📊 statistics

The causal structure of galactic astrophysics

This paper proposes applying causal discovery methods to astrophysics to overcome the limitations of traditional correlation-based analysis by inferring direct causal relationships, variable directions, and hidden confounders, demonstrating the approach's effectiveness on a dataset of approximately 450,000 nearby galaxies to distinguish between physical mechanisms that appear degenerate under standard correlation analysis.

Original authors: Harry Desmond, Joseph Ramsey

Published 2026-03-16

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Harry Desmond, Joseph Ramsey

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 figure out how a city works. You have a massive database containing information about every building: how tall it is, how many people live there, how much it costs, and what color the paint is.

If you just look at the numbers, you might notice a pattern: Tall buildings tend to have expensive paint.

A standard scientist might say, "Aha! Expensive paint causes buildings to be tall!" Or, "Tall buildings cause the paint to be expensive!" But they are stuck. They only see the correlation (the two things happening together), not the causation (which one actually pushes the other). Maybe a third factor, like "being in a rich neighborhood," causes both the height and the paint color.

This is the problem astronomers have faced for decades with galaxies. They see that big galaxies are bright, and bright galaxies are big, but they can't be sure why or which way the influence flows.

The New Detective Tool: Causal Discovery

This paper introduces a new detective tool called Causal Discovery. Instead of just asking, "Do these two things happen together?" it asks, "If I hold one thing constant, does the relationship between the others break?"

Think of it like this:

The Old Way (Correlation): You notice that whenever it rains, people carry umbrellas. You also notice that when people carry umbrellas, the ground gets wet. You might wrongly conclude that umbrellas cause the ground to get wet.
The New Way (Causal Discovery): You look at a day where it's raining, but you force everyone to put their umbrellas down (or look at a day where the ground is wet but it's not raining). If the ground stays dry when the umbrellas are down, you realize: Rain causes wet ground, and rain also causes umbrellas. The umbrellas didn't cause the wetness; they were just a side effect.

The Experiment: The "Galaxy City"

The authors took a massive dataset called the NASA Sloan Atlas, which contains information on about 450,000 nearby galaxies. They looked at seven key features for each galaxy:

Distance (Redshift)
Brightness (Luminosity)
Mass (How much stuff is in it)
Size (How big it is)
Shape (Is it a flat disk or a round ball?)
Recent Star Birth (How many new stars are being made right now?)
Thickness (Is it a thin disk or a fat sphere?)

They fed this data into a new, super-fast computer algorithm called FCIT (Fast Causal Inference with Targeted Testing). This algorithm is special because it can handle the messy, non-linear, and huge amounts of data that astrophysicists deal with, which older tools couldn't do.

What Did They Find?

The algorithm drew a "family tree" of cause and effect for galaxies. Here is the story it told, translated into simple terms:

1. The "Hidden Hand" of Distance
The algorithm correctly figured out that Distance (how far away a galaxy is) dictates how big it looks to us. This is just a trick of perspective (like a toy car looking small from far away). It also realized that distance tricks our measurements of brightness and mass because we only see the brightest, biggest galaxies far away. The algorithm successfully identified these "optical illusions."

2. Mass is the Boss
The most important finding is that Mass is the root cause.

Mass → Brightness: More mass means more stars, which means more light.
Mass → Size: Heavier galaxies tend to be physically larger.
Shape (Morphology) → Size: The shape of the galaxy (whether it's a flat disk or a round ball) directly influences how big it is, even if you know how bright it is.

3. The "Star Formation" Misconception
This is the coolest part. You might think that a galaxy's current burst of star formation (making new stars right now) is what makes the galaxy grow big.

The Algorithm says: NO.
It found that recent star formation does not directly control the galaxy's size.
Instead, star formation is just a side effect. It makes the galaxy brighter temporarily, but the size is determined by the long-term history of how the galaxy built up its mass and what shape it took over billions of years.

4. The "Missing Link" (The Mystery)
The algorithm drew some lines with question marks (circles) on them. This means, "We know these things are related, but we don't know which one causes the other, or maybe there's a secret third factor we didn't measure."

For example, the link between a galaxy's shape and its mass has a question mark. The authors suspect there is a "hidden variable" (like the amount of gas the galaxy ate or the environment it lives in) that is pulling the strings on both, but they didn't have that data in their list.

Why Does This Matter?

Before this paper, astronomers had to guess the direction of the arrows. They would say, "Maybe A causes B, or maybe B causes A, or maybe C causes both."

This paper is like finally getting a map that shows the direction of the arrows.

It proves that Mass drives Size, not the other way around.
It proves that Shape is a key player in determining size, not just a decoration.
It proves that Recent Star Birth is a passenger, not the driver, when it comes to galaxy size.

The Big Picture

Think of this as moving from looking at a shadow puppet show (where you just see shapes moving on a wall and guessing what's happening) to walking onto the stage and seeing the puppets, the strings, and the puppeteer.

The authors have built a new tool that allows us to stop just guessing about how the universe works and start actually testing our theories about how galaxies are born, grow, and change. It's a huge step forward, turning astronomy from a field of "looking at patterns" into a field of "understanding the machinery."

1. Problem Statement

Traditional astrophysics relies heavily on identifying statistical correlations between observed galaxy properties (e.g., mass, luminosity, size) to test physical theories. However, correlation does not imply causation.

Limitations of Current Methods: Correlation analysis cannot determine the direction of influence (e.g., does mass drive size, or vice versa?), nor can it distinguish between direct causal links and spurious correlations caused by confounding variables (unobserved factors influencing multiple observed variables).
The Gap: While causal discovery is well-established in fields like genomics and economics, it has seen limited application in astrophysics due to the scale of data (hundreds of thousands of objects), the highly non-linear nature of physical relationships, and the likelihood of unobserved confounders (e.g., dark matter halo mass, gas accretion history).
Goal: To develop and apply a causal discovery methodology capable of inferring the causal structure (directionality and confounding) of large-scale astrophysical datasets, thereby distinguishing between degenerate physical mechanisms that appear identical under standard correlation analysis.

2. Methodology

The authors propose a pipeline using the Fast Causal Inference with Targeted Testing (FCIT) algorithm, adapted for large-scale astrophysical data.

A. Data Source

Dataset: NASA Sloan Atlas (NSA) v1.0.1, containing $\sim 4.5 \times 10^5$ nearby low-redshift galaxies ( $z \lesssim 0.15$ ).
Variables: Seven key physical properties were selected:
1. Cosmological redshift ( $z_{dist}$ )
2. Absolute magnitude ( $r$ -band luminosity)
3. Recent star formation rate (normalized by 300 Myr average)
4. Stellar mass
5. Sérsic index (morphology)
6. Axis ratio (morphology/projection)
7. Angular half-light radius (size)
Preprocessing: Strict cuts were applied to remove anomalous objects, resulting in a clean sample of 445,763 galaxies.

B. The FCIT Algorithm

FCIT is a hybrid constraint-based and score-based algorithm designed to handle:

Latent Confounders: It does not assume "causal sufficiency" (that all relevant variables are observed). Instead, it outputs a Partial Ancestral Graph (PAG), which can represent directed edges (causal), bidirected edges (common unobserved cause), and ambiguous edges (circles).
Non-linearity: It utilizes basis-function expansions (Legendre polynomials) to model conditional dependencies, allowing for flexible, non-linear relationships without assuming Gaussian residuals.
Scalability: Unlike traditional algorithms that scale poorly ( $O(N^3)$ ), FCIT uses targeted testing and local conditioning to reduce computational complexity, enabling analysis of $\sim 10^5$ objects in minutes.

Key Steps in FCIT:

Skeleton Discovery: Iteratively removes edges by testing conditional independence using a likelihood ratio test (LRT) based on basis functions.
Targeted Blocking: Uses recursive blocking to identify conditioning sets that block inducing paths, avoiding combinatorial explosion.
Orientation: Applies orientation rules to identify unshielded colliders and propagate edge directions, resulting in a PAG.

C. Validation via Mock Data

To calibrate hyperparameters and assess reliability, the authors generated synthetic datasets using the Causal Perceptron Network (CPN):

Generation: Synthetic data was created using random multi-layer perceptrons (MLPs) to simulate realistic, non-linear causal structures with known ground-truth DAGs.
Calibration: The algorithm was tested on 200 mock datasets to optimize the truncation_limit (basis function order) and penalty_discount (BIC complexity penalty).
Result: Optimal performance (F1 score $\approx 0.9$ ) was achieved with truncation_limit = 14 and penalty_discount = 50.

3. Key Results

Applying the calibrated FCIT algorithm to the NSA data yielded a Partial Ancestral Graph (PAG) revealing the following causal structures:

A. Confirmed Causal Directions

Redshift ( $z$ ) $\to$ Apparent Size: Redshift determines apparent size due to angular diameter distance (observational effect).
Mass $\to$ Luminosity: Stellar mass causally drives absolute magnitude (more stars = more light).
Luminosity $\to$ Size: A directed edge exists from luminosity to size. This suggests that while mass drives luminosity, the observed size is fundamentally linked to the luminosity/mass assembly history, consistent with the mass-size relation ( $R \propto M^{0.4-0.55}$ ).
Recent Star Formation $\to$ Luminosity: Recent star formation activity directly brightens the galaxy (short-term effect).

B. Absence of Direct Links & Physical Interpretation

Star Formation $\nrightarrow$ Size (Direct): There is no direct edge from recent star formation to galaxy size once mass and morphology are accounted for.
- Interpretation: This disfavors models where instantaneous star formation efficiency drives size growth. Instead, it supports inside-out disk growth and hierarchical merging, where size is regulated by long-term mass assembly and structural evolution, with star formation being a secondary or correlated process.
Morphology as a Mediator: The Sérsic index ( $n$ ) has a directed edge to size ( $n \to \text{Size}$ ), indicating that morphology (bulge vs. disk dominance) causally influences size at fixed luminosity.

C. Latent Confounding (Ambiguous Edges)

Several edges, particularly between morphology, mass, and star formation, ended in circles ( $\circ$ ).

Meaning: These represent uncertainty in directionality or the presence of unobserved confounders.
Likely Causes: The authors identify potential latent variables such as halo mass, gas accretion history, feedback efficiency, and large-scale environment that jointly influence these observed variables.

4. Significance and Contributions

Methodological Breakthrough: This is one of the first successful applications of causal discovery to a dataset of $\sim 10^5$ astrophysical objects. It demonstrates that causal inference is computationally feasible for modern survey data.
Beyond Correlation: The study moves beyond "what correlates with what" to "what drives what." It successfully distinguishes between physical mechanisms that are degenerate under correlation analysis (e.g., distinguishing mass-driven growth from feedback-driven scenarios).
Theoretical Validation: The recovered PAG aligns with established galaxy formation theories (e.g., hierarchical merging, mass-size relations) while providing a rigorous, data-driven confirmation of causal directions.
Future Roadmap: The paper outlines how causal discovery can be used to:
- Test specific physical models by comparing their predicted DAGs against observational PAGs.
- Identify missing variables (confounders) to guide future observations (e.g., measuring gas mass or metallicity to resolve ambiguous edges).
- Move astrophysics from purely statistical fitting to causal model inference, effectively "inferring theories directly from data."

Conclusion

Desmond and Ramsey demonstrate that causal discovery is a powerful, necessary tool for modern astrophysics. By applying the FCIT algorithm to the NASA Sloan Atlas, they recovered a causal structure that validates hierarchical galaxy evolution models, identifies the role of morphology as a mediator, and explicitly highlights the limits of current data due to unobserved confounders. This work paves the way for causal discovery to become a mainstream technique for constraining physical theories in the era of big-data astronomy.