In Understanding Black-box Predictions via Influence Functions paper Appendix A, the authors provide a standard derivation for influence functions, however, I could not understand one of the steps. They mentioned that from the step $(12)$

$$\Delta_{\epsilon} \approx -[\nabla^2R(\hat{\theta}) + \epsilon\nabla^2L(z, \hat{\theta})]^{-1}[\nabla R(\hat{\theta}) + \epsilon\nabla L(z, \hat{\theta})]$$

to step $(13)$

$$ \Delta_\epsilon \approx -\nabla^2R(\hat{\theta})\nabla L(z, \hat{\theta})\epsilon $$

by dropping $o(\epsilon)$ terms. Also, we have that $\nabla R(\hat{\theta}) = 0$. Therefore, I could not get what it meant when they mentioned dropping $o(\epsilon)$ terms.

  • $\begingroup$ Please clarify your specific problem or provide additional details to highlight exactly what you need. As it's currently written, it's hard to tell exactly what you're asking. $\endgroup$
    – Community Bot
    Commented Oct 11, 2023 at 16:44


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.