Technical Guide

Testing whether an object is a number with uncertainty

The recommended way of testing whether value carries an uncertainty handled by this module is by checking whether value is an instance of UFloat, through isinstance(value, uncertainties.UFloat).

Pickling

The quantities with uncertainties created by the uncertainties package can be pickled (they can be stored in a file, for instance).

If multiple variables are pickled together (including when pickling NumPy arrays), their correlations are preserved:

>>> import pickle
>>> x = ufloat(2, 0.1)
>>> y = 2*x
>>> p = pickle.dumps([x, y])  # Pickling to a string
>>> (x2, y2) = pickle.loads(p)  # Unpickling into new variables
>>> y2 - 2*x2
0.0+/-0

The final result is exactly zero because the unpickled variables x2 and y2 are completely correlated.

However, unpickling necessarily creates new variables that bear no relationship with the original variables (in fact, the pickled representation can be stored in a file and read from another program after the program that did the pickling is finished: the unpickled variables cannot be correlated to variables that can disappear). Thus

>>> x - x2
0.0+/-0.14142135623730953

which shows that the original variable x and the new variable x2 are completely uncorrelated.

Comparison operators

Comparison operations (>, ==, etc.) on numbers with uncertainties have a pragmatic semantics, in this package: numbers with uncertainties can be used wherever Python numbers are used, most of the time with a result identical to the one that would be obtained with their nominal value only. This allows code that runs with pure numbers to also work with numbers with uncertainties.

The boolean value (bool(x), if x …) of a number with uncertainty x is defined as the result of x != 0, as usual.

However, since the objects defined in this module represent probability distributions and not pure numbers, comparison operators are interpreted in a specific way.

The result of a comparison operation is defined so as to be essentially consistent with the requirement that uncertainties be small: the value of a comparison operation is True only if the operation yields True for all infinitesimal variations of its random variables around their nominal values, except, possibly, for an infinitely small number of cases.

Example:

>>> x = ufloat(3.14, 0.01)
>>> x == x
True

because a sample from the probability distribution of x is always equal to itself. However:

>>> y = ufloat(3.14, 0.01)
>>> x == y
False

since x and y are independent random variables that almost always give a different value (put differently, x-y is not equal to 0, as it can take many different values). Note that this is different from the result of z = 3.14; t = 3.14; print z == t, because x and y are random variables, not pure numbers.

Similarly,

>>> x = ufloat(3.14, 0.01)
>>> y = ufloat(3.00, 0.01)
>>> x > y
True

because x is supposed to have a probability distribution largely contained in the 3.14±~0.01 interval, while y is supposed to be well in the 3.00±~0.01 one: random samples of x and y will most of the time be such that the sample from x is larger than the sample from y. Therefore, it is natural to consider that for all practical purposes, x > y.

Since comparison operations are subject to the same constraints as other operations, as required by the linear approximation method, their result should be essentially constant over the regions of highest probability of their variables (this is the equivalent of the linearity of a real function, for boolean values). Thus, it is not meaningful to compare the following two independent variables, whose probability distributions overlap:

>>> x = ufloat(3, 0.01)
>>> y = ufloat(3.0001, 0.01)

In fact the function (x, y) → (x > y) is not even continuous over the region where x and y are concentrated, which violates the assumption of approximate linearity made in this package on operations involving numbers with uncertainties. Comparing such numbers therefore returns a boolean result whose meaning is undefined.

However, values with largely overlapping probability distributions can sometimes be compared unambiguously:

>>> x = ufloat(3, 1)
>>> x
3.0+/-1.0
>>> y = x + 0.0002
>>> y
3.0002+/-1.0
>>> y > x
True

In fact, correlations guarantee that y is always larger than x: y-x correctly satisfies the assumption of linearity, since it is a constant “random” function (with value 0.0002, even though y and x are random). Thus, it is indeed true that y > x.

Linear propagation of uncertainties

Constraints on the uncertainties

This package calculates the standard deviation of mathematical expressions through the linear approximation of error propagation theory.

The standard deviations and nominal values calculated by this package are thus meaningful approximations as long as uncertainties are “small”. A more precise version of this constraint is that the final calculated functions must have precise linear expansions in the region where the probability distribution of their variables is the largest. Mathematically, this means that the linear terms of the final calculated functions around the nominal values of their variables should be much larger than the remaining higher-order terms over the region of significant probability (because such higher-order contributions are neglected).

For example, calculating x*10 with x = 5±3 gives a perfect result since the calculated function is linear. So does umath.atan(umath.tan(x)) for x = 0±1, since only the final function counts (not an intermediate function like tan()).

Another example is sin(0+/-0.01), for which uncertainties yields a meaningful standard deviation since the sine is quite linear over 0±0.01. However, cos(0+/-0.01), yields an approximate standard deviation of 0 because it is parabolic around 0 instead of linear; this might not be precise enough for all applications.

More precise uncertainty estimates can be obtained, if necessary, with the soerp and mcerp packages. The soerp package performs second-order error propagation: this is still quite fast, but the standard deviation of higher-order functions like f(x) = x³ for x = 0±0.1 is calculated as being exactly zero (as with uncertainties). The mcerp package performs Monte-Carlo calculations, and can in principle yield very precise results, but calculations are much slower than with approximation schemes.

NaN uncertainty

If linear error propagation theory cannot be applied, the functions defined by uncertainties internally use a not-a-number value (nan) for the derivative.

As a consequence, it is possible for uncertainties to be nan:

>>> umath.sqrt(ufloat(0, 1))
0.0+/-nan

This indicates that the derivative required by linear error propagation theory is not defined (a Monte-Carlo calculation of the resulting random variable is more adapted to this specific case).

However, even in this case where the derivative at the nominal value is infinite, the uncertainties package correctly handles perfectly precise numbers:

>>> umath.sqrt(ufloat(0, 0))
0.0+/-0

is thus the correct result, despite the fact that the derivative of the square root is not defined in zero.

Mathematical definition of numbers with uncertainties

Mathematically, numbers with uncertainties are, in this package, probability distributions. They are not restricted to normal (Gaussian) distributions and can be any distribution. These probability distributions are reduced to two numbers: a nominal value and an uncertainty.

Thus, both independent variables (Variable objects) and the result of mathematical operations (AffineScalarFunc objects) contain these two values (respectively in their nominal_value and std_dev attributes).

The uncertainty of a number with uncertainty is simply defined in this package as the standard deviation of the underlying probability distribution.

The numbers with uncertainties manipulated by this package are assumed to have a probability distribution mostly contained around their nominal value, in an interval of about the size of their standard deviation. This should cover most practical cases.

A good choice of nominal value for a number with uncertainty is thus the median of its probability distribution, the location of highest probability, or the average value.

Probability distributions (random variables and calculation results) are printed as:

nominal value +/- standard deviation

but this does not imply any property on the nominal value (beyond the fact that the nominal value is normally inside the region of high probability density), or that the probability distribution of the result is symmetrical (this is rarely strictly the case).

Differentiation method

The uncertainties package automatically calculates the derivatives required by linear error propagation theory.

Almost all the derivatives of the fundamental functions provided by uncertainties are obtained through analytical formulas (the few mathematical functions that are instead differentiated through numerical approximation are listed in umath_core.num_deriv_funcs).

The derivatives of mathematical expressions are evaluated through a fast and precise method: uncertainties transparently implements automatic differentiation with reverse accumulation. This method essentially consists in keeping track of the value of derivatives, and in automatically applying the chain rule. Automatic differentiation is faster than symbolic differentiation and more precise than numerical differentiation.

The derivatives of any expression can be obtained with uncertainties in a simple way, as demonstrated in the User Guide.

Tracking of random variables

This package keeps track of all the random variables a quantity depends on, which allows one to perform transparent calculations that yield correct uncertainties. For example:

>>> x = ufloat(2, 0.1)
>>> a = 42
>>> poly = x**2 + a
>>> poly
46.0+/-0.4
>>> poly - x*x
42+/-0

Even though x*x has a non-zero uncertainty, the result has a zero uncertainty, because it is equal to a.

If the variable a above is modified, the value of poly is not modified, as is usual in Python:

>>> a = 123
>>> print poly
46.0+/-0.4  # Still equal to x**2 + 42, not x**2 + 123

Random variables can, on the other hand, have their uncertainty updated on the fly, because quantities with uncertainties (like poly) keep track of them:

>>> x.std_dev = 0
>>> print poly
46+/-0  # Zero uncertainty, now

As usual, Python keeps track of objects as long as they are used. Thus, redefining the value of x does not change the fact that poly depends on the quantity with uncertainty previously stored in x:

>>> x = 10000
>>> print poly
46+/-0  # Unchanged

These mechanisms make quantities with uncertainties behave mostly like regular numbers, while providing a fully transparent way of handling correlations between quantities.

Python classes for variables and functions with uncertainty

Numbers with uncertainties are represented through two different classes:

1. a class for independent random variables (Variable, which inherits from UFloat),

2. a class for functions that depend on independent variables (AffineScalarFunc, aliased as UFloat).

Documentation for these classes is available in their Python docstring, which can for instance displayed through pydoc.

The factory function ufloat() creates variables and thus returns a Variable object:

>>> x = ufloat(1, 0.1)
>>> type(x)
<class 'uncertainties.Variable'>

Variable objects can be used as if they were regular Python numbers (the summation, etc. of these objects is defined).

Mathematical expressions involving numbers with uncertainties generally return AffineScalarFunc objects, because they represent mathematical functions and not simple variables; these objects store all the variables they depend on:

>>> type(umath.sin(x))
<class 'uncertainties.AffineScalarFunc'>