TauPolaris: reconstructing tau lepton polarimetric vectors with conditional normalizing flows
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: High Energy Physics: Phenomenology. The day in particle phenomenology: models, signatures and constraints.
Kai: I'm Kai.
Mira: And I'm Mira, here with Lev.
Kai: We're looking at the paper: "TauPolaris: reconstructing tau lepton polarimetric vectors with conditional normalizing flows".
Mira: Here's the gist of the paper. TauPolaris presents a new method for reconstructing tau lepton polarimetric vectors by estimating undetected neutrino momenta with a conditional normalizing flow, which is crucial because this technique improves upon existing methods by modeling the full conditional density rather than performing point regression <ref:2608.10961#pg6>.
Kai: This approach allows for the estimation of both the most likely configuration for each event and an estimate of its uncertainty, while respecting correlations between neutrino momentum components <ref:2608.10961#pg6>.
Mira: The core of TauPolaris involves using a conditional normalizing flow to model the probability density of the neutrino kinematics given a context vector learned from the event <ref:2608.10961#pg6>.
Kai: This flow expresses this non-trivial density using a series of invertible transformations mapping neutrino momenta to a latent variable that follows a D–dimensional Gaussian distribution <ref:2608.10961#pg6>.
Title and authors: Mira: Moving onto the summary, it explains how they set up this conditional normalizing flow to model those neutrino kinematics given a context vector learned from the event itself.
Kai: They take inputs like the four-momenta of all visible tau decay products, some displacement variables sensitive to the tau lifetime, and that missing transverse momentum.
Lev: So, these are the ingredients they feed into their conditioning network before it hits the flow model? What kind of inputs are we talking about in terms of dimensionality?
Mira: For fully hadronic decays, they have up to fifty two variables as inputs, but for semileptonic decays, it's thirty five inputs.
Kai: That's a lot of information being fed into the conditioning network that sets up the model before the flow does its main work. It’s a complex input feature set.
Lev: The architecture itself is also quite deep; they use nine invertible conditional flow blocks, and each one starts with an LU-decomposed linear mixing layer followed by two rational quadratic spline coupling transformations.
Mira: That structure is what allows the model to express that non-trivial density using these series of invertible transformations that map the neutrino momenta to a latent variable following a D dimensional Gaussian distribution.
Kai: So, in essence, they’re using this flow structure to represent the probability distribution of those neutrino momenta in a very structured way.
Lev: And they train this whole system by minimizing the negative log-likelihood of the generator-level neutrino kinematics under that learned density. That's standard generative modeling training.
Mira: The goal is to get a model that accurately reflects how these neutrinos behave kinematically given the tau decay context, which then lets us reconstruct those spin vectors.
The paper's summary: Kai: Now for the specific improvements they claim, it’s really about what this method actually does better than existing tools. They focus on modeling the full conditional density instead of just doing point regression.
Mira: That full density modeling is what lets them estimate both the most likely configuration *and* an estimate of its uncertainty, which is a big deal for spin measurements.
Lev: And this capability to model those correlations between neutrino momentum components directly addresses a major issue where simpler methods fail, which is why it’s better than point regression.
Kai: They show that this allows them to reproduce the generator-level distributions closely across all variables, including things like the reconstructed tau mass.
Mira: That's significant because reconstructing the tau mass itself is hard; it requires accurately modeling those correlations between different components of the neutrino momenta.
Lev: So, they are showing that if you model these correlations correctly in the neutrino estimation step, you get better accuracy on observables that depend on those correlated inputs.
Kai: And in terms of performance, they tested it on three million events and found a significant improvement over methods trained with a mean-squared-error loss.
Mira: They report an improvement of approximately forty percent compared to the stand-alone transformer, and another twenty percent compared to a sampled approach for angular variables that contribute to the spin density matrix.
The paper's improvements: Kai: So, wrapping up on "TauPolaris: reconstructing tau lepton polarimetric vectors with conditional normalizing flows." The main implication is that this method significantly boosts the resolution of tau spin observables compared to previous methods.
Mira: It seems like the ability to model the full conditional density provides a way to get not just a prediction, but a full picture including uncertainty, which is vital for high-precision measurements.
Lev: From an engineering standpoint, that improvement in resolution—that forty percent gain—means that when we take this tool and run it on real hardware, the resulting spin data will be much cleaner than what we could get from simpler regression techniques.
Kai: And it opens up avenues for using these reconstructed vectors to test deep physics questions, like entanglement measurements at the LHC.
Mira: They showed a significance of at least four point three sigma for distinguishing quantum entanglement in H to tau tau decays, which is a strong result.
Lev: It also helps with background suppression; they constructed new variables from those polarimetric vectors that can separate spin zero Higgs boson events from spin one Z boson events, and these are independent of the CP state of the Higgs boson.
Kai: And they also improved sensitivity to CP violation in H to tau tau decays by about eighteen percent when using these reconstructed polarimetric vectors for phiCP.
Mira: Overall, it's a tool that provides better inputs for searching for new physics related to CP structure in those tau decays. It's a solid piece of work.
Conclusion: Lev: From an engineering standpoint, the overhead of running this flow model on real hardware is where the real challenge lies, and getting that fidelity up to what they claim will be needed for actual experiments is something we have to worry about.
Kai: I agree, but the performance numbers they're showing on three million events against those old regression methods are compelling enough to keep us interested in making it work.
Mira: The uncertainty estimation is key here; if we can reliably use that event-level uncertainty for dynamic selection, that could optimize our cuts a lot.
Lev: If the AI system can actually use that predicted uncertainty to guide event selection instead of just using fixed cuts, then we’re looking at a real optimization in signal yield.
Kai: It sounds like the next step is moving from just reconstructing the vectors to using those uncertainties to dynamically select the best data for analysis.
Daniel Winterbottom, *, Lucas Russell †
Imperial College London · CERN
hep-ph, hep-ex
Submitted: 2026-08-11
Updated: 2026-11-13
Code: https://github.com/danielwinterbottom/TauPolaris
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 90/100
The gist: TauPolaris presents a new method for reconstructing tau lepton polarimetric vectors by estimating undetected neutrino momenta with a conditional normalizing flow…
Key concepts
- Conditional Normalizing Flow
- This is a machine learning model used to map complex input data (neutrino momenta) to a simpler latent variable distribution. It allows the system to model the probability density of the neutrino kinematics given event context, rather than just predicting a single value, thus capturing the full spread of possible configurations.
- Full Conditional Density
- Instead of estimating just one best guess for neutrino momentum, this technique models the entire probability distribution. This is crucial because it gives AI systems an estimate of uncertainty alongside the most likely configuration, which is necessary for accurately reconstructing spin observables.
- Tau Polarimetric Vectors
- These are vectors that describe the polarization state of a tau lepton. TauPolaris reconstructs these vectors by inferring the momenta of neutrinos that were not directly detected in the decay process, using sophisticated modeling techniques.
- MAP Approach
- Maximum A Posteriori (MAP) is a method used to find the most probable set of parameters (in this case, neutrino momenta) given observed data. TauPolaris uses this approach combined with the normalizing flow to find the optimal reconstruction of tau polarimetric vectors.
Terminology
Summary
TauPolaris presents a new method for reconstructing tau lepton polarimetric vectors by estimating undetected neutrino momenta with a conditional normalizing flow, which is crucial because this technique improves upon existing methods by modeling the full conditional density rather than performing point regression <ref:2608. p. 0 We present TauPolaris, a tool for reconstructing tau polarimetric
vectors by estimating the undetected neutrino momenta with a conditional normalizing flow. Rather
than performing a point regression, the method models the full conditional density of the neutrino
kinematics, providing both the most likely configuration for each event and an estimate of its
uncertainty. The spread of the
sampled density also gives a per-event uncertainty estimate. Crucially, because the flow models the joint density of the neutrino momentum components, any configuration drawn from it respects the correlations between them.Where the paper says it
Where the paper says it
Methodology Overview
The core of TauPolaris involves using a conditional normalizing flow to model the probability density of the neutrino kinematics given a context vector learned from the event <ref:2608.10961#pg6>. This flow expresses this non-trivial density using a series of invertible transformations mapping neutrino momenta to a latent variable that follows a D–dimensional Gaussian distribution <ref:2608. p. 3 Conditional Flow Architecture
The conditional normalizing flow models the probability density of the neutrino kinematics p(ν | c), conditional on a context vector c, which is a learned summary
of the event. The flow expresses this non-trivial density
using a series of invertible transformations f, mapping
ν to a latent variable z = f(ν | c), which follows a D–
dimensional Gaussian distribution N (z; 0,I). The GELU activation function
is used within the feed-forward layers in the conditioning
network. The conditioning network and the flow are trained simultaneously, minimizing the negative log-likelihood of
the generator-level neutrino kinematics under the learned
density:
L = −
1
N
X
N
i=1
log p(νi
| ci), (9)
where νi denotes the true neutrino kinematics of event i,
ci the context vector of the corresponding reconstructed
event, and the density is evaluated via Eq.Where the paper says it
Where the paper says it
Input Features and Architecture
The models take as inputs the four-momenta of all visible tau decay products, displacement variables sensitive to the finite tau lifetime, and the missing transverse momentum, which are combined by a transformer-based conditioning network <ref:2608.10961#pg6>. The input features include up to 52 variables for the fully hadronic (τhτh) model and 35 inputs for the semileptonic (τlτh) model <ref:2608. p. 3 Finally, flags are used to
determine the type of decay, by specifying the number of
π
± and π
0
s in τh decays and the light lepton flavor in τℓ
decays. The full set of input features is summarized in
Table II; there are a total of 52 input variables for the
τhτh model and 35 inputs for the τℓτh model. The flow transformation consists of nine conditional
flow blocks, depicted in Fig. Each block begins with an
LU-decomposed linear mixing layer [27] (except for the
first block), followed by two successive rational quadratic
spline coupling transformations [28], acting on alternating halves of the dimensions.Where the paper says it
Where the paper says it
Performance and Comparison
The trained models are evaluated on an independent test sample of 3 (2.5) million events in the τhτh (τlτh) channel, where the resolution achieved on the reconstructed spin observables significantly improves on that obtained from a stand-alone transformer regressor trained with a mean-squared-error loss <ref:2608. p. 0 Rather
than performing a point regression, the method models the full conditional density of the neutrino
kinematics, providing both the most likely configuration for each event and an estimate of its
uncertainty. The resolution of the reconstructed spin observables improves on that obtained from
networks trained with a mean-squared-error loss. The resolution is defined as the interquartile range (IQR) of the difference between the prediction and the generator-level truth. The resolution for
the TauPolaris MAP approach is improved by approximately 40% compared with the stand-alone transformer,
and 20% compared with the sampled approach, where
the improvement is quoted as the fractional reduction in
the IQR.Where the paper says it
Where the paper says it
Scientific Applications
The method has been demonstrated in three benchmark studies:
-
Entanglement measurements at the LHC, where it shows that the presence of quantum entanglement in H → τ τ decays can be distinguished from its absence with a significance of at least 4. p. 0
3σ <ref:2608. p. 0Where the paper says it
Using simulated LHC proton-proton collisions, including detector resolution effects, we demonstrate the method in three applications. For tau leptons produced in Higgs boson decays, we show that the presence of quantum entanglement can be distinguished from its absence with a significance of at least 4.
10961#pg6>.Where the paper says it
For tau leptons produced in Higgs boson decays, we show that the presence of quantum entanglement can be distinguished from its absence with a significance of at least 4. 3σ at the High-Luminosity LHC.
-
Improved measurements of CP violation in H → τ τ decays, demonstrating an 18% improvement in sensitivity to CP violation <ref:2608. p. 0
10961#pg6>.Where the paper says it
3σ at the High-Luminosity LHC. We demonstrate an 18% improvement in the sensitivity to CP violation in H → τ τ decays.
-
New variables for suppressing the Z → τ τ background in Higgs boson searches, constructed from the reconstructed polarimetric vectors that are independent of the CP state of the Higgs boson <ref:2608. p. 0
10961#pg6>.Where the paper says it
We demonstrate an 18% improvement in the sensitivity to CP violation in H → τ τ decays. Finally, we introduce new variables for suppressing the Z → τ τ background in Higgs boson searches.
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Conditional normalizing flows can be used to estimate
the full conditional density of the neutrino kinematics rather than a single point estimate,
providingboth the most likely configuration for each event and an estimate of its uncertainty.
This allows AI systems to not only predict a single neutrino momentum but also quantify the spread of possible configurations, which is crucial for spin measurements. -
The method can
reproduce the generator-level distributions closely across all variables,
including thereconstructed tau mass (lower left), which is difficult to regress as it requires accurate modeling of the correlations between different components of the neutrino momenta.
This enables AI systems to reconstruct physical observables like tau mass with better accuracy than methods that only predict individual components independently. -
The system can be used to determine CP violation in H -> tau tau decays with a
significance of at least 4.3 sigma at the High-Luminosity LHC,
and it can also improve sensitivity to CP violation byat least 18% using the phiCP reconstructed with the estimated polarimetric vectors.
This allows AI systems to perform high-precision searches for new physics related to CP structure in tau decays. -
The tool can be used as a
handle for suppressing background processes
such as Z -> tau tau, by constructing variables from reconstructed polarimetric vectors thatseparate spin-0 Higgs boson events from spin-1 Z boson events, and which are independent of the CP state of the Higgs boson.
This allows AI systems to design specific analysis cuts to effectively reduce irreducible backgrounds in Higgs searches. -
The system can provide an
event-level uncertainty on the reconstructed phiCP due to the estimation of the neutrinos,
which can be used foran optimized event selection, which we leave to future work,
potentially increasing sensitivity by selecting events where phiCP is well reconstructed. This allows AI systems to implement dynamic event selection based on predicted uncertainty rather than just a fixed cut. -
The performance comparison shows that the TauPolaris MAP resolution is
improved by approximately 40% compared with the stand-alone transformer
and20% compared with the sampled approach,
indicating that using a gradient-based maximization routine for finding the most probable neutrino configuration is superior for spin observable reconstruction.
Abstract
The kinematics of tau lepton decay products depend on the tau spin, giving access to the spin correlations and CP structure of the process that produced them. The optimal spin observable is the polarimetric vector, which points along the most likely direction of the tau spin. Determining it requires the momenta of the neutrinos produced in the tau decays, which escape detection and must therefore be inferred. We present TauPolaris, a tool for reconstructing tau polarimetric vectors by estimating the undetected neutrino momenta with a conditional normalizing flow. Rather than performing a point regression, the method models the full conditional density of the neutrino kinematics, providing both the most likely configuration for each event and an estimate of its uncertainty. The resolution of the reconstructed spin observables improves on that obtained from networks trained with a mean-squared-error loss. Using simulated LHC proton-proton collisions, including detector resolution effects, we demonstrate the method in three applications. For tau leptons produced in Higgs boson decays, we show that the presence of quantum entanglement can be distinguished from its absence with a significance of at least 4.3σ at the High-Luminosity LHC. We demonstrate an 18% improvement in the sensitivity to CP violation in H toττ decays. Finally, we introduce new variables for suppressing the Z toττ background in Higgs boson searches.
Sources
- Prospects of Measuring the Parity of Higgs Particles
- Analysis of the CP structure of the Yukawa coupling between the Higgs boson and $\tau$ leptons in proton-proton collisions at $\sqrt{s} =$ 13 TeV
- Measurement of the $CP$ properties of Higgs boson interactions with $\tau$-leptons with the ATLAS detector
- Analysis of the $C\!P$ structure of the Yukawa coupling between the Higgs boson and tau leptons in proton-proton collisions at $\sqrt{s}$ = 13.6 TeV
- Quantum tomography with $\tau$ leptons at the FCC-ee
- Constraining new physics in entangled two-qubit systems: top-quark, tau-lepton and photon pairs
- Quantum information and CP measurement in $H \to \tau^+ \tau^-$ at future lepton colliders
- Entanglement and Bell Nonlocality in $\tau^+ \tau^-$ at the LHC using Machine Learning for Neutrino Reconstruction
- Measurement of the $\tau$ lepton polarization in Z boson decays in proton-proton collisions at $\sqrt{s}$ = 13 TeV
- \nu-Flows: Conditional Neutrino Regression
- $\nu^2$-Flows: Fast and improved neutrino reconstruction in multi-neutrino final states with conditional normalizing flows
- Entanglement of Formation of an Arbitrary State of Two Qubits
- A comprehensive guide to the physics and usage of PYTHIA 8.3
- Extraction and validation of a new set of CMS PYTHIA8 tunes from underlying-event measurements
- Documentation of TauSpinner algorithms -- program for simulating spin effects in tau-lepton production at LHC
- Universal Interface of TAUOLA Technical and Physics Documentation
- DELPHES 3, A modular framework for fast simulation of a generic collider experiment
- Description and performance of track and primary-vertex reconstruction with the CMS tracker
- Performance of reconstruction and identification of $\tau$ leptons decaying to hadrons and $\nu_\tau$ in pp collisions at $\sqrt{s}=$ 13 TeV
- PyTorch: An Imperative Style, High-Performance Deep Learning Library
Related papers
- MadNIS at NLO
- The NANOGrav 15-year data set: A Limit on Continuous Spin Gravitons
- DIS-secting the universe with high-energy neutrinos
- Crowdsourcing Gravitational Waves from Superradiant Axions
- Classification of g-modes for neutron stars with a strong transition: Novel universal relation including slow stable hybrid stars
- Scalar axion field of toroidal electromagnetic pulses