648 IV.
Branches of Mathematics
important property it should satisfy is that all the sam- Sn
ple paths are continuous: that is, for fixed ω the point
Bt (ω) depends continuously on t. 2
Physical observations, as well as the contributions of
Einstein and Bachelier described above, suggested a few 1
other properties that Brownian motion should satisfy.
It then became a substantial mathematical problem to 0
n
prove that there existed a stochastic process with those
properties. Wiener was the first person to establish this, −1
which he did in 1923, and for this reason the mathemat-
ical concept of Brownian motion is sometimes called −2
the Wiener process.
The most famous names of probability theory in
the twentieth century, including kolmogorov [VI.88], Figure 1 The accumulated gain in coin tossing.
Lévy, Itô, and Doob, all made important contributions
to the study of Brownian motion. Detailed properties cise, the binomial distribution tells you that the prob-
of the sample paths have received particular attention, ability
that the number of heads after n tosses is k is
ever since the physicist Jean Perrin observed that these n
2−n k . If it is k, then Sn = k − (n − k) = 2k − n.) What
functions are nowhere differentiable (despite Wiener’s is more, if m > 0 then Sm+n − Sm = εm+1 + · · · + εm+n ,
later result that they were continuous). The nondif- which is also a sum of n of the εi , so the distribution
ferentiability of Brownian trajectories led Itô to intro- of Sm+n − Sm is the same as that of Sn . Note too that it
duce a differential calculus for functions of Brownian is independent of the values of S0 , S1 , . . . , Sm .
motion and more general stochastic processes. This The name “random walk” comes from the fact that
Itô stochastic calculus, which will be briefly presented we can think of the sequence S0 , S1 , S2 , . . . as taking
in section 4, has found many applications in many a succession of random steps, each of either 1 or −1.
different areas of modern probability theory. Brownian motion can be thought of as the limit of this
process as the number of steps gets larger and larger
2 Coin Tossing and Random Walks and the sizes of the steps get correspondingly smaller.
To see what “correspondingly” means here, we ap-
One of the easiest ways to understand Brownian motion
peal to the central limit theorem [III.71 §5], which
is via another important concept of probability: that
tells us about the limiting behavior of the distribution
of random walks. Suppose you were to play a game
of Sn when n gets large. Or rather, it tells us about the
where you repeatedly tossed a coin, winning €1 if √
distribution of (1/ n)Sn : the reason it is appropriate
it came up heads, and losing €1 if it came up tails. √ √
to divide by n is that n is the standard deviation
One could then define a sequence of random vari-
[III.71 §4] of Sn . This one can think of as its “typical
ables S0 , S1 , S2 , . . . , where Sn represented your total
size”: thus, when we divide by it, the “renormalized”
gain (which could well be negative) after n tosses of
distribution will have “typical size” 1 (and therefore we
the coin. Two simple properties of this sequence are
will get the same typical size for each n).
that S0 must be 0 and that Sn and Sn−1 always differ
The precise information that the central limit theo-
by 1. One can see this in figure 1, which plots a graph
rem gives us is that for any real numbers a and b with
of the sequence in the case where the coin tosses are √
a < b, the probability that a < (1/ n)Sn < b tends to
HTTTHTHHHTHHTH ….
b
A third property becomes clear if one defines another 1 2
√ e−x /2
dx
sequence of random variables ε1 , ε2 , . . . , representing 2π a
the outcome of each toss of the coin. These are inde- as n tends to ∞. That is, the limiting behavior of the
√
pendent, and each εn takes the value 1 with probability distribution of (1/ n)Sn is Gaussian with mean 0 and
1 1
2 and −1 with probability 2 . Moreover, for each n we standard deviation 1. Since the distribution of Sm+n −
can write Sn = ε1 + · · · + εn . The distribution of sums Sm is the same as that of Sn (as we saw earlier), this
of this kind depends in a very simple way on the well- also tells us the limiting behavior of the distribution of
√
known binomial distribution [III.71 §1]. (To be pre- (1/ n)(Sm+n − Sm ) for any m.
