Exploring the SMEFT landscape: Bayesian Model Selection for indirect discovery

This paper proposes a Bayesian model selection framework utilizing genetic algorithms to navigate the SMEFT landscape as a space of competing operator hypotheses, demonstrating that this approach yields more robust characterizations of Wilson coefficients and clearer identification of model correlations than traditional global fits when applied to LEP and LHC data.

The Gist

Imagine the Standard Model of particle physics as a massive, incredibly detailed instruction manual for how the universe works. For decades, this manual has been perfect. But scientists suspect there are missing pages or hidden chapters describing "New Physics" (like dark matter or why neutrinos have mass). The problem is, we can't see these new chapters directly yet.

Instead of looking for the new chapters directly, this paper proposes a new way to look for them indirectly: by checking if the existing instructions in the manual are slightly "off" when we run high-speed experiments at the Large Hadron Collider (LHC).

Here is a breakdown of the paper's approach and findings, using simple analogies.

1. The Old Way vs. The New Way

The Old Way (Global Fits):
Imagine you have a giant jigsaw puzzle with 52 different pieces that might be part of the picture. The traditional method tries to force all 52 pieces into the puzzle at once, even if most of them don't belong. It then asks, "How much does the picture change if we wiggle these pieces?"

• The Problem: If you try to move 52 pieces at once, the puzzle becomes so flexible that it can stretch to fit almost anything. A real, small "glitch" in the picture might get lost because the puzzle is so wobbly. It's like trying to hear a whisper in a room where everyone is shouting.

The New Way (Bayesian Model Selection):
This paper suggests we stop trying to fit all 52 pieces at once. Instead, we treat every possible combination of pieces as a different "hypothesis" or a different version of the puzzle.

• The Analogy: Imagine a detective trying to solve a crime. Instead of assuming every suspect is guilty at the same time, the detective tests specific groups: "Is it just Suspect A?" "Is it Suspect A and B?" "Is it just Suspect C?"
• The Tool: The authors use a "Genetic Algorithm." Think of this as a digital evolution process. The computer creates thousands of different "teams" of operators (pieces), tests how well they explain the data, and then "breeds" the best teams together, keeping the winners and discarding the losers. This allows the computer to efficiently find the specific combination of pieces that actually fits the data, without getting confused by the ones that don't.

2. The "Occam's Razor" Rule

The paper uses a statistical rule called Bayesian Model Selection. Think of this as a strict judge who loves simplicity.

• If a complex model (with many new pieces) only explains the data slightly better than a simple model (the Standard Model with no new pieces), the judge rejects the complex one.
• The judge only accepts a new piece if it provides a significant improvement in the explanation. This prevents the scientists from "overfitting"—creating a complex story just to explain random noise in the data.

3. The Results: The "Ghost" in the Machine

The authors ran this new method on a massive dataset from the LHC and the older LEP collider, looking at data from Higgs bosons, top quarks, and other particles.

• The Linear vs. Quadratic Trap:

  • Linear Analysis (The First Glance): When they looked at the data using a simple, straight-line approximation, they found a few "suspects" (specific particle interactions) that seemed to fit the data better than the Standard Model. It looked like there might be a hint of new physics.
  • Quadratic Analysis (The Second Look): However, the paper argues that the simple approximation was a trick. When they added the "squared" terms (a more accurate, curved mathematical description), the "suspects" vanished.
  • The Metaphor: It's like seeing a shadow in the corner of a room and thinking it's a monster. When you turn on the bright light (the more accurate math), you realize it was just a coat rack. The "improvement" seen in the first glance was an illusion caused by the math being too simple.

• The Verdict:
  After running the genetic algorithm and applying the strict "simplicity judge," the paper concludes: There is no statistically significant evidence for new physics. The Standard Model remains the best description of the data. The "ghost" was just a trick of the light.

4. Why This Method is Better

Even though the result was "nothing new found," the paper argues the method is a huge success for two reasons:

1. Sharper Focus: Because the method doesn't try to fit all 52 pieces at once, it can pinpoint exactly which pieces are supported by the data and which are not. It gives a much clearer picture of the "shape" of the data.
2. Mapping the Relationships: The paper creates a "correlation map." It shows which pieces of the puzzle tend to appear together in the winning models. This helps scientists understand which measurements are currently "flat" (where different pieces look the same) and which new experiments would be most valuable to break those ties.

Summary

The paper introduces a smarter, more efficient way to search for new physics by testing specific combinations of possibilities rather than guessing everything at once. When they applied this to the latest data from particle colliders, they found that the Standard Model still holds up perfectly. The "anomalies" that looked promising in simpler analyses were revealed to be mathematical artifacts. The authors conclude that while we haven't found new particles yet, this new "detective toolkit" is ready to find them the moment they appear.

Technical Summary

Technical Summary: Exploring the SMEFT Landscape: Bayesian Model Selection for Indirect Discovery

Problem Statement
The Standard Model Effective Field Theory (SMEFT) is the standard framework for indirect searches for New Physics (NP). However, traditional SMEFT analyses rely on "global fits," where a large number of Wilson coefficients are constrained simultaneously. The authors argue that this paradigm is ill-suited for discovery. In high-dimensional global fits, the marginalization over many weakly constrained directions dilutes genuine signals, potentially washing out statistically significant deviations that occupy narrow regions of operator space. Furthermore, the standard approach treats SMEFT as a single high-dimensional model rather than a structured space of competing hypotheses, where different subsets of operators correspond to distinct low-energy realizations of UV completions. The challenge is to develop a framework that can systematically navigate this vast, discrete model space ($2^d$ possible operator subsets) to identify which specific operator combinations are favored by data, rather than merely estimating parameters within a fixed, potentially incorrect, model.

Methodology
The paper proposes a framework based on Bayesian Model Selection (BMS) applied to the SMEFT operator space.

• Hypothesis Space: SMEFT is redefined as a space of competing models $\mathcal{M} = \{0, 1\}^d$, where each model is defined by a binary vector $\vec{b}$ indicating the presence or absence of specific dimension-six operators.
• Inference Level: The analysis performs inference at the model level, calculating the posterior probability $p(\vec{b}|D)$ for operator subsets, rather than just parameter posteriors within a fixed model. This allows for the calculation of marginal inclusion probabilities for individual operators and Bayesian Model Averaged (BMA) posteriors for Wilson coefficients.
• Computational Strategy: Given the astronomical size of the model space ($2^{52} \approx 4.5 \times 10^{15}$ for $d=52$ operators), exhaustive enumeration is impossible. The authors employ a Genetic Algorithm (GA) to navigate the discrete space. The GA evolves a population of candidate models using tournament selection, crossover, and mutation, concentrating evaluations in regions of high posterior support.
• Approximations:
  • Evidence Approximation: To avoid expensive high-dimensional integrations for Bayesian evidence, the Bayesian Information Criterion (BIC) is used as a tractable proxy for the log-evidence, penalizing model complexity.
  • Parameter Optimization: Within each evaluated model, Wilson coefficients are optimized to minimize $\chi^2$. At linear order, this is analytic; at quadratic order, numerical minimization is used.
  • Uncertainty Estimation: A Hessian approximation is used to estimate parameter uncertainties, justified by the regularizing effect of Renormalisation Group Evolution (RGE) operator mixing which suppresses multimodal structures.
• Data and Setup: The analysis utilizes a dataset of 417 points comprising LEP electroweak precision observables and LHC Run 2 measurements (Higgs, top-quark, diboson). It considers a basis of 52 CP-even dimension-six operators in the Warsaw basis with specific flavor symmetries. The analysis is performed at both linear and quadratic orders in Wilson coefficients, with one-loop RGE evolution included to connect the matching scale $\mu_0$ to experimental energies.

Key Contributions

1. Formalism: The paper formalizes SMEFT analysis as a structured hypothesis space problem, shifting the focus from parameter estimation to model comparison and selection.
2. Algorithmic Implementation: It demonstrates the feasibility of navigating the full combinatorial SMEFT space using a Genetic Algorithm coupled with information criteria, making BMS computationally tractable for $d \sim 50$.
3. BMA vs. Global Fit: The authors introduce and demonstrate the utility of BMA posteriors. Unlike global fits, which marginalize over all directions (often diluting signals), BMA weights predictions by the posterior probability of the models, leading to sharper characterizations of supported operators.
4. Correlation Structure: The framework extracts a posterior correlation matrix ($\phi_{ij}$) between binary inclusion variables. This identifies operator pairs that co-appear in high-posterior models (indicating joint sensitivity) or are mutually exclusive (indicating flat directions/degeneracies), providing a relational map of the model posterior.
5. Scale Dependence: The study systematically assesses the sensitivity of results to the matching scale $\mu_0$, treating it as a probe of the preferred NP scale rather than a fixed convention.

Results

• Global Evidence: Applying the framework to current LEP and LHC Run 2 data, the analysis finds no statistically significant evidence for NP. The Standard Model (SM) is favored with high posterior probability: $p(H_{SM}|D) = 99.5\%$ (linear) and $99.97\%$ (quadratic), corresponding to "decisive" evidence for the SM on the Jeffreys scale.
• Linear vs. Quadratic Order:
  • At linear order, several four-fermion operators (notably $c_{tG}$ and $c^8_{Qt}$) show high marginal inclusion probabilities ($\sim 90\%$). However, the authors argue this is an artifact of EFT truncation; the required Wilson coefficients are large enough that neglected quadratic terms are numerically significant.
  • At quadratic order, these preferences largely vanish. The inclusion of squared terms breaks degeneracies and disfavors the large coefficients needed to fit the data at linear order. The SM becomes the second-best model overall, and no operator subset clearly outperforms it.
• Operator Correlations: At linear order, $c_{tG}$ and $c^8_{Qt}$ are strongly positively correlated ($\phi \approx 0.6$); neither improves the fit sufficiently in isolation due to RGE mixing effects (specifically $c_{tG} \to c_{\phi G}$). Quadratic terms break this correlation. The operator $c^{1111}_{ll}$ (four-lepton) remains robustly preferred across both orders and scales.
• BMA vs. Global Fit: BMA credible intervals are substantially narrower than those from traditional global fits at linear order, indicating that targeted model selection avoids the dilution of signals. At quadratic order, for some operators, BMA intervals are slightly wider, reflecting the non-Gaussian nature of the likelihood when squared terms are included.
• Matching Scale ($\mu_0$): Results show dependence on $\mu_0$ (tested at 1 TeV and 10 TeV). While specific operator inclusion probabilities shift (e.g., $c_{tG}$ drops from 85% to 20% when moving from 1 TeV to 10 TeV in the quadratic fit due to mixing), the global conclusion (SM preference) remains robust.

Significance and Claims
The paper claims that its primary significance lies in providing a concrete picture of how an indirect discovery of NP could unfold. It argues that SMEFT should be treated as a space of competing hypotheses rather than a single model. By applying Bayesian Model Selection, the framework:

• Avoids the "dilution" of signals inherent in global fits.
• Provides a probabilistic language for discovery (via Bayes factors and model posteriors) that naturally accounts for the look-elsewhere effect.
• Offers a more refined characterization of potential signals through BMA and correlation matrices, identifying which operators are truly supported and which are degenerate.
• Demonstrates that current data, when analyzed with quadratic terms and RGE, show no deviation from the SM, and that apparent linear-order tensions are likely truncation artifacts.

The authors conclude that while the current dataset supports the SM, the framework is essential for future high-precision programs (like HL-LHC) to resolve degeneracies and concentrate posterior weight on the correct effective description as data accumulates. They identify the promotion of the matching scale $\mu_0$ to a continuous parameter of inference as a natural future extension.