Apparently, the totalOrder predicate from IEEE 754-2008 specifies that -NaN < -Inf < negative finite numbers < -0 < +0 < positive finite numbers < +Inf < +NaN.
So if floating point values are allowed, then I think that "-Nan" would be the obvious starting point ... coz you can't not get no lower than that ;-)
(Hopefully, indexes have to be real integer values.)