Worst case in interference analysis for medical interference

In its second order on reconsideration regarding the incentive auction released on June 19, 2015 (docket 12-268, pdf) the FCC noted that its analysis interference into wireless medical telemetry systems its work "is a worst case analysis and in most installations one or more of the parameters we assumed here will provide additional protection" (recon order at para 119).


Even this wasn't good enough for GE Healthcare, who filed a petition on July 28, 2015 asking the FCC to reconsider its reconsideration, saying "Due to the severe and wide-ranging negative consequences of interference to Channel 37 WMTS, the Commission's expressed intent to use a worst-case (i.e. minimum coupling loss) analysis in evaluating separation between Channel 37 WMTS and 600 MHz band mobile base stations is appropriate, but its adopted separation rules are not, in fact, based on a worst-case analysis, as the Commission appears to believe."

The trouble with worst case is that there is no worst case. That is: one can always imagine something worse. It’s not a sufficiently stable concept to be usable. This leads to oxymorons like the “realistic worst-case” GE HealthCare refers to in its petition. It leads to oxymorons like the “realistic worst-case” GE HealthCare refers to.

There’s even a term of art: RWCS, the Reasonable Worst Case Scenario, that even has a definition in the UK: a scenario "designed to exclude theoretically possible scenarios which have so little probability of occurring that planning for them would be likely to lead to disproportionate use of resources” (House of Commons Select Committee Report). (There's also the term "reasonably foreseeable worst case use scenarios" that's used in passing in IEC 60601.)

It’s related to the unbounded character of the maximum of a distribution.  It’s well known among statisticians, but apparently not that many spectrum folk, that the longer you sample a parameter with an unbounded distribution (e.g. a propagation path loss with a log-normal fading distribution), the larger the maximum you find will be.