Average of Sigma(n)/n

One of the simplest proofs of the fact that the average "order" of sigma(n)/n 
is equal to zeta(2) = pi^2/6 (i.e., the sum of the inverse squares) is by 
counting the lattice points in certain plane regions (c.f., Hardy and Wright),
but there are some other interesting approaches as well.  For example, by the 
Taylor series expansion of the natural log function, we have

                   / x    x^2   x^3   x^4       \
       ln(1-x) = -( --- + --- + --- + --- + ...  )            (1)
                   \ 1     2     3     4        /

for all |x| less than 1.  Now, if we consider the sum of the 
natural logs of 1-x, 1-x^2, 1-x^3, and so on, we have

    oo                  x    x^2   x^3   x^4   x^5   x^6
 -SUM  ln(1 - x^n)  =  --- + --- + --- + --- + --- + --- + ...
   n=1                  1     2     3     4     5     6


                       x^2   x^4   x^6   x^8   x^10   x^12
                     + --- + --- + --- + --- + ---- + ---- + ...
                        1     2     3     4     5      6


                       x^3   x^6   x^9   x^12   x^15   x^18
                     + --- + --- + --- + ---- + ---- + ---- + ...
                        1     2     3     4      5      6


                       x^4   x^8   x^12   x^16   x^20   x^24
                     + --- + --- + ---- + ---- + ---- + ---- + ...
                        1     2     3      4      5      6


                     +  etc...

Combining terms by powers of x gives the nice identity

             oo                     oo  sigma(n)
         - SUM  ln(1 - x^n)   =   SUM   -------- x^n             (2)
            n=1                    n=1      n

If we divide both sides of this relation by the geometric series, i.e., 
the sum of x^n for n=1 to oo, we have

                                  oo  sigma(n)
                                SUM   -------- x^n
     1-x   oo                    n=1      n
   - --- SUM  ln(1 - x^n)   =   ------------------              (3)
      x   n=1                         oo
                                    SUM   x^n
                                     n=1

Notice that the right side of this equation a geometrically weighted 
average of all the values of sigma(n)/n from n=1 to infinity.  Also,
as x approaches 1, this approaches the EVENLY weighted average out to
arbitrarily large n.  This suggests that it would be interesting to
evaluate the limit of the left side as x goes to 1.  This turns out 
to be sort of a delicate operation, because if the summation on the 
left side of (3) is evaluated for n=1 to N for any fixed N, the overall 
quantity on on the left side goes to zero as x goes to 1.  However, for 
any fixed x less than 1, the quantity converges on a non-zero value as 
N goes to infinity, and these limiting values converge on a certain 
value as x goes to 1.

Anyway, by making x sufficiently close to 1 we can make the ratios of 
successive terms in the left-hand summation arbitrarily close to 1, 
so the summation can be approached by the integral in the limit as
x goes to 1.  In other words
                                    oo
               oo                   /
    lim      SUM  ln(1 - x^n)   =   | ln(1 - x^n) dn             (4)
    x->1      n=1                   /
                                   n=1

To evaluate the integral, let's make the substitution q = x^n, where
q goes from x to 0 as n goes from 1 to infinity.  Noting that q can be 
written as exp(nln(x)) we have

          dq
          --  =  exp(n ln(x)) ln(x)  =  ln(x) x^n                (5)
          dn

which gives

                    1                 1
         dn  =  --------- dq   =   ------- dq
                ln(x) x^n          q ln(x)


With these substitutions the integral becomes

        oo                            0
        /                        1    /  ln(1-q)
        | ln(1 - x^n) dn   =   -----  |  ------- dq              (6)
        /                      ln(x)  /     q
       n=1                           q=x

Recalling the power series expansion of ln(1-q), we have the nice
indefinite integral
      
    /   ln(1-q)             / q    q^2   q^3   q^4       \
    |   ------- dq   =   - ( --- + --- + --- + --- + ...  )      (7)
    /      q                \1^2   2^2   3^2   4^2       /
     
Evaluating this from q=x to q=0 gives

      0
      /   ln(1-q)           x      x^2     x^3     x^4 
      |   ------- dq   =   ---  +  ---  +  ---  +  ---  +  ...  
      /      q             1^2     2^2     3^2     4^2 
     q=x

Substituting this back into equation (6), and from there back into 
equation (3), we arrive at the expression for the evenly weighted 
average of sigma(n)/n for all integers n:

      
               1 - x   /  x      x^2     x^3     x^4         \
   lim      - ------- (  ---  +  ---  +  ---  +  ---  +  ...  )
   x->1       x ln(x)  \ 1^2     2^2     3^2     4^2         /


Since the ratio of  -(1-x)  to  ln(x)  goes to 1 as x approaches 1,
and the numerators inside the parentheses also go to 1, we have the
result
                  oo  sigma(n)
                SUM   -------- x^n
                 n=1      n                oo   1
       lim      ------------------   =   SUM   ---
       x->1          oo                   n=1  n^2
                   SUM   x^n
                    n=1

which shows, as expected, that the average value of sigma(n)/n for all 
integers is zeta(2) = pi^2 / 6.  This approach is certainly much less
economical than the simple lattice-point derivation presented in Hardy 
and Wright, but it does provide the opportunity to relate some common 
power series expansions to arithmetic functions.  For example, from the 
expansions of (1 - 1/x)ln(1-x^n) we see that the ratio sigma(n)/n equals 
the sum of the numbers in the first n columns of the following array:

 1 -1/2 -1/6 -1/12 -1/20 -1/30 -1/42 -1/56 -1/72 -1/90 -1/110
     1   -1    1/2  -1/2   1/3  -1/3   1/4  -1/4   1/5  -1/5
          1    -1          1/2  -1/2         1/3  -1/3
                1    -1                1/2  -1/2
                      1    -1                      1/2  -1/2
                            1    -1
                                  1    -1
                                        1    -1
                                              1    -1
                                                    1    -1
                                                          1

This is related to the interesting fact that the series expansion of 
the left side of (3) is

        1       1         5          11         24
  1  +  - x  -  - x^2  +  -- x^3  -  -- x^4  +  -- x^5  - ...
        2       6         12         20         30

where the coefficient of x^k is sigma(k)/k - sigma(k-1)/(k-1).  This 
converges for all x less than 1, whereas for x=1 the partial sums 
are just the successive values of sigma(k)/k, so it never converges.  
Nevertheless, the limit of the convergent values as x approaches 1 
is (pi^2)/6.

Return to MathPages Main Menu