inconsistency in computation of differential of a map between manifolds

[Andreas Hermann]

Hello,

In the following code I define a differentiable map between two manifolds, then restrict it to a subset and then compute the matrix of its differential at a point in the subset (see also the attached jpynb file).

I have noticed that when I execute this code several times by inserting the cursor and pressing Shift+Enter I do not always get the same result. Namely, the first two rows of the matrix are sometimes swapped (try it for about 10 times).

reset()
M = Manifold(2, 'M', latex_name=r'\mathcal{M}', start_index=1)
M_cartesian.<u,v> = M.chart()
U = M.open_subset('U', coord_def={M_cartesian: (v!=0, u>0)})
U_cartesian = M_cartesian.restrict(U)
U_polar.<r,th> = U.chart(r'r:(0,+oo) th:(-pi,pi):\theta')
polar_to_cartesian = U_polar.transition_map(U_cartesian, [r*sin(th), r*cos(th)])
polar_to_cartesian.set_inverse(sqrt(u*u+v*v), atan2(v,u))
N = Manifold(3, 'N', latex_name=r'\mathcal{N}', start_index=1)
N_cartesian.<x,y,z> = N.chart()
Phi = M.diff_map(N, [u,v,u^2+v^2], name='Phi', latex_name=r'\Phi')
Phi_U=Phi.restrict(U)
p=M.point((2,1), name='p')
Phi_U.differential(p).matrix()

I would like to know: Where does this inconsistency come from?

In order to answer this question the following observations might be useful:

1. If I compute the differential of the unrestricted map (i.e. Phi.differential(p).matrix()) the problem does not appear.

2. If I insert the line "Phi_U.differential_functions()" before "Phi_U.differential(p).matrix()" the problem does not appear.

3. If in Jupyter notebook I click "Kernel" then "Restart & Run All" the problem does not appear.

Although methods 2 and 3 provide a workaround, it would certainly be great to have this inconsistency cleared. I would appreciate any help.

Best,

Andreas

[Andreas Hermann]

I am sorry, but this is a stupid mistake. The function defined via set_inverse() is not the inverse function since the first entry of atan2 is by definition the y-coordinate. If one swaps u and v in atan2 then everything is fine.

[Eric Gourgoulhon]

Le mardi 27 août 2019 16:51:20 UTC+2, Andreas Hermann a écrit :

I am sorry, but this is a stupid mistake. The function defined via set_inverse() is not the inverse function since the first entry of atan2 is by definition the y-coordinate. If one swaps u and v in atan2 then everything is fine.

 To avoid such mistake, you can run set_inverse() with the option verbose=True. Then the provided inverse transform is checked. In your original example

polar_to_cartesian.set_inverse(sqrt(u*u+v*v), atan2(v,u), verbose=True)

leads to




Check of the inverse coordinate transformation:
  r == r
  th == arctan2(r*cos(th), r*sin(th))
  u == v
  v == u

from which we recognize that there is a problem, since the last two lines should be u == u and v == v.

Maybe verbose=True should be the default?

Best wishes,

Eric.

[Simon King]

Hi Eric,

On 2019-08-27, Eric Gourgoulhon <egourg...@gmail.com> wrote:
> To avoid such mistake, you can run set_inverse() with the option
> verbose=True. Then the provided inverse transform is checked.

Is the option really called "verbose=True"? Not "check=True", like in
most other cases?

Best regards,
Simon

[Eric Gourgoulhon]

Hi Simon,

Le mercredi 28 août 2019 07:07:03 UTC+2, Simon King a écrit :

Is the option really called "verbose=True"? Not "check=True", like in
most other cases?

You are perfectly right: "check" would be more appropriate than "verbose" in this context.

I'll open a ticket for this and make "check=True" the default.

Best regards,

Eric.

[Eric Gourgoulhon]

Le mercredi 28 août 2019 21:29:29 UTC+2, Eric Gourgoulhon a écrit :

You are perfectly right: "check" would be more appropriate than "verbose" in this context.

I'll open a ticket for this and make "check=True" the default.

This is now  https://trac.sagemath.org/ticket/28422.

Eric.