The "Normal" Probability Distribution and the Central Limit Theorem

We now return to investigate the connection between the standard deviation and the "width" we defined earlier. In the process of doing this, we will come upon a very important property of the binomial distribution, and understand the deeper meaning of the standard deviation.

First we plot below the binomial probability distributions for three different cases:

(1) the probability for r wins in N=4 trials, when the probability for one win is p=1/2
(2) the probability for r wins in N=5 trials, when the probability for one win is p=1/3
(3) the probability for r wins in N=11 trials, when the probability for one win is p=1/10

In each case, the probability distribution plotted is obtained from our general formula

using the appropriate values of N and p, for r=0, 1, ...,N.

I have chosen these three cases because they all have a standard deviation of approximately unity. We can see this as follows. Using our earlier results that  and , we have for the three cases




Using our earlier results for the mean, µ1=p, and µN=N µN, we find the mean of the three cases are:

(1)  µ=4=2

(2)  µ =5=1.667

(3)  µ=11=1.1.

The graphs of the three probability distributions are:

Remember, the probability distribution is measured at integer values of r only, and is shown in the graph by the circle, square, and triangle symbols. The lines connecting these points are only guides to the eye to help distinguish the three different cases from each other.

Note the following about the three curves above. For case (1) with p=1/2, the curve is completely symmetric about its mean µ = 2; this is a consequence of the fact that for p = 1/2 the probability to win is the same as the probability to lose. For case (2) however, with µ =1.667, and more clearly so for case (3) with µ=1.1, the curves are not symmetric about their mean; this is a consequence of the fact that for any p ≠ 1/2, the probability to win is different from the probability to lose.

To see that the shapes of the three curves are indeed different, we now translate the curves so that the horizontal origin will lie at the mean value for each of the three distributions; that is, we plot the probability distributions PN(r) versus the distance from their mean values r - µ.

In these plots, the lack of symmetry about the mean for cases (2) and (3) can be more clearly seen. That the curves for the three cases have different shapes should not be surprising; they represent, after all, three completely different binomial distributions with different values of p and N.

However, consider now what happens when we increase N in each case, so that all the distributions have a larger value of the standard deviation, 2. We choose the cases:

(1) N=16, p=1/2 with mean µ = 16= 8 and standard deviation 

(2) N=18, p=1/3 with mean µ = 18= 6 and standard deviation 

(3) N=45, p=1/10 with mean µ = 45= 4.5 and 

We first plot these distributions versus the number of wins r:

Note that the curves now all appear symmetrical about their mean, with a characteristic "bell" shape. This becomes more obvious when we plot the distributions versus the distance from the mean, r - µ.

The peaks all lie essentially at the mean value, and the distributions are much more nearly symmetrical about the mean. But something more interesting than these observations is to be seen! The points for the three difference cases now all fall on very close to a single curve!

We can again increase N in each of the three cases so that now the standard deviations are all 4. We choose the cases:

(1) N=64, p=1/2, with µ = 64= 32 and 

(2) N=72, p=1/3, with µ = 72= 24 and 

(3) N=178, p=1/10, with µ = 178= 17.8 and .

We first plot these distributions versus the number of wins r:

We now replot these probability distributions versus the distance from their mean, r-µ:

Even more so than in the previous case with 2, we see that for 4, all the points lie on one single curve.

What we have just observed empirically is that as N increases, the binomial distribution approaches a universal curve which is characterized by just two quantities: its mean µ and its standard deviation . This universal curve is known as the normal probability distribution, or also the Gaussian probability distribution.

Thus for N sufficiently large, different binomial distributions, with different values of N and p, but with the same , will all have the exact identical shape, only with peaks centered at their respective means µ=Np. Hence when we replot these distributions shifting their peaks to the origin, or equivalently plotting them versus r - µ, they all fall on one universal curve.

That the shape of the large N binomial distribution depends only on , and not on the detailed parameters of the distribution N and p, is a surprising result that is extremely important! To prove this result mathematically, and write the equation which describes the normal distribution, is beyond the scope of this course. For now we just point out that the properties of this normal distribution have been very extensively studied. Among them are the facts that:

(1) The two points µ + and µ - locate the two values of r where the probability has decreased from its value at the peak µ by a factor of  = 0.7071, i.e. P(µ + )= P(µ - ) = 0.7071P(µ). Thus 2 is comparable to our empirically defined "width" which measured the distance between the values of n for which P(n) = 0.5P(µ). Since  measures a smaller decrease from P(µ) than the "width" did, 2 is somewhat smaller than the "width".

(2) If one adds up the probabilities for all the values of r between µ +  and µ - , this adds up to 0.6827. Thus the outcome of the experiment has a 68% chance to lie within one standard deviation of the mean. Similarly, if one adds up the probabilities for all values of r between µ + 2 and µ - 2, this adds up to 0.9545. Thus the outcome of the experiment has a 95% chance to lie within two standard deviations of the mean. Similarly, the outcome has a 99% chance to lie within 2.6 standard deviations of the mean. One can make a table of the probability to lie within an interval specified by so many standard deviations (or fractions of a standard deviation) from the mean:

Standard Deviations      Probability

         0.1                          0.0797

         0.5                          0.3829

         1.0                          0.6827

         1.5                          0.8664

         2.0                          0.9545

         2.5                          0.9876

         3.0                          0.9973

         3.5                          0.9995

         4.0                          0.9999

Thus virtually all the probability of the normal distribution is concentrated within a few standard deviations of the mean. This is why the standard deviation is such an important parameter for describing a probability distribution.

How big must N be for the binomial distribution to be well approximated by the normal distribution? As a rough rule of thumb, it must be true that the standard deviation  should be very much less than the mean µ. Even when this holds true, the normal distribution will be a good approximation for the binomial distribution only in the range of several standard deviations about the mean. The approximation will not be so good in the tails very far from the mean. The size of the range about the mean for which the normal is a good approximation increases as N increases.

What we have seen above regarding the large N behavior of the binomial distribution, is actually a much more general feature of probability distributions. When we consider the number of wins r in N trials of an experiment, we can view the number of wins r as being the sum of the number of wins in each individual trial. The number of wins in each individual trial is of course either 1 or 0. As we add more and more trials, the probability for the sum of the results (that is the binomial distribution) approaches the normal distribution. One can now imagine another type of "elementary" experiment in which there are more than two possible outcomes. As an example, the role of a die can have 6 outcomes, which we can label by the number on the face which lands on top. Suppose we roll the die N times, and we want the probability distribution for the sum of the faces which land on top. It turns out that for N sufficiently large, this is also a normal distribution! Virtually anytime one adds the results of many trials of an "elementary" experiment, then regardless of the probability distribution for the outcomes of the "elementary" experiment (provided it is not too pathological), if N is sufficiently large the result will be well approximated by a normal distribution with mean µN = Nµ1, and standard deviation N1, where µ1 and 1 are the mean and the standard deviation of the "elementary" experiment. This result is known as the Central Limit Theorem. The general proof of this theorem is beyond the scope of this course, however if you go back to our proof of the results µN = Nµ1 and N1 for the binomial distribution, you will see that they hold quite generally for the sum of any N "elementary" experiments (in the proof, we never had to explicitly use the binomial form for the distribution PN(n); hence the proof works for the sums of any arbitrary distribution P(n)).