21 April 2002
Dear Warren Smith
You are right of course. Please excuse me for being feebleminded;
I should have seen that I was working with an 
(mn+sm)*(ns+ms)*(sm+ms) tensor, not (2mn)*(2ns)*(2sm) 
which would correspond to independent (m,n,s)+(m,n,s).

Your "reward" [a cheque for $2.56] is enclosed. 
I hope you don't mind that BIG mistakes are worth no 
more than trivial typos... that's life! Anyway this 
serious error will be absent from the 10th printing 
of the 3rd edition, after it had survived more than 20 years 
before being noticed.

I'm extremely grateful for your help and your patience.

--Cordially,
Donald E. Knuth
Professor Emeritus of the Art of Computer Programming,
Stanford University, Stanford CA 94305.

cc: Victor Pan - Victor! I apologize for bothering you with 
something I should have been able to handle by myself. But thanks 
for helping to straighten me out.

==================================================================

[This letter is re me pointing out that Knuth's exercise
4.6.4-60b in the 3rd ed. of his "Art of computer programming" vol 2 
"seminumerical algorithms" (error present in all editions 1-3)
wrongly claimed to be able to solve two independent  mXn * nXs
matrix multiplication problems.
Knuth then said really he was right and we argued the 
point by mail for months before he sent me this
letter and a cheque for $2.56, one "hexadecimal dollar."  
This actually appears to be the longest-standing serious error in 
this classic book... at least, so far.  I spotted it while working
on my own papers on fast matrix algorithms, 2002.]