Uncertainties in arrays

The unumpy package

This package contains:

1. utilities that help with the creation and manipulation of NumPy arrays and matrices of numbers with uncertainties;

2. generalizations of multiple NumPy functions so that they also work with arrays that contain numbers with uncertainties.

While basic operations on arrays that contain numbers with uncertainties can be performed without it, the unumpy package is useful for more advanced uses.

Operations on arrays (including their cosine, etc.) can thus be performed transparently.

These features can be made available with

>>> from uncertainties import unumpy

Creation and manipulation of arrays and matrices


Arrays of numbers with uncertainties can be built from values and uncertainties:

>>> arr = unumpy.uarray([1, 2], [0.01, 0.002])
>>> print arr
[1.0+/-0.01 2.0+/-0.002]

NumPy arrays of numbers with uncertainties can also be built directly through NumPy, thanks to NumPy’s support of arrays of arbitrary objects:

>>> arr = numpy.array([ufloat(1, 0.1), ufloat(2, 0.002)])


Matrices of numbers with uncertainties are best created in one of two ways. The first way is similar to using uarray():

>>> mat = unumpy.umatrix([1, 2], [0.01, 0.002])

Matrices can also be built by converting arrays of numbers with uncertainties into matrices through the unumpy.matrix class:

>>> mat = unumpy.matrix(arr)

unumpy.matrix objects behave like numpy.matrix objects of numbers with uncertainties, but with better support for some operations (such as matrix inversion). For instance, regular NumPy matrices cannot be inverted, if they contain numbers with uncertainties (i.e., numpy.matrix([[ufloat(…), …]]).I does not work). This is why the unumpy.matrix class is provided: both the inverse and the pseudo-inverse of a matrix can be calculated in the usual way: if mat is a unumpy.matrix,

>>> print mat.I

does calculate the inverse or pseudo-inverse of mat with uncertainties.

Uncertainties and nominal values

Nominal values and uncertainties in arrays (and matrices) can be directly accessed (through functions that work on pure float arrays too):

>>> unumpy.nominal_values(arr)
array([ 1.,  2.])
>>> unumpy.std_devs(mat)
matrix([[ 0.1  ,  0.002]])

Mathematical functions

This module defines uncertainty-aware mathematical functions that generalize those from uncertainties.umath so that they work on NumPy arrays of numbers with uncertainties instead of just scalars:

>>> print unumpy.cos(arr)  # Cosine of each array element

NumPy’s function names are used, and not those from the math module (for instance, unumpy.arccos() is defined, like in NumPy, and is not named acos() like in the math module).

The definition of the mathematical quantities calculated by these functions is available in the documentation for uncertainties.umath (which is accessible through help() or pydoc).

Storing arrays in text format

Arrays of numbers with uncertainties can be directly pickled, saved to file and read from a file. Pickling has the advantage of preserving correlations between errors.

Storing instead arrays in text format loses correlations between errors but has the advantage of being both computer- and human-readable. This can be done through NumPy’s savetxt() and loadtxt().

Writing the array to file can be done by asking NumPy to use the representation of numbers with uncertainties (instead of the default float conversion):

>>> numpy.savetxt('arr.txt', arr, fmt='%r')

This produces a file arr.txt that contains a text representation of the array:


The file can then be read back by instructing NumPy to convert all the columns with uncertainties.ufloat_fromstr(). The number num_cols of columns in the input file (1, in our example) must be determined in advance, because NumPy requires a converter for each column separately. For Python 2:

>>> converters = dict.fromkeys(range(num_cols), uncertainties.ufloat_fromstr)

For Python 3, since numpy.loadtxt() passes bytes to converters, they must first be converted into a string:

>>> converters = dict.fromkeys(
        lambda col_bytes: uncertainties.ufloat_fromstr(col_bytes.decode("latin1")))

(Latin 1 appears to in fact be the encoding used in numpy.savetxt() [as of NumPy 1.12]. This encoding seems to be the one hardcoded in numpy.compat.asbytes().)

The array can then be loaded:

>>> arr = numpy.loadtxt('arr.txt', converters=converters, dtype=object)

Additional array functions: unumpy.ulinalg

The unumpy.ulinalg module contains more uncertainty-aware functions for arrays that contain numbers with uncertainties.

It currently offers generalizations of two functions from numpy.linalg that work on arrays (or matrices) that contain numbers with uncertainties, the matrix inverse and pseudo-inverse:

>>> unumpy.ulinalg.inv([[ufloat(2, 0.1)]])
array([[0.5+/-0.025]], dtype=object)
>>> unumpy.ulinalg.pinv(mat)
        [0.4+/-0.00161789987329]], dtype=object)