[Date Index] [Thread Index] [Author Index]
Re: Re: Re: "Assuming"
On 26 Feb 2008, at 13:43, Daniel Lichtblau wrote: > Mariano Su=E1rez-Alvarez wrote: >> On Feb 23, 7:34 am, Daniel Lichtblau <d... at wolfram.com> wrote: >> [...] > >>> [From David Cantrell:] >>> >>>>> [...] But I do know of a case where >>>>> Mathematica goes even further, removing a singularity at which the >>>>> function >>>>> is defined as a number: >>> >>>>> In:= FullSimplify[UnitStep[-x^2]] >>>>> Out= 0 >>> >>>>> despite the fact that correctly >>> >>>>> In:= UnitStep[-x^2] /. x -> 0 >>>>> Out= 1 >>> >>>>> Perhaps the simplification above is considered a bug, perhaps not. >>> >>>>> David >>> >>> A feature, really. That is, it's wrong, but FullSimplify can make >>> mistakes on measure zero sets. We do not generally regard this >>> phenomenon as a bug, though we reconsider on case by case basis. >> >> >> How does that `measure zero' allowance work in a context >> of something like >> >> Assuming[Element[x, Integers], FullSimplify[something]] >> >> ? >> >> -- m > > I've seen cases where the FullSimplify[something] result differs from > something on a finite set of integers. This motivated me several > months > ago to alter assumptions of integrality, to reality (realness? > realhood?), in processing of Integrate. > > Daniel Lichtblau > Wolfram Research > Formally speaking, reasonable measures (e.g. Radon measures) are either diffuse or Dirac measures (or linear combinations of these). However, for the former, the entire set of integers has measure zero, and for the latter certain finite sets will have a non-zero measure. So none of these seems to fit the intended meaning of "measure zero". More seriously; I think the intended meaning is that in Simplify[thing1] -> thing2 thing1 and thing2 should both be functions of some variable that is defined on an uncountable set, then they may be are allowed to differ for a finite number of values. But, if the functions are defined only on countable sets (e.g. the set of all integers, as is the case with many number theoretic functions) then the failure of thing2 to be equal to thing1 on a finite set could be very serious. I think in such situations the "set of measure zero" should really be the empty set, or perhaps in really exceptional cases a set that contains no more than a single point. Andrzej Kozlowski