Causal Discovery with Instantaneous Temporal Effects

415 concepts211 papers1 proposal

In the study of dynamic systems, the "gold standard" for inferring causality is temporal precedence: a cause must happen before its effect. This principle underlies Granger causality and most time-series models, which assume that an influence requires at least one discrete time step to manifest. However, this assumption breaks down when the speed of the underlying physical process outpaces the sampling rate of our sensors.

When data is collected at a coarse resolution—such as a camera capturing 30 frames per second for a high-speed collision, or a monthly economic report tracking daily market shifts—multiple causal events appear to occur simultaneously. This phenomenon, often called temporal aliasing or temporal aggregation bias, creates "instantaneous" or "contemporaneous" effects that baffle traditional algorithms.

The Sampling Rate Gap

The core challenge is a mismatch between the causal frequency (how fast the system evolves) and the sampling frequency (how often we observe it). As Sergey Plis and David Danks argue in Rate-Agnostic (Causal) Structure Learning, undersampling leads to the identification of spurious causal links or the complete omission of genuine dependencies.

If a cause and effect are bundled into the same observation window, the temporal arrow disappears. Phillip Lippe et al. describe this in iCITRIS: Causal Representation Learning for Instantaneous Temporal Effects as a "chicken-and-egg" problem: without a measurable time lag, it is theoretically impossible to distinguish whether $X$ causes $Y$, $Y$ causes $X$, or a hidden "confounder" causes both.

Recovering Directionality Without Lags

To overcome the loss of temporal ordering, researchers have moved beyond simple time-lags to exploit other statistical and structural signatures:

  • Distribution Shifts and Non-stationarity: Weiran Yao et al. demonstrate in Temporally Disentangled Representation Learning that natural changes in data distributions over time can be used to identify latent causal variables and their instantaneous relationships. Similarly, Niklas Pfister et al. leverage the principle of invariance—the idea that a causal mechanism remains stable even when the environment changes—to detect instantaneous effects with statistical confidence in Invariant Causal Prediction for Sequential Data.
  • Interventional Cues: When observational data is insufficient, interventions (like an agent's actions in a video) can break the symmetry between simultaneous variables. Phillip Lippe et al. show that knowing which variable was targeted by an intervention allows a model to orient instantaneous edges that would otherwise be ambiguous.
  • Statistical Signatures of Simultaneity: Surprisingly, the presence of simultaneous events can sometimes improve identifiability. Jie Qiao argues in Structural Hawkes Processes for Learning Causal Structure from Discrete-Time Event Sequences that the specific patterns of co-occurrence in low-resolution data provide unique statistical fingerprints that help recover the true causal graph.

Beyond Summary Graphs

Traditional time-series analysis often uses "summary graphs," which collapse all temporal information into a single edge. However, as Charles K. Assaad et al. point out in Discovery of extended summary graphs in time series, these graphs fail to distinguish between a relationship that happens within a time step and one that persists across them.

By explicitly modeling contemporaneous edges, frameworks like those proposed by Daniel Malinsky and Peter Spirtes become robust to varying sampling rates. This is critical in fields like neuroimaging; Yongjie Zhu et al. show in Second-order instantaneous causal analysis of spontaneous MEG that using second-order statistics (autocorrelations) can reliably infer causal directions in brain activity even when the sampling rate is too slow to capture the physical propagation of signals.

Why It Matters

Ignoring instantaneous effects isn't just a theoretical oversight; it leads to model misspecification. In high-speed dynamical systems—from robotics to high-frequency trading—the assumption of a strictly positive temporal delay is often a fiction of the measurement device. By developing models that can "see through" the temporal blurring of low-resolution data, researchers are creating world models that are more robust to the limitations of their own sensors.

History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_temporal_resolution_mismatch_4dfcbea7,
  title = {temporal resolution mismatch},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/temporal-resolution-mismatch/rdr_4dfcbea73ed04ec0ac6573e055245dfa},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_temporal_resolution_mismatch_4dfcbea7_v1,
  title = {temporal resolution mismatch},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/temporal-resolution-mismatch/rdr_4dfcbea73ed04ec0ac6573e055245dfa/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake