Change default of repeat detection tolerance in KTrajectory

[fzimmermann89]

mrpro/src/mrpro/data/KTrajectory.py

Line 72 in 8d24ebb

repeat_detection_tolerance: float | None = 1e-8,

What does it mean to have a trajectory where all the kz's only differ by less than 1?

I think we set the default when our trajectories could also be scaled to -pi..pi
If I am not mistaken, they are now always scaled to the encoding matrix size, correct?

So we could increase this to 1.0?

[fzimmermann89]

@ckolbPTB @schuenke Or am i wrong with the scaling?

[schuenke]

hmm... close to 1 probably. It might be that we expect a step of 1, but, due to rounding etc, only get 0.99. This should not be treated as "no step". So something > 0.5 and < 1.0 IMO.

[fzimmermann89]

Is it more likely if we have a Trajektory with two values (-0.499,0.499) in kz that it really has 2 different values, or that it is a scaling issue?

[fzimmermann89]

this is not the step, right?
isnt the logic if (max(kz)-min(kz)) < tolerance assume singleton

[ckolbPTB]

To me the detection tolerance should be related to numerical accuracy and rounding errors so if the step we expect is in the order of 1 then a default detection tolerance of 0.01 or 0.1 seems reasonable. By setting it too high I am afraid that we mask errors somewhere else in the code (as Patrick's example showed).

[fzimmermann89]

There really is no reason to have anything less than 1 there.
If a dimensions does not at least have a difference of 1 between min and max, it must be singleton.
This is IMHO the better fix than conitionally rounding depending on the encoding matrix size, as done in #551.

The Fourier Op will not make any sense with steps less than 1 between min and max -- what should the result be?
Afterthing that is less than one is numerical artifact. So we can ignore it.

[fzimmermann89]

And just removing these values would make issues even more obvious -- wrong scaling of a trajectory calculator, i.e. scaling that is incompatible with FourierOp (which assumes -enc_matrix//2 .. enc_matrix//2 scaling) would then result in a singleton Trajectory --printing it out would makeit really obvious that the trajectory is wrong.