⚛️ phenomenology

RIDDLE: Revealing Irregularities through Distributional Discrepancies in Latent-space via Stein-Witness Estimation

RIDDLE is a novel, data-driven method for resonant anomaly detection that utilizes a Stein-witness function in a latent space to effectively distinguish signal events from background distributions without requiring an explicit signal model, while successfully preserving the background dijet-mass spectrum.

Original authors: Ali Eren Simsek

Published 2026-10-23

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Ali Eren Simsek

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the vast, high-energy collisions that occur inside particle accelerators, physicists are constantly hunting for something new. They are looking for signs of particles that have never been seen before, often appearing as tiny, sharp spikes in a massive, smooth hill of data. This hill is made of ordinary particles produced by well-understood forces, while the spike would represent a rare, exotic event. The challenge is that the hill is so enormous and the spike so small that finding it is like trying to spot a single, specific grain of sand on a beach while the wind is constantly shifting the dunes. If the method used to sift through the sand accidentally reshapes the beach itself, it can create fake peaks that look like discoveries but are actually just artifacts of the search. To find new physics, scientists need a way to separate the rare signal from the common background without distorting the landscape they are standing on.

A researcher has developed a new tool called RIDDLE to solve this exact problem. Instead of trying to guess what the new particle looks like in advance, RIDDLE acts as a keen observer that learns what "normal" looks like and then flags anything that deviates from that pattern. The method was tested using simulated data from the Large Hadron Collider, specifically looking at collisions that produce two jets of particles. In these simulations, the researcher created a scenario where a new, heavy particle was hiding among millions of ordinary collisions. They wanted to see if RIDDLE could find the hidden particle without messing up the smooth shape of the background data.

The core of the method relies on a clever trick involving a "witness." Imagine the data from the collisions as a complex cloud of points in a multi-dimensional space. The researcher first used data from the edges of the search area—regions where they knew no new particles existed—to build a detailed map of what the background cloud should look like. This map served as a reference. Then, they trained a computer program, a neural network, to act as a witness. This witness was not told which points were the new particles and which were the background; it was only shown the mixed data from the center of the search area. Its job was to learn the subtle differences between the mixed data and the background map it had built from the edges. By analyzing how the witness reacted to the data, the system could assign a score to every event. A high score meant the event looked very different from the expected background, suggesting it might be a signal.

What makes this approach particularly powerful is that it does not require the scientist to know the exact properties of the new particle they are looking for. Traditional methods often need a specific model of what the signal looks like to be effective. RIDDLE, however, learns purely from the data itself. It identifies irregularities by detecting discrepancies in the distribution of the data, rather than by matching a pre-defined template. The researcher tested this by training the system on a mixture of ordinary background events and a small number of simulated signal events, without telling the system which was which. The system successfully learned to distinguish the signal from the background, performing almost as well as if it had been given a perfect, pre-made map of the background from the signal region itself.

The study also checked to ensure that the method was robust. One major concern in these searches is "mass sculpting," a phenomenon where the process of selecting interesting events accidentally creates a fake peak in the background data. If the selection method changes the shape of the background hill, it becomes impossible to tell if a peak is real or just an artifact. The researcher tested RIDDLE under various conditions, including adding extra information about the angles between the jets and even deliberately shifting the data to create artificial correlations with the mass. In every case, RIDDLE managed to separate the signal from the background effectively while leaving the shape of the background data largely untouched. The smooth, falling curve of the background remained stable, confirming that the method was not creating false alarms.

The results showed that RIDDLE could identify the hidden signal with high precision, retaining a large fraction of the true signal events while rejecting the vast majority of the background noise. When compared to other methods that had access to more information, such as knowing exactly which events were signals during the training phase, RIDDLE performed remarkably close to those idealized scenarios. This suggests that the method is highly effective even when working with unlabeled data, where the true nature of the events is unknown. The study concludes that RIDDLE offers a robust, data-driven way to enhance the search for new resonances, combining strong discrimination power with the safety of preserving the background shape. It represents a significant step forward in the ability to find the needle in the cosmic haystack without reshaping the haystack itself.

Technical Summary of RIDDLE: Revealing Irregularities through Distributional Discrepancies in Latent-space via Stein-Witness Estimation

Problem Statement
Searches for narrow resonances in the dijet invariant mass ( $m_{jj}$ ) at the LHC are limited by the large, smoothly falling Quantum Chromodynamic (QCD) multijet background. While auxiliary observables (e.g., jet masses, substructure) can distinguish signal-like events, correlations between these observables and $m_{jj}$ can distort the background mass spectrum during event selection. This "mass sculpting" can create artificial structures resembling resonances, complicating the search. An effective anomaly detection method must therefore achieve strong signal-background discrimination while preserving the smooth shape of the background mass spectrum to allow for subsequent resonance searches.

Methodology: RIDDLE
The paper introduces RIDDLE (Revealing Irregularities through Distributional Discrepancies in Latent-space via Stein-Witness Estimation), a data-driven method for resonant anomaly detection that operates without an explicit signal model or labeled signal events.

Conditional Background Reference:
- A background reference is inferred from dijet-mass sidebands (regions outside the candidate signal region).
- A neural network mapping $T_\phi$ transforms auxiliary jet observables ( $x$ ) into a latent representation ( $z$ ). This mapping is conditioned on the dijet mass ( $m$ ) to account for variations in QCD jet observables across the mass spectrum.
- A differentiable conditional density model $q(z|m)$ is fitted to the mapped sideband events to serve as the background reference.
Scalar Stein-Witness Function:
- Instead of a binary classifier, RIDDLE optimizes a scalar neural network witness $u_\theta(z)$ with a single output.
- The training objective utilizes the Langevin–Stein operator $T_q$ . For a test function $f$ , the operator is defined as $T_q f = \nabla_z \cdot f + f \cdot \nabla_z \log q(z|m)$ .
- In RIDDLE, the test field is defined as the gradient of the scalar witness, $f = \nabla_z u_\theta$ . The resulting scalar Stein response is:
  $\mathcal{A}_q u_\theta(z, m) = \Delta_z u_\theta(z) + \nabla_z u_\theta(z) \cdot s_q(z, m)$
  where $s_q(z, m) = \nabla_z \log q(z|m)$ is the background density score.
- Stein's Identity guarantees that the mean response under the background reference is zero (under suitable smoothness and boundary conditions). The witness is trained to maximize the mean response on unlabeled signal-region (SR) data, thereby revealing distributional discrepancies between the SR data and the sideband-derived background.
Training Loss and Anomaly Scoring:
- The primary loss minimizes the negative mean Stein response on the SR data, regularized by an $L_2$ penalty on the witness gradient.
- An auxiliary "tail-ranking" term is added to improve the separation of high-scoring events, which are critical for background rejection.
- The final anomaly score $S(z, m)$ is constructed from the negative witness output, the squared gradient magnitude, and the Stein response. These components are normalized using percentiles from generated background events and calibrated to produce a score indicating how unusual an event is relative to the background at a specific mass.

Benchmarks
The method is evaluated against two benchmarks sharing the same Stein-witness construction but differing in information access:

Idealized Anomaly Detection (IAD): Uses the same unlabeled SR mixture for training but fits the background reference to an independent sample of pure QCD events from the SR (rather than sidebands). This tests the limit of sideband-based estimation.
Supervised: Trains the witness on known signal events and fits the reference on known background events. This represents the upper bound of performance when signal labels are available.

Key Results
The method was evaluated using simulated dijet events from the LHC Olympics dataset ( $W' \to XY \to (q\bar{q})(q\bar{q})$ signal).

Discrimination Performance:
- Using four auxiliary observables (two jet masses, two $N$ -subjettiness ratios), RIDDLE achieved an Area Under the Curve (AUC) of 0.885, comparable to IAD (0.890) and approaching the Supervised benchmark (0.968).
- The maximum Significance Improvement Characteristic (SIC) for RIDDLE was 14.56, nearly identical to IAD (14.45) and Supervised (14.61).
- RIDDLE demonstrated strong signal retention at fixed background acceptances (e.g., 55.2% signal efficiency at 0.5% background efficiency).
Background Mass Stability:
- Crucially, the anomaly-score selection preserved the shape of the QCD dijet-mass spectrum. The $\chi^2/\text{ndof}$ test for mass sculpting showed that RIDDLE's selection resulted in distortions comparable to random selection, confirming that the method does not introduce artificial resonance-like structures.
Robustness Tests:
- Angular Information: Adding the interjet angular separation ( $\Delta R$ ) as a fifth observable maintained RIDDLE's performance, though the Supervised benchmark benefited more significantly from this additional information.
- Mass-Correlated Features: Shifting jet-mass observables to introduce explicit dependence on $m_{jj}$ did not degrade RIDDLE's discrimination or mass-shape stability, demonstrating robustness against feature correlations.
- Lower Signal Fractions: RIDDLE retained useful discrimination capabilities even as the injected signal fraction in the training sample was reduced, outperforming the need for explicit signal labels.

Significance and Claims
The paper claims that RIDDLE offers a robust, data-driven approach for enhanced resonance searches. Its primary significance lies in its ability to:

Combine effective anomaly discrimination with background-shape preservation, addressing the critical challenge of mass sculpting in resonance searches.
Operate without an explicit signal model or signal labels during training, relying solely on unlabeled signal-region events and sideband-derived background references.
Achieve performance comparable to idealized benchmarks (IAD) and approaching supervised limits, validating the utility of sideband-derived references for signal-background separation in the studied conditions.

The author concludes that RIDDLE highlights the potential of Stein-witness estimation for resonant anomaly detection, providing a method that is both sensitive to new physics signals and safe for use in background estimation strategies.

Technical Summary of RIDDLE: Revealing Irregularities through Distributional Discrepancies in Latent-space via Stein-Witness Estimation

More like this