How is the MD2 hash function S-table constructed from Pi?

For fun, I'm learning more about cryptography and hashing. I'm implementing the MD2 hash function following RFC 1319 (https://www.rfc-editor.org/rfc/rfc1319). I'll preface by saying I know there are libraries, I know this is an old hash, and I do not intend to use this in anything real-world, so don't anyone freak out.

In learning about S-tables I notice MD2 uses an S-table based on Pi. What I don't understand, is how these numbers relate to Pi. And oddly, I can't find anywhere that describes it. The RFC just says:

This step uses a 256-byte "random" permutation constructed from the digits of pi.

In the reference code, there is the following comment:

Permutation of 0..255 constructed from the digits of pi. It gives a "random" nonlinear byte substitution operation.

Looking at the first few bytes:

41, 46, 67

I cannot see how you get these numbers from Pi. I've tried looking at the binary representation of Pi, and these numbers do not seem to match up with the first bytes of pi.

My usual Googling just leads me back to the RFC or other implementations which seem to copy the above comment exactly. No where can I find where the construction of this table from Pi is explained.

• Are you asking how are these numbers $\pi$ or why are these numbers taken from $\pi$? Nov 25, 2013 at 12:29
• Interesting question, I don't know the answer yet (and couldn't find it with some googling). Might be actually not a nothing-up-my-sleeve-number, but only pretending to be one ;-) Nov 25, 2013 at 20:10
• The authors of Network Security: Private Communications in a Public World cast some doubt on whether this table is actually based on Pi. Nov 26, 2013 at 0:30
• Also note that RC2 (in RFC2268) also has a permutation supposedly based on the digits of $\pi$ in its key schedule. The final table in that RFC is also unexplained (and used in other algorithms from RSA as well). These were also never explained, AFAIK. Dec 10, 2013 at 16:06
• I would say the proper answer to this question would be a concise enough explanation that I could implement an algorithm that took an input of pi, and output a result that was this S-table. Aug 1, 2014 at 5:17

I sent an email to Ron Rivest and got an answer back.

The digits of $\pi$ are used as a sort of random number generator that is used in the Durstenfeld shuffle (see also Knuth vol 3, sec 3.4.2).

Below is some pseudocode adapted from the description and code he sent me.


S = [0, 1, ..., 255]
digits_Pi = [3, 1, 4, 1, 5, 9, ...] # the digits of pi

def rand(n):
x = next(digits_Pi)
y = 10

if n > 10:
x = x*10 + next(digits_Pi)
y = 100
if n > 100:
x  = x*10 + next(digits_Pi)
y = 1000

if x < (n*(y/n)): # division here is integer division
return x % n
else:
# x value is too large, don't use it
return rand(n)

for i in 2...256: #inclusive
j = rand(i)
tmp = S[j]
S[j] = S[i-1]
S[i-1] = tmp


The next function just steps through the digits of pi. You'll notice that rand does not use some of it's possible outputs. Dr. Rivest says that using these values would not give a uniform distribution as they are too large.

• For what it’s worth: C proggy that creates the MD2 S-Box Aug 15, 2014 at 4:47
• @e-sushi: I was pointed that the above rot to 404. If you post an updated link in a comment below, I'll merge it in the original comment so as to keep the date.
– fgrieu
Sep 20, 2022 at 18:51

Based on mikeazo's pseudocode, here's a fully working program in Python 3 that will generate the S-table:

#!/usr/bin/env python3

import io

# We need 722 decimal digits of pi, including the integer part (3).
pi = io.StringIO('3'
'1415926535897932384626433832795028841971693993751058209749445923078164062'
'8620899862803482534211706798214808651328230664709384460955058223172535940'
'8128481117450284102701938521105559644622948954930381964428810975665933446'
'1284756482337867831652712019091456485669234603486104543266482133936072602'
'4914127372458700660631558817488152092096282925409171536436789259036001133'
'0530548820466521384146951941511609433057270365759591953092186117381932611'
'7931051185480744623799627495673518857527248912279381830119491298336733624'
'4065664308602139494639522473719070217986094370277053921717629317675238467'
'4818467669405132000568127145263560827785771342757789609173637178721468440'
'901224953430146549585371050792279689258923542019956112129021960864034418')

# Generate a pseudorandom integer in interval [0,n) using decimal
# digits of pi (including the leading 3) as a seed.
def pi_prng(n):
while True:
# based on n, decide how many of digits to work with
if   n <=   10: x, y = int(pi.read(1)),   10
elif n <=  100: x, y = int(pi.read(2)),  100
elif n <= 1000: x, y = int(pi.read(3)), 1000
else: raise ValueError(f'Given value of n ({n}) is too big!')

# Compute the largest integer multiple of n not larger than y.
# If x is smaller than that, we can safely return it modulo n,
# otherwise we need to try again to avoid modulo bias.
if x < (n * (y // n)): return x % n

# Fischer-Yates/Durstenfeld shuffling algorithm, except counting up
# XXX Does counting up bias the results?
S = list(range(256))
for i in range(1,256):
# generate pseudorandom j such that 0 ≤ j ≤ i
j = pi_prng(i+1)
S[j], S[i] = S[i], S[j]

# Print the S-table as shown on Wikipedia.
for i in range(16):
prefix = '{ ' if i ==  0 else '  '
suffix = ' }' if i == 15 else ','
row = S[i*16:i*16+16]
print(prefix + ', '.join(map(lambda s: '0x%02X' % s, row)) + suffix)