Recently I came across the terms signal efficiency and background rejection (for example in this paper), but I couldn't find an understandable definitions to those terms and how they are used. Specifically, I couldn't find a mathematical definition.

I'd appreciate an explanation with an example or a reference.

  • $\begingroup$ Would Signal Processing be a better home for this question? $\endgroup$ – Emilio Pisanty Jul 26 '17 at 15:37
  • $\begingroup$ @EmilioPisanty Probably not. Those are (loosely defined) terms of art in data analysis. Arguable Cross Validated would be a good home as they should apply to any analysis, but I was intending to answer from a particle physics prospective. $\endgroup$ – dmckee --- ex-moderator kitten Jul 26 '17 at 15:48

This answer is framed in terms of particle physics, but the basic idea can be generalized to any analysis where the presence or absence of the thing you are trying to study must be deduced from the raw data.

Many particle physics detectors collect data in a very general sense, and to study something in particular you need to select from all the myriad recorded events that those that correspond to the physics you want to study.

That is, you are going to partition the whole data set into those events that you believe represent your physics and those that do not. Obviously you can make two classes of errors in that partition:1

  • You can include events in your study group that represent some physics other that what you wanted (false positive)
  • You can leave some even that does represent your physics out of your study group (false negative)

The term signal efficiency is usually defined as the fraction of the desired events that you actually get (i.e. it goes down as the false negative rate goes up). The term background rejection is usually defined as the fraction of events that don't belong in your sample that are excluded from your sample (i.e. it goes down as the false positive rate goes up).

1 And it is generally the case that the harder you work to prevent one class of errors the more of the other class you are going to get, so there is a trade-off here.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.