June 2nd: Amsterdam Causality Meeting with Saber Salehkaleybar and Junhyung Park

[p.a.boeken]

Dear all,

This is an announcement of the next Amsterdam Causality Meeting, which will take place on Tuesday June 2nd. 

Date & time: June 2nd, 14.30-17.30
Location: VU, New University Building, room NU-4B47.

Schedule:

14.30-15.30: Saber Salehkaleybar (Leiden University), Causal Inference in Linear Systems: From Acyclic Models to Feedback Dynamics
15.30-16.30: Junhyung Park (ETH Zürich), Causal Spaces: A Measure-Theoretic Axiomatisation of Causality.

16.30-17.30: Drinks at Grand Café Living

Please find the abstracts below this message.

If you're interested in this event or in the seminar series, please check our website. For announcements regarding upcoming meetings, you can also register to our Google group.

This meeting is financially supported by the ELLIS unit Amsterdam.

Best wishes,

Philip Boeken, Giovanni Cinà, Sara Magliacane, Joris Mooij and Stéphanie van der Pas


Abstracts:

Causal Inference in Linear Systems: From Acyclic Models to Feedback Dynamics

Saber Salehkaleybar (Leiden University)

Linear systems are prevalent across science and engineering, and often provide interpretable first-order models of more complex phenomena. In this talk, I will present recent identifiability results for causal inference in linear systems, moving from acyclic structural causal models (SCMs) to cyclic ones and continuous-time stochastic dynamics. I will first discuss how non-Gaussianity and higher-order moments enable causal effect identification in acyclic linear SCMs with latent confounding, including proxy, instrumental-variable, and heterogeneous-environment settings. I will then turn to feedback systems. Cyclic linear SCMs provide one way to model feedback, and I will discuss distributional equivalence in cyclic linear non-Gaussian models. Finally, I will consider linear stochastic differential equations as another framework for modeling feedback through continuous-time interactions. When only the stationary distribution is observed, exact causal-effect strengths are generally not identifiable without additional assumptions. I will instead focus on edge-sign identifiability, which asks when the sign of a direct causal effect is uniquely determined by the stationary covariance matrix. I will present criteria for determining whether an edge sign is identifiable, non-identifiable, or partially identifiable.

Causal Spaces: A Measure-Theoretic Axiomatisation of Causality.

Junhyung Park (ETH Zürich)

Mathematical theories are founded upon simple axioms that isolate the essential structure of an object or a concept. Abstract algebra starts with the group axioms, linear algebra the vector space axioms, and topology the axioms of open sets. Most relevantly, Kolmogorov's axioms of probabilities based on measure theory have established probability theory as a branch of pure mathematics in the last century. The late 20th century saw the introduction of two hugely influential mathematical frameworks of causality, namely, the structural causal models of Judea Pearl and the potential outcomes of Donald Rubin. Their value, especially in causal inference from data, have been proved beyond doubt. However, these, and other frameworks of causality, were not designed as minimal axiomatic foundations, in the spirit of the above mathematical theories. To fill this gap, we introduce causal spaces, with two minimal axioms capturing the essence of interventions, namely, (i) doing nothing changes nothing (trivial intervention), and (ii) when X is given value x, X has value x (interventional determinism). We show that the prior formalisms can be recovered as special cases of causal spaces, and that a rich and fruitful mathematical theory, with clear causal semantics, can arise from these minimal axioms.