blog/2026/08/06/measurement-uncertainty-and-measured-constants/

[giscus[bot]]

mp-units/latest/blog/2026/08/06/measurement-uncertainty-and-measured-constants/

The quantities and units library for C++

https://mpusz.github.io/mp-units/latest/blog/2026/08/06/measurement-uncertainty-and-measured-constants/

[FrancoisCarouge]

This post reminds me of interval arithmetic: https://en.wikipedia.org/wiki/Interval_arithmetic
You hint the solution to better propagation, correlation, is to use an uncertainty library, as a quantity representation type, I assume?

[mpusz]

Related, but they answer different questions. Interval arithmetic propagates guaranteed bounds and combines them worst case, so absolute values add. uncertain<T> propagates a standard uncertainty statistically, first order and in quadrature, which is the GUM formula and gives a coverage estimate rather than a bound. What both share is the dependency problem: every operand is treated as independent, which is why x - x is not zero in either case.

And yes, your assumption is right. uncertain<T> is not privileged in any way. It just satisfies RepresentationOf, so an interval type or a full covariance-tracking one plugs into the same slot without any changes to the library. If such a type also wants the declared uncertainty of measured constants that show up in unit conversion factors, it opts in with an ADL-findable fold_conversion_uncertainty(value, relative_uncertainty) and combines it with whatever the value already carries.