A few days ago I wrote about U-statistics, statistics which can be expressed as the average of a symmetric function over all combinations of elements of a set. V-statistics can be written as an average of over all products of elements of a set.

Let S be a statistical sample of size n and let h be a symmetric function of r elements. The average of h over all subsets of S with r elements is a U-statistic. The average of h over the Cartesian product of S with itself r times

is a V-statistic.

As in the previous post, let h(x, y) = (x − y)²/2. We can illustrate the V-statistic associated with h with Python code as before.

    import numpy as np
    from itertools import product

    def var(xs):
        n = len(xs)
        h = lambda x, y: (x - y)**2/2
        return sum(h(*c) for c in product(xs, repeat=2)) / n**2

    xs = np.array([2, 3, 5, 7, 11])
    print(np.var(xs))
    print(var(xs))

This time, however, we iterate over product rather than over combinations. Note also that at the bottom of the code we print

   np.var(xs)

rather than

   np.var(xs, ddof=1)

This means our code here is computing the population variance, not the sample variance. We could make this more explicit by supplying the default value of ddof.

   np.var(xs, ddof=0)

The point of V-statistics is not to calculate them as above, but that they could be calculated as above. Knowing that a statistic is an average of a symmetric function is theoretically advantageous, but computing a statistic this way would be inefficient.

U-statistics are averages of a function h over all subsamples of S of size r without replacement. V-statistics are averages of h over all subsamples of size r with replacement. The difference between sampling with or without replacement goes away as n increases, and so V-statistics have the same asymptotic properties as U-statistics.

Related posts

• Unbiased versus consistent
• Unbiased estimators can be terrible
• Estimating standard deviation from range

John Tukey developed his so-called g-and-h distribution to be very flexible, having a wide variety of possible values of skewness and kurtosis. Although the reason for the distribution’s existence is its range of possible skewness and values, calculating the skewness and kurtosis of the distribution is not simple.

Definition

Let φ be the function of one variable and four parameters defined by

A random variable Y has a g-and-h distribution if it has the same distribution as φ(Z; a, b, g, h) where Z is a standard normal random variable. Said another way, if Y has a g-and-h distribution then the transformation φ^-1 makes the data normal.

The a and b parameters are for location and scale. The name of the distribution comes from the parameters g and h that control skewness and kurtosis respectively.

The transformation φ is invertible but φ^-1 does not have a closed-form; φ^-1 must be computed numerically. It follows that the density function for Y does not have a closed form either.

Special cases

The g distribution is the g-and-h distribution with h = 0. It generalizes the log normal distribution.

The limit of the g-and-h distribution as g does to 0 is the h distribution.

If g and h are both zero we get the normal distribution.

Calculating skewness and kurtosis

The following method of computing the moments of Y comes from [1].

Define f by

Then the raw moments of Y are given by

Skewness is the 3rd centralized moment and kurtosis is the 4th centralized moment. Equations for finding centralized moments from raw moments are given here.

Related posts

• Generalized normal distribution
• Burr distribution
• Computing skewness and kurtosis in one pass

[1] James B. McDonald and Patrick Turley. Distributional Characteristics: Just a Few More Moments. The American Statistician, Vol. 65, No. 2 (May 2011), pp. 96–103

A symmetric function is a function whose value is unchanged under every permutation of its arguments. The previous post showed how three symmetric functions of the sides of a triangle

• a + b + c
• ab + bc + ac
• abc

are related to the perimeter, inner radius, and outer radius. It also mentioned that the coefficients of a cubic equation are symmetric functions of its roots.

This post looks briefly at symmetric functions in the context of statistics.

Let h be a symmetric function of r variables and suppose we have a set S of n numbers where n ≥ r. If we average h over all subsets of size r drawn from S then the result is another symmetric function, called a U-statistic. The “U” stands for unbiased.

If h(x) = x then the corresponding U-statistic is the sample mean.

If h(x, y) = (x − y)²/2 then the corresponding U-function is the sample variance. Note that this is the sample variance, not the population variance. You could see this as a justification for why sample variance as an n−1 in the denominator while the corresponding term for population variance has an n.

Here is some Python code that demonstrates that the average of (x − y)²/2 over all pairs in a sample is indeed the sample variance.

    import numpy as np
    from itertools import combinations

    def var(xs):
        n = len(xs)
        bin = n*(n-1)/2    
        h = lambda x, y: (x - y)**2/2
        return sum(h(*c) for c in combinations(xs, 2)) / bin

    xs = np.array([2, 3, 5, 7, 11])
    print(np.var(xs, ddof=1))
    print(var(xs))

Note the ddof term that causes NumPy to compute the sample variance rather than the population variance.

Many statistics can be formulated as U-statistics, and so numerous properties of such statistics are corollaries general results about U-statistics. For example U-statistics are asymptotically normal, and so sample variance is asymptotically normal.