As reported in gh-10722, solve_triangular can be quite slow. Replacing
with direct call to dtrsv is faster.

…tran order

…asis

…efficiently

…dancies

…B in Fortran order"

This reverts commit f065d85 to undo
changes to doc/scipy-sphinx-theme.

…an order

This commit adds the tests f07vef and v07vsf from the NAG examples for
real and complex cases respectively. These are available at:

    * https://www.nag.com/numeric/fl/nagdoc_latest/html/f07/f07vef.html
    * https://www.nag.com/numeric/fl/nagdoc_latest/html/f07/f07vsf.html

The function being tested still takes the input matrix `ab` not `a` as
prescribed in #gh-11190.

This commit adds a Python wrapper for the set of LAPACK functions
?tbtrs. This solves a triangular system of the form:

    A * X = B or A**T * X = B,

The signature differs from the signature proposed in gh-11190 as the
matrix `a` is expected, rather than `ab`. This was chosen to follow the
documented signature used in LAPACK.

The test tests whether the system Ax = b is calculated correctly for
each of float, double, complex float and complex double type values.
Previously the value of b was constrained to being a real value inside a
complex datatype (i.e. a + 0j). This has now been fixed such that b may
cover all potential complex numbers.

This commit adds test for the transpose and conjugate transpose
arguments these correspond to the systems of:

    A.T x = b and A.H x = b

respectively. Some complexity was added in the parameterization as the
complex conjugate does not exist for real data types (float32 and
float64).

?tbtrs checks for singular matrices by identifying any zero elements in
the diagonal. If the i-th diagonal element of A is singular then `info`
should be equal to i. Note that this uses Fortran indexing.

This commit tests the precondition that the leading dimension of b (ldb)
is greater than the minimum allowed order of matrix A (n + 1).

This PR changes definitions of the leading dimensions to

``MAX(1, shape(?, n)``

This prevents segfaults in cases where the leading
dimension is 0.

This kind of setting would be used a lot in numerical simulations, like numerical PDE, I was confused a bit on how to get it, so maybe someone else would be confused too. It would be nice to have such an example for them so they can easily plug in and run.

Critical sections (indptr, indices and data array creations) are now
implemented in Cython.

…S tests.

Added a tutorial.  Reference kstwo from multiple locations.
Added default shape parameters for testing, and added to list of fails_cmplx.

…thod.

…_fit.py.

Change stats.kstwo's momtype from the default of 1, which uses self.ppf(),
to 0, which uses self.pdf() when computing the distribution's moments.
Added kolmogni(n, q) in _ksstats.py to calculate the ppf/isf for kstwo.

Fixed a general bug in rv_continuous._ppf_single(): If self.a is 0, it
incorrectly used -10 as the left-hand endpoint of the seach interval passed to
optimize.brentq.
If a distribution is excluded from the basic entropy calculation in
test_continuous_basic.py, also exclude it from the private_entropy, vec_entropy
and entropy_vect_scale entropy calculations.

Lower n=100 to n=10 in the example parameters for kstwo in _distr_params.py.
Inside TestKSTwo:
- Add test_isf_of_sf() and test_isf_of_sf_sqrtn() methods.
- Keep CDFs and SFs away from 1 when doing round-trip cdf/ppf sf/isf tests, as
these are subject to more error due to the extreme flatness of the function at
the end-points. (Computing the sf(isf(n, 1-p, n), n) for a CDF probability p
close to 1 works fine, as does cdf(ppf(1-p, n), n).)

All this reduces the computation time in in the kstwo tests by a factor of ~10.

A commit message had been inserted as a comment during rebasing. Now removed.

The table is from "Computing the Two-Sided Kolmogorov-Smirnov Distribution"
by Simard & L'Ecuyer, with values computed using SageMath 500-bit arithmetic.

…entropy.)

…rst.

Removed special case handling from _kolmogn_DMTW as handled in _kolmogn.
Simplified _kolmogn_PelzGood as subtractive cancellation not an issue for the
 extra terms in K2, K3.
Added _log_nfactorial_div_n_pow_n(n) to compute log(n!/n**n).
Removed unneeded comments and/or code notived by reviewer.
Removed _kolmogn_LCK as not used.
Made some function docstrings compliant.

5 unused imports, 5 unused assignments;
7 in test code, 3 in main library.

Replaced "_ = func(...)"  with  "func(...)"

…nterpolate.

…pecial.

…ial.

MAINT: Fixed several unused imports and unused assignments in scipy/special

MAINT: Fixed several unused imports in scipy/fft.

MAINT: Removed unused imports; fixed unused assignments in scipy.sparse.

This was needed when _argcheck could set the .a and .b attributes. Since it
no longer does, the call is not needed. The attributes are set in _get_support
which is, literally, the next line.

MAINT: Removed unused imports; fixed unused assignments in scipy/spatial

MAINT: stats: remove an _argcheck call

This test previously failed due to a numpy issue (see gh-9617). This fix
had been implemented in numpy version 1.4, which is now the miniumum
version used in SciPy.

This test could not be run previously due to a numpy bug (see gh-9617)
as the minimum numpy version (1.14) required by SciPy now includes the
fix, we are able to add the requried test.

The term "matrix" is ambiguous and can refer to either a two-dimensional
array or a numpy `matrix` object (which is deprecated). The changed
error message can confuse users to think they need the latter, when the
former will do.

I also added more detail about the inputs provided by the user, as
additional feedback when things go wrong.

ENH: optimize: dense redundancy removal routine optimizations

ENH: linalg: Python wrapper for ?tbtrs

MAINT: Removed unused imports; fixed unused assignments in scipy/stats.

MAINT: Small change to `optimize.nnls` error messages.

MAINT: Fixed several unused imports and unused assignments from linalg

DOC: sparse: An ex. for creating arbitrary size tri-diagonal

MAINT: Update sytrd/hetrd tests

This fixes an issue that would worst case cause infeasible
solutions to be produced by maximum_bipartite_matching as the
value of the `x`-array would end up containing edges that do
not belong to the input graph.

The root cause of the issue was a piece broken logic in the DFS
part of the Hopcroft--Karp implementation concerned with keeping
track of the augmenting path. We fix the issue by replacing
this logic with a more standard approach of marking the parents
of each vertex on the path (or rather, its grandparent, as we
still only deal explicitly with one partition at a time).

A regression test has been added to showcase an example of an
input that would previously lead to an infeasible solution.

FIX: Allow non-native byte order inputs to scipy.fft

ENH: Added kstwo probability distribution for two-sided one-sample Kolmogorov-Smirnov tests.

…11587)

BUG: Differential evolution with LinearConstraint with sparse matrix

MAINT: Replace xrange with range

BUG: fix augmentation being broken in maximum_bipartite_matching

MAINT: add rvs method to argus in scipy.stats

…10014 (#11527)

* ENH: improve docstring of weibull_min_gen and weibull_max_gen function to fix #10014.

* minor fix of docstrings.

* fix docstring based on review.

* fix docstring based on code review:

* fixed URL.

ENH: improve sparse.lil.tocsr performance

Closes gh-11058. Add `log_softmax` to `special/_logsumexp.py` and
unit tests for it to special/tests/test_log_softmax.py.

MAINT: Remove Python2 module init

MAINT: Fixed several unused imports and unused assignments in scipy/interpolate