Physicists Teach Machines to Draw the Perfect Histogram

Every discovery in particle physics rests on an unglamorous act: sorting millions of collisions into boxes. When researchers at the Large Hadron Collider search for a Higgs boson decay or a hint of new physics, they do not examine each collision individually. Instead, they compress the data into histograms—counts of events falling into bins defined by a discriminant, often the output of a machine-learning classifier—and compare those counts against theoretical predictions in a statistical fit. The trouble is that the placement of the bin boundaries has traditionally been a matter of hand-tuning, intuition, and habit. A new study published in The European Physical Journal C argues that this crucial step deserves the same algorithmic rigor as the classifiers themselves, and it delivers two ready-to-use tools that learn the optimal bins directly.

The work, carried out by Johannes Erdmann, Nitish Kumar Kasaraguppe Veerappa Gowda and Florian Mausolf of RWTH Aachen University, tackles a problem that quietly limits the sensitivity of many analyses. In multi-class classification tasks, the standard practice is to assign each event to the class with the highest classifier score—a rule known as argmax—and then apply simple equidistant binning to the resulting one-dimensional scores. Every such projection discards information about correlations between the score dimensions, and every fixed bin boundary may carve the data in a way that dilutes the statistical power of the final fit. The Aachen team asked a deceptively simple question: what if the bins themselves were treated as trainable parameters, optimized to maximize the expected signal significance of the analysis?

Their answer is a binning model built from a Gaussian Mixture Model, or GMM. Each bin is represented by a Gaussian component in the discriminant space, characterized by a mean vector, a covariance matrix and a mixture weight. An event is assigned to the component that gives it the highest probability density. Because the boundaries between components emerge from comparing multiple Gaussian densities rather than from any single ellipse, the resulting bins can take on highly non-linear, even convoluted shapes—something no rectangular grid could achieve. To keep the mathematics tractable, the covariance matrices are parameterized through a Cholesky factorization, which guarantees they remain positive definite while all parameters stay unconstrained and trainable. A damping hyperparameter on the off-diagonal elements keeps the training stable in practice.

On top of this flexible model, the researchers built two complementary optimization engines. The first, dubbed GATO for Gradient-based cATegorization Optimizer, exploits differentiable programming. Hard bin assignment is not differentiable, so GATO replaces it with a temperature-scaled softmax over the component scores: at high temperature the assignment is smooth and gradients flow easily, while as the temperature is annealed toward zero the assignment approaches the true step-like binning. Combined with a cosine annealing schedule for the learning rate and the RMSprop optimizer, this trick lets gradient descent carve out near-optimal categories. The second engine, BOBR—Bayesian Optimization of Bin boundaRies—takes a fundamentally different route. It treats the bin configuration as a black-box function to be maximized, using a Tree-structured Parzen Estimator as implemented in the Optuna framework, balancing exploration of the parameter space with exploitation of promising regions.

Both methods optimize the Asimov significance, a standard asymptotic formula for the expected statistical sensitivity of a counting experiment, summed in quadrature over all bins. Crucially, the team also introduced penalty terms that encode practical wisdom from real analyses. Bins with almost no expected background events are statistically treacherous, and bins whose background prediction carries a large relative uncertainty from Monte Carlo statistics can destabilize a fit. The optimization therefore penalizes configurations that fall below a minimum background yield or exceed a maximum relative uncertainty, steering the solution toward bins that are not only sensitive but also robust. The authors note that the framework could be extended further to nuisance-aware objectives that account for systematic uncertainties, for example through differentiable Fisher-information-based losses.

To benchmark the methods, the team constructed toy datasets mimicking typical collider discriminants: three-dimensional features sampled from multivariate Gaussians representing one or two signal processes and five backgrounds, with weights reproducing the imbalanced yields of real analyses. From these they built a one-dimensional binary likelihood-ratio discriminant and a three-class softmax-like output. In the one-dimensional case, where GATO can alternatively represent bin boundaries directly as sigmoid-smoothed cut values, both GATO and BOBR achieved very similar significances for a given number of bins—and both clearly outperformed equidistant binning, which would need substantially more bins to match the optimized sensitivity. The optimization histories tell two different stories: BOBR’s objective fluctuates as it probes different configurations, while GATO converges smoothly and steadily toward its optimum.

The three-class case is where the approach reveals its real power. Here the discriminant lives on a simplex, and the researchers optimized bins directly in the two-dimensional plane of the two signal scores, using the geometric mean of the two signal significances as a class-balanced figure of merit. In a scenario where the two signals and the background overlap substantially, the GATO GMM approach achieved roughly 20 percent higher significance than argmax classification with either equidistant or subsequently optimized one-dimensional binning, at every tested bin count. Even k-means clustering of the score space, an unsupervised alternative explored in earlier work, needed far more bins to catch up. In a second scenario with more localized signal regions, the projected approaches became competitive—but the multi-dimensional method never lost ground, and it consistently reached a target sensitivity with fewer bins.

The decisive test came on the FAIR Universe H→ττ dataset, a public benchmark based on simulated proton–proton collisions representative of ATLAS experiment data, featuring a Higgs signal and three backgrounds with 28 features per event. The team trained a four-class neural network classifier and fed its softmax scores into the binning optimizers, evaluating the expected H→ττ significance on five million independent events. Once again the GMM-based GATO optimization delivered the highest significance across all bin counts, improving on argmax categorization with one-dimensional Bayesian bin optimization by approximately 15 percent. BOBR with the GMM parameterization also beat the argmax baseline, though as the number of parameters grows with each additional bin, the gradient-based approach increasingly outpaces Bayesian optimization in these moderately high-dimensional search spaces.

What makes this work likely to spread quickly through the community is its packaging. Both methods are released as lightweight Python plugins—gato-hep on PyPI and bobr-hep on GitHub—designed to drop into existing analysis workflows. Unlike end-to-end differentiable pipelines such as INFERNO or neos, which retrain the classifier itself around an inference-aware loss, GATO and BOBR operate post-training on fixed classifier scores, meaning analysts can adopt them without touching their carefully validated machine-learning setups. At the same time, GATO can be embedded as one module in a fully differentiable analysis chain, opening the door to jointly optimizing classifier, binning and statistical inference. For a field where every fraction of a percent in significance brings the next discovery closer, teaching the histogram to learn may prove one of the most quietly consequential ideas in modern particle physics.

Subject of Research: Differentiable and Bayesian optimization of multi-dimensional histogram binning for signal significance in high-energy physics analyses

Article Title: Learning to bin: differentiable and Bayesian optimization for multi-dimensional discriminants in high-energy physics

Article References: Erdmann, J., Kasaraguppe Veerappa Gowda, N. K., & Mausolf, F. (2026). Learning to bin: differentiable and Bayesian optimization for multi-dimensional discriminants in high-energy physics. The European Physical Journal C, 86(8), Article 1014. https://doi.org/10.1140/epjc/s10052-026-16255-1

Image Credits: AI Generated

DOI: 10.1140/epjc/s10052-026-16255-1

Keywords: high-energy physics, histogram binning, Gaussian Mixture Model, differentiable programming, Bayesian optimization, signal significance, machine learning classifiers, binned likelihood fits, Higgs boson, FAIR Universe dataset, GATO, BOBR