Profiling systematic uncertainties in Simulation-Based Inference with Factorizable Normalizing Flows

This paper proposes a general Simulation-Based Inference framework utilizing Factorizable Normalizing Flows and an amortized training strategy to efficiently profile systematic uncertainties while simultaneously extracting multivariate distributions of interest, overcoming the computational bottlenecks of traditional unbinned likelihood fits.

The Gist

Imagine you are a detective trying to solve a crime, but the crime scene is a chaotic mess of millions of tiny clues (data points) scattered across a vast, multi-dimensional landscape. In the world of particle physics (like at the Large Hadron Collider), scientists are constantly trying to figure out the "true" laws of nature hidden inside this mess.

The problem is that their tools are often blunt. Traditionally, they would chop this landscape into a grid (like a histogram) to count the clues. But this is like trying to understand a painting by only counting the number of red pixels in each square inch—you lose the brushstrokes, the gradients, and the fine details.

Furthermore, there are "ghosts" in the machine. These are systematic uncertainties: tiny, annoying errors in the equipment, the simulation software, or our understanding of the universe that can shift the clues around. To get a reliable answer, scientists have to "profile" these ghosts—essentially asking, "What if the scale was off by 1%? What if the temperature was different?" Doing this for every single clue is computationally impossible with old methods.

This paper proposes a brilliant new way to solve this using AI and a concept called "Factorizable Normalizing Flows." Here is how it works, broken down into simple analogies:

1. The "Shape-Shifting" Map (The Distribution of Interest)

Instead of just counting how many clues are in a box, the authors want to learn the exact shape of the clue distribution.

• The Analogy: Imagine you have a lump of clay (the raw data from the experiment) and a perfect, smooth sculpture (the theoretical prediction). Your job is to stretch, squash, and twist the clay until it perfectly matches the sculpture.
• The Innovation: Instead of just guessing the final shape, they use a special AI (a Normalizing Flow) that acts like a magical, reversible map. It learns exactly how to stretch the clay to match the sculpture. Because it's reversible, they can also go backward: take the sculpture and turn it back into the clay. This allows them to measure the "shape" of the data without losing any information.

2. The "Ghost" Problem (Systematic Uncertainties)

Now, imagine that every time you try to stretch the clay, a mischievous ghost (a nuisance parameter) sneaks in and pushes the clay slightly to the left or right. There are hundreds of these ghosts.

• The Old Way: To handle this, scientists used to make a separate clay model for every possible position of every ghost. If you had 100 ghosts, you'd need to build millions of models. It was slow, expensive, and impossible to scale.
• The New Way (Factorizable Normalizing Flows): The authors realized that these ghosts don't need to be modeled as one giant, tangled mess. They can be factored.
  • Think of the clay stretching as a recipe. The old way tried to learn one giant recipe for "Ghost A + Ghost B + Ghost C."
  • The new way says: "Let's learn a tiny, separate recipe for Ghost A, a tiny one for Ghost B, and so on."
  • Then, when you want to see what happens if Ghost A is strong and Ghost B is weak, you just mix those specific tiny recipes together. This prevents the "combinatorial explosion" (the math getting too big to handle).

3. The "Amortized" Training (Learning Once, Using Forever)

This is the real magic trick. Usually, to figure out how the ghosts affect the result, you have to stop, change the ghosts, re-run the whole simulation, and re-stretch the clay. This takes forever.

• The Analogy: Imagine you are training a chef to cook a meal.
  • Old Method: You tell the chef, "Cook for a spicy day." They cook. Then you say, "Now cook for a salty day." They have to start over from scratch. Then "Sour day." Start over again.
  • New Method (Amortized): You train the chef once by feeding them thousands of different scenarios in a single day: "Here is a spicy day, here is a salty day, here is a mix of both." The chef learns a universal rule for how to adjust the seasoning based on the weather.
  • The Result: Once the training is done, if you ask the chef, "What if it's 50% spicy and 20% salty?" they can answer instantly without re-cooking. They have "amortized" (spread out) the cost of learning over the whole training session.

4. The "Principal Components" (Finding the Real Culprits)

Sometimes, the ghosts work together in confusing ways. The paper introduces a way to untangle them.

• The Analogy: Imagine a choir where everyone is singing slightly off-key. It's hard to tell who is the worst singer. The authors use a mathematical trick (like a spotlight) to find the "Principal Modes."
• They identify the specific combinations of ghosts that actually move the needle the most. It's like realizing that while there are 50 singers, only 3 of them are actually causing the song to sound bad. This makes it much easier to understand and fix the problem.

Why Does This Matter?

This paper is a game-changer for High Energy Physics because:

1. It keeps all the details: No more chopping data into bins and losing information.
2. It handles the mess: It can deal with hundreds of uncertainties without the computer crashing.
3. It's fast: Once trained, it can instantly tell you how uncertainties affect your results, saving months of computing time.

In short, they built a smart, flexible, and reversible map that learns how to correct experimental errors on the fly, allowing scientists to see the true shape of the universe with unprecedented clarity.

Technical Summary

1. Problem Statement

Unbinned likelihood fits are the gold standard in High Energy Physics (HEP) for extracting maximum information from continuous data distributions. However, their application is severely hindered by two main challenges:

• Computational Cost of Systematic Profiling: Traditional methods model systematic uncertainties (nuisance parameters, $\nu$) by creating discrete template variations (e.g., $\pm 1\sigma$ histograms). In high-dimensional feature spaces, binning is intractable, and interpolating between templates becomes computationally prohibitive. Current Simulation-Based Inference (SBI) methods often require retraining models for every systematic variation or training expensive conditional density estimators that must learn the full joint dependence on all nuisance parameters simultaneously, leading to a "curse of dimensionality."
• Limitation to Scalar Parameters: Most existing SBI methods focus on estimating scalar parameters of interest (POIs), such as signal strength. They struggle to measure full Differential Distributions (functional measurements) or correct for data-simulation mismodeling without resorting to binning.

2. Methodology

The authors propose a framework combining Normalizing Flows (NFs) with a novel architecture called Factorizable Normalizing Flows (FNF) to perform unbinned profile likelihood fits.

A. Core Framework: Distributions of Interest (DoI)

Instead of optimizing scalar parameters, the method treats the measurement target as a Distribution of Interest (DoI), defined as a learnable, invertible transformation $T_\phi$.

• The observed data distribution is modeled as a transformation of a reference (nominal) simulation density: $p_{data}(y) = p_{nom}(T_\phi(y)) \cdot |\det \nabla T_\phi|$.
• This allows the fit to learn the optimal "morphing" of the observable space to match data, effectively performing a functional measurement (e.g., differential cross-sections) without binning.

B. Handling Systematics: Factorizable Normalizing Flows (FNF)

To address the scalability of nuisance parameters, the authors introduce FNF. Instead of learning a single complex conditional density $p(y|\nu)$, the systematic effect is modeled as a parametric deformation of the nominal density.

• Factorization: The transformation $T_\nu$ (mapping distorted data back to nominal space) is decomposed into independent contributions from each nuisance parameter $\nu_k$.
• Quadratic Expansion: The scale ($s$) and shift ($t$) parameters of the affine coupling layers are modeled as a sum of independent linear and quadratic terms for each nuisance parameter:
  $$s_j(\nu) = \sum_k (\alpha_j^{(k)} \nu_k + \beta_j^{(k)} \nu_k^2)$$
  $$t_j(\nu) = \sum_k (\gamma_j^{(k)} \nu_k + \delta_j^{(k)} \nu_k^2)$$
• Architecture: The coefficients ($\alpha, \beta, \gamma, \delta$) are output by separate, lightweight Masked Multi-Layer Perceptrons (MLPs) conditioned on the input features. This ensures the model scales linearly with the number of nuisance parameters, avoiding combinatorial explosion.

C. Amortized Training Strategy

A critical innovation is the amortized profiling strategy, which shifts the computational cost of profiling from the inference stage to the training stage.

1. Step 1 (Global Fit): Jointly optimize the DoI transformation ($T_\phi$) and nuisance parameters ($\nu$) to find the global best-fit point.
2. Step 2 (Amortized Systematic Training): Fix the nominal transformation and train the systematic component ($T_\psi$) by sampling nuisance configurations $\nu_j$ from their prior distributions during training.
   • The network learns the conditional dependence $T_\psi(\nu)$ in a single optimization pass.
   • Once trained, the model can instantaneously evaluate the likelihood for any configuration of nuisance parameters, bypassing the need for repetitive retraining or scanning during the likelihood minimization.

D. Uncertainty Quantification

The framework enables Orthogonal Decomposition of the uncertainty space. By computing the Hessian of the negative log-likelihood with respect to $\nu$ and performing eigen-decomposition, the authors identify "principal modes" of systematic variation. This allows analysts to visualize the dominant systematic effects on the DoI transformation, decoupling correlated nuisance parameters.

3. Key Contributions

1. Factorizable Normalizing Flows (FNF): A novel architecture that models systematic uncertainties as additive, quadratic deformations in the latent space of normalizing flows. This enables tractable modeling of high-dimensional systematic effects with linear scaling in the number of nuisances.
2. Amortized Profiling: A training strategy that learns the full response of the likelihood to nuisance parameters in one go. This eliminates the computational bottleneck of traditional profile likelihood scans in unbinned analyses.
3. Functional Measurement (DoI): Extending SBI targets from scalar parameters to full invertible transformations, enabling unbinned, differential measurements and data-simulation correction without binning artifacts.
4. Interpretability: The factorized structure allows for the direct interpretation of individual systematic contributions and the orthogonal decomposition of complex uncertainty correlations.

4. Experimental Results

The method was validated on a synthetic dataset mimicking a high-energy physics measurement with two classes of events, kinematic variables, and two distinct sources of systematic uncertainty (scaling and shape distortion).

• Global Fit: The model successfully learned the global best-fit transformation, accurately correcting the distorted data distributions to match the nominal simulation across all kinematic bins.
• Likelihood Scans: The method produced well-defined likelihood profiles for nuisance parameters, correctly identifying best-fit values and confidence intervals.
• Amortized Profiling: After the amortized training step, the model could instantly evaluate the likelihood across the nuisance space. The resulting profiles reflected the systematic uncertainties on the DoI, showing shifts in best-fit values compared to the nominal fit.
• Orthogonal Decomposition: The PCA-based analysis successfully identified the principal components of the systematic variations, visualizing how specific combinations of nuisance parameters drive the uncertainty in the measured distribution.

5. Significance and Outlook

This work bridges a critical gap between advanced machine learning inference and rigorous statistical analysis in HEP.

• Scalability: It makes unbinned likelihood fits feasible for analyses with hundreds of nuisance parameters, a task previously considered computationally intractable due to template interpolation costs.
• Precision: By avoiding binning, it preserves the full information content of the data, potentially leading to more precise measurements of differential cross-sections and new physics searches.
• Unfolding: The framework offers a path toward "systematic-aware unfolding," where detector effects and systematic uncertainties are corrected simultaneously in a high-dimensional, continuous space.
• Reinterpretability: The invertible nature of the DoI transformation allows theoretical physicists to test new models against experimental results without re-running full detector simulations, provided they can map their predictions through the learned systematic variations.

In conclusion, the paper presents a robust, scalable, and interpretable framework for next-generation HEP analyses, transforming the profiling of systematic uncertainties from a computational bottleneck into an efficient, learned component of the inference pipeline.