Handling of negative mass2/tau2 and boosts: Add warnings or exceptions?

[meliache]

I reconstruct a virtual W four-vector from the difference of reconstructed Bsig and D* four-vectors and was surprised that sometimes, the result of the mass2 method is negative.

p = vector.obj(px=-1.1764122315216456, py=0.7286259979009628, pz=1.8719534137922902, E=2.3187584307488134)
print(f"{p.mass=}\n{p.mass2=}\n{p.mass**2=}")

which prints

p.mass=-0.2059381133516135
p.mass2=-0.04241050653082201
p.mass**2=0.042410506530822005

Well, the reconstructed vector is obviously unphysical and when I looked at the code I saw that mass2 is an alias for tau2 and it just returns just E**2 - mag**2, so this makes sense when the vector is unphysical and thus the magnitude is higher than the energy. I just wrongly assumed it would return the same as mass**2 and at first I thought this breaks the principle of least surprise.

I looked at ROOT GenVector, which you try to be convention-compatible with, and it seems it does the same for M2() , so it seems this is not a bug but a feature and for ROOT users it keeps in line with expectations.. But GenVector seems to throw an exception when e.g. calculating Beta() with a negative M2(), while vector just quietly returns NaN's.

Suggestion:
I think it would be nice to throw some exception when trying to boost with a time-like vector or at least print some warnings like numpy does when dividing by zero. And maybe have an optional keyword argument to disable exceptions. Then, the user would need to be more explicit to ignore these cases. Though I'm not sure if that's feasible from a design-perspective, maybe just returning NaN's is enough 🤷

[jpivarski]

Considering that these are being applied to arrays of vectors, rather than an individual vector, raising exceptions is more problematic than setting NaN. (What if a huge dataset has just one such value, due to accumulated round-off?) That's why NumPy sets NaN—but with warnings (by default).

Vector suppresses these warnings because some of them are intermediate to the calculation:

vector/src/vector/_compute/lorentz/beta.py

Lines 122 to 133 in 0b9bf7e

	with numpy.errstate(all="ignore"):
	return v._wrap_result(
	_flavor_of(v),
	function(
	v.lib,
	*v.azimuthal.elements,
	*v.longitudinal.elements,
	*v.temporal.elements
	),
	returns,
	1,
	)

But perhaps there should be a vector.errstate that does the same thing as NumPy's, but only when final results contain NaNs? Just deciding whether there are any NaNs in the output would require an additional pass over the data, with np.isnan.