Suggestion structuring to make the calculations more understandable:

1. Calculate desired figure height and width in pixel
2. Create figure (needs calculation of figure size in inches)
3. Add axes (needs calculation of dimensions in figure coordinates)

A and B are purposefully different widths, so I don't think this is correct.

You are right. The numbers / expressions are not correct. Still the idea (which possibly can be communicated more explicitly) is that you can place everything at exact pixel positions. And to do that, you have to follow the three steps above.

Possibly now correct. - I don't have an interpreter at hand, so can't check, but I hope you get the idea.

I do get the idea, but not convinced pre-calculating the figure dimensions buys you anything. Also, imagine we were to add a third or fourth image in there, we would need to change the code in two places.

Just for context, I originally wrote as a for-loop to allow flexibility, but backed off on that as perhaps too magical.

Just for context, I originally wrote as a for-loop to allow flexibility, but backed off on that as perhaps too magical.

👍 Examples must show the pattern, not necessarily a generic solution if that makes the code harder to understand.

I had difficulty following your code. Reflecting on it, the reasons were

1. lines 100-115 define a lot of variables, which are sometimes
   a) tersely named (figh)
   b) cryptic data structuresposB = [left, figh - buffer - ny, nx, ny]
   c) a mixture of definitions for later use (e.g.figh) and temporary local variables that are reassigned (left,nx, `ny´)
2. It's unclear that these are all in pixels
3. Transformation calculations to inches (for figsize) and figure units (for add_axes) are hard to understand, because both the input units (see 2.) and the resulting units are not clear.

I tried a more minimal refactor of your code to alleviate these points, but I ended up with almost the same code as I've proposed above 🤷‍♂️.

I'm a bit torn on the motivation for this: You can get away without by using an appropriate figsize (which is much simpler).

The advantages I see is

1. pixel-perfect positioning (if you need that)
2. in this case 1:1 mapping from data points to visual pixels.

Maybe these should be highlighted more.

Optional: Slightly more concise wording: