mp-units/latest/blog/2025/01/15/bringing-quantity-safety-to-the-next-level/

[giscus[bot]]

mp-units/latest/blog/2025/01/15/bringing-quantity-safety-to-the-next-level/

The quantities and units library for C++

https://mpusz.github.io/mp-units/latest/blog/2025/01/15/bringing-quantity-safety-to-the-next-level/?h=

[doocman]

I see another benefit of doing it this way: we also simplify external users of the code. For example, I am using Eigen as a raw linear algebra library and wrap it up in classes that incorporates mp-units in an attempt to achieve type-safety. For example, a 'GaussianRandomVariable' is by itself a vector of means and a covariance matrix. Adding unit-safety for this does mean when I incorporate Eigen that I essentially have to create a 'GaussianRandomQuantity' and a 'GaussianRandomQuantityPoint' that are my analogues to 'mp_units::quantity' and 'mp_units::quantity_point'. Having the 'point-vs-delta' inside the reference itself would potentially slash the number of classes needed by the user in half.

[mpusz]

@FrancoisCarouge, I will publish a new article on mp-units blog about Linear Algebra and quantities soon. I would like to advertise your TypedLinearAlgebra project there. The draft can be found here. I hope it is correct and you are OK with that?

[FrancoisCarouge]

This is exciting. I am favorable to this article and the content looks appropriate. Thank you for the publicity.

[doocman]

@mpusz Nice, I'll look into it! However, the example covers 'vectors', but when I want to represent "random variables", I also need (square) variance-covariance matrices. They would be a "quantity of squared reference", regardless if the random variable represents a quantity or a quantity point.
I also feel like this touches on the broader subject of wheter one should make a 'struct of quantities' or a 'unit-aware struct'. To demonstrate what I mean we can take the random-variable example. My current approach is that I have a 'raw' type that has no units, it is just essentially a struct with a vector (the multivariate mean) and a matrix (the variance-covariance values). Then I have a separate 'wrapper' struct that wraps the raw type with a bit more unit-aware APIs. But you could imagine making the random-variable-type unit/reference aware immediately by templating the mean and CV-rep and enforce with concepts that they have compatible units and sizes. And the last alternative could even be to add all necessary customization points to allow a 'quantity' with the random variable as the rep type. Personally I see pros and cons with all these approaches, and I wonder if there is a recommendation from the library how to conceptually think about this and similar circumstances.

[mpusz]

@doocman You can find answers to most of your questions in https://mpusz.github.io/mp-units/latest/blog/2026/06/23/units-meet-linear-algebra-two-approaches-two-problems/. If something is not covered there, please let me know, and I will be happy to help.

[FrancoisCarouge]

@doocman This sample of a Kalman filter in 2D with six states, constant acceleration dynamic model, no control, may give you a few ideas for the covariance.