⚡ electrical engineering

Rethinking Chronological Causal Discovery with Signal Processing

This paper investigates how the sensitivity of causal discovery methods to sampling rate and window length mismatches with underlying event timing can be better understood and addressed through concepts from signal processing.

Original authors: Kurt Butler, Damian Machlanski, Panagiotis Dimitrakopoulos, Sotirios A. Tsaftaris

Published 2026-02-24

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Kurt Butler, Damian Machlanski, Panagiotis Dimitrakopoulos, Sotirios A. Tsaftaris

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

The Big Idea: The "Blind Watchmaker" Problem

Imagine you are a detective trying to figure out who is causing what in a busy kitchen. You have a video camera, but you can only take a photo once every 10 seconds.

The Reality: The chef chops an onion (Event A), and 2 seconds later, the pot starts boiling (Event B).
Your Camera: Because you only snap a photo every 10 seconds, you might see the pot boiling in one photo and the chopped onion in the next. You might think, "Oh, the boiling caused the onion to appear!" or you might miss the connection entirely because the action happened between your photos.

This paper argues that Causal Discovery (using computers to figure out cause-and-effect from data) is failing because we often take our "photos" (data points) at the wrong speed or look at the wrong amount of history.

The Two Main Mistakes

The authors say that when scientists try to find cause-and-effect in time-series data (like heart rates, stock prices, or brain waves), they usually mess up in two ways:

1. The "Window" is Too Small or Too Big (The Memory Problem)

Imagine you are trying to guess what your friend is thinking based on their facial expressions.

Window Length (Q): This is how many seconds of "history" you look at.
The Mistake: If you only look at their face for 1 second (a tiny window), you might miss the slow smile that started 5 seconds ago. But if you look at 1 hour of video (a huge window), you get confused by everything that happened before the event you care about.
The Paper's Finding: The computer needs to look at the exact right amount of history to see the cause. If the window is too short, it misses the cause. If it's too long, it gets confused by noise. There is a "Goldilocks zone" for every specific system.

2. The "Camera Speed" is Wrong (The Sampling Rate Problem)

This is about how often you record data.

The Nyquist Analogy: Think of a spinning fan. If you take a photo of a fan spinning very fast, but your camera is too slow, the fan might look like it's standing still, or spinning backward. This is called "aliasing."
The Paper's Finding: If you record data too slowly (low sampling rate), the computer thinks things are happening instantly or in the wrong order. It might think "A caused B" when actually "B caused A," or it might think they are unrelated when they are actually linked.
The Twist: Sometimes, slowing down the data (downsampling) actually helps the computer find the pattern, but only if you do it in a very specific way. If you just guess, you break the logic.

The "Signal Processing" Solution

The authors are experts in Signal Processing (the math behind how we handle sound, radio, and images). They are telling the "Causal AI" community:

"You are trying to build a map of cause-and-effect, but you are using a blurry, low-resolution map. You need to apply the rules of signal processing to fix your camera settings before you try to draw the map."

They suggest that just as audio engineers have rules for how fast you must record sound to avoid distortion (the Nyquist rate), causal scientists need a "Causal Nyquist Rate." This would be a rule that tells you: "To figure out that A causes B, you must record data at least X times per second and look back Y seconds in history."

What They Tested

They took popular AI methods (like Granger Causality, Transfer Entropy, and newer Deep Learning models) and ran them through a simulation:

They created a fake world where X definitely causes Y with a specific delay.
They changed the "camera speed" (sampling rate) and the "memory window" (how much history the AI looked at).
The Result: The AI got it right only when the settings were perfect. If they changed the settings even a little bit, the AI started saying "X causes Y" when it didn't, or missed the connection entirely.

The Takeaway for Everyone

We are building powerful AI to understand the world (from how drugs affect the body to how the economy works). But this paper warns us: Garbage In, Garbage Out.

If we don't understand the physics of how we are recording our data (the timing and the speed), no amount of fancy AI math will save us. We need to treat our data collection like a high-quality recording studio, not just a quick snapshot.

In short: To find the truth about cause and effect, you have to tune your "camera" and your "memory" perfectly, or the AI will just tell you a convincing lie.

1. Problem Statement

The paper addresses a critical gap in Chronological Causal Discovery (CCD): the sensitivity of causal inference algorithms to the sampling rate ( $f_S$ ) and window length ( $Q$ ) of the input time series data.

The Core Issue: Most causal discovery methods assume discrete-time data faithfully represents underlying continuous-time physical or biological processes. However, in reality, the timing of data recording often mismatches the timing of the actual events.
The Consequence: If the sampling rate is too low (undersampling) or the analysis window length ( $Q$ ) does not align with the true propagation delay of the causal influence, causal discovery methods may fail to detect relationships or infer spurious ones.
The Gap: While signal processing theory (e.g., Nyquist-Shannon) dictates sampling requirements for signal reconstruction, there is no established theoretical framework for the sampling requirements necessary for accurate causal inference.

2. Methodology

The authors approach the problem by bridging Causal Machine Learning (Causal ML) and Signal Processing.

A. Theoretical Framework

Generative Model: They define the data-generative process as $x_t = F(x_{t-1}, \dots, x_{t-Q}, w_t)$ , where $Q$ is the window length (memory) and $w_t$ is exogenous noise.
Graph Topologies: They distinguish between two output types:
- Window Graphs: Directed graphs over nodes $x_{i,t}$ and their lags (used by PCMCI, DYNOTEARS).
- Summary Graphs: Aggregated graphs showing existence of relationships without specific lag details (used by Granger Causality, Transfer Entropy, CCM).
Hyperparameters: The study focuses on two key parameters:
1. Window Length ( $Q$ ): The number of time lags considered.
2. Sampling Rate ( $f_S$ ) / Downsampling Factor ( $k$ ): The rate at which the continuous process is observed.

B. Experimental Setup

The authors benchmarked both classical and modern causal discovery methods using a simulated bivariate system ( $x \to y$ ):

Data Generation: $x$ is a stochastic process; $y$ is generated by applying a delay filter to $x$ (approx. 50 samples delay) with added autoregressive noise.
Algorithms Tested:
- Classical: Granger Causality (GC) with variance reduction and F-statistic, Transfer Entropy (TE), Convergent Cross Mapping (CCM).
- Modern Causal ML: PCMCI, DYNOTEARS, VARLiNGAM.
Evaluation Metric: F1-score (harmonic mean of Precision and Recall) comparing the inferred causal graph against the ground truth.

3. Key Contributions

Empirical Demonstration of Sensitivity: The paper provides robust empirical evidence that both classical (GC, TE) and modern (PCMCI, DYNOTEARS) methods are highly sensitive to hyperparameter selection ( $Q$ and $k$ ).
Identification of "Detection Windows":
- Window Length ( $Q$ ): Causality is only detectable when $Q$ matches the true propagation delay. If $Q$ is too small, the delay is missed; if $Q$ is too large, the signal is diluted by noise.
- Sampling Rate: Excessive downsampling (low $f_S$ ) causes performance to deteriorate significantly. The authors show that a "heuristic detection window" exists where the downsampling factor $k$ aligns the true delay with the available lags in the model.
Signal Processing Perspective: The authors argue that causal discovery should be viewed through the lens of signal processing. They propose that just as the Nyquist rate is required for signal reconstruction, a "Causal Nyquist Rate" may be required for valid causal inference, though defining this for non-band-limited systems (like high-pass filters) remains an open challenge.
Algorithm-Specific Robustness: The study highlights that the "optimal" hyperparameters are not universal; they depend heavily on the specific algorithm used.

4. Key Results

Granger Causality & Transfer Entropy:
- In the simulation with a 50-sample delay, GC and TE failed to detect causality when $Q < 50$ .
- Performance peaked when $Q \approx 50$ (matching the true delay) and degraded significantly for $Q \gg 50$ .
- TE was strictly dependent on matching the window to the propagation delay.
Downsampling Effects:
- When downsampling the data (increasing $k$ ), the performance of GC dropped unless the new sampling rate preserved the delay within the model's lag structure.
- A "detection window" was identified where the downsampling factor allowed the true delay to fall within the first few lags ( $q=1$ to $q=5$ ) of the downsampled model.
Modern ML Methods (PCMCI, DYNOTEARS, VARLiNGAM):
- These methods exhibited similar sensitivity. In the "X causes Y" scenario, F1 scores varied drastically across different combinations of $Q$ and $k$ .
- In the "X independent of Y" scenario, incorrect hyperparameter choices led to high false-positive rates (inferring causality where none existed).
Neuroscience Case Study: The paper references a real-world neuroscience study where causal strength between brain regions was only detectable when the window length matched the hypothesized 8ms biological delay (at 1ms sampling), reinforcing the simulation results.

5. Significance and Future Directions

Relevance to Healthcare AI: Since many causal discovery applications (e.g., in healthcare or biology) rely on irregular or fixed-interval recordings, the mismatch between recording intervals and biological event times is a major source of error. This work warns practitioners that "black-box" causal inference without tuning sampling parameters is unreliable.
Call to Action for Signal Processing: The authors call for the signal processing community to develop a theoretical framework for causal inference, specifically:
- Defining a Causal Nyquist Rate based on the spectral properties of causal mechanisms.
- Developing methods to handle non-band-limited causal filters (e.g., high-pass systems).
Practical Implication: Researchers cannot assume a default $Q=1$ or arbitrary sampling rates. Hyperparameter selection must be treated as a critical part of the causal discovery pipeline, potentially requiring domain knowledge about the system's time scales.

In conclusion, the paper asserts that temporal robustness is currently a weakness in causal discovery. By applying signal processing concepts, the authors demonstrate that accurate causal inference is contingent upon aligning the observation parameters (sampling rate and window length) with the physical dynamics of the underlying system.