Re: Today's evil Bayesian math problem

From: Eliezer S. Yudkowsky (sentience@pobox.com)
Date: Mon Sep 15 2003 - 10:39:06 MDT

  • Next message: Randy S: "Extropians are not rich"

    Hal Finney wrote:
    >
    > I imagine a more complex proof could show that the probability of white
    > on both of any two rounds is constant, and presumably this could be
    > generalized to any n rounds. But I don't see how to prove all this
    > in a simple way.

    If you were to prove that for any N sequential samples from this pot, the
    probability of getting a sequence with M red chips and (N - M) white chips
    is a constant regardless of the ordering of white and red chips, it would
    prove both constant P(Wk) and constant P(Wj|Wk), j != k.

    The above property, that an outcome containing M red chips and (N - M)
    white chips has a constant probability regardless of the ordering of red
    and white chips, is known as exchangeability.

    Suppose the sequence is exchangeable. No matter how long the sequence
    goes on, P(W1) and P(W1&W2) is constant, since this is computed in the
    first two draws. P(W2) = P(W1), because P(R1&W2) + P(W1&W2) = P(W1&R2) +
    P(W1&W2). P(W1&W2) = P(W2&W3) because P(W1&W2&W3) + P(W1&W2&R3) =
    P(R1&W2&W3) + P(W1&W2&W3). The same holds for arbitrary P(Wi) and P(Wj|Wk).

    Actually, if a sequence is exchangeable, we can swap any i, j != k in the
    joint distribution, and properties such as P(Wi), P(Wi|Wk), P(Wi&Wk), and
    so on, will remain constant.

    To show that the sequence is exchangeable, consider any R, W of initial
    red and white chips in the bowl, with a total number of chips N. The
    probability of the sequence RWWRR is then given by:

    R W W+1 R+1 R+2
    - * --- * --- * --- * ---
    N N+1 N+2 N+3 N+4

    Changing the ordering of the R and W chips will not change the total
    factors in the numerator and the denominator.

    The entire proof is exactly analogous to the case of a bowl where, instead
    of *adding* one chip of the color drawn, we draw a chip and discard it.
    If we started with 5 red and 3 white chips, and drew a white chip, we
    would be down to 5 red and 2 white chips, and so on, until we ran out of
    chips after 8 rounds. Obviously, the probability of drawing a red or
    white chip on any round N is the same as on round 1, if you have no
    information about the chips previously drawn; and the probability of
    drawing two red chips on any two rounds is the same; and the probability
    you drew a red chip on round 2 if you drew a red chip on round 4, is the
    same as the probability of drawing a red chip on round 2 if you drew a red
    chip on round 1, since in either case there is one less red chip to go
    around. You can imagine the chips "scattered" over N holes,
    simultaneously, rather than being placed in sequence - this is what it
    means for the sequence to be exchangeable. Swapping the labels of "round
    1" to "round 3" has no effect on how the chips land in the holes.

    If you start out with R red chips and W white chips, and draw without
    replacement, the probability of getting a particular sequence is:

    R W W-1 R-1 R-2
    - * --- * --- * --- * ---
    N N-1 N-2 N-3 N-4

    Which is what makes this distribution exchangeable:

    > Evil Hint #3:
    >
    > The classical version of the problem has a bowl where we sample without replacement - take the chip, and throw it away. Suppose there are M red chips and N chips. If you don't know what you got on the first round, the probability of getting a red chip on the second round is still M/N. The probability of getting a red chip on the second round, given that you got a red chip on the fourth round, is just (M-1)/(N-1), since there's one less red chip to go around. We can think of sprinkling the chips over holes, instead of drawing them in successive rounds, and it doesn't matter if you swap around the labels on the holes.
    >
    > The evil version of the bowl, instead of running out of the chips already sprinkled, has more and more of them, like a P2P file-sharing network. But that doesn't change anything important.

    The math is also not affected if you instead add three chips of the color
    drawn, or if you subtract two chips of the color drawn, and so on.

    Christian Rovner has an interesting way to visualize this, which is that
    the drawing procedure always leaves the bowl the same shade of pink
    afterward. If you remove the chip drawn on each round, you're removing a
    chip with a 3/8 probability of being white, from a bowl with a 3/8
    probability of producing white chips. If you add a chip of the color
    drawn on each round, you're adding a chip with a 3/8 probability of being
    white, to a bowl with a 3/8 probability of producing white chips. If you
    know that you drew a white chip on the first round, then on future rounds
    with unknown draws you're adding chips with a 4/9 probability of being
    white. And so on. It doesn't matter whether you *know* that the pot has
    5 red and 3 white chips, as on the first round, or if you're working with
    some probability distribution of mixes that add up to 3/8; it's all the
    same shade of pink. The drawing procedure leaves the pinkness unchanged,
    both for the main distribution, and for any differently shaded
    subdistributions that you want to consider separately.

    See also Chapter 3 of E.T. Jaynes's "Probability Theory: The Logic of
    Science". (I did not get the problem from there; I made it up after
    reading it. Though I am sure it is not original.)

    -- 
    Eliezer S. Yudkowsky                          http://singinst.org/
    Research Fellow, Singularity Institute for Artificial Intelligence
    


    This archive was generated by hypermail 2.1.5 : Mon Sep 15 2003 - 10:50:04 MDT