COMPUTING SCIENCE
Rumours and Errours
Brian Hayes
Back to the Stacks
Not all of my confusion was cleared up by the discovery of this
error. In particular, the algorithm that I now knew to be incorrect
for the Daley-Kendall model still seemed to give the right answer
for the Maki-Thompson rules. To make sense of this situation, I
finally went back to the library to find out what the original
authors had said.
Daley and Kendall are Daryl J. Daley, now of the Australian National
University, and David G. Kendall, a distinguished statistician and
probabilist at the University of Cambridge. Their paper, published
in 1965, is a model of lucid exposition, which would have spared me
all my stumbling in the dark—and for that reason I'm glad I
didn't see it sooner. The correct calculation of probabilities is
stated very clearly (there's a factor of 1/2 in the expression for
the spreader-spreader interaction). Furthermore, the origin of the
mysteriously precise number 0.203188 is made plain. Those six digits
come not from a discrete-event simulation like the ones I had
designed but from a continuous, differential-equation version of the
model. The number θ is a solution of the equation:
θ e 2(1-
θ) = 1.
(This equation brings us back almost to the Lambert W
function, WeW
.)
Maki and Thompson are Daniel P. Maki and Maynard Thompson of Indiana
University, who discussed rumors in a 1973 textbook,
Mathematical Models and Applications. They described
the rumor-passing process in terms of telephone calls, and they
limited their attention to calls placed by spreaders; because of
this asymmetry, only the middle row of the matrix in Figure 4 enters
into the calculation, and my first program was in fact a correct
implementation of their model. (At least I got something
right.) It is almost a coincidence that Maki and Thompson arrive at
the same value of θ as Daley and Kendall: Their
spreader-spreader interactions are twice as likely but have only
half the effect.
The paper by Belen and Pearce that launched me on this adventure
also deserves a further comment. The phrase "general initial
conditions" in their title refers to rumors initiated not by a
single spreader but by many. One might guess that with enough
spreaders, the rumor must surely permeate the entire society, but
Belen and Pearce show otherwise. Measuring the fraction of those
originally ignorant who remain ignorant when the rumor has finished,
they find that this fraction actually increases along with the
number of initial spreaders, tending to a maximum of 1/e,
or about 0.368. In other words, as more people spread the news, more
people fail to hear it. The reason is simply that the multitude of
spreaders quickly stifle one another.
By now the mathematics of rumors has acquired a vast literature.
Variant models track competing rumors and counter-rumors or allow
people to meet more than two at a time. Still more models study the
progress of rumors through networks or lattices rather than
structureless mixed populations. I have not yet had a chance to make
any errors in exploring these systems.
» Post Comment