There are a total of 14 mathtext image miscompares, all due totest_mathtext_rendering[*-mathtext-*-81] i.e. the tests for\left(X\right)_a^b.

The reason is thatcm andstixsans do not expand brackets the waystix,dejavusans anddejavuserif do.

The brackets are not high enough in case ofstixsans, and not low enough in case ofcm. So, the shifts calculated for one font may not look good when used on another.

I used shifts which (I hope) would make fractions more closely resemble what LaTeX does. I think the result is reasonable.

Would love to know what experts think.

This looks like really good work, bt I feel un-able to properly review it. Who is our resident expert in mathtext?

I thought I'd revive this, since I now have a better understanding of TeX's algorithm. I implemented the calculations described in the book (with a few shortcuts). However, the CM font behaves weirdly: it reports a wrong x-height, which throws off the calculations.

import matplotlib.pyplot as plt(fig, ax) = plt.subplots()ax.text(.1, .5, r'$\mathrm{x}\alpha_{i+1}^j$', size=50)plt.rc('mathtext', fontset='cm')ax.text(.5, .5, r'$\mathrm{x}\alpha_{i+1}^j$', size=50)plt.show()

DejaVu Sans correctly reports the x-height as 38 (I verified this by counting the pixels 😅), whereas CM reports it as 59, which is clearly wrong, as this image shows.

(The icons are missing because I am using a WSL installation of Matplotlib on Linux.)

Which means that even though the algorithm is correct, the alignment is incorrect for CM. Should we just shelve this?

I found the problem. DejaVu Sans does not store the x-height, so it is actually calculated in theget_xheight function. Whereas CM does, but the value it provides is incorrect.

If I force the "poor man's x-height" calculation for CM, the result is reasonably good.

However, making that change on themain branch breaks 126 Mathtext tests (incidentally, all of them are CM rendering tests).

Shortcuts I took for the algorithm

• The default values ofshift_up andshift_down depend on two font constants,supdrop andsubdrop, respectively. However, Mathtext usessubdrop for both (as the comment in theFontConstantsBase class mentions), so I did the same.
• The actual value ofshift_up depends on the current style (display, text, script or script-script), and may be calculated usingsup1,sup2 orsup3. However, superscripts and subscripts in Mathtext do not know about the current style (Mathtext only has thesup1 constant); they are simply one size smaller than the nucleus. I've also used onlysup1.

Closing. Superseded by#22852.