Introduction
In experimental particle physics, measurements at the Large Hadron Collider (LHC) are inherently distorted by detector effects. This "reconstruction level" (reco) data is a smeared, often lower-resolution version of the underlying "parton level" (truth) physics. The process of reversing these effects to recover fundamental physics is known as unfolding.
Traditional unfolding methods, typically based on binned matrices and histograms, struggle to scale as the number of variables (dimensions) increases. Furthermore, many modern machine learning-based unfolding techniques suffer from "prior bias," where the neural network "remembers" the specific physics model used during its training. If the real data follows a different theory than the simulation, the unfolded result may be systematically incorrect. The Simulation-Prior Independent Neural Unfolding Procedure (SPINUP) addresses these challenges by providing a one-step, prior-independent framework capable of handling high-dimensional phase spaces.
The SPINUP Methodology
The core philosophy of SPINUP is forward unfolding. Instead of attempting to learn an inverse mapping from reconstruction to truth, SPINUP models the forward detector response and optimizes a truth-level distribution to match observed data.
The framework consists of several key technical components:
- Transfer Network: A conditional generative model (using the "Transfermer" architecture) that explicitly learns the detector response $p(x_{reco}|x_{part})$. Once trained on simulation, this model is frozen to act as a surrogate for the detector.
- Unfolding Network: A generative model that represents the unknown "true" distribution. Its parameters are optimized to minimize the difference between the "folded" prediction and the actual observed data.
- Neural Importance Sampling (NIS): Calculating likelihoods in high-dimensional spaces is computationally intensive. SPINUP uses an auxiliary network to focus calculations on the most relevant parts of the phase space, ensuring efficient training.
- Ensembling: By training multiple versions of the network, the authors use the spread of results to estimate the uncertainty caused by information loss during the detector's blurring process.
By comparing the folded distribution directly to observed data, SPINUP ensures that the unfolding network remains independent of the simulation's initial truth-level assumptions. It only cares about matching the observed data through the learned forward model.
Experimental Validation and Results
The authors validated SPINUP across four distinct scenarios, ranging from simple 1D Gaussian toy models to complex 7-dimensional phase spaces involving associated Higgs and single-top production ($tHj$).
In the $tHj$ and jet substructure benchmarks, SPINUP recovered truth-level distributions with percent-level accuracy across the entire phase space. A critical metric for the method is "closure" at the reco level: the unfolded result, when passed back through the detector model, should perfectly match the input data.
Beyond simple distribution recovery, SPINUP was used to accurately extract fundamental parameters, such as the CP-phase $α$ of the top-Higgs coupling. The measurement remained robust even when the number of observed events was significantly reduced, demonstrating the method's statistical efficiency.
Discussion and Limitations
While SPINUP offers a powerful tool for prior-independent unfolding, it is not without limitations. The method is highly dependent on the accuracy of the forward Transfer Network; if the simulation used to train it does not reflect the real detector, the unfolding will be biased. Additionally, neural networks tend to prefer smooth functions, which can lead to the underestimation of sharp, narrow mass resonances. Future work may involve incorporating physics-inspired priors to help the network recognize such features without reintroducing model bias.
Conclusion
SPINUP represents a significant advancement in high-dimensional unfolding for particle physics. By prioritizing forward modeling and prior independence, it allows researchers to recover underlying physics distributions with high precision, providing a robust path for precision measurements at the LHC and future colliders.
Summary updated