Stolarsky mean

In mathematics, the Stolarsky mean is a mean that generalizes the logarithmic mean. It was introduced by Kenneth B. Stolarsky in 1975.

Definition

For two positive real numbers x and y, the Stolarsky mean is defined as:

Sp(x,y)={x,if x=y,(xp−ypp(x−y))1/(p−1),otherwise.

Derivation

It is derived from the mean value theorem, which states that a secant line, cutting the graph of a differentiable function f at (x,f(x)) and (y,f(y)), has the same slope as a line tangent to the graph at some point ξ in the interval [x,y].

∃ξ∈[x,y] f′(ξ)=f(x)−f(y)x−y

The Stolarsky mean is obtained by

ξ=[f′]−1(f(x)−f(y)x−y)

when choosing f(x)=xp.

Special cases

Generalizations

Additional parameters

  • The standard two-parameter version extends the one-parameter case with α=p, β=1: Sα,β(x,y)={x,if x=y,(β(xα−yα)α(xβ−yβ))1/(α−β),otherwise. (Note that the names of the parameters can vary, and some authors reverse their order.) This extra parameter allows for additional special cases; e.g.:
    • The rth power mean has α = r and β = 2r.
    • The harmonic mean has α = -2 and β = -1.
  • This can be generalized further to have any number of pairs of parameters a1,…,ak,b1,…,bk (where Σb = 0?) by making the result (xa1−ya1a1)b1×⋯×(xak−yakak)bk. The two-parameter case results from a1=α, a2=β, b1=1α−β, b2=1β−α.

More than two variables

There are various different approaches for extending the mean to n variables, such as:

  • The x-value where the n-1st derivative equals the divided difference (a generalization of slope) of the values on the pth-degree monic power function: using the mean value theorem for divided differences, Sp(x1,…,xn)=f(n−1)−1((n−1)!⋅f[x1,…,xn]) where f(x)=xp, or for the two-parameter version, c such that g[x1,…,xn]f[x1,…,xn]=g(n)(c)f(n)(c) where f(x)=xβ and g(x)=xα, using ln(x) to replace zero divided by zero if α or β is in ℤ∩[0,n).
  • (∫(∑i=1nuixiβ)α−ββd𝐮)1α−β, where the integral is over a Dirichlet measure, and if it is flat (all concentration parameters equal 1), this is equal to ((∑i=1nxiα+(n−2)β∏j∈ℤ∩[1,n]∖{i}(xiβ−xjβ))∏i=2n(i−1)βα+(i−2)β)1α−β.
  • Based on a ratio of logarithmic means (for a particular way of extending it): [(β/α)n−1∑i=1nxiα/∏j∈ℤ∩[1,n]∖{i}(ln⁡(xi)−ln⁡(xj))∑i=1nxiβ/∏j∈ℤ∩[1,n]∖{i}(ln⁡(xi)−ln⁡(xj))]1α−β.
  • ((βα)2(∑i=1npixi2α)−∏i=1nxipi2α(∑i=1npixi2β)−∏i=1nxipi2β)12α−2β where the standard two-variable case has p1 = p2 = 1/2.