> On 20 Dec 2019, at 23:45, Gilles Sadowski <[email protected]> wrote: > > Le sam. 21 déc. 2019 à 00:05, Alex Herbert <[email protected] > <mailto:[email protected]>> a écrit : >> >> Looking at the c++ standard [1] we are missing this function: >> >> norm() = a^2 + b^2 >> >> The field norm of the complex, or the square of the absolute. An example on >> C++ reference states that this is faster for comparing magnitudes for >> ranking as it avoids the sqrt() required in abs(). >> >> z.abs() > y.abs() == z.norm() > y.norm() >> >> >> I suggest this is added to comply with the standard. > > +1 > >> >> >> It seems odd to me that the constructor ofPolar throws an exception. It does >> this when the magnitude (rho) for the complex is negative. However if the >> magnitude is NaN it will not throw an exception and will end up returning >> NaN. I think this should be changed to return NaN for negative magnitude and >> avoid exceptions. This is basically stating that the polar representation >> you used is invalid and so in the Cartesian representation it will be (nan, >> nan). >> >> The C++ standard on this was previously vague and was clarified in [2]: >> >> “ >> -?- Requires: rho shall be non-negative and non-NaN. theta shall be finite. >> >> -9- Returns: The complex value corresponding to a complex number whose >> magnitude is rho and whose phase angle is theta. >> “ >> >> The assumption is that abs(polar(rho, theta)) == rho. >> >> If this cannot be ensured then polar(rho, theta) is undefined and we return >> NaN. >> >> >> Note that if theta is finite and rho is non-negative and non-nan: >> >> x = rho * Math.cos(theta) >> y = rho * Math.sin(theta) >> >> In the event that sin(theta) is zero (i.e. theta is zero) then inf * 0 is >> NaN. In this case the complex could either be: >> >> (Inf, nan) or (inf, 0) >> >> I have tried 2 c++ implementations and both return (inf, nan). The c++ >> <complex> header I have found would return (inf, 0). The same header also >> corrects if cos(theta) is zero however in Java cos(pi/2) is not zero so this >> is not an issue. >> >> Note: >> If the result is (inf, nan) then abs((inf, nan)) should return inf to >> satisfy the contract abs(polar(rho, theta)) == rho. This is currently true >> as abs() uses Math.hypot(x, y) which will return positive infinity if either >> argument is infinite. So I do not think it matters. An infinite is infinite >> even when the other part is nan. >> >> >> I suggest we update ofPolar to not throw exceptions and return NAN unless >> theta is finite and rho is non-negative and non-nan. > > +0 > Not sure how useful it is to instantiate a useless object.
It would not be instantiated. It can use the singleton NaN. But I do see your point. Something has gone wrong so raise it as a problem. Looking at various implementations the polar(rho, theta) method is used when computing results in polar coordinates to then return a result. For example this is used in various sqrt() implementations as the equivalent of: Complex.ofPolar(Math.sqrt(z.abs()), z.arg() / 2) In this case it would not be nice to throw an exception for a complex z which is NaN. Instead return NaN as the sqrt of NaN. I thought it would be more in line with the c++ standard to not throw exceptions. I also think that returning NaN is more in line with how java.util.Math handles invalid cases. > > Regards, > Gilles > >> >> Alex >> >> >> [1] http://www.cplusplus.com/reference/complex/ >> <http://www.cplusplus.com/reference/complex/> >> <http://www.cplusplus.com/reference/complex/ >> <http://www.cplusplus.com/reference/complex/>> >> [2] https://cplusplus.github.io/LWG/issue2459 >> <https://cplusplus.github.io/LWG/issue2459> >> <https://cplusplus.github.io/LWG/issue2459 >> <https://cplusplus.github.io/LWG/issue2459>> >> >> > > --------------------------------------------------------------------- > To unsubscribe, e-mail: [email protected] > <mailto:[email protected]> > For additional commands, e-mail: [email protected] > <mailto:[email protected]>
