Below is a MWE of a figure that includes two legends, one in two-column format and the other in one-column format. As written, the horizontal alignment between the two legends depends on details such as whether I generate the figure in a Jupyter notebook or with a script. I’ve experimented with transformations and have been unable to find a solution that robustly centers the legend keys of the single-column legend at the horizontal center of the two legend keys in the two-column legend. Any suggestions?
fig, ax = plt.subplots()
x = np.linspace(0, 1)
m = np.arange(1, 4)
lgnd_loc = (0.1, 0.1)
lbl = []
ln_pos = []
ln_neg = []
for k in m:
ln_pos += ax.plot(x, k * x)
ln_neg += ax.plot(x, -k * x)
lbl.append(f"m = {k:d}")
ln_zero = ax.plot(x, np.zeros_like(x), label=f"m = 0")
# Create first part of the legend
lgnd_2col = ax.legend(
title="Pos. Neg.",
alignment="left",
handles=[*zip(ln_pos, ln_neg)],
labels=lbl,
frameon=False,
fontsize="small",
handler_map={tuple: HandlerTuple(ndivide=None, pad=1.0)},
handlelength=6.0,
loc="lower left",
bbox_to_anchor=lgnd_loc,
borderpad=0,
)
# Add first part of legend as an artist to create second part
lgnd_2col_art = ax.add_artist(lgnd_2col)
# Create second part of the legend
lgnd_loc_new = (lgnd_loc[0] - ax.transAxes.inverted().transform((60, 0))[0], lgnd_loc[1])
lgnd_1col = ax.legend(
alignment="center",
handles=ln_zero,
ncols=1,
frameon=False,
fontsize="medium",
loc="upper left",
bbox_to_anchor=lgnd_loc_new,
borderpad=0,
)
plt.show()
Thanks for the suggestion—I would prefer to have all of the keys the same size, though. I discovered I could do this by creating an empty handle with ln_empty = ax.plot([], []) and replacing the ln_zero in your example with (ln_empty, ln_zero). I was then able to center it by adjusting xdata for ln_zero.
That said, in my actual application I wanted the extra space between the last legend entry and the rest, and I also wanted better control over the column headings than the title parameter allowed. In the end, I used two calls to legend, one for the positive and negative values and one for the zero value, one call to fig.canvas.draw() to get intermediate positioning information, a call to lgnd_loc_new.set_bbox_to_anchor(...) to reposition the second legend with respect to the first, and finally two calls to text to place the column labels for the first legend. The call to fig.canvas.draw() feels a bit clunky, but I couldn’t find a better way to get the necessary legend positioning information.
The legend guide and the transformation tutorial were both helpful, but neither provided enough guidance for the issues I encountered with this problem. If anyone knows of a good reference for this kind of thing, please let me know.
