This is a transcript of the most recent discussions with @blakelyd and @rayshrish about the logic of injection recovery. We need to implement these methods to be able to do the nondetections both of binaries and disks.
False Negative Rate
simulate binaries from a grid of parameter values in sep, theta, con
1k noise realizations
at true params for each, calculate Laplace evidence
at each separation, report azimuthally averaged contrast at which false negative rate is 99% etc
by linearity this should be scalable like Absil but think about this
False Positive Rate
Draw 1k noise realizations
for each, fit binary in grid using vmap max likelihood
evaluate evidence of binary
as a function of evidence, what is the global false positive rate?
ROC Curve
Using false positive and false negative calculations above, we should be able to calibrate the model evidence threshold in each parameter bin that gives us desired false pos / neg
Contrast Limit (Ruffio)
in each sep, theta cell calculate optimum companion flux
find best Gaussian fit (Hessian), truncate to positive companion flux
This is a transcript of the most recent discussions with @blakelyd and @rayshrish about the logic of injection recovery. We need to implement these methods to be able to do the nondetections both of binaries and disks.
False Negative Rate
False Positive Rate
ROC Curve
Contrast Limit (Ruffio)