Appendix: Technical Details

This technical appendix explains the design of Measurements.jl package, how it propagates the uncertainties when performing calculations, and how you can contribute by providing new methods for mathematical operations.

The Measurement Type

Measurement is a composite parametric type, whose parameter is the AbstractFloat subtype of the nominal value and the uncertainty of the measurement. Measurement type itself is subtype of AbstractFloat, thus Measurement objects can be used in any function taking AbstractFloat arguments without redefining it, and calculation of uncertainty will be exact.

In detail, this is the definition of the type:

immutable Measurement{T<:AbstractFloat} <: AbstractFloat

The fields represent:

  • val: the nominal value of the measurement
  • err: the uncertainty, assumed to be standard deviation
  • tag: a unique identifier, it is used to identify a specific measurement in the list of derivatives. This is automatically created with rand. The result of mathematical operation will have this field set to NaN because it is not relevant for non independent measurements.
  • der: the list of derivates with respect to the independent variables from which the expression comes. Derivatives is a lightweight dictionary type. The keys are the tuples (val, err, tag) of all independent variables from which the object has been derived, while the corresponding value is the partial derivative of the object with respect to that independent variable.

As already explained in the “Usage” section, every time you use one of the constructors

measurement(value, uncertainty)
value ± uncertainty

you define a new independent measurement. This happens because these contructors generate a new random and (hopefully) unique tag field, that is used to distinguish between really equal objects and measurements that only by chance share the same nominal value and uncertainty. For these reasons,

julia> x = 24.3 ± 2.7
24.3 ± 2.7

julia> y = 24.3 ± 2.7
24.3 ± 2.7

will produce two independent measurements and they will be treated as such when performing mathematical operations. In particular, you can also notice that they are not egal

julia> x === y

If you instead intend to make y really the same thing as x you have to use assignment:

julia> x = y = 24.3 ± 2.7
24.3 ± 2.7

julia> x === y

Thanks to how the Julia language is designed, support for complex measurements, arbitrary precision calculations and array operations came with practically no effort during the development of the package. As explained by Steven G. Johnson, since in Julia a lot of nonlinear functions are internally implemented in terms of elementary operations on the real and imaginary parts it was natural to make the type subtype of Real in order to easily work with complex measurements. In particular, it was then chosen to select the AbstractFloat type because some functions of complex arguments (like sqrt and log) take Complex{AbstractFloat} arguments instead of generic Complex{Real}, and any operation on a Measurement{R} object, with R subtype of Real different from AbstractFloat, would turn it into Measurement{F}, with F subtype of AbstractFloat, anyway.


One must carefully take care of correlation between two measurements when propagating the uncertainty for an operation. Actually, the term “correlation” may refer to different kind of dependences between two or more quantities, what we mean by this term in Measurements.jl is explained in the “Usage” section of this manual.

Dealing with functional correlation between Measurement objects, when using functions with arity larger than one, is an important feature of this package. This is accomplished by keeping inside each Measurement object the list of its derivatives with respect to the independent variables from which the quantity comes. This role is played by the der field. This dictionary is useful in order to trace the contribution of each measurement and propagate the uncertainty in the case of functions with more than one argument.

The use of the list of derivatives has been inspired by Python package uncertainties, but the rest of the implementation of Measurements.jl is completely independent from that of uncertainties package, even though it may happen to be similar.

Uncertainty Propagation

For a function \(G(a, b, c, \dots)\) of real arguments with uncertainties \(a = \bar{a} \pm \sigma_{a}\), \(b = \bar{b} \pm \sigma_{b}\), and \(c = \bar{c} \pm \sigma_{c}\), ..., the linear error propagation theory prescribes that uncertainty is propagated as follows:

\[\begin{split}\sigma_G^2 = \left( \left.\frac{\partial G}{\partial a}\right\vert_{a = \bar{a}} \sigma_a \right)^2 + \left( \left.\frac{\partial G}{\partial b}\right\vert_{b = \bar{b}} \sigma_b \right)^2 + \left( \left.\frac{\partial G}{\partial c}\right\vert_{c = \bar{c}} \sigma_c \right)^2 + \cdots \\ + 2 \left(\frac{\partial G}{\partial a}\right)_{a = \bar{a}} \left(\frac{\partial G}{\partial b}\right)_{b = \bar{b}} \sigma_{ab} + 2 \left(\frac{\partial G}{\partial a}\right)_{a = \bar{a}} \left(\frac{\partial G}{\partial c}\right)_{c = \bar{c}} \sigma_{ac} \\ + 2 \left(\frac{\partial G}{\partial b}\right)_{b = \bar{b}} \left(\frac{\partial G}{\partial c}\right)_{c = \bar{c}} \sigma_{bc} + \dots\end{split}\]

where the \(\sigma_{ab}\) factors are the covariances defined as

\[\sigma_{ab} = \text{E}[(a - \text{E}[a])(b - \text{E}[b])]\]

\(E[a]\) is the expected value, or mean, of \(a\). If uncertainties of the quantities \(a\), \(b\), \(c\), ..., are independent and normally distributed, the covariances are null and the above formula for uncertainty propagation simplifies to

\[\sigma_G^2 = \left( \left.\frac{\partial G}{\partial a}\right\vert_{a = \bar{a}} \sigma_a \right)^2 + \left( \left.\frac{\partial G}{\partial b}\right\vert_{b = \bar{b}} \sigma_b \right)^2 + \left( \left.\frac{\partial G}{\partial c}\right\vert_{c = \bar{c}} \sigma_c \right)^2 + \cdots\]

In general, calculating the covariances is not an easy task. The trick adopted in Measurements.jl in order to deal with simple functional correlation is to propagate the uncertainty always using really independent variables. Thus, dealing with functional correlation boils down to finding the set of all the independent measurements on which an expression depends. If this set is made up of \(\{x, y, z, \dots\}\), it is possible to calculate the uncertainty of \(G(a, b, c, \dots)\) with

\[\sigma_G^2 = \left( \left.\frac{\partial G}{\partial x}\right\vert_{x = \bar{x}} \sigma_x \right)^2 + \left( \left.\frac{\partial G}{\partial y}\right\vert_{y = \bar{y}} \sigma_y \right)^2 + \left( \left.\frac{\partial G}{\partial z}\right\vert_{z = \bar{z}} \sigma_z \right)^2 + \cdots\]

where all covariances due to functional correlation are null. This explains the purpose of keeping the list of derivatives with respect to independent variables in Measurement objects: by looking at the der fields of \(a\), \(b\), \(c\), ..., it is possible to determine the set of independent variables. If other types of correlation (not functional) between \(x\), \(y\), \(z\), ..., are present, they should be treated by calculating the covariances as shown above.

For a function of only one argument, \(G = G(a)\), there is no problem of correlation and the uncertainty propagation formula in the linear approximation simply reads

\[\sigma_G = \left\vert \frac{\partial G}{\partial a} \right\vert_{a = \bar{a}} \sigma_a\]

even if \(a\) is not an independent variable and comes from operations on really independent measurements.

For example, suppose you want to calculate the function \(G = G(a, b)\) of two arguments, and \(a\) and \(b\) are functionally correlated, because they come from some mathematical operations on really independent variables \(x\), \(y\), \(z\), say \(a = a(x, y)\), \(b = b(x, z)\). By using the chain rule, the uncertainty on \(G(a, b)\) is calculated as follows:

\[\sigma_G^2 = \left( \left(\frac{\partial G}{\partial a}\frac{\partial a}{\partial x} + \frac{\partial G}{\partial b}\frac{\partial b}{\partial x}\right)_{x = \bar{x}} \sigma_x \right)^2 + \left( \left(\frac{\partial G}{\partial a}\frac{\partial a}{\partial y}\right)_{y = \bar{y}} \sigma_y \right)^2 + \left( \left(\frac{\partial G}{\partial b}\frac{\partial b}{\partial z}\right)_{z = \bar{z}} \sigma_z \right)^2\]

What Measurements.jl really does is to calulate the derivatives like \(\partial a/\partial x\) and \(\partial G/\partial x = (\partial G/\partial a)(\partial a/\partial x) + (\partial G/\partial b)(\partial b/\partial x)\), and store them in the der field of \(a\) and \(G\) respectively in order to be able to perform further operations involving these quantities.

This method is also described in [GIO16].

Defining Methods for Mathematical Operations

Measurements.jl defines new methods for mathematical operations in order to make them accept Measurement arguments. The single most important thing to know about how to define new methods in the package is the Measurements.result. This function, not exported because it is intended to be used only within the package, takes care of propagating the uncertainty as described in the section above. It has two methods: one for functions with arity equal to one, and the other for any other case. This is its syntax:

result(val::Real, der::Real, a::Measurement)

for functions of one argument, and

result(val, der, a)

for functions of two or more arguments, in which der and a are the collections (tuples, arrays, etc...) of the same length. The arguments are:

  • val: the nominal result of the operation \(G(a, \dots)\);
  • der: the partial derivative \(\partial G/\partial a\) of a function \(G = G(a)\) with respect to the argument \(a\) for one-argument functions or the tuple of partial derivatives with respect to each argument in other cases;
  • a: the argument(s) of \(G\), in the same order as the corresponding derivatives in der argument.

In the case of functions with arity larger than one, der and a tuples must have the same length.

For example, for a one-argument function like \(\cos\) we have

cos(a::Measurement) = result(cos(a.val), -sin(a.val), a)

Instead, the method for subtraction operation is defined as follows:

-(a::Measurement, b::Measurement) =
    result(a.val - b.val, (1, -1), (a, b))

Thus, in order to support Measurement argument(s) for a new mathematical operation you have to calculate the result of the operation, the partial derivatives of the functon with respect to all arguments and then pass this information to Measurements.result function.