 |
Universität Augsburg
|
|
|
Universität Augsburg
|
|
Herr Matthias Schneider
spricht am
Dienstag, 31. März 2026
um
10:00 Uhr
im
Raum 1009 (L1)
über das Thema:
| Abstract: |
| This thesis addresses the problem of parameter estimation including causal direction identification in statistical graphical models based on directed acyclic graphs (DAGs). Traditional approaches based on covariance matrices suffer from a fundamental limitation known as covariance equivalence. This limitation makes it impossible to determine causal directions using only observational data. To address this issue, this thesis introduces an alternative approach based on the empirical third moment tensor. It further assumes nonzero skewness in the noise of the underlying model. This work shows that the symmetric CP decomposition of the third moment tensor uniquely identifies the causal coefficient matrix up to permutation and scaling of its rows. Furthermore, this thesis introduces a theoretically grounded, stable algorithm to recover the true permutation and scaling. Finally, numerical experiments validate the proposed approach by analyzing the stability of the algorithm and the accuracy of the estimated causal coefficient matrix as functions of the sample size and other parameters such as the dynamic range of the coefficients and the scaling of the noise. |
| Hierzu ergeht herzliche Einladung. |
| Prof. Dr. André Uschmajew |