User Guide

Basic setup

Basic mathematical operations involving numbers with uncertainties only require a simple import:

>>> from uncertainties import ufloat

The ufloat() function creates numbers with uncertainties. Existing calculation code can usually run with no or little modification and automatically produce results with uncertainties.

The uncertainties module contains other features, which can be made accessible through

>>> import uncertainties

The uncertainties package also contains sub-modules for advanced mathematical functions, and arrays and matrices.

Creating numbers with uncertainties

Numbers with uncertainties can be input either numerically, or through one of many string representations, so that files containing numbers with uncertainties can easily be parsed. Thus, x = 0.20±0.01 can be expressed in many convenient ways, including:

>>> x = ufloat(0.20, 0.01)  # x = 0.20+/-0.01
>>> from uncertainties import ufloat_fromstr
>>> x = ufloat_fromstr("0.20+/-0.01")
>>> x = ufloat_fromstr("(2+/-0.1)e-01")  # Factored exponent
>>> x = ufloat_fromstr("0.20(1)")  # Short-hand notation
>>> x = ufloat_fromstr("20(1)e-2")  # Exponent notation
>>> x = ufloat_fromstr(u"0.20±0.01")  # Pretty-print form
>>> x = ufloat_fromstr("0.20")  # Automatic uncertainty of +/-1 on last digit

Each number created this way is an independent (random) variable (for details, see the Technical Guide).

More information can be obtained with pydoc uncertainties.ufloat and pydoc uncertainties.ufloat_fromstr (“20(1)×10-2” is also recognized, etc.).

Basic math

Calculations can be performed directly, as with regular real numbers:

>>> square = x**2
>>> print square
0.040+/-0.004

Mathematical operations

Besides being able to apply basic mathematical operations to numbers with uncertainty, this package provides generalizations of most of the functions from the standard math module. These mathematical functions are found in the uncertainties.umath module:

>>> from uncertainties.umath import *  # Imports sin(), etc.
>>> sin(x**2)
0.03998933418663417+/-0.003996800426643912

The list of available mathematical functions can be obtained with the pydoc uncertainties.umath command.

Arrays of numbers with uncertainties

It is possible to put numbers with uncertainties in NumPy arrays and matrices:

>>> arr = numpy.array([ufloat(1, 0.01), ufloat(2, 0.1)])
>>> 2*arr
[2.0+/-0.02 4.0+/-0.2]
>>> print arr.sum()
3.00+/-0.10

Thus, usual operations on NumPy arrays can be performed transparently even when these arrays contain numbers with uncertainties.

More complex operations on NumPy arrays and matrices can be performed through the dedicated uncertainties.unumpy module.

Correlated variables

Correlations between variables are automatically handled whatever the number of variables involved, and whatever the complexity of the calculation. For example, when x is the number with uncertainty defined above,

>>> square = x**2
>>> print square
0.040+/-0.004
>>> square - x*x
0.0+/-0
>>> y = x*x + 1
>>> y - square
1.0+/-0

The last two printed results above have a zero uncertainty despite the fact that x, y and square have a non-zero uncertainty: the calculated functions give the same value for all samples of the random variable x.

Thanks to the automatic correlation handling, calculations can be performed in as many steps as necessary, exactly as with simple floats. When various quantities are combined through mathematical operations, the result is calculated by taking into account all the correlations between the quantities involved. All of this is done completely transparently.

Printing

Numbers with uncertainties can be printed conveniently:

>>> print x
0.200+/-0.010

The resulting form can generally be parsed back with ufloat_fromstr() (except for the LaTeX form).

The nominal value and the uncertainty always have the same precision: this makes it easier to compare them.

More control over the format can be obtained (in Python 2.6+) through the usual format() method of strings:

>>> print 'Result = {:10.2f}'.format(x)
Result =       0.20+/-      0.01

(Python 2.6 requires '{0:10.2f}' instead, with the usual explicit index. In Python 2.5 and earlier versions, str.format() is not available, but one can use the format() method of numbers with uncertainties instead: 'Result = %s' % x.format('10.2f').)

All the float format specifications are accepted, except those with the n format type. In particular, a fill character, an alignment option, a sign or zero option, a width, or the % format type are all supported.

When uncertainties must choose the number of significant digits on the uncertainty, it is defined with the Particle Data Group rounding rules (these rules keep the number of digits small, which is convenient for reading numbers with uncertainties, and at the same time prevent the uncertainty from being displayed with too few digits):

>>> print 'Automatic number of digits on the uncertainty: {}'.format(x)
Automatic number of digits on the uncertainty: 0.200+/-0.010
>>> print x
0.200+/-0.010

It is possible to control the number of significant digits of the uncertainty by adding the precision modifier u after the precision (and before any valid float format type like f, e, the empty format type, etc.):

>>> print '1 significant digit on the uncertainty: {:.1u}'.format(x)
1 significant digit on the uncertainty: 0.20+/-0.01
>>> print '3 significant digits on the uncertainty: {:.3u}'.format(x)
3 significant digits on the uncertainty: 0.2000+/-0.0100
>>> print '1 significant digit, exponent notation: {:.1ue}'.format(x)
1 significant digit, exponent notation: (2.0+/-0.1)e-01
>>> print '1 significant digit, percentage: {:.1u%}'.format(x)
1 significant digit, percentage: (20+/-1)%

The usual float formats with a precision retain their original meaning (e.g. .2e uses two digits after the decimal point): code that works with floats produces similar results when running with numbers with uncertainties.

A common exponent is automatically calculated if an exponent is needed for the larger of the nominal value (in absolute value) and the uncertainty (the rule is the same as for floats). The exponent is generally factored, for increased legibility:

>>> print x*1e7
(2.00+/-0.10)e+06

When a format width is used, the common exponent is not factored:

>>> print 'Result = {:10.1e}'.format(x*1e-10)
Result =    2.0e-11+/-   0.1e-11

(Using a (minimal) width of 1 is thus a way of forcing exponents to not be factored.) Thanks to this feature, each part (nominal value and standard deviation) is correctly aligned across multiple lines, while the relative magnitude of the error can still be readily estimated thanks to the common exponent.

Formatting options can be added at the end of the format string: S for the shorthand notation, L for a LaTeX output, P for pretty-printing:

>>> print '{:+.1uS}'.format(x)  # Sign, 1 digit for the uncertainty, shorthand
+0.20(1)
>>> print '{:L}'.format(x*1e7)  # Automatic exponent form, LaTeX
\left(2.00 \pm 0.10\right) \times 10^{6}

The pretty-printing mode uses “±” and superscript exponents: the default output is such that print '{:.2e}'.format(x) yields “(2.00+/-0.10)e-01”, whereas the pretty-printing mode in print u'{:.2eP}'.format(x) yields “(2.00±0.10)×10-1”. Note that the pretty-printing mode implies using Unicode format strings (u'…' in Python 2, but simply '…' in Python 3).

These formatting options can be combined (when meaningful).

An uncertainty which is exactly zero is always formatted as an integer:

>>> print ufloat(3.1415, 0)
3.1415+/-0
>>> print ufloat(3.1415e10, 0)
(3.1415+/-0)e+10
>>> print ufloat(3.1415, 0.0005)
3.1415+/-0.0005
>>> print '{:.2f}'.format(ufloat(3.14, 0.001))
3.14+/-0.00
>>> print '{:.2f}'.format(ufloat(3.14, 0.00))
3.14+/-0

All the digits of a number with uncertainty are given in its representation:

>>> y = ufloat(1.23456789012345, 0.123456789)
>>> print y
1.23+/-0.12
>>> print repr(y)
1.23456789012345+/-0.123456789
>>> y
1.23456789012345+/-0.123456789

More information on formatting can be obtained with pydoc uncertainties.UFloat.__format__ (customization of the LaTeX output, etc.).

Global formatting

It is sometimes useful to have a consistent formatting across multiple parts of a program. Python’s string.Formatter class allows one to do just that. Here is how it can be used to consistently use the shorthand notation for numbers with uncertainties:

class ShorthandFormatter(string.Formatter):

    def format_field(self, value, format_spec):
        if isinstance(value, uncertainties.UFloat):
            return value.format(format_spec+'S')  # Shorthand option added
        # Special formatting for other types can be added here (floats, etc.)
        else:
            # Usual formatting:
            return super(ShorthandFormatter, self).format_field(
                value, format_spec)

frmtr = ShorthandFormatter()

print frmtr.format("Result = {0:.1u}", x)  # 1-digit uncertainty

prints with the shorthand notation: Result = 0.20(1).

Access to the uncertainty and to the nominal value

The nominal value and the uncertainty (standard deviation) can also be accessed independently:

>>> print square
0.040+/-0.004
>>> print square.nominal_value
0.04
>>> print square.n  # Abbreviation
0.04
>>> print square.std_dev
0.004
>>> print square.s  # Abbreviation
0.004

Access to the individual sources of uncertainty

The various contributions to an uncertainty can be obtained through the error_components() method, which maps the independent variables a quantity depends on to their contribution to the total uncertainty. According to linear error propagation theory (which is the method followed by uncertainties), the sum of the squares of these contributions is the squared uncertainty.

The individual contributions to the uncertainty are more easily usable when the variables are tagged:

>>> u = ufloat(1, 0.1, "u variable")  # Tag
>>> v = ufloat(10, 0.1, "v variable")
>>> sum_value = u+2*v
>>> sum_value
21.0+/-0.223606797749979
>>> for (var, error) in sum_value.error_components().items():
...     print "{}: {}".format(var.tag, error)
...
u variable: 0.1
v variable: 0.2

The variance (i.e. squared uncertainty) of the result (sum_value) is the quadratic sum of these independent uncertainties, as it should be (0.1**2 + 0.2**2).

The tags do not have to be distinct. For instance, multiple random variables can be tagged as "systematic", and their contribution to the total uncertainty of result can simply be obtained as:

>>> syst_error = math.sqrt(sum(  # Error from *all* systematic errors
...     error**2
...     for (var, error) in result.error_components().items()
...     if var.tag == "systematic"))

The remaining contribution to the uncertainty is:

>>> other_error = math.sqrt(result.std_dev**2 - syst_error**2)

The variance of result is in fact simply the quadratic sum of these two errors, since the variables from result.error_components() are independent.

Comparison operators

Comparison operators behave in a natural way:

>>> print x
0.200+/-0.010
>>> y = x + 0.0001
>>> y
0.2001+/-0.01
>>> y > x
True
>>> y > 0
True

One important concept to keep in mind is that ufloat() creates a random variable, so that two numbers with the same nominal value and standard deviation are generally different:

>>> y = ufloat(1, 0.1)
>>> z = ufloat(1, 0.1)
>>> print y
1.00+/-0.10
>>> print z
1.00+/-0.10
>>> y == y
True
>>> y == z
False

In physical terms, two rods of the same nominal length and uncertainty on their length are generally of different sizes: y is different from z.

More detailed information on the semantics of comparison operators for numbers with uncertainties can be found in the Technical Guide.

Covariance and correlation matrices

Covariance matrix

The covariance matrix between various variables or calculated quantities can be simply obtained:

>>> sum_value = u+2*v
>>> cov_matrix = uncertainties.covariance_matrix([u, v, sum_value])

has value

[[0.01, 0.0,  0.01],
 [0.0,  0.01, 0.02],
 [0.01, 0.02, 0.05]]

In this matrix, the zero covariances indicate that u and v are independent from each other; the last column shows that sum_value does depend on these variables. The uncertainties package keeps track at all times of all correlations between quantities (variables and functions):

>>> sum_value - (u+2*v)
0.0+/-0

Correlation matrix

If the NumPy package is available, the correlation matrix can be obtained as well:

>>> corr_matrix = uncertainties.correlation_matrix([u, v, sum_value])
>>> corr_matrix
array([[ 1.        ,  0.        ,  0.4472136 ],
       [ 0.        ,  1.        ,  0.89442719],
       [ 0.4472136 ,  0.89442719,  1.        ]])

Correlated variables

Reciprocally, correlated variables can be created transparently, provided that the NumPy package is available.

Use of a covariance matrix

Correlated variables can be obtained through the covariance matrix:

>>> (u2, v2, sum2) = uncertainties.correlated_values([1, 10, 21], cov_matrix)

creates three new variables with the listed nominal values, and the given covariance matrix:

>>> sum_value
21.0+/-0.223606797749979
>>> sum2
21.0+/-0.223606797749979
>>> sum2 - (u2+2*v2)
0.0+/-3.83371856862256e-09

The theoretical value of the last expression is exactly zero, like for sum - (u+2*v), but numerical errors yield a small uncertainty (3e-9 is indeed very small compared to the uncertainty on sum2: correlations should in fact cancel the uncertainty on sum2).

The covariance matrix is the desired one:

>>> uncertainties.covariance_matrix([u2, v2, sum2])

reproduces the original covariance matrix cov_matrix (up to rounding errors).

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:

>>> (u3, v3, sum3) = uncertainties.correlated_values_norm(
...     [(1, 0.1), (10, 0.1), (21, 0.22360679774997899)], corr_matrix)
>>> print u3
1.00+/-0.10

The three returned numbers with uncertainties have the correct uncertainties and correlations (corr_matrix can be recovered through correlation_matrix()).

Making custom functions accept numbers with uncertainties

This package allows code which is not meant to be used with numbers with uncertainties to handle them anyway. This is for instance useful when calling external functions (which are out of the user’s control), including functions written in C or Fortran. Similarly, functions that do not have a simple analytical form can be automatically wrapped so as to also work with arguments that contain uncertainties.

It is thus possible to take a function f() that returns a single float, and to automatically generalize it so that it also works with numbers with uncertainties:

>>> wrapped_f = uncertainties.wrap(f)

The new function wrapped_f() accepts numbers with uncertainties as arguments wherever a Python float is used for f(). wrapped_f() returns the same values as f(), but with uncertainties.

With a simple wrapping call like above, uncertainties in the function result are automatically calculated numerically. Analytical uncertainty calculations can be performed if derivatives are provided to wrap().

More details are available in the documentation string of wrap() (accessible through the pydoc command, or Python’s help() shell function).

Miscellaneous utilities

It is sometimes useful to modify the error on certain parameters so as to study its impact on a final result. With this package, the uncertainty of a variable can be changed on the fly:

>>> sum_value = u+2*v
>>> sum_value
21.0+/-0.223606797749979
>>> prev_uncert = u.std_dev
>>> u.std_dev = 10
>>> sum_value
21.0+/-10.00199980003999
>>> u.std_dev = prev_uncert

The relevant concept is that sum_value does depend on the variables u and v: the uncertainties package keeps track of this fact, as detailed in the Technical Guide, and uncertainties can thus be updated at any time.

When manipulating ensembles of numbers, some of which contain uncertainties while others are simple floats, it can be useful to access the nominal value and uncertainty of all numbers in a uniform manner. This is what the nominal_value() and std_dev() functions do:

>>> print uncertainties.nominal_value(x)
0.2
>>> print uncertainties.std_dev(x)
0.01
>>> uncertainties.nominal_value(3)
3
>>> uncertainties.std_dev(3)
0.0

Finally, a utility method is provided that directly yields the standard score (number of standard deviations) between a number and a result with uncertainty: with x equal to 0.20±0.01,

>>> x.std_score(0.17)
-3.0

Derivatives

Since the application of linear error propagation theory involves the calculation of derivatives, this package automatically performs such calculations; users can thus easily get the derivative of an expression with respect to any of its variables:

>>> u = ufloat(1, 0.1)
>>> v = ufloat(10, 0.1)
>>> sum_value = u+2*v
>>> sum_value.derivatives[u]
1.0
>>> sum_value.derivatives[v]
2.0

These values are obtained with a fast differentiation algorithm.

Additional information

The capabilities of the uncertainties package in terms of array handling are detailed in Uncertainties in arrays.

Details about the theory behind this package and implementation information are given in the Technical Guide.