It's pretty easy to just run "Save as PDF" on each chapter individually, and then stitch them together with:

    pdftk $(ls -tr *.pdf) cat output DeepLearningBook.pdf

Out of respect for the authors' contract, I won't post the resulting file here, but anyone can reproduce it with about 5 minutes of work.

In case anyone else has trouble with the "Save as PDF" part, I did it successfully using Firefox on OSX: from the print menu, choose "show details" and change all headers/footers to "--blank--" so there's no ugly URLs/dates in the corners, then press PDF -> "Save as PDF". (In Chrome, on the other hand, saving/printing as PDF chopped each page into quarters for me...) At least in the one chapter I've saved so far, all the math notation renders just like in the HTML version.

That would give you a PDF version of the web page, right? Wouldn't you still have problems with "some things like subscript expressions do not render correctly"?

Unfortunately, according to the site, the HTML pages themselves have problems, such as missing parens and incorrect math symbols.

Doubt it. We talked to MIT Press about publishing our book and they were extremely hostile to having any way for our readers to produce an un-DRM'd PDF, which is how our book is currently distributed.

Actually, they were mostly just very hostile and unpleasant. This is among the many reasons why we're probably staying independent and printing the book ourselves - so that the digital version stays unencumbered among other things.

Ugh, MIT Press, the same organization that won't allow easy distribution of things like the Wizard Book.