Derivatives and Roundoff Errors

Event Sponsor: 
Mathematics and Computer Science Division LANS Seminar
Start Date: 
Apr 12 2017 - 3:00pm
Building 240/Room 1404-1405
Argonne National Laboratory
Paul Hovland
Speaker(s) Title: 
Argonne National Laboratory, MCS

There is a long tradition of using derivatives computed via automatic/algorithmic differentiation (AD) to estimate the impact of roundoff errors in numerical computation (going back at least to Linnainma in 1976). I’ll review some of the AD-based roundoff error estimation methods, describe in more detail the so-called CENA method due to Langlois, and discuss the merits of using forward- versus reverse-mode AD. I’ll finish up with some speculation on other uses for AD in the context of approximate and multi-precision computation.

Miscellaneous Information: 

Coffee and goodies will be served.

Please click below to add this event to your calendar.



Upcoming Seminars:
April 19, 2017, "Statistical methods for simulation and prediction of space-time geophysical data" Julie Bessac, Postdoctoral Appointee, MCS/ANL.
April 26, 2017, "Hybrid quantum-classical computation for chemistry and materials" Jarrod McClean, Alvarez Fellow, Lawrence Berkeley National Laboratory.