… when covariance is singular (scipy#5288)

Co-authored-by: Matt Haberland <mhaberla@calpoly.edu>
Co-authored-by: Tirth Patel <tirthasheshpatel@gmail.com>

…uld print convergence info from final iteration. (scipy#13322)

* BUG: optimize: fixed 13321
A bug in the way the `iprint` and `disp` arguments were handled
made it impossible to run the code with `iprint=0`, since it was
overwritten from the `disp` handling.

Co-authored-by: Matt Haberland <mhaberla@calpoly.edu>

* BUG: stats: Reformulate loggamma._rvs to handle c << 1.

Closes scipygh-11094.

Several tests in test_morestats.py call loggamma.rvs to generate
test data.  These tests have hardcoded expected results that
depend on the values in the input data.  Because the stream
of variates generated by loggamma.rvs has changed, those tests
will fail if the input data is generated by loggamma.rvs.  Instead,
the data that was generated by a new function that takes the log of random variates from `stats.gamma.rvs`.

Co-authored-by: Matt Haberland <mhaberla@calpoly.edu>

* BUG: fix powell evaluated outside limits

Co-authored-by: Joseph T. Iosue <jtiosue@gmail.com>

…5381)

The documentation of `rv_histogram` was ambiguous about what it expected as input when bin sizes vary: counts or density? Adding to the confusion, the default behavior of `rv_histogram` was to assume that the input was a density, whereas the default of `np.histogram` is to produce bin counts. To resolve this, this PR adds a parameter `density` so that the user can specify whether the input is to be treated as counts or probability density. Because the default is different from that of `np.histogram`, this also warns the user to pass `density` explicitly.

Co-authored-by: Matt Haberland <mhaberla@calpoly.edu>

Caused by an OOB access due to uninitialized local variables.

Fixes scipy#3691.

…ero (scipy#16460)

MAINT: stats.ttest_ind: randomized permutation pvalue should not be zero  

Co-authored-by: Matt Haberland <mhaberla@calpoly.edu>

* Fix bug scipy#14589 where arguments to objective function aren't passed properly

Co-authored-by: J. J. Ramsey <jjramsey@pobox.com>

…erpn

…rity

With numpy 1.22.x f2py, dimension of isuppz in syevr is mistranslated.
Avoid this by explicitly specifying operator priority with parenthesis.

Fixes scipy#16527

…py#16588)

Co-authored-by: Pamphile Roy <roy.pamphile@gmail.com>

to guard against /detect the integer overflow

Co-authored-by: peterbell10 <peterbell10@live.co.uk>

Also cast integers to npy_intp, because cannot use size_t in
PyArray_SimpleNew et al (it expects signed integers)

Co-authored-by: Matthew Brett <matthew.brett@gmail.com>

The file maybe used to be autogenerated, but has since been updated manually
many times over by several people on several occasions. The autogeneration
script is also long gone, too.

For completeness, the original version (or a close approximation to it)
is now available at https://github.com/matthew-brett/multipack ---
this repository content is from a Waybackmachine archive of Travis Oliphant's
website, found by Matthew Brett. The archive was trawled on September 29 2000.
(https://web.archive.org/web/20000929143650/http://oliphant.netpedia.net/packages/multipack-0.7.tar.gz)

Co-authored-by: Matthew Brett <matthew.brett@gmail.com>

Make sure that the product mx*my is <= min(max NPY_INTP, fortan int)
to avoid an overflow on both C and Fortran sides on 32- and 64-bit
systems with and without HAVE_ILP64 (which make Fortran ints 32- or 64-bit, too)

When a sparse matrix was indexed with an argument that was effectively
empty (e.g. `0:0:2`, `[]`, `[False, False]`), the dtype of the matrix
was lost and the result became float64.

Closes scipygh-16656.

…a} (scipy#16701)

* MAINT: stats.boltzmann: correct _shape_info typo

…0) and k=1 (scipy#16780)

Co-authored-by: Jake Bowhay <60778417+j-bowhay@users.noreply.github.com>

…py#16872)

Make sure that `nuest` parameter is large enough when `s=0` and
`pole_continuity=(True, True)`. This value translates to
`iopt(2) = iopt(3) = 1` in the FITPACK routine `spgrid.f`.
This routine contains the following comment:

c  nuest : integer. unchanged on exit.
c  nvest : integer. unchanged on exit.
c          on entry, nuest and nvest must specify an upper bound for the
c          number of knots required in the u- and v-directions respect.
c          these numbers will also determine the storage space needed by
c          the routine. nuest >= 8, nvest >= 8.
c          in most practical situation nuest = mu/2, nvest=mv/2, will
c          be sufficient. always large enough are nuest=mu+6+iopt(2)+
c          iopt(3), nvest = mv+7, the number of knots needed for
c          interpolation (s=0).

Therefore, we set in the f2py wrappers `nuest = mu+6+2`, which is
the maximum possible value, since each `iopt` array element is at most
one.

The test of the _compute_lwork() work-around for a bug in
LAPACK versions older than 3.10.1 is simplified.  Instead
of calling svd() in a subprocess with a large input array
and checking that it doesn't crash Python, we check that
the function sgesdd_lwork() returns the correct value.  It
is the incorrect value that resulted in the Python crash,
so this is a simpler test of the same issue.

[skip ci]

Change `lamv` implementation to use intended branching for assignment
to variable `cs` and avoid useless assignment to implicitly typed
variable `elsecs`.

See scipy#17104
See lfortran/lfortran#779

Signed-off-by: Sebastian Ehlert <28669218+awvwgk@users.noreply.github.com>

[skip ci]

When SCIPY_XSLOW is defined, the tests `test_riemann_zeta` and `test_zetac`
(that compare the results of `zeta` and `zetac` to the results computed
with `mpmath`) are run with 5000 points, and it turns out that the input
value 1 is included in that case.  The SciPy functions returns `inf`, but
`mpmath.zeta(1)` raises an exception, and that caused the tests to fail.

The fix is to wrap the `mpmath` function to return `mpmath.inf` when the
input is 1.  After fixing that, both tests then failed because some inputs
result in a relative error of roughly 2e-13, so I bumped the tolerance up
to 5e-13.

[ci skip]

Co-authored-by: GavinZhang <zhanggan@cn.ibm.com>
Co-authored-by: Ralf Gommers <ralf.gommers@gmail.com>

All could using `#include <thread>` needs to have this `thread_dep`
dependency. HiGHS code was missing it, and for `stats/_qmc_cy.pyx`
it was incorrectly specified.

It's wrong in the distutils build as well, but let's ignore that
since it's much harder to change and we'll get rid of that build
soon.

See scipygh-17193 for the type of compile error this causes.

[ci skip]

…escription (scipy#17204)

DOC: stats.mode: add versionadded tag and correct order of keepdims description

Shebangs do not work this way, but Meson could usually paper over the
differences by noticing that the last element is "python3" and rewiring
it to use its own sys.executable. However, this failed on WSL, likely
because platform quirks meant the script was directly executable, while
simultaneously having a shebang that was a fatal error.

Use a standard shebang indicating this script would like to run with
some form of python3 (which is exactly what it needs).

Fixes scipy#17020

* fix missing import from cherry-picked patch

After the change in numpy/numpy#22385, numpy
raises a deprecation warning with calls such as np.int8(5000) and
np.uint32(-1).  This change avoids such calls in the tests.