Refactor functions in matrixtools.py #442
Conversation
| if not is_hermitian(mx, tol): | ||
| return False | ||
| if attempt_cholesky: | ||
| try: | ||
| _ = _spl.cholesky(mx) | ||
| return True # Cholesky succeeded | ||
| except _spl.LinAlgError: | ||
| pass # we fall back on eigenvalue decomposition | ||
| evals = _np.linalg.eigvalsh(mx) |
There was a problem hiding this comment.
The property of being positive semidefinite is generally ascribed to Hermitian matrices, since the value of the quadratic form
The new implementation starts with a check that the matrix is Hermitian. If we do have a Hermitian matrix then the fastest way to check if it's positive definite is a Cholesky decomposition. However, Cholesky will fail on matrices that are only positive semidefinite, so there is no point in trying Cholesky if we expect is_pos_def will regularly handle singular matrices. I've set the default behavior for attempt_cholesky to False. If we skip Cholesky or if Cholesky fails then we use Hermitian eigendecomposition.
There was a problem hiding this comment.
A couple of things:
- Should we ensure the docstring is consistent? It says PD in the first line, but return says PSD.
- Both implementations are clearly for PSD. Should the function name be
is_pos_semidefinstead? Pro is that it is more accurate, con is that it is an API change. While this is in matrixtools and therefore could have moderate/high user exposure, it seems like the only use is inis_valid_density_mxbelow (which also seems like it has low usage), so maybe the API change would go mostly unnoticed?
There was a problem hiding this comment.
In the context of quantum information I find people are lax about positive definite versus positive semidefinite, so I think the function name is fine. But I'll definitely change the docstring to be consistent about really checking that a matrix is PSD.
| def remove_dependent_cols(mx, tol=1e-7): | ||
| """ | ||
| Removes the linearly dependent columns of a matrix. | ||
|
|
||
| Parameters | ||
| ---------- | ||
| mx : numpy.ndarray | ||
| The input matrix | ||
|
|
||
| Returns | ||
| ------- | ||
| A linearly independent subset of the columns of `mx`. | ||
| """ | ||
| last_rank = 0; cols_to_remove = [] | ||
| for j in range(mx.shape[1]): | ||
| rnk = _np.linalg.matrix_rank(mx[:, 0:j + 1], tol) | ||
| if rnk == last_rank: | ||
| cols_to_remove.append(j) | ||
| else: | ||
| last_rank = rnk | ||
| #print("Removing %d cols" % len(cols_to_remove)) | ||
| return _np.delete(mx, cols_to_remove, axis=1) | ||
|
|
||
|
|
There was a problem hiding this comment.
This was called in one function (below), which I reimplemented to rely on independent_columns.
rileyjmurray
changed the title
Get the latest fixes for the ComplementPOVMEffect.to_vector() calls.
Forgot to pull latest develop before attempting the previous merge.
There was a problem hiding this comment.
Thanks for the great effort @rileyjmurray, as always! Also, I greatly appreciate your preemptive comments on some of the more technical changes, it makes it much easier to follow. :)
There is only one real action item here, for an additional unit test for the attempt_cholesky codepath in is_pos_def (and a small docstring update). However, I've also made some non-blocking comments/questions. I 👍'd any of your comments that I read and agreed with but had no active comments or action items for.
| # I wanted to remove the implementations of vec and unvec and just in-line equivalent | ||
| # code in the few places they were used. However, the fact that they're included in this | ||
| # __init__.py file suggests that they might be used outside of pyGSTi itself. | ||
| # |
There was a problem hiding this comment.
We can check with @kevincyoung and @bruzic3, but I think I'm OK with inlining these in our source, deprecating the functions in pyGSTi, and then assisting/enforcing the relevant changes in the downstream code so that they can be removed if they are inefficient/not best practice.
| if not is_hermitian(mx, tol): | ||
| return False | ||
| if attempt_cholesky: | ||
| try: | ||
| _ = _spl.cholesky(mx) | ||
| return True # Cholesky succeeded | ||
| except _spl.LinAlgError: | ||
| pass # we fall back on eigenvalue decomposition | ||
| evals = _np.linalg.eigvalsh(mx) |
There was a problem hiding this comment.
A couple of things:
- Should we ensure the docstring is consistent? It says PD in the first line, but return says PSD.
- Both implementations are clearly for PSD. Should the function name be
is_pos_semidefinstead? Pro is that it is more accurate, con is that it is an API change. While this is in matrixtools and therefore could have moderate/high user exposure, it seems like the only use is inis_valid_density_mxbelow (which also seems like it has low usage), so maybe the API change would go mostly unnoticed?
| @@ -1372,6 +1424,9 @@ def _findx(a, inds, always_copy=False): | |||
| return a_inds | |||
|
|
|||
|
|
|||
| # TODO: reevaluate the need for this function. It seems like we could just in-line @ | |||
There was a problem hiding this comment.
Generally in favor of this. Not a requirement for approval for this PR though, can punt to future.
There was a problem hiding this comment.
Cool. I'll try to make this change then.
This PR resolves #429. I've provided detailed comments explaining my changes to
matrixtools.py. I've also removed some commented-out code.