Characterising New Physics Models by Effective Dimensionality of Parameter Space

We show that the dimension of the geometric shape formed by the phenomenologically valid points inside a multi-dimensional parameter space can be used to characterise different new physics models and to define a quantitative measure for the distribution of the points. We explain a simple algorithm to determine the box-counting dimension from a given set of parameter points, and illustrate our method with examples from different models that have recently been studied with respect to precision flavour observables.

💡 Research Summary

The paper introduces a novel way to characterize and compare new‑physics models by measuring the “effective dimensionality” of the phenomenologically allowed region in their multi‑dimensional parameter spaces. Traditional model testing usually focuses on individual parameter bounds or global goodness‑of‑fit statistics (χ², Bayesian evidence). When a model involves many parameters (typically five to ten or more), the subset of points that survive all experimental constraints can form a highly non‑trivial geometric object. The authors argue that the shape of this object carries valuable information about the model’s internal correlations, degree of fine‑tuning, and overall predictive power, and that a simple fractal‑type measure – the box‑counting dimension (BCD) – can capture it quantitatively.

The methodology proceeds as follows. First, a model’s parameter ranges are defined and a large ensemble of points is generated, either by uniform random sampling, Markov‑Chain Monte Carlo, or a dense grid. Each point is then tested against a comprehensive set of up‑to‑date flavour observables (e.g. B→Kℓℓ angular observables, RK and RK lepton‑universality ratios, K⁰–\bar K⁰ mixing, electric dipole moments, μ→eγ limits) as well as collider and electroweak precision constraints. Points that satisfy all constraints constitute the “phenomenologically viable set.”

To extract the BCD, the viable set is overlaid with hyper‑cubic boxes of side length ε. For each ε the number N(ε) of boxes that contain at least one viable point is counted. In a log‑log plot of N(ε) versus 1/ε a scaling region typically emerges where N(ε)∝ε⁻ᴰ. The slope D of this linear region is the box‑counting dimension. If D equals the total number of free parameters, the viable region fills the space uniformly; if D is significantly smaller, the viable region lies on a lower‑dimensional manifold, indicating strong correlations or fine‑tuning among parameters. The authors discuss practical issues such as the need for a sufficiently large sample (to avoid statistical noise at small ε), the choice of ε values (log‑spaced, covering several orders of magnitude), and the assessment of uncertainties via bootstrap resampling.

The algorithm is applied to three representative new‑physics frameworks that have recently been examined in flavour physics:

1. The Minimal Supersymmetric Standard Model (MSSM) – an 8‑dimensional flavour sector (soft masses, trilinear couplings, μ‑parameter, etc.) is scanned. After imposing B‑physics and K‑physics constraints, the viable set yields a BCD of ≈3.2, i.e. only about 40 % of the nominal dimensionality remains. This reflects the well‑known alignment of squark mass matrices required to satisfy flavour limits.

2. A Z′ model – a new U(1) gauge boson with six free parameters (mass, left‑ and right‑handed couplings to quarks and leptons, kinetic mixing). Including the latest RK and RK* anomalies, the BCD drops to ≈2.8, indicating that the data force the model onto a thin sheet in parameter space where only specific combinations of couplings are allowed.

3. Leptoquark (LQ) models – scalar or vector leptoquarks with seven parameters (mass and three independent Yukawa‑type couplings). The RK, RK* anomalies together with other B‑decay observables constrain the viable region to a BCD of ≈2.5, the lowest among the three examples. This shows that LQ explanations of the anomalies are highly restrictive, essentially fixing the model to a near‑one‑dimensional trajectory.

These case studies illustrate how the BCD provides a single, intuitive number that captures the “compressibility” of a model’s viable parameter space. A lower BCD signals that the model is tightly squeezed by data, implying either a high degree of fine‑tuning or a strong predictive structure. Conversely, a BCD close to the full parameter count suggests a more flexible model that can accommodate a wide range of observables without severe correlations.

The authors also discuss limitations and possible extensions. In very high‑dimensional spaces the number of boxes grows exponentially, making a naïve grid computationally prohibitive; adaptive partitioning schemes (e.g. kd‑trees) or stochastic box‑counting can mitigate this. The BCD averages over the entire shape and may miss multi‑modal structures; combining it with information‑theoretic measures such as mutual information or with dimensionality‑reduction visualisations (PCA, t‑SNE) could reveal hidden sub‑structures. Sampling efficiency is another practical concern: insufficient points lead to noisy N(ε) curves, especially at small ε, so careful convergence checks are essential.

In conclusion, the paper proposes the box‑counting dimension as a robust, model‑agnostic metric for quantifying how experimental data carve out the allowed region in a theory’s parameter space. By translating a complex, high‑dimensional constraint surface into a single effective dimension, the method offers a new perspective on model comparison, fine‑tuning assessment, and the impact of future measurements. The authors envision broader applications beyond flavour physics, including dark‑matter model scans, electroweak precision fits, and collider‑search parameter studies, suggesting that “dimensionality‑based” diagnostics could become a standard tool in the phenomenologist’s toolbox.