Hello community,
here is the log from the commit of package python-uncertainties for
openSUSE:Factory checked in at 2019-06-03 18:57:34
++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
Comparing /work/SRC/openSUSE:Factory/python-uncertainties (Old)
and /work/SRC/openSUSE:Factory/.python-uncertainties.new.5148 (New)
++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
Package is "python-uncertainties"
Mon Jun 3 18:57:34 2019 rev:5 rq:707107 version:3.1.1
Changes:
--------
---
/work/SRC/openSUSE:Factory/python-uncertainties/python-uncertainties.changes
2018-10-31 13:20:39.835100971 +0100
+++
/work/SRC/openSUSE:Factory/.python-uncertainties.new.5148/python-uncertainties.changes
2019-06-03 18:57:36.296379125 +0200
@@ -1,0 +2,6 @@
+Mon Jun 3 09:24:16 UTC 2019 - Tomáš Chvátal <[email protected]>
+
+- Update to 3.1.1:
+ * No upstream changelog provided
+
+-------------------------------------------------------------------
Old:
----
uncertainties-3.0.3.tar.gz
New:
----
uncertainties-3.1.1.tar.gz
++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
Other differences:
------------------
++++++ python-uncertainties.spec ++++++
--- /var/tmp/diff_new_pack.KPMgQv/_old 2019-06-03 18:57:38.800378196 +0200
+++ /var/tmp/diff_new_pack.KPMgQv/_new 2019-06-03 18:57:38.836378182 +0200
@@ -1,7 +1,7 @@
#
# spec file for package python-uncertainties
#
-# Copyright (c) 2018 SUSE LINUX GmbH, Nuernberg, Germany.
+# Copyright (c) 2019 SUSE LINUX GmbH, Nuernberg, Germany.
#
# All modifications and additions to the file contributed by third parties
# remain the property of their copyright owners, unless otherwise agreed
@@ -18,16 +18,21 @@
%{?!python_module:%define python_module() python-%{**} python3-%{**}}
Name: python-uncertainties
-Version: 3.0.3
+Version: 3.1.1
Release: 0
Summary: Uncertainties on the Quantities Involved (aka "Error
Propagation")
License: BSD-3-Clause
Group: Development/Languages/Python
-URL: http://pythonhosted.org/uncertainties/
+URL: https://github.com/lebigot/uncertainties/
Source:
https://files.pythonhosted.org/packages/source/u/uncertainties/uncertainties-%{version}.tar.gz
+BuildRequires: %{python_module nose}
+BuildRequires: %{python_module numpy}
BuildRequires: %{python_module setuptools}
BuildRequires: fdupes
BuildRequires: python-rpm-macros
+# Package uses 2to3 to generate python3 compatible code
+BuildRequires: python3-testsuite
+BuildRequires: python3-tools
BuildArch: noarch
%if 0%{?suse_version}
Recommends: python-numpy
@@ -41,12 +46,19 @@
%prep
%setup -q -n uncertainties-%{version}
+# crazy directory layout
+rm -r uncertainties-py23
+mv uncertainties-py27 uncertainties
%build
%python_build
%install
%python_install
+%python_expand %fdupes %{buildroot}%{$python_sitelib}
+
+%check
+%python_expand PYTHONPATH=%{buildroot}%{$python_sitelib} $python setup.py
nosetests -v
%files %{python_files}
%license LICENSE.txt
++++++ uncertainties-3.0.3.tar.gz -> uncertainties-3.1.1.tar.gz ++++++
diff -urN '--exclude=CVS' '--exclude=.cvsignore' '--exclude=.svn'
'--exclude=.svnignore' old/uncertainties-3.0.3/PKG-INFO
new/uncertainties-3.1.1/PKG-INFO
--- old/uncertainties-3.0.3/PKG-INFO 2018-10-28 23:11:47.000000000 +0100
+++ new/uncertainties-3.1.1/PKG-INFO 2019-05-30 20:02:29.000000000 +0200
@@ -1,6 +1,6 @@
Metadata-Version: 2.1
Name: uncertainties
-Version: 3.0.3
+Version: 3.1.1
Summary: Transparent calculations with uncertainties on the quantities
involved (aka error propagation); fast calculation of derivatives
Home-page: http://uncertainties-python-package.readthedocs.io/
Author: Eric O. LEBIGOT (EOL)
@@ -108,6 +108,10 @@
Main changes:
+ - 3.1: Variables built through a correlation or covariance matrix, and
that
+ have uncertainties that span many orders of magnitude are now
+ calculated more accurately (improved ``correlated_values()`` and
+ ``correlated_values_norm()`` functions).
- 3.0: Massive speedup for some operations involving large numbers of
numbers with uncertainty, like ``sum(ufloat(1, 1) for _ in xrange(100000))``
(this is about 5,000 times faster than before).
- 2.4.8: Friendlier completions in Python shells, etc.: internal
functions should not appear anymore (for the user modules: ``uncertainties``,
``uncertainties.umath`` and ``uncertainties.unumpy``). Parsing the shorthand
notation (e.g. ``3.1(2)``) now works with infinite values (e.g. ``-inf(inf)``);
this mirrors the ability to print such numbers with uncertainty. The Particle
Data Group rounding rule is applied in more cases (e.g. printing 724.2±26.2 now
gives ``724±26``). The shorthand+LaTeX formatting of numbers with an infinite
nominal value is fixed. ``uncertainties.unumpy.matrix`` now uses ``.std_devs``
instead of ``.std_devs()``, for consistency with floats with uncertainty
(automatic conversion of code added to ``uncertainties.1to2``).
- 2.4.7: String formatting now works for ``(-)inf+/-...`` numbers.
@@ -261,6 +265,7 @@
Classifier: Programming Language :: Python :: 3.4
Classifier: Programming Language :: Python :: 3.5
Classifier: Programming Language :: Python :: 3.6
+Classifier: Programming Language :: Python :: 3.7
Classifier: Programming Language :: Python :: Implementation :: Jython
Classifier: Programming Language :: Python :: Implementation :: PyPy
Classifier: Topic :: Education
@@ -271,7 +276,7 @@
Classifier: Topic :: Software Development :: Libraries
Classifier: Topic :: Software Development :: Libraries :: Python Modules
Classifier: Topic :: Utilities
-Provides-Extra: all
-Provides-Extra: optional
-Provides-Extra: tests
Provides-Extra: docs
+Provides-Extra: tests
+Provides-Extra: optional
+Provides-Extra: all
diff -urN '--exclude=CVS' '--exclude=.cvsignore' '--exclude=.svn'
'--exclude=.svnignore' old/uncertainties-3.0.3/README.rst
new/uncertainties-3.1.1/README.rst
--- old/uncertainties-3.0.3/README.rst 2018-10-28 23:09:33.000000000 +0100
+++ new/uncertainties-3.1.1/README.rst 2019-05-19 22:33:18.000000000 +0200
@@ -10,6 +10,9 @@
:target: https://codecov.io/gh/lebigot/uncertainties/
.. image::
https://readthedocs.org/projects/uncertainties-python-package/badge/?version=latest
:target:
http://uncertainties-python-package.readthedocs.io/en/latest/?badge=latest
+.. image:: https://img.shields.io/pypi/v/uncertainties.svg
+ :target: https://pypi.org/project/uncertainties/
+
This is the ``uncertainties`` Python package, which performs **transparent
calculations with uncertainties** (aka "error propagation"):
diff -urN '--exclude=CVS' '--exclude=.cvsignore' '--exclude=.svn'
'--exclude=.svnignore' old/uncertainties-3.0.3/doc/index.rst
new/uncertainties-3.1.1/doc/index.rst
--- old/uncertainties-3.0.3/doc/index.rst 2018-10-28 23:09:33.000000000
+0100
+++ new/uncertainties-3.1.1/doc/index.rst 2019-05-19 22:33:18.000000000
+0200
@@ -138,7 +138,7 @@
conda install -c conda-forge uncertainties
-If you have `pip <hhttps://pypi.org/project/pip/`_, you can try to install
+If you have `pip <https://pypi.org/project/pip/>`_, you can try to install
the latest version with
.. code-block:: sh
diff -urN '--exclude=CVS' '--exclude=.cvsignore' '--exclude=.svn'
'--exclude=.svnignore' old/uncertainties-3.0.3/doc/user_guide.rst
new/uncertainties-3.1.1/doc/user_guide.rst
--- old/uncertainties-3.0.3/doc/user_guide.rst 2018-10-28 23:09:33.000000000
+0100
+++ new/uncertainties-3.1.1/doc/user_guide.rst 2019-05-19 22:33:18.000000000
+0200
@@ -134,8 +134,8 @@
.. index:: correlations; detailed example
-Correlated variables
-====================
+Automatic correlations
+======================
Correlations between variables are **automatically handled** whatever
the number of variables involved, and whatever the complexity of the
@@ -350,10 +350,16 @@
Use of a correlation matrix
---------------------------
-Alternatively, correlated values can be defined through a
-*correlation* matrix (the correlation matrix is the covariance matrix
-normalized with individual standard deviations; it has ones on its
-diagonal), along with a list of nominal values and standard deviations:
+Alternatively, correlated values can be defined through:
+
+- a sequence of nominal values and standard deviations, and
+- a *correlation* matrix between each variable of this sequence
+ (the correlation matrix is the covariance matrix
+ normalized with individual standard deviations; it has ones on its
+ diagonal)—in the form of a NumPy array-like object, e.g. a
+ list of lists, or a NumPy array.
+
+Example:
>>> (u3, v3, sum3) = uncertainties.correlated_values_norm(
... [(1, 0.1), (10, 0.1), (21, 0.22360679774997899)], corr_matrix)
diff -urN '--exclude=CVS' '--exclude=.cvsignore' '--exclude=.svn'
'--exclude=.svnignore' old/uncertainties-3.0.3/setup.py
new/uncertainties-3.1.1/setup.py
--- old/uncertainties-3.0.3/setup.py 2018-10-28 23:09:33.000000000 +0100
+++ new/uncertainties-3.1.1/setup.py 2019-05-30 20:02:20.000000000 +0200
@@ -42,7 +42,7 @@
# Common options for distutils/setuptools's setup():
setup_options = dict(
name='uncertainties',
- version='3.0.3',
+ version='3.1.1',
author='Eric O. LEBIGOT (EOL)',
author_email='[email protected]',
url='http://uncertainties-python-package.readthedocs.io/',
@@ -160,6 +160,10 @@
Main changes:
+- 3.1: Variables built through a correlation or covariance matrix, and that
+ have uncertainties that span many orders of magnitude are now
+ calculated more accurately (improved ``correlated_values()`` and
+ ``correlated_values_norm()`` functions).
- 3.0: Massive speedup for some operations involving large numbers of numbers
with uncertainty, like ``sum(ufloat(1, 1) for _ in xrange(100000))`` (this is
about 5,000 times faster than before).
- 2.4.8: Friendlier completions in Python shells, etc.: internal functions
should not appear anymore (for the user modules: ``uncertainties``,
``uncertainties.umath`` and ``uncertainties.unumpy``). Parsing the shorthand
notation (e.g. ``3.1(2)``) now works with infinite values (e.g. ``-inf(inf)``);
this mirrors the ability to print such numbers with uncertainty. The Particle
Data Group rounding rule is applied in more cases (e.g. printing 724.2±26.2 now
gives ``724±26``). The shorthand+LaTeX formatting of numbers with an infinite
nominal value is fixed. ``uncertainties.unumpy.matrix`` now uses ``.std_devs``
instead of ``.std_devs()``, for consistency with floats with uncertainty
(automatic conversion of code added to ``uncertainties.1to2``).
- 2.4.7: String formatting now works for ``(-)inf+/-...`` numbers.
@@ -320,6 +324,7 @@
'Programming Language :: Python :: 3.4',
'Programming Language :: Python :: 3.5',
'Programming Language :: Python :: 3.6',
+ 'Programming Language :: Python :: 3.7',
'Programming Language :: Python :: Implementation :: Jython',
'Programming Language :: Python :: Implementation :: PyPy',
'Topic :: Education',
diff -urN '--exclude=CVS' '--exclude=.cvsignore' '--exclude=.svn'
'--exclude=.svnignore' old/uncertainties-3.0.3/uncertainties-py23/__init__.py
new/uncertainties-3.1.1/uncertainties-py23/__init__.py
--- old/uncertainties-3.0.3/uncertainties-py23/__init__.py 2018-10-28
23:08:27.000000000 +0100
+++ new/uncertainties-3.1.1/uncertainties-py23/__init__.py 2019-05-30
19:48:32.000000000 +0200
@@ -225,7 +225,7 @@
from uncertainties.core import __all__ # For a correct help(uncertainties)
# Numerical version:
-__version_info__ = (3, 0, 3)
+__version_info__ = (3, 1, 1)
__version__ = '.'.join(map(str, __version_info__))
__author__ = 'Eric O. LEBIGOT (EOL) <[email protected]>'
diff -urN '--exclude=CVS' '--exclude=.cvsignore' '--exclude=.svn'
'--exclude=.svnignore' old/uncertainties-3.0.3/uncertainties-py23/core.py
new/uncertainties-3.1.1/uncertainties-py23/core.py
--- old/uncertainties-3.0.3/uncertainties-py23/core.py 2018-10-28
23:08:27.000000000 +0100
+++ new/uncertainties-3.1.1/uncertainties-py23/core.py 2019-05-30
19:48:32.000000000 +0200
@@ -150,28 +150,84 @@
The numbers with uncertainties returned depend on newly
created, independent variables (Variable objects).
- If 'tags' is not None, it must list the tag of each new
- independent variable.
-
nom_values -- sequence with the nominal (real) values of the
numbers with uncertainties to be returned.
- covariance_mat -- full covariance matrix of the returned
- numbers with uncertainties (not the statistical correlation
- matrix, i.e., not the normalized covariance matrix). For
- example, the first element of this matrix is the variance of
- the first returned number with uncertainty.
+ covariance_mat -- full covariance matrix of the returned numbers with
+ uncertainties. For example, the first element of this matrix is the
+ variance of the first number with uncertainty. This matrix must be a
+ NumPy array-like (list of lists, NumPy array, etc.).
+
+ tags -- if 'tags' is not None, it must list the tag of each new
+ independent variable.
"""
+ # !!! It would in principle be possible to handle 0 variance
+ # variables by first selecting the sub-matrix that does not contain
+ # such variables (with the help of numpy.ix_()), and creating
+ # them separately.
+
+ std_devs = numpy.sqrt(numpy.diag(covariance_mat))
+
+ # For numerical stability reasons, we go through the correlation
+ # matrix, because it is insensitive to any change of scale in the
+ # quantities returned. However, care must be taken with 0 variance
+ # variables: calculating the correlation matrix cannot be simply done
+ # by dividing by standard deviations. We thus use specific
+ # normalization values, with no null value:
+ norm_vector = std_devs.copy()
+ norm_vector[norm_vector==0] = 1
+
+ return correlated_values_norm(
+ # !! The following zip() is a bit suboptimal: correlated_values()
+ # separates back the nominal values and the standard deviations:
+ zip(nom_values, std_devs),
+ covariance_mat/norm_vector/norm_vector[:,numpy.newaxis],
+ tags)
+
+ __all__.append('correlated_values')
+
+ def correlated_values_norm(values_with_std_dev, correlation_mat,
+ tags=None):
+ '''
+ Return correlated values like correlated_values(), but takes
+ instead as input:
+
+ - nominal (float) values along with their standard deviation, and
+ - a correlation matrix (i.e. a normalized covariance matrix).
+
+ values_with_std_dev -- sequence of (nominal value, standard
+ deviation) pairs. The returned, correlated values have these
+ nominal values and standard deviations.
+
+ correlation_mat -- correlation matrix between the given values, except
+ that any value with a 0 standard deviation must have its correlations
+ set to 0, with a diagonal element set to an arbitrary value (something
+ close to 0-1 is recommended, for a better numerical precision). When
+ no value has a 0 variance, this is the covariance matrix normalized by
+ standard deviations, and thus a symmetric matrix with ones on its
+ diagonal. This matrix must be an NumPy array-like (list of lists,
+ NumPy array, etc.).
+
+ tags -- like for correlated_values().
+ '''
+
# If no tags were given, we prepare tags for the newly created
# variables:
if tags is None:
- tags = (None,) * len(nom_values)
+ tags = (None,) * len(values_with_std_dev)
+ (nominal_values, std_devs) = numpy.transpose(values_with_std_dev)
+
+ # We diagonalize the correlation matrix instead of the
+ # covariance matrix, because this is generally more stable
+ # numerically. In fact, the covariance matrix can have
+ # coefficients with arbitrary values, through changes of units
+ # of its input variables. This creates numerical instabilities.
+ #
# The covariance matrix is diagonalized in order to define
# the independent variables that model the given values:
-
- (variances, transform) = numpy.linalg.eigh(covariance_mat)
+ (variances, transform) = numpy.linalg.eigh(correlation_mat)
# Numerical errors might make some variances negative: we set
# them to zero:
@@ -187,43 +243,19 @@
Variable(0, sqrt(variance), tag)
for (variance, tag) in zip(variances, tags))
+ # The coordinates of each new uncertainty as a function of the
+ # new variables must include the variable scale (standard deviation):
+ transform *= std_devs[:, numpy.newaxis]
+
# Representation of the initial correlated values:
values_funcs = tuple(
AffineScalarFunc(
value,
- LinearCombination(dict(itertools.izip(variables, coords))))
- for (coords, value) in zip(transform, nom_values))
+ LinearCombination(dict(zip(variables, coords))))
+ for (coords, value) in zip(transform, nominal_values))
return values_funcs
- __all__.append('correlated_values')
-
- def correlated_values_norm(values_with_std_dev, correlation_mat,
- tags=None):
- '''
- Return correlated values like correlated_values(), but takes
- instead as input:
-
- - nominal (float) values along with their standard deviation, and
-
- - a correlation matrix (i.e. a normalized covariance matrix,
- normalized with individual standard deviations).
-
- values_with_std_dev -- sequence of (nominal value, standard
- deviation) pairs. The returned, correlated values have these
- nominal values and standard deviations.
-
- correlation_mat -- correlation matrix (i.e. the normalized
- covariance matrix, a matrix with ones on its diagonal).
- '''
-
- (nominal_values, std_devs) = numpy.transpose(values_with_std_dev)
-
- return correlated_values(
- nominal_values,
- correlation_mat*std_devs*std_devs[numpy.newaxis].T,
- tags)
-
__all__.append('correlated_values_norm')
###############################################################################
diff -urN '--exclude=CVS' '--exclude=.cvsignore' '--exclude=.svn'
'--exclude=.svnignore'
old/uncertainties-3.0.3/uncertainties-py23/test_uncertainties.py
new/uncertainties-3.1.1/uncertainties-py23/test_uncertainties.py
--- old/uncertainties-3.0.3/uncertainties-py23/test_uncertainties.py
2018-10-28 23:08:27.000000000 +0100
+++ new/uncertainties-3.1.1/uncertainties-py23/test_uncertainties.py
2019-05-30 19:48:32.000000000 +0200
@@ -2128,7 +2128,8 @@
both their nominal value and uncertainty are equal (up to the
given precision).
- m1, m2 -- NumPy matrices.
+ m1, m2 -- NumPy arrays.
+
precision -- precision passed through to
uncertainties.test_uncertainties.numbers_close().
"""
@@ -2151,6 +2152,7 @@
if not numbers_close(elmt1.std_dev,
elmt2.std_dev, precision):
return False
+
return True
@@ -2273,6 +2275,44 @@
assert numbers_close(corr_matrix[0,0], 1)
assert numbers_close(corr_matrix[1,2], 2*v.std_dev/sum_value.std_dev)
+ ####################
+
+ # Test of numerical robustness despite wildly different
+ # orders of magnitude (see
+ # https://github.com/lebigot/uncertainties/issues/95):
+ cov = numpy.diag([1e-70, 1e-70, 1e10])
+ cov[0, 1] = cov[1, 0] = 0.9e-70
+ cov[[0, 1], 2] = -3e-34
+ cov[2, [0, 1]] = -3e-34
+ variables = uncert_core.correlated_values([0]*3, cov)
+
+ # Since the numbers are very small, we need to compare them
+ # in a stricter way, that handles the case of a 0 variance
+ # in `variables`:
+ assert numbers_close(
+ 1e66*cov[0,0], 1e66*variables[0].s**2, tolerance=1e-5)
+ assert numbers_close(
+ 1e66*cov[1,1], 1e66*variables[1].s**2, tolerance=1e-5)
+
+ ####################
+
+ # 0 variances are a bit special, since the correlation matrix
+ # cannot be calculated naively, so we test that there is no
+ # specific problem in this case:
+
+ cov = numpy.diag([0, 0, 10])
+ nom_values = [1, 2, 3]
+ variables = uncert_core.correlated_values(nom_values, cov)
+
+ for (variable, nom_value, variance) in zip(
+ variables, nom_values, cov.diagonal()):
+
+ assert numbers_close(variable.n, nom_value)
+ assert numbers_close(variable.s**2, variance)
+
+ assert arrays_close(
+ cov,
+ numpy.array(uncert_core.covariance_matrix(variables)))
def test_correlated_values_correlation_mat():
'''
diff -urN '--exclude=CVS' '--exclude=.cvsignore' '--exclude=.svn'
'--exclude=.svnignore' old/uncertainties-3.0.3/uncertainties-py27/__init__.py
new/uncertainties-3.1.1/uncertainties-py27/__init__.py
--- old/uncertainties-3.0.3/uncertainties-py27/__init__.py 2018-10-28
23:00:31.000000000 +0100
+++ new/uncertainties-3.1.1/uncertainties-py27/__init__.py 2019-05-30
19:47:40.000000000 +0200
@@ -225,7 +225,7 @@
from .core import __all__ # For a correct help(uncertainties)
# Numerical version:
-__version_info__ = (3, 0, 3)
+__version_info__ = (3, 1, 1)
__version__ = '.'.join(map(str, __version_info__))
__author__ = 'Eric O. LEBIGOT (EOL) <[email protected]>'
diff -urN '--exclude=CVS' '--exclude=.cvsignore' '--exclude=.svn'
'--exclude=.svnignore' old/uncertainties-3.0.3/uncertainties-py27/core.py
new/uncertainties-3.1.1/uncertainties-py27/core.py
--- old/uncertainties-3.0.3/uncertainties-py27/core.py 2018-10-28
23:00:31.000000000 +0100
+++ new/uncertainties-3.1.1/uncertainties-py27/core.py 2019-05-30
19:47:40.000000000 +0200
@@ -143,28 +143,84 @@
The numbers with uncertainties returned depend on newly
created, independent variables (Variable objects).
- If 'tags' is not None, it must list the tag of each new
- independent variable.
-
nom_values -- sequence with the nominal (real) values of the
numbers with uncertainties to be returned.
- covariance_mat -- full covariance matrix of the returned
- numbers with uncertainties (not the statistical correlation
- matrix, i.e., not the normalized covariance matrix). For
- example, the first element of this matrix is the variance of
- the first returned number with uncertainty.
+ covariance_mat -- full covariance matrix of the returned numbers with
+ uncertainties. For example, the first element of this matrix is the
+ variance of the first number with uncertainty. This matrix must be a
+ NumPy array-like (list of lists, NumPy array, etc.).
+
+ tags -- if 'tags' is not None, it must list the tag of each new
+ independent variable.
"""
+ # !!! It would in principle be possible to handle 0 variance
+ # variables by first selecting the sub-matrix that does not contain
+ # such variables (with the help of numpy.ix_()), and creating
+ # them separately.
+
+ std_devs = numpy.sqrt(numpy.diag(covariance_mat))
+
+ # For numerical stability reasons, we go through the correlation
+ # matrix, because it is insensitive to any change of scale in the
+ # quantities returned. However, care must be taken with 0 variance
+ # variables: calculating the correlation matrix cannot be simply done
+ # by dividing by standard deviations. We thus use specific
+ # normalization values, with no null value:
+ norm_vector = std_devs.copy()
+ norm_vector[norm_vector==0] = 1
+
+ return correlated_values_norm(
+ # !! The following zip() is a bit suboptimal: correlated_values()
+ # separates back the nominal values and the standard deviations:
+ zip(nom_values, std_devs),
+ covariance_mat/norm_vector/norm_vector[:,numpy.newaxis],
+ tags)
+
+ __all__.append('correlated_values')
+
+ def correlated_values_norm(values_with_std_dev, correlation_mat,
+ tags=None):
+ '''
+ Return correlated values like correlated_values(), but takes
+ instead as input:
+
+ - nominal (float) values along with their standard deviation, and
+ - a correlation matrix (i.e. a normalized covariance matrix).
+
+ values_with_std_dev -- sequence of (nominal value, standard
+ deviation) pairs. The returned, correlated values have these
+ nominal values and standard deviations.
+
+ correlation_mat -- correlation matrix between the given values, except
+ that any value with a 0 standard deviation must have its correlations
+ set to 0, with a diagonal element set to an arbitrary value (something
+ close to 0-1 is recommended, for a better numerical precision). When
+ no value has a 0 variance, this is the covariance matrix normalized by
+ standard deviations, and thus a symmetric matrix with ones on its
+ diagonal. This matrix must be an NumPy array-like (list of lists,
+ NumPy array, etc.).
+
+ tags -- like for correlated_values().
+ '''
+
# If no tags were given, we prepare tags for the newly created
# variables:
if tags is None:
- tags = (None,) * len(nom_values)
+ tags = (None,) * len(values_with_std_dev)
+ (nominal_values, std_devs) = numpy.transpose(values_with_std_dev)
+
+ # We diagonalize the correlation matrix instead of the
+ # covariance matrix, because this is generally more stable
+ # numerically. In fact, the covariance matrix can have
+ # coefficients with arbitrary values, through changes of units
+ # of its input variables. This creates numerical instabilities.
+ #
# The covariance matrix is diagonalized in order to define
# the independent variables that model the given values:
-
- (variances, transform) = numpy.linalg.eigh(covariance_mat)
+ (variances, transform) = numpy.linalg.eigh(correlation_mat)
# Numerical errors might make some variances negative: we set
# them to zero:
@@ -180,43 +236,19 @@
Variable(0, sqrt(variance), tag)
for (variance, tag) in zip(variances, tags))
+ # The coordinates of each new uncertainty as a function of the
+ # new variables must include the variable scale (standard deviation):
+ transform *= std_devs[:, numpy.newaxis]
+
# Representation of the initial correlated values:
values_funcs = tuple(
AffineScalarFunc(
value,
- LinearCombination(dict(itertools.izip(variables, coords))))
- for (coords, value) in zip(transform, nom_values))
+ LinearCombination(dict(zip(variables, coords))))
+ for (coords, value) in zip(transform, nominal_values))
return values_funcs
- __all__.append('correlated_values')
-
- def correlated_values_norm(values_with_std_dev, correlation_mat,
- tags=None):
- '''
- Return correlated values like correlated_values(), but takes
- instead as input:
-
- - nominal (float) values along with their standard deviation, and
-
- - a correlation matrix (i.e. a normalized covariance matrix,
- normalized with individual standard deviations).
-
- values_with_std_dev -- sequence of (nominal value, standard
- deviation) pairs. The returned, correlated values have these
- nominal values and standard deviations.
-
- correlation_mat -- correlation matrix (i.e. the normalized
- covariance matrix, a matrix with ones on its diagonal).
- '''
-
- (nominal_values, std_devs) = numpy.transpose(values_with_std_dev)
-
- return correlated_values(
- nominal_values,
- correlation_mat*std_devs*std_devs[numpy.newaxis].T,
- tags)
-
__all__.append('correlated_values_norm')
###############################################################################
diff -urN '--exclude=CVS' '--exclude=.cvsignore' '--exclude=.svn'
'--exclude=.svnignore'
old/uncertainties-3.0.3/uncertainties-py27/test_uncertainties.py
new/uncertainties-3.1.1/uncertainties-py27/test_uncertainties.py
--- old/uncertainties-3.0.3/uncertainties-py27/test_uncertainties.py
2018-10-28 23:00:31.000000000 +0100
+++ new/uncertainties-3.1.1/uncertainties-py27/test_uncertainties.py
2019-05-30 19:47:40.000000000 +0200
@@ -2132,7 +2132,8 @@
both their nominal value and uncertainty are equal (up to the
given precision).
- m1, m2 -- NumPy matrices.
+ m1, m2 -- NumPy arrays.
+
precision -- precision passed through to
uncertainties.test_uncertainties.numbers_close().
"""
@@ -2155,6 +2156,7 @@
if not numbers_close(elmt1.std_dev,
elmt2.std_dev, precision):
return False
+
return True
@@ -2277,6 +2279,44 @@
assert numbers_close(corr_matrix[0,0], 1)
assert numbers_close(corr_matrix[1,2], 2*v.std_dev/sum_value.std_dev)
+ ####################
+
+ # Test of numerical robustness despite wildly different
+ # orders of magnitude (see
+ # https://github.com/lebigot/uncertainties/issues/95):
+ cov = numpy.diag([1e-70, 1e-70, 1e10])
+ cov[0, 1] = cov[1, 0] = 0.9e-70
+ cov[[0, 1], 2] = -3e-34
+ cov[2, [0, 1]] = -3e-34
+ variables = uncert_core.correlated_values([0]*3, cov)
+
+ # Since the numbers are very small, we need to compare them
+ # in a stricter way, that handles the case of a 0 variance
+ # in `variables`:
+ assert numbers_close(
+ 1e66*cov[0,0], 1e66*variables[0].s**2, tolerance=1e-5)
+ assert numbers_close(
+ 1e66*cov[1,1], 1e66*variables[1].s**2, tolerance=1e-5)
+
+ ####################
+
+ # 0 variances are a bit special, since the correlation matrix
+ # cannot be calculated naively, so we test that there is no
+ # specific problem in this case:
+
+ cov = numpy.diag([0, 0, 10])
+ nom_values = [1, 2, 3]
+ variables = uncert_core.correlated_values(nom_values, cov)
+
+ for (variable, nom_value, variance) in zip(
+ variables, nom_values, cov.diagonal()):
+
+ assert numbers_close(variable.n, nom_value)
+ assert numbers_close(variable.s**2, variance)
+
+ assert arrays_close(
+ cov,
+ numpy.array(uncert_core.covariance_matrix(variables)))
def test_correlated_values_correlation_mat():
'''
diff -urN '--exclude=CVS' '--exclude=.cvsignore' '--exclude=.svn'
'--exclude=.svnignore' old/uncertainties-3.0.3/uncertainties.egg-info/PKG-INFO
new/uncertainties-3.1.1/uncertainties.egg-info/PKG-INFO
--- old/uncertainties-3.0.3/uncertainties.egg-info/PKG-INFO 2018-10-28
23:11:47.000000000 +0100
+++ new/uncertainties-3.1.1/uncertainties.egg-info/PKG-INFO 2019-05-30
20:02:28.000000000 +0200
@@ -1,6 +1,6 @@
Metadata-Version: 2.1
Name: uncertainties
-Version: 3.0.3
+Version: 3.1.1
Summary: Transparent calculations with uncertainties on the quantities
involved (aka error propagation); fast calculation of derivatives
Home-page: http://uncertainties-python-package.readthedocs.io/
Author: Eric O. LEBIGOT (EOL)
@@ -108,6 +108,10 @@
Main changes:
+ - 3.1: Variables built through a correlation or covariance matrix, and
that
+ have uncertainties that span many orders of magnitude are now
+ calculated more accurately (improved ``correlated_values()`` and
+ ``correlated_values_norm()`` functions).
- 3.0: Massive speedup for some operations involving large numbers of
numbers with uncertainty, like ``sum(ufloat(1, 1) for _ in xrange(100000))``
(this is about 5,000 times faster than before).
- 2.4.8: Friendlier completions in Python shells, etc.: internal
functions should not appear anymore (for the user modules: ``uncertainties``,
``uncertainties.umath`` and ``uncertainties.unumpy``). Parsing the shorthand
notation (e.g. ``3.1(2)``) now works with infinite values (e.g. ``-inf(inf)``);
this mirrors the ability to print such numbers with uncertainty. The Particle
Data Group rounding rule is applied in more cases (e.g. printing 724.2±26.2 now
gives ``724±26``). The shorthand+LaTeX formatting of numbers with an infinite
nominal value is fixed. ``uncertainties.unumpy.matrix`` now uses ``.std_devs``
instead of ``.std_devs()``, for consistency with floats with uncertainty
(automatic conversion of code added to ``uncertainties.1to2``).
- 2.4.7: String formatting now works for ``(-)inf+/-...`` numbers.
@@ -261,6 +265,7 @@
Classifier: Programming Language :: Python :: 3.4
Classifier: Programming Language :: Python :: 3.5
Classifier: Programming Language :: Python :: 3.6
+Classifier: Programming Language :: Python :: 3.7
Classifier: Programming Language :: Python :: Implementation :: Jython
Classifier: Programming Language :: Python :: Implementation :: PyPy
Classifier: Topic :: Education
@@ -271,7 +276,7 @@
Classifier: Topic :: Software Development :: Libraries
Classifier: Topic :: Software Development :: Libraries :: Python Modules
Classifier: Topic :: Utilities
-Provides-Extra: all
-Provides-Extra: optional
-Provides-Extra: tests
Provides-Extra: docs
+Provides-Extra: tests
+Provides-Extra: optional
+Provides-Extra: all
diff -urN '--exclude=CVS' '--exclude=.cvsignore' '--exclude=.svn'
'--exclude=.svnignore'
old/uncertainties-3.0.3/uncertainties.egg-info/requires.txt
new/uncertainties-3.1.1/uncertainties.egg-info/requires.txt
--- old/uncertainties-3.0.3/uncertainties.egg-info/requires.txt 2018-10-28
23:11:47.000000000 +0100
+++ new/uncertainties-3.1.1/uncertainties.egg-info/requires.txt 2019-05-30
20:02:28.000000000 +0200
@@ -1,8 +1,8 @@
[all]
-nose
numpy
sphinx
+nose
[docs]
sphinx
