Skip to content

More Rigorous Error Analysis #15

Description

@emerali

Vanilla Jackknife/Bootstrap assume IID samples, and thus underestimate standard errors (this is true even for simple estimators which are just means). There are Bootstrap methods specifically for time-series data which could be useful (e.g. Stationary Bootstrap).
There may be a way to turn one of these into a "streaming" bootstrap

Also: autocorrelation time we're computing seems to be larger than what's computed by other packages. Should try calculating autocorrelation time using a binning analysis instead as it may be more "stable" than the FFT based method we're using.

Should compute autocorrelation time for all observables in order to get accurate error bars.

Useful Papers:

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions