Consider $X, Y$ and suppose you have some i.i.d. observations of $Y$ and $X + \text{do}(Y)$ with observation noise $\epsilon_0$ and $\epsilon_1$ (Gaussian). So the observations of the latter should not influence out estimate of $Y$.

Is there some nice way of expressing this when doing the analysis (I've made this problem very simple, but consider doing regression or bayesian updates).

So you have some log-likelihood like this (for example)

$$ J = \frac{(\hat{Z}_0 - Y)^2}{\epsilon_0^2} + \frac{(\hat{Z}_1 - \text{do}(Y) - X)^2}{\epsilon_1^2} $$

Are there any tricks to using the DAG matrix to effectively do a stop gradient?

This post seems related? Mathematical notation for suppressing differentiation

Another way of asking this question is suppose you have a standard algorithm for dealing with the non-causal situation here (consider a filter or online mean) is there a simple way besides recursively solving the system, to change that solution into a causal one?

  • $\begingroup$ What are you trying to do with the log-likelihood? As in, what does the model you have in mind currently do, and what should it be able to do after making it "causal"? $\endgroup$
    – Scriddie
    Jun 6, 2023 at 12:48


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Browse other questions tagged or ask your own question.