[Bug]: Alignment of minus sign when using LaTeX

[johannfaouzi]

Bug summary

There seems to be a different alignment for the minus sign depending on if the string contains only the minus sign or more characters when LaTeX is used.

Code for reproduction

import matplotlib.pyplot as plt
plt.rcParams.update({"text.usetex": True})

plt.text(0, 0, '$-3$', fontsize=40)
plt.text(0.2, 0, '$-$', fontsize=40)
plt.text(0.28, 0, '$3$', fontsize=40)
plt.ylim(-0.1, 0.1);
plt.axis('off');

Actual outcome

Expected outcome

Additional information

No response

Operating system

macOS Sonoma 14.6.1

Matplotlib Version

3.9.2

Matplotlib Backend

module://matplotlib_inline.backend_inline

Python version

3.10.13

Jupyter version

6.5.4

Installation

conda

[oscargus]

Matplotlib just grabs the generated image, cropped to minimal size, and position it in the correct location.

Although it may seem silly that this example doesn't work, one may also imagine much more complicated equations with multiple fractions etc, where it is clear that there is no way to really align two different equations as if they were written as one. That information is simply missing in the output Matplotlib gets.

[timhoffm]

That information is simply missing in the output Matplotlib gets.

One could add that information by adding a constant prefix including letters with descenders and ascenders, e.g. "gH", and then strip away the prefix.

[QuLogic]

I thought we did measure something with lp somewhere, but not sure if that is linked to this code path.

[anntzer]

Actually the information is already available (#16476), and indeed the bug is not present in any backend except agg; the problem arises because, even though the dvi file correctly places the minus sign well above the intended baseline position, the dvi parser cannot report the raster image of the minus sign as having "negative" depth (likely ultimately because in _output(), _font._height_depth_of(g) doesn't know about these negative depths). Other backends (including mplcairo's raster backend) don't have this issue because they simply read the baseline and glyph coordinates from the dvi file and redraw the glyphs themselves, which is probably the better approach in any case (see also part of https://github.com/matplotlib/matplotlib/issues/20469#issuecomment-864470954 as well as the paragraph I've deleted at https://github.com/matplotlib/matplotlib/pull/28810/files.