The
Law of Large Numbers is
a theorem that describes large collections of numbers or
observations that are subject to independent and identically
distributed random variation, such as the result of performing
the same measurement a large number of times. The average of
the results obtained from a large number of trials should be
close to the actual long-term value, and will tend to become
closer as more trials are performed. It is an important idea
because it guarantees stable long-term results for the
averages of some random events. This is why gambling casinos
are able to make money; their games are designed to give the
casino a small advantage in the long run but highly variable
results in the short term, guaranteeing plenty of (noisy)
winners, which encourages the gamblers, but even a greater
number of (usually quiet) losers. And that is why investors in
the stock market often make money in the long run, despite the
unpredictable day-to-day variation, up one day and down the
next, and why it is so hard to see *climate *change in
the much wilder short-term hot and cold day-to-day and
year-to-year swings in the *weather*. *Short term*
is closer and easier; *long term *is harder to see from
here.

But "The average ... will tend to become
closer as more trials are performed" does *not* mean that the
average becomes *steadily
and irreversibly* closer. In fact, the average can wander
around quite a bit. Take the example above, which shows the running
average of a set of normally distributed independent
random numbers with a population mean of 1.000 and a standard
deviation of 1.000, as more and more numbers from that
population are averaged, up to 1000. (This is generated by the
Matlab script RunningAverage.m, shown on the left). Note that the
average wanders around, reaching and crossing over the true
population average twice in this case before ending up near
1.0 after 1000 points are accumulated. But if you ran this
script again, the final average may *not* be so close to
1.0. In fact, the predicted standard deviation of the average
of 1000 random numbers is reduced by a factor of 1/sqrt(1000),
which is about 0.031, or 3% relative, meaning that most
results will fall within 6% of the true average of
1.000, that is, between 0.94 and 1.06.

**The ****uncertainty of uncertainty.**
The situation is even worse if you wish to estimate the *standard
deviation* of a population from small samples. The Matlab
script RunningStandardDeviation.m simulates this for the same population in
the

As shown in the graph above, the sample
standard deviation wanders around alarmingly for small samples
and only settles down slowly. Even worse, the standard
deviation for very small samples is biased down, often returning values far lower than
the population standard deviation.

There is a well-documented tendency for
people to *overestimate*
the quality of small numbers of observations, sometimes
referred to as hasty
generalization, or insensitivity
to
sample size, or the gambler's
fallacy. This is
related to the field of study of a famous pair of
psychologists named Amos Tversky and Daniel Kahneman, who
collaborated in a long-running study of human cognitive biases
in the 1970s. They formulated a hypothesis that people tend to
believe in a false "Law
of
Small Numbers", the
name they coined for the mistaken belief that a small sample
drawn from a large population is representative of that large
population. We would like to believe that scientists are
immune to these foibles and that they always think logically
and correctly. But scientists are only human, so it is
important to be aware of this tendency, particularly when a
small sample of data supports your favorite hypothesis. It is
tempting to stop there, "while you are ahead". This is called
"confirmation
bias". Don't do it.

Of course in many practical experimental
measurements, you may really be constrained to a rather small
number of repeated measurements. There may be a fixed number
of data points and no possibility of gathering more. Or the
cost, in money or in time, of gathering more data may be
excessive, even in a laboratory environment. For example, the
process of calibrating an analytical instrument
for quantitative measurement may involve the preparation
and measurement of several standard samples or solutions of
known composition. If the calibration curve (the relationship
between instrument reading and sample composition) is
non-linear, it takes several different standards to define the
curve. You have to consider not only cost of preparing many
standards but also the cost of cleaning up and safely storing
or disposing of the (potentially hazardous) chemicals
afterwards. The bottom line is, if you are limited to a small
number of data points, do not over-represent the precision of
your results. To use the 3-sigma
rule to determine uncertainty ranges for a set of data,
the distribution must be normal (Gaussian) and you need to
know the standard deviation. The problem is that, for small
sets of data, *both are uncertain*.

This page is part of "A Pragmatic Introduction to Signal
Processing", created and maintained by Prof. Tom O'Haver ,
Department of Chemistry and Biochemistry, The University of
Maryland at College Park. Comments, suggestions and questions
should be directed to Prof. O'Haver at toh@umd.edu. Updated July, 2022.