Timeline for Optimal Monte Carlo Trace Estimator
Current License: CC BY-SA 4.0
6 events
| when toggle format | what | by | license | comment | |
|---|---|---|---|---|---|
| Aug 16, 2022 at 3:52 | comment | added | Sandeep Silwal | This paper will be very interesting to you :) arxiv.org/pdf/2010.09649.pdf | |
| Aug 1, 2022 at 13:05 | history | edited | Nawaf Bou-Rabee | CC BY-SA 4.0 |
changed the title to something more effective
|
| Aug 1, 2022 at 12:58 | comment | added | Nawaf Bou-Rabee | Since $q^*$ has a density relative to a Gaussian reference measure (proportional to $e^{\log(x^T A x)}$), it seems like a preconditioned MCMC method (in the same spirit as elliptical slice sampling) might work well. The convergence of the pMCMC method will, however, depend crucially on the properties of the Hessian of $U(x) = - \log(x^T A x)$. | |
| Aug 1, 2022 at 11:40 | comment | added | Sebastian Nowozin | @NawafBou-Rabee, yes, you are correct, Hutchinson's original paper only used Rademacher vectors. The question only relates to the Gaussian trace estimator. | |
| Aug 1, 2022 at 11:25 | comment | added | Nawaf Bou-Rabee | Doesn't the standard Hutchinson trace estimator use a random vector $X$ whose components are i.i.d. Rademacher random variables? I thought the corresponding estimator based on Gaussians is called the Gaussian trace estimator. | |
| Aug 1, 2022 at 11:10 | history | asked | Sebastian Nowozin | CC BY-SA 4.0 |