Thanks! I did not realize there were different compression levels. So I discovered optipng. If I scan to PBM, that can be fed into optipng set for max compression:

$ convert sample_scan.pbm sample_IM.png; # ImageMagick for comparison
$ optipng -out sample_optipng_default.png sample_scan.pbm; # run with defaults (moderate effort, -o2)
$ optipng --strip all -o7 -out sample_optipng_min.png sample_scan.pbm; # max effort (smallest result)
$ pdfjam --a4paper -o sample_optipng_min.pdf sample_optipng_min.png
$ cjb2 -lossless sample_scan.pbm sample_lossless.djvu; # DjVu for comparison
$ du -h …
24K     sample_lossless.djvu
68K     sample_optipng_min.png
68K     sample_optipng_default.png
72K     sample_optipng_min.pdf
80K     sample_IM.png

The tightest PNG compression w/metadata stripped (or even just the optipng default) is much smaller than ImageMagick’s output. The PDF produced by pdfjam is bigger than the png it contains (rightfully so). So I’m pleased with this. I’ll replace ImageMagic with optipng. And if a scanned doc is just 1 page, I won’t even bother with pdfjam and just keep the png.

I thought PNG was inherently lossless but the pngquant tool proves there is a lossy compression algo. Although I suppose anything can be lossy with image processing done independently. I did not fiddle with pngquant because it seems to only be relevant to color reduction, which makes no sense on a bitonal image.

DjVu is a clear winner. ⅓ the size and still lossless. But there is no LaTeX pkg for DjVu so the smaller size obviously comes with a big trade off. But then we must consider that a DjVu file can be directly converted to PDF (LaTeX importable) using ddjvu. So we have:

$ ddjvu -format=pdf -quality=150 sample_lossless.djvu sample_ddjvu_lossless.pdf; # 150→highest possible quality
$ pdfimages -all sample_ddjvu_lossless.pdf innards; # because we want to see what’s inside that thing
$ du -h …
24K     sample_lossless.djvu
44K     sample_ddjvu_lossless.pdf
44K     innards-000.ccitt
4.0K    innards-000.params

$ cat innards-000.params; # to get an idea of what the CCITT is (fax or whatever)
-4 -P -X 2464 -B -M

The PDF size (from DjVu) is still smaller than the smallest PNG. And we can see that ddjvu did not do anything foolish like stuff a jpg in the PDF. It’s not fun to work with CCITT files, but I suppose that would be rare. In the end, DjVu seems like a winner for archiving because ddjvu can always make a PDF that is still smaller than a PNG, and LaTeX can work with that.