 9, 10, 11, 12, 13]. For collider applications, phase-space geometry has also informed anomaly detection, through topology-aware latent spaces [14] and event-level observables for weakly supervised searches [15]. More recently, relative representations constructed from massless phase-space and Wasserstein distances have been combined with particle-identification information for binary tagging of individual boosted jets [16]. For event classification purposes, the matrix element tells us the distribution of events in this high dimensional space. Therefore, a natural question is: does the event distribution in phase space carry information useful for classification?
Ref. [7] studied this exact question in massless phase space by using lightcone coordinates on an open domain [5], which follow from a spinor-frame description [6]. That classifier has already been used in a CMS search for microscopic black holes and electroweak sphalerons, including BlackMax signals against QCD multijets [17, 18]. Alongside other recent distance-based approaches (including the energy mover’s distance [19], linearized and unbalanced optimal transport [20, 21], spectral metrics [22], and multiscale constructions for complete events [23]), distances based on the geometry of phase space complement conventional observables for event classification and anomaly detection [24, 25]. The massless approximation used by the earlier phase-space classifier [7] discards clustered-object masses and becomes less accurate at coarse resolution, motivating a comparison that retains them throughout the resolution scan.
We start from the known spherical topology of fixed-mass phase space [26, 27]. We present a simple radial map that assigns each momentum configuration to a unit vector, defining an angular distance. Including masses among the coordinates extends this construction to events with different reconstructed masses. The resulting radial representation, its convex energy–momentum counterpart, and two energy-weighted spherical representations provide the four comparisons studied here, illustrated in Figure 1. After removing each event’s overall invariant-mass scale, these representations retain the same four-vector kinematics but weight differences in energy sharing, momenta, and masses differently. Their classification performance can therefore differ even though they preserve the same normalized event information.
Figure 1. Schematic construction of event distances from phase-space geometry. For , where is center-of-mass energy, fixed-mass phase space has spherical topology. Momentum is decomposed into polar coordinates . Allowing reconstructed masses to vary leads to a convex representation and several distinct spherical representations of collider events.
The resolved multiplicity of a collider event can greatly exceed that of the hard scattering: showering and hadronization populate additional momenta, while exclusive clustering resolves the final state into a chosen number of pseudojets. Fully hadronic four-top production is a demanding example: the top-pair-plus-jets background already contains real tops and substantial radiation, so a top tag alone cannot distinguish the two processes [28]. Semiclassical black-hole signals provide a complementary case, motivating comparisons of many-particle kinematics without assuming a fixed resonance-decay chain [18]. Additional radiation can itself help distinguish signal from background: an ATLAS search for R-parity-violating gluino decays gains sensitivity by requiring a jet beyond the six-quark decay topology [29]. Finer exclusive- resolutions can expose more of this decay and radiation structure, while at coarser resolutions some of it is retained in pseudojet masses. The comparison should therefore accommodate the masses produced by clustering and remain computationally tractable as the resolved multiplicity increases.
At high multiplicity, explicit minimization over all particle permutations is computationally prohibitive. We use the Hungarian algorithm [30] to efficiently find the one-to-one matching of particles between two events that minimizes a chosen matching cost with their orientations held fixed, followed by alignment in the transverse plane. This sequential procedure approximates the joint minimum over assignment and alignment. We use pairwise distances to train a support vector machine (SVM) on labeled signal and background events. The same four-vector preprocessing and event comparisons can be used to train classifiers for different signal and background processes, without constructing process-specific discriminating observables.
We study how the event representation and resolved multiplicity affect discrimination in three particle-level benchmarks: versus QCD dijets, fully hadronic four-top versus +jets, and conditioned semiclassical black-hole signals versus QCD multijets. In the massless –QCD comparison, the updated prescription using Hungarian assignment substantially reduces the earlier method’s loss of discrimination at high multiplicity. Among the complete pipelines compared, the energy-weighted spherical representation that treats all particles symmetrically achieves the highest discriminating performance at the largest tested multiplicity in all three benchmarks. Its discrimination improves or approaches a plateau as the resolution increases, with four-top and black hole reaching broad plateaus near and , respectively. Rankings at lower multiplicity and the useful resolution remain process dependent.
Section § 2 develops the physical event space from energy and momentum conservation, and variable reconstructed masses, establishing the basis for spherical and convex comparisons. Section § 3 defines the four mass-aware representations and their associated metrics. Section § 4 specifies the normalization, assignment–alignment procedure, and supervised classifier, together with the massless benchmark and common evaluation protocol. Section § 5 assesses the resulting discrimination as a function of resolved multiplicity. The appendices collect numerical and computational details, phase-space measure treatment, formal topology and metricity proofs, and spinor-helicity details.
From phase space to reconstructed events
Our goal is to compare collider events through their structure in phase space. We first establish its topology: its global shape, including how it is connected and whether it has holes or boundaries. We begin with fixed particle masses, then enlarge the space to accommodate the varying masses of reconstructed objects. We also describe the same events in different coordinates, providing complementary views of this enlarged space. This common topological foundation leaves freedom in how differences between events are measured, motivating the distances introduced in § 3. Particle matching and alignment are addressed in § 4.
Fixed-mass phase space
We can define phase space as the set of on-shell momenta that satisfy energy and momentum conservation rather than in the traditional fashion through the phase-space measure. In the center-of-mass (COM) frame, with total momentum , for particles, , and , define -particle fixed-mass phase space as
where are the momentum 3-vectors. Equation (1) includes when , and the particle masses are fixed. The motivation for proposing this change in perspective is that correct parametrizations of an integral may discard essential topological information. Removing a measure-zero region like from the integration domain in does not change its value, but it has topological consequences. The phase-space measure and the treatment of zero momentum are specified in Appendix § 9.
The global topology of phase space is particularly simple and well known [26, 27]. Early event generation literature used several of its features,1 even when it isn’t directly stated. Topological equivalence (denoted ) is defined by whether two spaces are homeomorphic, meaning we can construct a bijective map from a space to another that behaves continuously in both directions. This map is called a homeomorphism. There are other equivalent ways to show topological equality. In this section, we will construct an explicit homeomorphism from a unit sphere onto phase space. Essentially, we are about to show that a ping-pong ball and a balloon are topologically equivalent, but for a higher dimensional shell. Further, a homeomorphism is useful to define a distance on the target that holds on the original space.
Let where . These are the normalized momentum components with respect to the concatenated vector magnitude, meaning is a polar decomposition, where the radius is in principle allowed to vary with the angular direction. Therefore, lives on a unit sphere in the -dimensional momentum conserving vector space
This is a -dimensional unit sphere. Let be the unique solution of the energy conservation constraint at each ,
Then, the maps from the unit sphere to phase space, and vice versa
are mutually inverse homeomorphisms. The resulting topology is a -dimensional sphere:
Intuitively, along each momentum-space ray, the total energy increases monotonically from and reaches at the unique radius . Thus the spherical direction specifies the complete momentum configuration up to the scale fixed by energy conservation. We can confirm the dimension of the sphere is correct: we started with momentum components, and we have applied 4 conservation laws. For , the shell is perfectly round, and its radius is where is the Källén function [31]. For , the shell is smoothly rounded, albeit ovoidal, when all masses are positive, but can develop sharp features where massless particles have zero momentum, as illustrated schematically in Figure 2.
Figure 2. Schematic illustration of fixed-mass phase-space manifold geometry. A shell with massless particles can develop sharp features.
This homeomorphism transports the round metric, a simple angle on
to phase space
The induced angle is a genuine metric on each fixed-mass shell and provides a prototype for the radial comparison with variable masses. The continuity of Eq. (4) is physically clear but we formally show it in Appendix § 10.§ 1, as well as the metric statement. In principle, the radius could jump between momentum configurations, and we check formally that does not happen. In principle, there are infinitely many other metrics we could define in this space.
Another particularly natural choice would be the shortest path directly on the shell, in the momentum ambient space, equipped with its Euclidean inner product. Due to the geometry, this would require a geodesic calculation in general. Thus, for event comparison, we use a radial map analogous to Eq. (4): normalizing the combined momentum vector identifies each configuration with a unique point on a unit sphere, so the angle between two such vectors defines a distance on the shell. We next extend the comparison to events with different reconstructed masses.
Adjoining the fixed-mass shells
Fixed-mass phase space is appropriate when the identities and masses of the final-state particles are fixed. Reconstructed collider events instead consist of jets or pseudojets whose invariant masses fluctuate from event to event. To compare such events directly, we therefore enlarge phase space by treating each reconstructed mass as a coordinate.
Promoting the masses from fixed parameters to nonnegative coordinates joins the separate fixed-mass shells into one fixed- event space. Concatenate the into with , and define an energy-based norm on this subspace, and the standard Euclidean norm
The signed shell and physical sector are
Negative are only bookkeeping coordinates for the signed extension. Each fixed-mass phase space is a slice in this enlarged space. The available kinetic energy is . At the threshold , each fixed-mass shell reduces to the single configuration with all spatial momenta zero. Masses and momentum directions therefore cannot be chosen independently throughout this space.
The signed shell has spherical topology as well, directly because is a norm. We are fixing the length of all with respect to this norm, which gives another radial homeomorphism that we don’t explicitly need. Its dimension, , comes from the dimension of fixed-mass phase space plus promoted mass variables. For and ,
Here denotes the closed -dimensional ball. Restricting a single coordinate cuts the sphere into a closed hemisphere, which is topologically a ball. All remaining impose further orthant restrictions in the dimensional ambient space, but do not change the topology. Appendix § 10.§ 2 gives the complete proof underlying this topology.
, henceforth the collider dome, will be our primary classification space on which to define angular distances. Specifically, normalize the complete mass–momentum vector:
so lies in the section of the unit sphere. At fixed this map is injective, with image a spherical sector of dimension . Radial normalization induces the angular comparison defined in § 3.
Convex energy–momentum coordinates
We constructed a space suitable to host momenta with variable nonnegative masses, . However, its ball topology immediately provides us with an additional opportunity: we can perform a very familiar non-linear transformation on the whole space to turn it into a convex shape. This convex shape can be equipped with a straight line distance, which will turn out to be a particularly simple choice. These two spaces can then be directly compared in discrimination performance.
Replacing each mass coordinate by the energy makes the total energy constraint linear. The overall dimension still remains , the masses remain non-fixed, but the points are now represented by rather than . Requiring real nonnegative masses is equivalent to for nonnegative energies, so the same events are described by
The maps
are continuous inverses. They are diffeomorphisms2 on the all-massive regions, while the inverse has the expected square-root behavior at a massless block. Although preserves the event information, Euclidean distances in generally differ from angular distances in normalized . Unlike the block-norm shell, is convex. If and are allowed and , then conservation is affine and
Thus the entire chord lies in . This region is closed and bounded, has dimension after imposing energy–momentum conservation, and contains , in its interior relative to those constraints. It is therefore a convex realization of the same topology. This convex realization is called a spectrahedron, which is shown explicitly in Appendix § 11.§ 2.
For two blocks at , conservation leaves and , so
This is a four-dimensional solid bicone; restricting the momenta to one fixed axis gives a filled diamond. The bicone in the convex panel of Figure 1 serves as a schematic of the convex representation for higher-multiplicity events.
The spectrahedron’s straight-line distance is defined in § 3. Straight-line and angular minimizing paths generally do not map into one another under the change of representation.
Connection to massless phase-space and spinor frames
We have two spaces suitable for event comparison; this last section does not introduce another one. Here, we make a quick but necessary connection to the spinor literature.
Setting all masses to zero connects the event space to the spinor and lightcone descriptions used in the earlier classifier. For and , the massless event space can be written as
Here consists of pairs of orthonormal complex vectors, which we call normalized spinor frames. Dividing by removes each particle’s little-group phase (a kinematic redundancy), leaving the physical momentum shell [5, 6]. The lightcone coordinates used in the earlier classifier [7] describe the region where for every particle as . This region contains all events except those with particles satisfying . The product describes only this open region: at its boundary, the lightcone coordinates break down, and some angular degrees of freedom become equivalent or collapse. Appendix § 11.§ 1 gives the variables and explains how they relate to the full, closed shell. The same spinor connection extends to nonzero masses, as developed in Appendix § 11. We will only use the product space for comparison with the earlier classifier in § 5.
Figure 3 summarizes this section and the connections between the constructions.

Figure 3. The circle of spaces used in the paper. The arrows record coordinate equivalences, varying the mass parameters, and taking determinant slices; the natural metrics carried by the representations need not agree under these maps. On the massless determinant slice, the spinor-frame quotient reduces to , which is homeomorphic to the massless fixed shell . The spinor frames do not play a role in classification.
Representations for event comparison
The mass–momentum and energy–momentum descriptions retain the same events but assign different distances to them. We compare four choices: angles between radially normalized mass–momentum vectors, Euclidean distances between energy–momentum vectors in the event rest frame, and angles between two types of energy-weighted features. Each representation retains the full event at fixed scale on its stated domain. Their classification performance can nevertheless differ because their distances weight kinematic differences differently. The metric and pseudometric statements are proved in Appendix § 10.
Radial and convex comparisons
From now on, let denote two collider events. With , radial normalization of the signed shell gives
At fixed this is a genuine labeled metric on and the physical sector . We call this branch radial spherical.
The convex representation carries the particularly simple Euclidean energy–momentum metric
It is a genuine labeled metric. On each fixed- ball, and likewise on the variable-scale positive cone, Eq. (19) is the intrinsic length distance induced by the Euclidean line element; its unique constant-speed minimizing segment is
Thus the convex energy–momentum branch uses the exact straight-line geodesic metric of the convex collider ball. Related momentum-space event distances, minimized over particle assignments, have been used for negative-weight reduction in Monte Carlo samples [32].
The radial and convex constructions thus define distinct metrics on the same fixed- physical event space. The former compares directions in mass–momentum coordinates, while the latter measures Euclidean separation in energy–momentum coordinates; both retain the reconstructed masses. Their coordinate relation is given in Eqs. (13) and (14).
Moving-frame feature representations
The radial spherical distance in Eq. (18) follows naturally from the radial homeomorphism. It normalizes the concatenated mass–momentum vector by a single overall norm, giving greater weight to the most energetic objects. The topology alone does not require this weighting. To give subleading kinematic structure greater relative influence, we construct two alternative spherical representations that first divide each object’s momentum and mass by the square root of its energy, before applying a common normalization. Define
At a zero-energy block, set . This is continuous because as . Write and similarly , , , and . These variables obey
The three spatial component vectors therefore lie in , which varies with the event’s energy fractions. We place them in a fixed subspace by choosing a nonnegative unit reference vector and applying the Householder reflection
where is the identity. The denominator is strictly positive for nonnegative and . The transformed vectors , , and lie in . The reflection acts in particle-label space and leaves unchanged; its orthogonality preserves the unit norm of .
Democratic representation. To treat every object symmetrically, choose
Using this reference in the transformed vectors above, define
Because all components of are equal, relabeling the input objects simply relabels the output features. At fixed , this map is a homeomorphism from onto the full spherical sector , including zero-energy limits. Up to coordinate ordering, this is the same sector as the radial spherical map’s image. The two maps retain the same event information while generally assigning different distances.
Pivoted representation. The pivoted variant tests a comparison in which the highest- object in each event serves as a distinguished reference, with these leading objects always matched to one another. For this variant, restrict to
and place the unique highest- object in slot . Choose
The transformed vectors now satisfy . Dropping these three zero entries gives
This map is injective on the positive-energy labeled domain with a unique pivot already in slot . Highest- ties are excluded from these metric statements, since choosing a pivot there can introduce ambiguity or discontinuity.
For either moving-frame representation, the angular distance is
These are metrics on the fixed- labeled domains specified above (Corollary C.6). On the closed democratic image, the angle is also the shortest-path distance within the spherical sector; no corresponding claim is made for the pivoted image.
Table 1 summarizes the representations. The classification pipeline combines these with eventwise scale normalization, particle assignment, and transverse alignment. The resulting pairwise similarities provide the input to the supervised classifier specified in the following section.
| Representation | Compared vector | Distance | Domain |
| Radial spherical | Angular | ||
| Convex energy–momentum | Euclidean | ||
| Democratic | Angular | ||
| Pivoted | Angular |
Table 1. Representations and labeled distances at fixed .
Scalable comparison on reconstructed-event space
We have equipped our collider spaces with distances so we are in principle ready for classification. First, we introduce the simulated samples, event selections, and laboratory-frame clustering. We consider three benchmark scenarios: versus QCD dijets, four-top production versus + jets, and semi-classical black hole signals versus QCD multijets. Next, we adapt the fixed- constructions of § 2 and § 3 to events with varying invariant mass by boosting each retained system to its center-of-momentum frame and normalizing its four-momenta to . We then describe the sequential assignment–alignment procedure that approximates joint minimization over particle permutations and transverse rotations and reflections, together with its implementation for each representation. Finally, we convert these comparisons into similarities for a support vector classifier and specify the training and evaluation protocol.
Event generation and preprocessing
The input consists of stable final-state particles, with invisible species excluded by default and no constituent transverse-momentum threshold. The acceptance cuts in Table 2 and Table 3 define . Accepted particles are clustered in the laboratory frame with exclusive , , using FastJet [33]. Each scan uses one event cohort and the same class balancing at every ; the retained four-vectors are boosted to their center-of-momentum frame and divided by their invariant mass as in Eq. (33), with the massless-branch preprocessing specified in § 4.3.1. No zero padding is used: the selected cohorts support the maximum scanned multiplicity, apart from fewer than of events with fewer than 30 accepted particles, which were discarded. Comparisons at other multiplicities or with other jet definitions can use the same prescription whenever the chosen common multiplicity is supported by the sample.
| versus QCD | versus jets | |
| Collider energy | TeV | TeV |
| Signal | showered | MadSpin-decayed, all-hadronic |
| Background | , and their antiparticles | MadSpin-decayed, all-hadronic MLM-merged , with including |
| Generated events | 10000 signal; 10000 background | 25000 signal; 50000 background |
| Matrix-element selection | dijet GeV; inclusive | GeV for both classes |
| Particle-level selection | GeV, | GeV, |
| Events after selection | 5826 signal; 6975 background | 11357 signal; 19089 background |
| Balanced classifier sample | 5000 events per class | 11357 events per class |
| Multiplicity scan |
Table 2. Generation and particle-level selections for the two 14 TeV benchmarks. The balanced sample sizes are per-class caps applied before the stratified training–test split.
| BlackMax B1 signal | QCD background | |
| Collider energy | TeV | TeV |
| Hard process | BlackMax B1 | , and their antiparticles |
| Common selection | TeV, | TeV, |
| Particle multiplicity | at least 30 accepted particles | at least 30 accepted particles |
| conditioning | TeV, | TeV, |
| After all cuts | 6907 | 8033 |
| Balanced classifier cap | 6907 | 6907 |
Table 3. Selected BlackMax B1 versus QCD sample. The final row gives the per-class cap used for classification.
Standard Model samples
The two Standard Model benchmarks are versus QCD dijets and four-top production versus jets, both at TeV. The leading-order hard events were generated with MadGraph5_aMC@NLO 3.5.13 [9] and showered and hadronized with Pythia 8.3.16 [34]. We work at particle level, providing a common basis for future detector-response and pileup studies. Table 2 summarizes the samples and selections.
The top-pair comparison follows the kinematic benchmark of Ref. [7], using inclusive showered with unrestricted top decays. The dijet generation additionally requires GeV, , , and leading and subleading parton transverse momenta of at least 150 and 100 GeV. This dataset’s classification task provides a relatively simple setting for comparison with the earlier method; unlike in the other two tasks, invisible particles were retained in this sample. In the four-top comparison, sums transverse momenta of the non-incoming matrix-element final-state particles before the common on-shell all-hadronic MadSpin decays [35]. Only the background is -MLM matched, with matching scales 20 GeV at matrix-element level and 30 GeV in the shower. Increasing resolves additional prongs and radiation and therefore changes the event representation.
BlackMax B1 black hole versus QCD multijet sample
Extra spatial dimensions allow microscopic black holes to be produced at LHC energies [36, 37]. At 13 TeV, we compare the tensionless, nonrotating BlackMax B1 scenario [38] with leading-order QCD multijets. The signal has Planck mass TeV, mass TeV and extra dimensions; string-ball production, graviton emission, and the modeled mass, momentum, and angular-momentum losses are disabled. The background combines with four light quark flavors, using MLM matching scales of 300 and 450 GeV at matrix-element and shower level, respectively. Its generation requires GeV, , and TeV. Both samples are showered with Pythia 8 using the CP5 tune [39].
The unconditioned samples are strongly separated by global observables. To reduce this separation, a common window is followed by restrictions on the invariant mass and linearized transverse sphericity, as summarized in Table 3.
Here is the invariant mass of the sum of the laboratory-frame objects, and
These cuts define one event cohort, held fixed throughout the multiplicity scan. Residual one-variable discrimination is reported in Appendix § 8.§ 2.
Eventwise invariant-mass normalization
After particle selection and jet reconstruction, the invariant mass of the retained system varies from event to event, even at fixed resolved multiplicity . To compare these events using the fixed- constructions above, we boost each retained system to its COM frame and divide all its four-momenta by its invariant mass. This gives a common normalization, removing the overall invariant-mass scale while retaining mass ratios and relative kinematics.
For each event at exclusive multiplicity , define
Let boost the retained system to its center-of-momentum frame, , and define normalized (barred) momenta as
Figure 4 shows the distribution of for a representative signal cohort. The retained-system invariant mass is broadly distributed at every displayed , with a median that increases from 340 GeV at to 675 GeV at . Because the hadron-collider exclusive- history includes beam assignments, the four-vector sum of the retained jets, and hence , also need not be independent of . The division in Eq. (33), as well as the center-of-momentum boost, selects a common unit- representative.
Figure 4. Unit-area distributions of the retained-system invariant mass for the indicated exclusive- multiplicities, using the first 5,000 accepted events from the generated signal cohort defined in Table 2. These distributions illustrate the eventwise scale variation removed by Eq. (33).
All numerical comparisons below use the representative. This normalization sends to , or equivalently places their energy–momentum coordinates in . A fixed-mass shell is sent to . Because these mass ratios vary across events, the normalized sample lies in the variable-mass collider spaces rather than one fixed-mass sphere.
While the fixed-scale labeled distances remain metrics on the normalized image, their pullbacks to the COM event family are pseudometrics, identifying events that differ only by a common positive rescaling. This means that events that only vary by an overall scale are distance zero. One can imagine that this affects events that are close to an overall scale apart as well: events with similar normalized kinematics can lie close together even when their invariant-mass scales differ substantially. Thus, the loss of scale information affects event proximity beyond exactly rescaled copies. For example, let contain four equal-energy massless particles arranged in two orthogonal back-to-back pairs, and obtain by rescaling the pairs by and , respectively. Although , their normalized distances are
Thus, approximately rescaled events can have arbitrarily small distances regardless of their overall scale separation. The probability of this for a given pair goes roughly like
and over all the pairs across ,
and so we leave studying this particular issue deeply to further work.
Exact assignment target versus implemented dissimilarity
Events described only by reconstructed four-momenta have no preferred particle labeling, while the beam axis singles out the longitudinal direction but no preferred azimuthal orientation. Following [7], we compare events up to allowed particle permutations and transverse rotations and reflections, represented by . Exhaustively minimizing over all assignments and their transverse alignments becomes impractical at high multiplicity, motivating the scalable sequential approximation described below.
For normalized events , , let denote the chosen labeled distance and the allowed permutation group: for unrestricted matching or when a distinguished pivot is held fixed. The exact distance between symmetry orbits and is
This is a metric on the corresponding normalized representation-orbit space, as established in Appendix § 10.§ 4. This applies to all § 3 representations.
We approximate this joint minimum in two steps. First, the Hungarian algorithm [30] selects an assignment in the input frame,
where is the cost of matching block in to block in . The representation-specific costs are given in § 4.3.1. We then minimize over transverse alignment with this assignment held fixed:
The Hungarian step is exact for the chosen fixed-frame additive objective, and the step is exact for the selected assignment. However, the fixed-frame optimal assignment need not minimize the distance after alignment. Their composition therefore gives a polynomial-time dissimilarity3 on the input representatives, without the metric guarantee of the exact joint minimum.
For a fixed assignment in a spherical representation, collect transverse coordinates into and retain real longitudinal and mass-like coordinates .4 Define
The two quantities encode the rotational and reflection branches of . The exact fixed-assignment cosine is
Appendix A.3 quantifies the approximation for all four mass-aware branches at , comparing the sequential procedure with exhaustive minimization over each branch’s allowed permutations and exact transverse alignment.
Compared representations
We compare five representations built from the same reconstructed events. Four retain reconstructed masses: the convex energy–momentum representation uses a Euclidean distance, while the radial and two moving-frame representations use angular distances. The massless lightcone representation provides a benchmark against the earlier method [7]. For each representation, we specify the cost matrix used in the assignment–alignment procedure described above. The massless benchmark is also evaluated using its original rank ordering.
For either moving-frame branch, write .
Radial spherical. Use the normalized physical blocks , with on physical data, in Eq. (18). The fixed-frame assignment cost is . Hungarian assignment first computes the additive permutation target exactly; the subsequent alignment gives the implemented two-stage distance in Eq. (39).
Pivoted moving frame. Use from Eq. (29), hold the highest- block fixed, and assign the remaining blocks using . This defines the pivoted branch.
Democratic moving frame.
Use the permutation-equivariant map defined by Eq. (26), with Householder vector , and assign all transformed blocks with . This same convention is used in all three production studies.
Convex energy–momentum.
For normalized events , a fixed assignment, and , define . Then
The fixed-frame additive cost
is minimized exactly by Hungarian assignment. For this branch set , , and . At fixed , minimizing the transverse term is equivalent to maximizing the transverse cross term . The labeled distance is the straight-chord distance of Eq. (19); the implemented assignment–alignment composition remains sequential.
Massless lightcone benchmark (Rank-ordered and Hungarian).
For both massless comparisons, we follow the event representation and preprocessing of Ref. [7], using the original implementation. After exclusive- clustering, each block is projected to the massless shell by replacing . The resulting massless event is then boosted to its own COM frame and normalized by its total COM energy before the simplex–sphere coordinates of Refs. [5, 7] are constructed.
The legacy baseline orders the blocks by decreasing . Both massless branches use the plus lightcone convention, , for the simplex coordinates. The legacy curves use sphere-to-simplex weights . The full implementation and coordinate definitions for are given in Ref. [7].
The massless lightcone–Hungarian branch leaves the massless projection, boost, normalization, lightcone convention, weighting, and final transverse alignment unchanged. It replaces rank ordering with fixed-frame Hungarian assignment and uses the highest- block as its pivot. The legacy baseline instead uses the lowest- block, which occupies the final slot under descending ordering. On the non-pivot labels, the assignment cost is
After selecting with this cost, we evaluate the same fixed-assignment simplex–sphere distance as in the legacy implementation, including exact minimization over the common transverse action, as in Eq. (39). The massless lightcone-Hungarian branch is included primarily to benchmark the new representations against the approach of Ref. [7]. To limit the computational cost of this comparison, we fixed separately for each task to the value yielding the best performance in pilot runs, without repeating the full analysis for all three values, .
Classifier and precomputed similarities
For each distance matrix , the implementation forms
with , and supplies it as a precomputed kernel to a support vector classifier [40, 41]. The metric properties of the idealized event comparisons do not by themselves guarantee positive semidefiniteness of these Gaussian similarities. A sufficient condition, and a necessary one for such a guarantee at every , is conditional negative semidefiniteness5 of . Squared geodesic distances on curved manifolds do not generally satisfy this condition [42].
We have not systematically verified positive semidefiniteness of the training kernels produced by the implemented assignment–alignment comparisons; for some diagnostic cases, particularly at low , negative eigenvalues were observed. A particularly extreme instance of this was for an democratic sphere kernel where . Because the Hessian of the SVC dual minimization objective has the same eigenvalues as the kernel matrix, an indefinite kernel can make the optimization nonconvex, so the solver’s stopping criterion need not certify a global optimum [41]. Nevertheless, classification performance was evaluated by applying the trained classifiers to unseen, held-out events. The reported test AUCs (Area Under receiver operating characteristic Curve) therefore measure the discrimination actually achieved by the fitted models, including any effects of convergence to a suboptimal solution. The absence of a convexity guarantee leaves open whether a better value of the training objective could have been obtained, but it does not invalidate the evaluation of the resulting classifier on independent test data. Global optimality of the training solution is not required for this assessment of predictive performance.
The data are split with stratification into 70% training and 30% test sets. For each outer training split, set
The common bandwidth grid is
The massless lightcone–Hungarian weight is for and BlackMax and for four-top; the massless grids have task and branch specific extensions.6 For each outer split, five independently shuffled three-fold cross-validations, with inner seeds 0–4, each select a pair by maximizing the mean validation AUC over its three folds; each selected pair is refitted on the full outer training sample and evaluated on the same held-out test sample. For a given outer seed, the reported AUC is the held-out test AUC of the candidate with the highest mean validation AUC among these five inner runs. In the extremely rare case of validation AUC ties, the two or more test AUCs are computed and averaged. Each plotted point is the mean of these selected test AUCs over outer seeds that generate new stratified training–test splits, including a new held-out sample. Its error bar is the standard error across outer splits, to quantify variation across repeated outer splits.
Once trained, the SVM classifies a new event by comparing it with a subset of the training events called support vectors. The distance to each support vector is converted into a kernel similarity using the selected , and these similarities are combined with learned coefficients and a bias to obtain a decision score. With signal assigned the positive label, the sign determines the predicted class, while larger scores indicate greater signal likeness. Under the positive semidefinite kernel assumption, this score is proportional to the signed distance from the learned decision hyperplane in the associated feature space.
Classification results
The same reconstructed event is compared here through several coordinate and feature representations of the unit- collider spaces. The convex energy–momentum, radial, and energy-adapted moving-frame constructions retain the complete normalized massive event on their stated domains, with the moving frames taking values in spheres; the massless lightcone branch connects to the earlier pipeline. The section first compares the legacy and updated massless prescriptions, after which all five representations are compared on the two Standard Model samples and the conditioned BlackMax sample. Since the coordinate or feature map, treatment of mass coordinates, and assignment cost vary together, the AUC curves rank the complete implemented representations. The democratic moving frame has the highest mean AUC at the endpoint of all three scans, with process-dependent crossovers and saturation as the resolution is increased.
versus QCD dijets
The benchmark is a natural choice to compare the present construction to the event-level classification of Ref. [7]. Figure 5 compares the legacy lightcone ordering with the updated massless prescription, which combines Hungarian assignment with a highest- pivot and subsequent transverse alignment. The main change is the behavior at high multiplicity. The legacy curves reach their maxima around – and then lose discrimination, whereas the updated branch reaches a broad maximum around – and retains most of that performance through . At matched weight , the endpoint gain is AUC percentage points over the 12 common outer splits. Thus, the combined update to assignment and pivot choice substantially improves the use of finely resolved events within the massless construction.
Figure 5. Mean test AUC versus exclusive- multiplicity for the showered sample and QCD dijets, comparing the ordering prescription of Ref. [7] at with fixed-frame Hungarian assignment in the same lightcone parametrization. Points are means over the training–test splits indicated in the legend, and bars are standard errors.
The comparison of all five pipelines in Figure 6 shows a further gain from the moving-frame representations. All branches improve rapidly as the event is first resolved into more blocks. The pivoted frame has the highest mean at , and the democratic frame leads at every scanned multiplicity from onward. Much of its discrimination is already present at modest resolution: its mean AUC is at . After a broad intermediate-resolution region, it improves further to at . This comparison uses ten common outer splits across all five branches.7
Figure 6. Mean test AUC versus exclusive- multiplicity for the showered sample and QCD dijets, comparing the five branches defined in § 4.3.1. All branches use ten outer splits. Points are outer-split means and bars are standard errors.
The high-resolution ordering distinguishes all five representations. The pivoted frame remains second throughout the scan from onward, and the radial spherical branch also benefits from finer resolution. From onward, all three massive spherical branches exceed both the convex and massless branches in mean AUC. The convex energy–momentum comparison saturates near its intermediate-resolution maximum, while the massless branch settles below its earlier peak. At , the democratic advantage over the pivoted frame is AUC percentage points, and the gaps to the convex and massless branches are approximately four points. The progression from the assignment comparison to the complete-pipeline comparison therefore identifies two useful improvements: greater stability with increasing multiplicity in the updated massless prescription, and stronger discrimination from the mass-aware spherical representations at finer resolution.
Fully hadronic four-top versus merged +jets
We next consider a substantially more challenging event-level benchmark: fully hadronic four-top production against a merged top-pair-plus-jets background. The selected samples and common particle-level processing are summarized in Table 2.
The +jets sample is a challenging background with substantial signal-like structure: it already contains real tops, heavy flavor, hadronic decays, and hard radiation. The benchmark therefore measures full-pipeline discrimination in this more closely matched process-level setting; dedicated scalar baselines would quantify the remaining coarse-observable contribution separately. In the merged background, up to two additional hard emissions are described at matrix-element level, with MLM vetoes preventing double counting with the shower. Since fully hadronic four-top production has an intrinsic twelve-parton core before showering, it motivates testing finer exclusive- resolutions over the wider range .
The main feature of Figure 7 is the separation between the two moving frames and the other three branches as the resolution increases. At the coarsest resolutions, the radial branch leads at and the convex branch at –. The democratic frame takes the lead at , followed by the pivoted frame in second place from onward. Both moving frames reach broad plateaus around – and sustain their discrimination through . The democratic branch attains its largest scanned mean, , at and ends at . Its advantage is therefore already present at intermediate resolution and persists across the remainder of the scan. The notable dip and error bar in the pivot branch are due to two of the five outer splits yielding markedly lower AUCs than the others. This behavior is consistent with increased sensitivity to the choice of pivot near the twelve-parton decay multiplicity, where several resolved hard prongs may have comparable transverse momenta.
Figure 7. Mean test AUC versus exclusive- multiplicity for fully hadronic versus MLM-merged . The curves compare the convex energy–momentum, massless lightcone–Hungarian, radial spherical, democratic moving-frame, and pivoted moving-frame representations. Points are means over five repeated 70:30 outer training–test splits shared by all branches, and bars are their standard errors.
The radial, convex, and massless branches instead lose performance beyond their intermediate-resolution maxima. The radial branch remains the strongest of these three at high multiplicity. The convex comparison, despite its early lead, shows the largest subsequent decline and falls below the massless branch from onward. By , the democratic frame exceeds the convex and massless branches by and AUC percentage points, respectively, while its advantage over the pivoted frame is points. This benchmark makes the resolution dependence particularly clear: the moving frames preserve their classification performance as the event is resolved at multiplicities well above the hard decay multiplicity, whereas the other representations favor a coarser comparison.
This benchmark also reaches the highest multiplicities considered here. At these resolutions, reduced fidelity of the sequential assignment–alignment approximation could contribute to the decline observed in some branches. Comparing both distance fidelity and classification performance with more accurate joint minimization would help determine the contribution of the approximation to this resolution dependence.
The comparison with the radial branch also connects directly to the motivation in § 3.2. Dividing each object’s momentum and mass by the square root of its energy gives subleading kinematic structure greater relative influence than radial normalization.
Their sustained performance at high multiplicity is consistent with that motivation. The radial and democratic maps preserve the same normalized kinematic information and the same spherical-sector topology, so their different AUC curves illustrate the importance of how this information enters the event comparison. Controlled variations of the weighting, mass treatment, and allowed assignments would quantify their individual contributions.
Conditioned BlackMax B1 versus QCD multijets
Massless phase-space-distance SVM classification has already been applied to BlackMax signals in the CMS black-hole and sphaleron analysis [18]. The conditioned BlackMax B1 sample defined in § 4.1.2 provides a BSM-versus-four-flavor-QCD stress test at substantially higher visible energy than the Standard Model benchmarks. The conditioning defines one event cohort, which is then held fixed and reclustered across . Before the conditioning, substantial scalar separation remains even within the common window. The additional restrictions reduce it to the residual structure quantified here. On the conditioned sample before balancing, the largest orientation-adjusted one-variable AUC is 0.7701 for the laboratory-frame block-energy sum , followed by 0.7586 for and 0.7471 for ; the summed mass of the blocks gives 0.5279. These direct one-variable scores provide scalar reference points for this comparison; a matched multivariate scalar model is a natural next comparison.
The moving frames lead throughout this scan, with the democratic branch first and the pivoted branch second at every multiplicity in Figure 8. Both continue to improve beyond the resolutions at which the convex and massless branches reach their largest means. The democratic curve forms a broad high-resolution plateau, with a maximum of at and an endpoint of at . The pivoted frame draws closer at the largest multiplicities; their paired endpoint difference is AUC percentage points. The two moving frames consequently provide comparable high-resolution performance on this sample, with a consistent lead in mean AUC for the democratic construction.
Figure 8. Mean test AUC versus exclusive- multiplicity for BlackMax B1 versus four-flavor QCD multijets. The curves compare the convex energy–momentum, massless lightcone–Hungarian, radial spherical, democratic moving-frame, and pivoted moving-frame representations. Points are means over the five common 70:30 outer splits, and bars are their standard errors.
The remaining branches show three distinct responses to finer resolution. The convex comparison settles into an early plateau, the massless comparison declines after its maximum, and the radial comparison gains again at high multiplicity to finish between the moving frames and the other two branches. At , the democratic frame exceeds the convex and massless branches by and AUC percentage points, respectively. Together with the Standard Model benchmarks, this shows a recurring advantage of the moving-frame pipelines across the selected samples, accompanied by process-dependent gains from finer resolution: continued improvement for , and broad plateaus for four-top and conditioned BlackMax.
Conclusions and future outlook
We have extended phase-space-based event comparison to variable reconstructed masses and scalable particle assignment for supervised binary classification. The known spherical topology of fixed-mass phase space supplies the starting point. An explicit radial map defines angular distances, and treating masses as coordinates allows events from different fixed-mass shells to be compared. The radial, convex, pivoted, and democratic representations retain the normalized four-vectors on their stated domains while assigning different event distances. In particular, the radial and democratic constructions give distinct homeomorphisms of the closed labeled collider space onto the same spherical sector, preserving its kinematic information and topology. The appendices relate these representations to spinor frames and the physical phase-space measure.
Eventwise normalization removes the common invariant-mass scale. The fixed-scale labeled metrics remain metrics on the normalized spaces; when applied to variable-scale events, these are pseudometrics. Exact simultaneous symmetry reduction gives metrics on the corresponding normalized event-orbit spaces, including for both moving frames, while the implementation uses scalable Hungarian assignment followed by conditional transverse alignment. In the matched massless comparison, the updated prescription improves the endpoint AUC over legacy ordering by percentage points at .
Across the complete pipelines, the democratic moving frame has the largest mean endpoint AUC in all three samples: for versus QCD at , for four-top versus merged + jets at , and for conditioned BlackMax B1 versus QCD at . The pivoted frame is second at each endpoint, with the BlackMax difference less well resolved by the split variation. The democratic construction provides a smooth, permutation-equivariant identification of momentum-closure subspaces, with its linear weight by energy fractions that may give subleading prongs greater influence. This interpretation alone does not explain the classification ranking.
More generally, the construction provides a direct route from pseudojet four-vectors to event classification. Given a labeled training sample, the normalization, representation, assignment, and classifier require no particle identities, hypothesized decay chain, or process-specific feature engineering. Exclusive is one way to supply these inputs; other jet definitions can use the same framework at a common supported multiplicity. The four-vectors are the only kinematic inputs to the classifier, allowing the prescription to be transferred between event definitions and signal hypotheses.
The method does not require very high values of to be effective, and zero padding was not studied. Finer resolution exposes additional event structure, but its classification benefit depends on the sample and representation. The democratic branch already reaches a mean AUC of 0.9397 at and improves further through ; four-top and conditioned BlackMax instead exhibit broad plateaus near and –30, respectively. The useful multiplicity should therefore be chosen within the range supported by the sample, balancing validation performance and computational cost.
A central next question is which geometric differences between signal and background populations produce the discrimination. Relating their separation to energy sharing, angular structure, and reconstructed masses could give the classifier a more direct physical interpretation. Cross-process AUC differences also reflect the selections: the BlackMax benchmark deliberately suppresses several otherwise powerful global discriminants, so its AUC cannot be compared with the value as a measure of intrinsic process similarity.
A more speculative direction is to tailor the representation to known features of particular processes, analogous to the use of physically motivated reference geometries in optimal-transport observables [43, 44]. The labeled collider space is homeomorphic to a closed -dimensional ball, leaving considerable freedom in the choice of representation and comparison metric. The spherical-sector and convex spectrahedral representations illustrate how the same kinematic information can support different notions of event similarity. Even with a fixed metric on the target space, different representation maps can induce different event distances and need not yield the same classification performance. Knowledge of kinematic constraints or of how signal and background populate phase space could guide further representations that emphasize features relevant for discrimination. Such constructions could incorporate process information as optional guidance for the geometry, extending the framework beyond the choices studied here, which require neither particle identities nor assumed decay chains.
Improved minimization is a complementary direction. Alternating permutation assignment with transverse alignment, possibly from several initializations, could improve the joint minimum approximation while retaining scalable individual steps. Repeating these steps for every event pair nevertheless increases the computational cost. Controlled mass ablations remain useful for separating mass information from frame and assignment effects; here finite resources were devoted to comparing substantially different representations and assignment prescriptions. Multiclass SVMs and comparisons involving several signal and background processes would extend the same event distances beyond the binary benchmarks studied here.
Acknowledgments
We thank Liam Brennan and Danyi Zhang for discussions of high-multiplicity collider samples; Tamas Almos Vami for these discussions as well as a careful reading of the manuscript and helpful comments; and Yu-Tse (Alan) Lee for discussions of geometry. We used computational facilities purchased with funds from the National Science Foundation (CNS-1725797) and administered by the Center for Scientific Computing. The Center for Scientific Computing is supported by the California NanoSystems Institute and the Materials Research Science and Engineering Center (NSF DMR 2308708) at the University of California, Santa Barbara. The authors used OpenAI’s ChatGPT and Codex to assist with proofreading the draft for correctness, consistency, and typographical errors, refining prose, making minor equation rearrangements, generating and editing LaTeX code for diagrams, tables, and figure layouts, and preparing job-submission scripts and running existing analysis code on the computing cluster. The authors take full responsibility for the final manuscript and its results.
Numerical results and comparison details
Endpoint AUCs and split counts
Table 4 gives the mean test AUC and outer-split standard error at the endpoint of each complete-pipeline scan.
| Branch | , | Four-top, | BlackMax, |
| Pivoted moving frame | |||
| Democratic moving frame | |||
| Radial spherical | |||
| Convex energy–momentum | |||
| Massless lightcone–Hungarian |
Table 4. Mean test AUC and outer-split standard error at the endpoint of the three classification scans.
Table 5 gives the corresponding assignment-control endpoints. At matched , the Hungarian branch improves over legacy ordering by across their 12 common outer splits.
| Assignment branch | Number of outer splits | Test AUC at |
| Legacy lightcone-ordering, | 10 | |
| Legacy lightcone-ordering, | 12 | |
| Legacy lightcone-ordering, | 12 | |
| Massless lightcone–Hungarian, | 12 |
Table 5. Mean test AUC and outer-split standard error at the assignment-scan endpoint.
| Democratic minus | , | Four-top, | BlackMax, |
| Pivoted | |||
| Convex energy–momentum | |||
| Massless lightcone–Hungarian |
Table 6. Paired mean endpoint AUC differences and standard errors across common outer splits.
Residual scalar separation in BlackMax
Table 7 reports direct rank-based one-variable AUCs on the full conditioned sample before class balancing, with score orientation chosen separately for each variable to give an AUC of at least . Before the , conditioning, but after the common window, acceptance, and multiplicity requirements of § 4.1.2, the values for , , and are 0.9641, 0.9591, and 0.9517, respectively. The conditioning therefore substantially reduces their separation. The laboratory-frame energy effective multiplicity is
| Single variable | Direct AUC after conditioning |
| at | 0.7701 |
| 0.7586 | |
| 0.7471 | |
| Energy fraction carried by the five leading blocks | 0.7090 |
| Laboratory-frame | 0.7062 |
| Particle-level | 0.5429 |
| Sum of the block masses | 0.5279 |
Table 7. Residual one-variable separation in the conditioned BlackMax B1–QCD sample. These values use each scalar directly as a score and provide one-variable reference points.
evaluated on the blocks before the center-of-momentum boost.
Assignment performance
We compare all four § 3 branches on the same 2,500 pairs of normalized events at each . The reference enumerates every allowed permutation, minimizing over both sectors of separately for each. There are candidates for the pivoted branch and otherwise. Figure 9 shows the distance ratio ; Figure 10 shows , where is the selected permutation’s rank after alignment, with rank one best. At , the median distance ratios range from 1.12 to 1.22 across the four branches, with appreciable upper tails. The median rank percentile decreases with multiplicity in every branch and lies within the best 1% of the allowed permutations at . The democratic and radial branches have similar median ratios despite different permutation-rank percentiles. These diagnostics provide complementary measures of assignment quality: the percentile ranks describe relative position within the factorially growing candidate set, while the distance ratios quantify the cost of the two-stage approximation.
Figure 9. Empirical cumulative distributions of the distance ratio for the four mass-aware branches. Colors indicate the exclusive multiplicity .
Figure 10. Permutation-rank percentiles for the same comparisons. Points show medians, bands the 5th–95th percentile range, and dotted curves the rank-one floor .
Exact Hungarian assignment ties are resolved according to solver ordering. On tie loci, event-exchange symmetry, , requires a tie rule compatible with exchanging the events.
Computational performance
For events with pseudojets each, a dense distance matrix requires pair comparisons and matrix storage. Hungarian assignment gives a worst-case operation count of . The implementation prepares each event’s features once and distributes the pair comparisons across CPU workers. Benchmarks of the democratic spherical branch approach the expected quadratic dependence on event count: at , the increases from 4096 to 8192 and then 16384 events multiply the runtime by 3.89 and 3.95, respectively.
These benchmarks used 40 physical CPU cores on a node with two 24-core Intel Xeon Platinum 8268 processors at 2.90 GHz8, one numerical-library thread per worker, and single-precision inputs and matrix storage. A complete 16384 × 16384 matrix at took 7.97 min. At fixed in the four-top sample, the times were 2.65, 4.70, and 13.76 s for , 30, and 60, respectively. This measured multiplicity dependence is substantially milder than the worst-case bound over the tested range. These component times are medians of three fresh computations from preprocessed events in memory, including feature preparation and matrix construction but excluding worker startup, persistence, and classification.
On the same hardware configuration, a scan of 8000 events constructed and stored all 16 matrices at in 13.54 min, including distance diagnostics. The full analysis from preprocessed events, including classification for one outer split and kernel eigenvalue diagnostics, took 44.25 min (with optional diagnostics dominating the time). These wall times demonstrate that classification studies over a broad multiplicity range are practical on a single multicore compute node.
The LIPS measure
This appendix expresses the Lorentz-invariant phase-space measure in the radial coordinates of § 2.1 and explains the inclusion of zero-momentum configurations. Throughout, , , and .
Soft particles
For total momentum , define and use
Integrating over each leaves the on-shell factor . For a massless particle, , so
The apparent Cartesian singularity at zero momentum is therefore locally integrable.
The same conclusion holds after imposing total four-momentum conservation, including simultaneous soft limits. Let be the set of all zero-energy legs at a given configuration, and its complement. Writing and , the measure factorizes as
To obtain this, we repeatedly used the two-body recursive formula [31] and integrated out the intermediate coordinates.9 At , the remaining subsystem is above threshold, and at least two of its particle velocities differ: equal velocities would force all spatial momenta to vanish in the COM frame, contradicting . The four conservation constraints therefore have full rank in the remaining momenta, so the remaining measure has a continuous, strictly positive density in suitable local coordinates for small .
Each soft factor is proportional to , whose integral over is positive and of order . Their product is locally integrable, assigns zero measure to the soft strata, and gives positive measure to every neighborhood of each such configuration. These configurations therefore belong to the support of LIPS. Retaining them in the closed phase space of Eq. (1) leaves ordinary LIPS integrals unchanged. Amplitudes or measurement functions may still diverge on approach; such divergences concern the integrand rather than a delta-atom at zero momentum.
The measure in radial coordinates
An orthogonal Jacobi transformation separates the COM coordinate from the relative momentum coordinates, with unit Jacobian. Since
integrating the momentum-conservation delta function leaves times Euclidean volume on the momentum-conserving subspace. Writing with , as in § 2.1, this volume is , where is the round area measure on the unit sphere.
With
Eq. (B.1) becomes
The unique positive solution of is the radial function defined in Eq. (3). On the soft-free locus, where all ,
Integrating the remaining delta function therefore gives
where and . This expresses LIPS as a generally nonuniform weight times the round sphere measure. The formula holds almost everywhere; at zero-energy configurations, the measure is understood through the locally integrable extension established above.
For , and , so Eq. (B.5) reduces to the standard two-body result
where is the Källén function [31].
Topology, coordinate relations, and exact metrics
This appendix supplies the arguments deferred in § 2 and § 3. For the fixed-mass radial map, the main text explains why each momentum direction determines exactly one event; we complete that argument by checking continuity. While continuity is obvious from a physics perspective, for completeness we include a proof. In principle, a single radial solution could jump between momentum configurations. For the variable-mass space, we make the radial rescaling explicit and verify that imposing nonnegative masses gives a closed ball.
The main reconstruction problem arises in the democratic representation. § 3.2 shows that its features lie in the stated spherical sector. Here we state that every point of that sector represents exactly one labeled event and that reconstruction is continuous, including at zero energy. These results support the claim that radial and democratic coordinates retain the same fixed-scale event information while defining different distances. We then treat the restricted pivoted representation and prove the metric statements for exact symmetry reduction.
Fixed-mass radial topology
Fix , masses , and center-of-mass energy . To slow down the argument in the main text, let’s define the momentum conserving subspace explicitly
and thus phase space is
It is empty for and is a single point for , so we assume . Equip with the inherited Euclidean inner product , so we can define the unit sphere through its norm
The radius is defined by solving the energy constraint . Having exactly one solution in each direction does not, by itself, show that nearby directions give nearby radii. Thus, we need to show continuity. The proof below establishes this: the radii remain bounded, and any limiting radius must solve the limiting energy equation, whose solution is unique.
Theorem C.1 (Sphere topology of fixed-mass phase space). Fix , , and . The closed shell in Eq. (C.2) is homeomorphic to . The mutually inverse maps are Eq. (4), where is the unique positive solution of
Proof. For fixed , is continuous and strictly increasing, , and . Thus every ray meets the shell exactly once. This unique radial intersection makes bijective. To check continuity, let . Since and , every subsequence of the radii has a convergent subsubsequence. Joint continuity of and uniqueness of its root force every such limit to be , so . Hence is continuous. Its source is compact and its target is Hausdorff,10 so the compact-to-Hausdorff closed-map lemma makes a homeomorphism [45].11 □
The homeomorphism gives two immediate consequences. The first recovers the spherical topology of massless phase space, discussed in § 2.4, with an explicit expression for the radial function. The second verifies that the angular distance introduced in Eq. (6) defines a metric compatible with the original phase-space topology.
Corollary C.2 (Massless shell). For and , the massless shell has topology , with radius given by
Corollary C.3 (Fixed-mass radial metric). For , and , the distance in Eq. (6) is a metric inducing the subspace topology on .
Proof. The map is a homeomorphism onto by Theorem C.1. The stated distance is its pullback of the round angular metric. □
This fixed-mass metric normalizes only the spatial momenta. For nonzero masses it need not equal the restriction of the mass–momentum radial metric in Eq. (18).
Adjoining the fixed-mass shells
Next, we supply the complete proofs for the collider topologies in the body. The proof of the sphere space is already substantially in the text: we are fixing the norm of a set and thus we end up with a sphere. The collider dome proof is omitted from the main text as this is physically quite simple: we start by restricting a single mass coordinate to , meaning we are cutting a sphere in two hemispheres, and discarding one of them. The remaining restrictions then select a convex portion of this ball with nonempty interior, which is itself topologically a closed ball. The proof below makes this argument precise and verifies that our low-dimensional projection intuition carries over.
To distinguish a variable signed coordinate from the fixed mass parameters of , write for the mass coordinate in this appendix. On the physical sector, is the coordinate denoted in the main text; a fixed-mass slice imposes . Define
with
Here is the Euclidean norm denoted in Eq. (7), whereas is the sum of the Euclidean block norms and is itself a norm on . Write . In this notation,
Note that this enlarged space includes the zero-momentum configuration , but indeed the unphysical sector is automatically excluded. In general, the masses and the momenta cannot be chosen independently in this space which motivates a careful proof of the topologies.
Theorem C.4 (Topology of the collider dome and signed space.). For and , the signed shell is homeomorphic to and its nonnegative-mass sector is homeomorphic to the closed ball .
Proof. For , with , set
Then is a homeomorphism from to , with inverse . This proves the sphere statement.
Radial normalization identifies the physical sector with
Project this spherical sector onto . Its image is
The projection is a homeomorphism, since the omitted coordinate is recovered continuously as
The image is a compact convex body with nonempty interior in the -dimensional space , and is therefore homeomorphic to .
The spherical orthant’s massless faces and zero-block loci correspond to the lower-rank strata in the spinor description of Appendix § 11.§ 2.
The democratic representation on the closed collider space
The democratic representation in § 3.2 was introduced to change the relative weighting of energetic and subleading objects while retaining the event’s kinematic information. The equations there show that every physical event maps continuously into a fixed spherical sector. What remains is to show that this sector is filled completely, that no two different labeled events have the same features, and that reconstruction from the features is continuous. These results allow us to take the metric on the unit sphere and map it back to the collider space and invert it, defining a distance directly on the collider dome.
Fix and . Write an event as , with . The signed and physical shells are
Both include zero-energy blocks. Such a block necessarily has and .
Set and define
Let be the matrix with rows . These variables obey
In matrix notation, the democratic Householder reflection and feature map are
Writing the three transformed spatial columns as a matrix is only a reordering of the coordinates in the democratic feature vector. Define its reference space and physical spherical sector by
Here is the normalized mass vector defined in Eq. (C.8), and denotes the Frobenius norm, which is the root of the sum of the squares of the entries of a matrix
The formal argument for the democratic homeomorphism is very technical and uninteresting. The central challenge with surjectivity is the recovery of the energy fraction square roots when starting from the targets. Concretely, given an or , to recover , one needs the Householder involution which requires the energy fractions to be built. With it, and are sufficient to recover the inverse coordinates. To show it, one can use Brouwer’s fixed point theorem [45] on a simplex of trial energy fractions in . Injectivity can be shown by contradiction: assume that two distinct events produce the same features. Their mass-shell conditions and momentum conservation yield a positivity identity that forces their energy fractions to agree. The inverse reflection then gives identical momenta and masses, so the two events must coincide.12 Injectivity in particular is crucial to take our metric on the unit sphere and define a metric on our collider dome.
For the pivoted representation, the same reconstruction argument applies after restoring the omitted zero spatial row and replacing the democratic reference by . Injectivity follows from the same positivity argument, with the assumption keeping the relevant inequalities strict. The pivot variant can only be shown to be injective in the restricted pivot domain which is sufficient for metricity. We will thus use the following statements without proof.
At fixed and , the democratic representation defines homeomorphisms
Thus every point of the closed physical spherical sector represents exactly one labeled event, and both the feature map and its inverse are continuous, including at zero energy.
For the pivoted representation, let consist of positive-energy events whose unique highest- object occupies slot . The map
is injective and is the restriction of a smooth map on the positive-energy domain. It is equivariant under : permuting the non-pivot objects or applying a common transverse rotation or reflection produces the corresponding transformation of the features.
These statements allow us to establish metricity of the associated distances. The event comparisons in the main text also account for the allowed particle permutations and transverse rotations and reflections. We next show that exact minimization over these symmetries gives a metric on the corresponding orbit space.
Metrics under exact symmetry reduction
The importance of this section is to show that the symmetry minimizations do not destroy the metric properties. We first show a general statement covering any compact group and apply it to our particular metrics.
For any map , the pullback is a pseudometric, and is a metric precisely when is injective. If is a homeomorphism onto its image, this metric also induces the original topology of . These elementary observations and the following lemma give the metric statements for all four mass-aware representations.
Let denote the orbit of with respect to our symmetry group. This is the set of events obtained from by the allowed relabelings and transverse transformations.
Lemma C.5 (Exact compact-group quotient). Let be a metric space and let a compact group act continuously on by isometries. Then
is a metric inducing the quotient topology on .
Proof. Compactness and continuity imply that the minimum is attained. Isometric invariance makes the definition independent of the representatives, and replacing by gives the symmetry axiom. If , attainment gives for some , so . If and are minimizers, then
which proves the triangle inequality. For the orbit projection ,
Thus is continuous and open for the metric topology, which is exactly the quotient topology. No freeness of the action is required.
Lastly, we show that after exact assignment and alignment, all our metrics remain metrics. We are simply applying the lemma to our particular metrics here to finalize the formalization.
Corollary C.6 (Representation metrics and exact symmetry reduction). At fixed and , the radial, democratic, and convex distances in Eqs. (18), (30), and (19)
are metrics inducing the original topology of the closed physical collider space. The convex statement uses the identification . Let . For each of these distances,
is a metric inducing the corresponding quotient topology. The democratic map also induces a homeomorphism
The pivoted angular distance is a metric on and on its feature image. Its exact symmetry reduction is a metric on both orbit spaces for , with the distinguished slot held fixed. All statements also hold with the transverse factor omitted.
Proof. The radial and democratic maps are homeomorphisms onto the same closed spherical sector. Pulling back the round angular metric therefore gives metrics with the stated topology. The convex map and its inverse are continuous, and the Euclidean metric on induces its subspace topology. The pivoted metric follows from the statements in Section C.3.
Radial normalization and the convex map commute with relabeling and transverse orthogonal transformations. For the democratic reference, for every permutation matrix, so . Permutations act on the rows of and the components of , whereas transverse transformations act on the spatial columns of and preserve the energies. The zero-block convention respects both actions. The democratic map is therefore equivariant, as is its inverse.
For the pivoted representation, every allowed permutation fixes slot , so and hence
Together with the same transverse equivariance, this establishes equivariance of the pivoted map under .
The target actions are orthogonal and hence isometric. The quotient metrics follow from Lemma C.5. Equivariance of the democratic homeomorphism and its inverse gives mutually inverse continuous maps on the orbit spaces, proving Eq. (C.18).
Remark C.7 (Intrinsic distance in convex coordinates). Every finite-length curve joining has length at least their Euclidean distance, while convexity ensures that the straight segment attains this bound. It is the unique constant-speed minimizing curve. The same argument applies to the convex positive-scale cone .
These results establish metricity of the exact joint minimum in Eq. (37). They do not establish metricity of the sequential procedure in § 4.3, whose approximation performance is assessed in Appendix § 8.§ 3.
Spinor frames and the collider spectrahedron
This appendix relates spinor-helicity descriptions to the event spaces introduced in the main text. Its central result identifies the collider spectrahedron, written in matrix form, with a quotient of normalized Gram factors. We also specify the open domain covered by the earlier classifier’s simplex–sphere parametrization and explain its relation to the closed massless phase space.
Massless spinor frames and lightcone coordinates
The spinor construction relates the original massless representation to the variable-mass event space. Write , use signature , and associate to each momentum the Hermitian matrix13
Here denotes the Pauli matrices and . For a nonzero future-directed massless momentum, this matrix is positive semidefinite of rank one and factors as
Here is a column spinor and denotes componentwise complex conjugation on the real momentum slice. The particlewise little-group action is . Zero momentum is represented by .
Proposition D.1 (Massless spinor-frame space). For massless momenta with nonnegative energies and total momentum , where , the normalized spinor frames form
The physical momentum shell, including zero-momentum blocks, is the topological orbit space
Proof. Collect the spinors into . Momentum conservation is . Hence has two orthonormal columns and lies in ; conversely, every such frame defines a conserved massless configuration. The momentum map is unchanged by independent row phases. Two frames give the same positive-semidefinite matrices of rank at most one precisely when they differ by those phases, with stabilizers allowed at zero rows. The induced map from the compact orbit space to the Hausdorff momentum shell is therefore a homeomorphism.
This is the known Stiefel description [5, 6]. The Stiefel manifold has real dimension ; removing the phase directions gives the expected momentum-space dimension on the soft-free locus. For , frames with zero rows make the torus action nonfree, so the quotient map does not define a principal bundle globally.
For massless momenta, define and, on the open locus for every ,
They satisfy
Writing and , these constraints give and
As varies, these orthogonal complements form a complex rank- vector bundle over the open simplex. The base is contractible, so the bundle and its unit-sphere bundle are trivial. The simplex–sphere variables of Refs. [5, 7] therefore give a full-measure product parametrization on this locus, of the form after the little-group phases are removed. At the compact boundary the chosen lightcone section degenerates and part of the angular fiber is identified or collapses, as is also implicit in the compact spherical description of Ref. [27].
Example D.2 (Two-particle lightcone parametrization). For two massless particles, the regular lightcone parametrization gives . At either endpoint the transverse radius vanishes and the angular circle collapses to a single beam-collinear pole. Thus
Gram factors and the collider spectrahedron
For a future-timelike momentum, the bispinor is positive definite of rank two and factors as
The determinant records the mass,
The same factorization includes massless and zero-momentum blocks through rank-deficient .
In the center-of-momentum frame, define the normalized bispinors
Their eigenvalues are , so the physical energy inequalities and momentum conservation become
Consequently,
After eliminating the last block, the right-hand side is the single linear matrix inequality
Thus is a bounded spectrahedron, identified with the convex set of labeled -outcome qubit POVMs [46].
To obtain the corresponding spinor frame, set , so , and stack the adjoints:
The normalization in Eq. (D.11) is exactly , so .
Blockwise transformations , with , give all and only alternative Gram factorizations of the same tuple, including rank-deficient blocks. Indeed, if , polar decomposition and extension of a partial isometry to a unitary in equal finite dimensions give and with . Hence . Conversely, right multiplication by a unitary leaves unchanged.14 The induced continuous bijection from the compact orbit space to the Hausdorff spectrahedron is therefore a homeomorphism:
If denotes a row block, then induces . Over the all-massive relative interior the action is free, giving
At a rank-one block the stabilizer is a , and at a zero block it is , giving the corresponding orbit-type stratification.
A fixed-mass shell is recovered by imposing
Thus
Its topology is established in Appendix § 10.§ 1. The massless restriction has blocks of rank at most one, with rank one on the soft-free locus, and reduces to the description above. Allowing the determinants to vary gives the collider spectrahedron , which is homeomorphic to the collider dome through the coordinate map in Eq. (13).
References
- [1]M.E. Peskin and D.V. Schroeder, An Introduction to quantum field theory, Addison-Wesley, Reading, USA (1995), 10.1201/9780429503559.
- [2]M. Srednicki, Quantum Field Theory, Cambridge University Press (2007), 10.1017/CBO9780511813917.
- [3]M.D. Schwartz, Quantum Field Theory and the Standard Model, Cambridge University Press (3, 2014), 10.1017/9781139540940.
- [4]Particle Data Group collaboration, Review of Particle Physics, Int. J. Mod. Phys. A 41 (2026) 2630011.
- [5]A.J. Larkoski and T. Melia, Covariantizing phase space, Phys. Rev. D 102 (2020) 094014.DOI
- [6]B. Henning and T. Melia, Conformal-helicity duality & the Hilbert space of free CFTs, 1902.06747.
- [7]T. Cai, J. Cheng, N. Craig, G. Koszegı and A.J. Larkoski, The phase space distance between collider events, JHEP 09 (2024) 054 [2405.16698].arxiv.org/abs/2405.16698
- [8]R. Kleiss, W.J. Stirling and S.D. Ellis, A new monte carlo treatment of multiparticle phase space at high energies, Computer Physics Communications 40 (1986) 359.
- [9]J. Alwall, R. Frederix, S. Frixione, V. Hirschi, F. Maltoni, O. Mattelaer et al., The automated computation of tree-level and next-to-leading order differential cross sections, and their matching to parton shower simulations, JHEP 07 (2014) 079.DOI
- [10]B. Henning and T. Melia, Constructing effective field theories via their harmonics, Phys. Rev. D 100 (2019) 016015 [1902.06754].DOI
- [11]J. Shu, M.-L. Xiao and Y.-H. Zheng, Constructing the general partial wave and renormalization in effective field theory, Phys. Rev. D 107 (2023) 095040 [2111.08019].DOI
- [12]P.P. Srivastava and G. Sudarshan, Multiple production of pions in nuclear collisions, Phys. Rev. 110 (1958) 765.DOI
- [13]Z. Bogorad, I. Elsharkawy, Y. Kahn, A.J. Larkoski and N. Levi, Generative models on phase space, 2026. 10.48550/arXiv.2604.02415.arxiv.org/abs/2604.02415
- [14]V.S. Ngairangbam, B. Rozwoda, K. Sakurai and M. Spannowsky, Enhancing anomaly detection with topology-aware autoencoders, Mach. Learn. Sci. Tech. 6 (2025) 045051 [2502.10163].arxiv.org/abs/2502.10163
- [15]L. Brennan, T.A. Vami, O. Amram, S. Sekhar, Y. Takahashi, L. Moureaux et al., Weekly supervised anomaly detection with event-level variables, Phys. Rev. D 112 (2025) 055040 [2504.13249].DOI
- [16]T. Cai and H. Li, A unified geometric framework for understanding collider event manifolds, 2610.09320.arxiv.org/abs/2610.09320
- [17]CMS collaboration, Search for black holes and sphalerons using novel machine learning techniques at CMS, PHEP 2026 (2026) 21 [2511.10662].arxiv.org/abs/2511.10662
- [18]CMS Collaboration, A search for microscopic black holes, string balls, and sphalerons in proton-proton collisions at √s = 13 TeV, JHEP 08 (2026) 098 [2604.10732].
- [19]P.T. Komiske, E.M. Metodiev and J. Thaler, The metric space of collider events, Phys. Rev. Lett. 123 (2019) 041801 [1902.02346].DOI
- [20]T. Cai, J. Cheng, N. Craig and K. Craig, Linearized optimal transport for collider events, Phys. Rev. D 102 (2020) 116019 [2008.08604].DOI
- [21]T. Cai, J. Cheng, K. Craig and N. Craig, Which metric on the space of collider events?, Phys. Rev. D 105 (2022) 076003 [2111.03670].DOI
- [22]A.J. Larkoski and J. Thaler, A spectral metric for collider geometry, JHEP 08 (2023) 107 [2305.03751].arxiv.org/abs/2305.03751
- [23]T. Cai, N. Craig, K. Craig and X. Lin, Multiscale optimal transport for complete collider events, Phys. Rev. D 112 (2025) 036021 [2501.10681].arxiv.org/abs/2501.10681
- [24]D. Guest, K. Cranmer and D. Whiteson, Deep learning and its application to LHC physics, Ann. Rev. Nucl. Part. Sci. 68 (2018) 161 [1806.11484].DOI
- [25]A.J. Larkoski, I. Moult and B. Nachman, Jet substructure at the large hadron collider: A review of recent advances in theory and machine learning, Phys. Rept. 841 (2020) 1 [1709.04464].DOI
- [26]S.N. Sokolov, On the parametrization of the phase space of N relativistic particles, Theoretical and Mathematical Physics 7 (1971) 485.DOI
- [27]J. Batson, C.G. Haaf, Y. Kahn and D.A. Roberts, Topological obstructions to autoencoding, JHEP 04 (2021) 280 [2102.08380].DOI
- [28]CMS Collaboration, Evidence for four-top quark production in proton-proton collisions at √s = 13 TeV, Phys. Lett. B 844 (2023) 138076 [2303.03864].DOI
- [29]ATLAS Collaboration, A search for R-parity-violating supersymmetry in final states containing many jets in pp collisions at √s = 13 TeV with the ATLAS detector, JHEP 05 (2024) 003 [2401.16333].DOI
- [30]H.W. Kuhn, The hungarian method for the assignment problem, Naval Research Logistics Quarterly 2 (1955) 83.
- [31]Particle Data Group, “Review of particle physics: Kinematics.” Particle Data Group review, 2025.
- [32]J.R. Andersen and A. Maier, Unbiased elimination of negative weights in Monte Carlo samples, Eur. Phys. J. C 82 (2022) 433 [2109.07851].DOI
- [33]M. Cacciari, G.P. Salam and G. Soyez, FastJet user manual: (for version 3.0.2), The European Physical Journal C 72 (2012) 1896.
- [34]C. Bierlich, S. Chakraborty, N. Desai, L. Gellersen, I. Helenius, P. Ilten et al., A comprehensive guide to the physics and usage of PYTHIA 8.3, SciPost Phys. Codeb. 2022 (2022) 8 [2203.111601].arxiv.org/abs/2203.11601
- [35]P. Artoisenet, R. Frederix, O. Mattelaer and R. Rietkerk, Automatic spin-entangled decays of heavy resonances in monte carlo simulations, JHEP 03 (2013) 015 [1212.3460].DOI
- [36]N. Arkani-Hamed, S. Dimopoulos and G. Dvali, The hierarchy problem and new dimensions at a millimeter, Phys. Lett. B 429 (1998) 263 [hep-ph/9803315].DOI
- [37]S. Dimopoulos and G. Landsberg, Black holes at the LHC, Phys. Rev. Lett. 87 (2001) 161602 [hep-ph/0106295].DOI
- [38]D.-C. Dai, G. Starkman, D. Stojkovic, C. Issever, E. Rizvi and J. Tseng, BlackMax: A black-hole event generator with rotation, recoil, split branes, and brane tension, Physical Review D 77 (2008) 076007.DOI
- [39]CMS collaboration, Extraction and validation of a new set of CMS PYTHIA8 tunes from underlying-event measurements, Eur. Phys. J. C 80 (2020) 4 [1903.12179].DOI
- [40]C. Cortes and V. Vapnik, Support-vector networks, Machine Learning 20 (1995) 273.
- [41]C.-C. Chang and C.-J. Lin, LIBSVM: A library for support vector machines, ACM Transactions on Intelligent Systems and Technology 2 (2011) 27:1.
- [42]A. Feragen, F. Lauze and S. Hauberg, Geodesic exponential kernels: When curvature and linearity conflict, in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 3032–3042, 2015, DOI.arxiv.org/abs/1411.0296
- [43]D. Ba, A.S. Dogra, R. Gambhir, A. Tasissa and J. Thaler, SHAPER: Can you hear the shape of a jet?, JHEP 06 (2023) 195 [2302.12266].DOI
- [44]C. Cesarotti and M. LeBlanc, A field guide to event-shape observables using optimal transport, JHEP 12 (2025) 014 [2409.13150].DOI
- [45]J.M. Lee, Introduction to Topological Manifolds, vol. 202 of Graduate Texts in Mathematics, Springer, New York, 2 ed. (2011), 10.1007/978-1-4419-7940-7.
- [46]J. Watrous, The Theory of Quantum Information, Cambridge University Press (2018), 10.1017/978131なる6848142.DOI
- [47]N. Arkani-Hamed, T.-C. Huang and Y.-t. Huang, Scattering amplitudes for all masses and spins, JHEP 11 (2021) 070 [1709.04891].arxiv.org/abs/1709.04891
- [48]R. Kleiss and W. Stirling, Cross sections for the production of an arbitrary number of photons in electron-positron annihilation, Physics Letters B 179 (1986) 159.DOI








