Causal Inference and Structure Learning in the Frequency Domain

This direction is merged into another direction: non-linear causal discovery. Its last version stays readable and citable here.

Version 1 of 114 papers24 conceptsfrozen 2026-10-04

In classical causal inference, we often deal with static snapshots: does smoking cause cancer? In the real world, however, systems are rarely static. They are dynamical. The state of a power grid, a neural circuit, or a financial market at this moment is a complex function of its own history and the history of its neighbors.

Traditionally, researchers have tackled these "time-series" problems in the time domain, looking at lagged correlations. But this approach faces a wall of complexity: as the time-lags grow, the number of parameters explodes, and the computational cost of identifying "who influences whom" becomes prohibitive.

A growing body of research suggests a more elegant path: shifting the entire problem into the frequency domain. By analyzing the spectral properties of a system, we can decouple complex temporal interactions into independent causal equations, turning a daunting dynamic problem into a collection of simple, static ones.

The Intuition: From Convolutions to Multiplications

In the time domain, a causal relationship is represented by a convolution. If variable $X$ influences $Y$, the value of $Y$ today is a weighted sum of $X$’s values over many previous time steps. Mathematically, this is messy.

As Mishfad Shaikh Veedu, James Melbourne, and Murti Salapaka demonstrate in Causal Structure Recovery of Linear Dynamical Systems: An FFT based Approach, the Fast Fourier Transform (FFT) changes the game. In the frequency domain, these convolutions become point-wise multiplications.

This transformation allows us to treat each frequency component as an independent "snapshot." At any single frequency, the relationship between variables mirrors a static Directed Acyclic Graph (DAG). This means we can take the entire toolkit of "static" causal inference—like Judea Pearl’s back-door and front-door criteria—and apply it directly to dynamical systems.

Decoupling and Macro-Causality

One of the most powerful features of the frequency domain is decoupling. In a complex system, a driver might exert a massive influence over long-term cycles (low frequencies) while being completely irrelevant to short-term fluctuations (high frequencies).

Dominik Janzing, Bernhard Schölkopf, and Paul K. Rubenstein explore this in Structural causal models for macro-variables in time-series. They show that interventions in the frequency domain are naturally targeted: manipulating a specific frequency in a cause only influences that same frequency in the effect. This provides a formal justification for "macro-causality," where we care about long-term averages or oscillatory patterns rather than moment-to-moment noise.

This scale-specific analysis is vital for fields like climatology. In Asymptotic Uncertainty in the Estimation of Frequency Domain Causal Effects for Linear Processes, Nicolas-Domenic Reiter and Jakob Runge provide the statistical framework to determine if a causal link at a specific frequency—such as an 11-year solar cycle—is a genuine physical driver or just a statistical fluke.

Efficiency: Logarithmic vs. Quadratic

The shift to the spectral domain isn't just a theoretical convenience; it is a computational necessity for large-scale systems.

In the time domain, recovering causal structure typically scales quadratically with the maximum time-lag. In contrast, Mohammed Tuhin Rana and colleagues show in Frequency-Domain Better than Time-Domain for Causal Structure Recovery in Dynamical Systems on Networks that spectral methods reduce this complexity to logarithmic scales. In practical terms, this means mapping the causal structure of a network with dozens of nodes can drop from hours of computation to mere seconds.

Enforcing Causality in Neural Architectures

This spectral perspective is also revolutionizing how we build efficient neural networks. Transformers are powerful but slow because their attention mechanism is $O(L^2)$. Researchers are now using frequency-domain "mixing" to achieve $O(L \log L)$ efficiency, but they face a hurdle: the Fourier Transform is inherently non-causal (it looks at the whole sequence at once).

To solve this, BINGZHENG GAN and the team behind Caracal: Causal Architecture via Spectral Mixing introduced frequency-domain causal masking. Similarly, Alexander Moreno in SKI to go Faster suggests enforcing causality by modeling only the real part of a frequency response and deriving the imaginary part via the Hilbert transform.

Even in edge-device applications like speech enhancement, Leyan Yang and Jing Lu show in UL-UNAS that "frequency folding" can provide global spectral context while maintaining the strict temporal causality required for real-time processing.

A Unified Framework: Structural Equation Processes

The culmination of this direction is the Structural Equation Process (SEP) framework. Developed by Andreas Gerhardus, Jonas Wahl, and others in Causal Inference on Process Graphs, Part I, this framework treats entire stochastic processes as single nodes in a graph.

By parameterizing the "spectral density" of these processes as matrices of rational functions, the authors prove that the algebraic structure of a time-evolving system in the frequency domain is identical to the covariance matrix of a static linear model. This bridge allows us to resolve infinite feedback loops and complex time-lags using simple, elegant linear algebra.

By looking at the "rhythm" of data rather than just its sequence, we gain a clearer, faster, and more mathematically robust view of the causal forces shaping dynamical systems.

Papers in this version

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_causal_inference_and_structure_learning__c5475842,
  title = {Causal Inference and Structure Learning in the Frequency Domain},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/causal-inference-and-structure-learning-in-the-frequency-domain/rdr_c5475842e7c440659a51163a1d90e361},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-22},
}

This exact version (version 1, frozen text):

@misc{lacuna_causal_inference_and_structure_learning__c5475842_v1,
  title = {Causal Inference and Structure Learning in the Frequency Domain},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/causal-inference-and-structure-learning-in-the-frequency-domain/rdr_c5475842e7c440659a51163a1d90e361/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-22},
}