🔭 astrophysics

DegenDetector: Symbolic Recovery of Parameter Degeneracies in Bayesian Posteriors

The paper introduces DegenDetector, a framework that combines mutual information screening with alternating symbolic regression to automatically identify and express complex parameter degeneracies in Bayesian posteriors as interpretable, closed-form symbolic equations without requiring domain-specific input.

Original authors: Chaipat Tirapongprasert, Matthew Ho

Published 2026-07-10

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Chaipat Tirapongprasert, Matthew Ho

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

Imagine you are a detective trying to solve a mystery, but instead of fingerprints, you are looking at a cloud of data points that represent the secrets of the universe. In physics, scientists often use a method called "Bayesian inference" to figure out the values of different knobs and dials (parameters) that control how the universe works. Usually, they get a big, messy cloud of possible answers.

Sometimes, these clouds have a weird shape. Instead of being a round ball where every knob can be adjusted independently, the cloud squishes into a long, thin tube or a flat sheet. This is called a degeneracy. It means that if you turn one knob, you must turn another knob in a very specific way to keep the universe looking the same. It's like a seesaw: if you push one side down, the other has to go up. If you don't know this rule, your measurements might look super precise, but they are actually misleading because the knobs are secretly tied together.

The problem is that standard tools are like looking at a 3D object through a 2D window. You can see that the dots are clustered, but you can't see the shape of the tube they are hiding in, and you certainly can't write down the exact math equation that describes the seesaw rule.

Enter DegenDetector, a new digital detective tool created by researchers at Columbia University. Here is how it works, using a playful analogy:

The Detective's Two-Step Dance

Step 1: The "Who's Talking to Whom?" Scan
First, the tool looks at all the knobs to see which ones are gossiping with each other. It uses a math trick called Mutual Information to measure how much knowing the value of one knob reduces the mystery of another. If two knobs are totally independent, they don't talk (score is zero). If they are locked in a tight dance, they talk a lot (high score). The tool ranks every possible group of knobs to find the ones that are most likely to be the culprits.

Step 2: The "Guess the Equation" Game
Once it finds a suspicious group of knobs, it tries to guess the exact rule that ties them together. It doesn't just guess a simple straight line; it tries to find complex, closed-form equations (like $x^3 + y + z = 0$ ).

To do this, it plays a game of "hot and cold" called Alternating Optimization. Imagine you are trying to solve a puzzle where you have to guess three different functions at once. Instead of guessing all three at the same time (which is a nightmare), the tool guesses one, holds it steady, then guesses the second, holds it, and so on, cycling back and forth until the picture snaps into focus. It uses a smart algorithm called PySR to find the simplest, most elegant math sentence that fits the data.

The Proof: Did it Work?

The researchers didn't just hope it worked; they put it through a rigorous training camp with four different types of tricky puzzles:

The "S-Curve": A complex, wiggly shape. The tool found the exact equation: $(x^3 - 3x) + y + z = 0$ . It matched the data with a score of 0.9931 (where 1.0 is perfect).
The "Banana": A curved shape that looks like a fruit. The tool found the rule: $x^2 + y^2 = 0.5$ . It scored 0.996.
The "Trig": A shape involving waves (sines and cosines). The tool guessed: $2\sin(x) + \cos(y) + z = 1$ . It scored 0.983.
The "Cubic": A shape with a sharp curve. The tool found the rule: $1/(10 - x)^3$ . It scored 0.997.

In all these cases, the tool successfully separated the "bad guys" (the degenerate knobs) from the innocent bystanders (the independent knobs) and wrote down the exact math rule that described their relationship.

The Real-World Test: The Planck Telescope

To see if this works on real physics, the team fed it data from the Planck 2018 mission, which mapped the Cosmic Microwave Background (the afterglow of the Big Bang). They had seven cosmological knobs to check.

Without telling the tool any physics beforehand, it spotted the famous "horizon-angle degeneracy." It figured out that the expansion rate of the universe ( $H_0$ ) and the matter density ( $\Omega_m$ ) are tied together. It wrote down the equation:
$123.97 \log H_0 + 42.07 \log \Omega_m = \text{const}$

When the researchers crunched the numbers, the ratio of the coefficients was about 2.947. This is incredibly close to the expected value of 3 (which comes from the physics of how light travels through the expanding universe). The tool matched the real data with a score of 0.98.

What It Can't Do (Yet)

It's important to remember what this tool doesn't do. The researchers explicitly state that their method assumes the relationship between the knobs can be broken down into separate parts (like $f(x) + g(y) + h(z)$ ). If the relationship is a tangled knot that can't be separated this way, the tool might struggle. They suggest that future versions might need to learn more complex, multi-dimensional shapes, but for now, it sticks to separable equations.

The Bottom Line

DegenDetector is a new framework that turns a messy, confusing cloud of data points into a clear, readable math sentence. It doesn't just tell you that parameters are linked; it tells you exactly how they are linked. In their simulations and with real Planck data, it successfully recovered the hidden rules with high precision (scores above 0.98), proving that we can now automatically uncover the secret "seesaw" rules of the universe without needing a human to guess the formula first.

Technical Summary: DegenDetector

Problem Statement

In Bayesian inference for physical models, parameter degeneracy is a persistent challenge where observational data constrains a specific combination of parameters more tightly than the individual parameters themselves. This causes posterior samples to concentrate along a lower-dimensional manifold within the parameter space. While standard diagnostic tools like corner plots can indicate the existence of correlations, they fail to reveal the underlying functional form of these relationships. This lack of interpretability leads to downstream consequences, such as deceptively narrow error bars for coupled combinations and confidence intervals that are artifacts of projection rather than reflections of the data. Detecting these degeneracies is particularly difficult in high-dimensional spaces (more than two parameters) because they do not manifest in pairwise projections.

Methodology

The authors introduce DEGENDETECTOR, an automated framework designed to identify degenerate parameter subsets and express their relationships as closed-form symbolic equations. The pipeline operates in two primary stages:

1. Pinpointing Candidate Parameters

To avoid combinatorial explosions when searching for degenerate tuples among $M$ parameters, the method utilizes Mutual Information (MI) to rank parameter combinations.

Pairwise Screening: The algorithm computes pairwise MI scores using a k-nearest-neighbor estimator to measure statistical dependence.
Tuple Ranking: For a target coupling depth $k$ , all parameter tuples are enumerated and assigned a score equal to the sum of their constituent pairwise MI values. Tuples with higher aggregated MI are prioritized as they imply stronger, more significant degeneracies.

2. Symbolic Surface Fitting

Once a candidate tuple is identified, the degeneracy is modeled as the level set of a separable function:
$g_1(\theta_{j_1}) + g_2(\theta_{j_2}) + \dots + g_k(\theta_{j_k}) = c$
where $g_\ell$ are univariate functions to be discovered and $c$ is a real constant.

Alternating Optimization: The framework fits the component functions and the constant using an alternating optimization strategy. It cycles through one component at a time, holding others constant, effectively decomposing the $k$ -dimensional fitting problem into $k$ independent one-dimensional symbolic regression problems.
Symbolic Regression: The component functions are discovered using PySR (Python Symbolic Regression).
Normalization and Scaling: Parameters are normalized via z-score to ensure equal contribution regardless of physical units. Upon convergence, SymPy is used to simplify the expressions and restore them to the original scale.
Log-Space Extension: A feature called LOGDEGEN allows for the detection of multiplicative degeneracies by applying a coordinate transformation ( $\tilde{\theta}_i = \log \theta_i$ ) before running the pipeline.

Diagnostics and Evaluation

To evaluate the quality of the fit without the asymmetry inherent in standard $R^2$ (which requires designating one variable as dependent), the authors define an orthogonal $R^2$ ( $R^2_\perp$ ). This metric calculates the mean squared perpendicular distance from the posterior samples to the fitted surface. The system provides diagnostic visualizations, including MI heatmaps, corner plots with overlaid contours, and true-versus-predicted plots for component functions.

Key Results

The framework was validated on synthetic posteriors with known analytical degeneracies and applied to real-world data:

Synthetic Benchmarks: DEGENDETECTOR was tested on four distinct regimes:
- S-curve: A non-monotone degeneracy involving three parameters ( $x, y, z$ ).
- Banana: A 3D extension of the Rosenbrock function.
- Trig: A periodic degeneracy involving sine and cosine functions.
- Cubic: A test on posteriors with mismatched parameter scales (anisotropic priors).
  In all cases, the algorithm successfully distinguished the degenerate parameters from nuisance parameters and recovered functional forms with $R^2_\perp > 0.98$ .
Planck 2018 Application: The method was applied to posterior chains from the Planck 2018 CMB measurements involving seven cosmological parameters. Without prior knowledge of the physics, the LOGDEGEN feature identified the $(H_0, \Omega_m)$ pair and recovered the relationship:
$123.97 \log H_0 + 42.07 \log \Omega_m = \text{const}$
This corresponds to a ratio of $\approx 2.95$ , which is consistent with the established physical degeneracy $\Omega_m h^3 \approx \text{const}$ (where the exponent is empirically $\sim 3$ ). The fit achieved an $R^2_\perp \approx 0.98$ .

Significance and Claims

The paper claims that DEGENDETECTOR fills a critical gap in Bayesian analysis by providing interpretable, closed-form symbolic equations for parameter degeneracies, rather than just indicating their existence.

Automation and Domain Independence: The framework operates without domain-specific input, making it applicable to complex physical phenomena where likelihoods are evaluated via simulation-based inference (SBI) or MCMC.
Scalability: By combining MI screening with alternating symbolic regression, the method scales to high-order parameter spaces where standard diagnostic tools fail.
Utility: The resulting equations can inform reparameterization strategies to decorrelate parameters, thereby improving the efficiency of inference and the interpretability of results.

The authors acknowledge that the primary limitation of the current method is the assumption of separability in the functional form. They suggest that future work could extend the framework to non-separable forms via multivariate symbolic regression.