Algorithmic definition of means acting on positive numbers and operators

Miklós PálfiaI would like to thank Dénes Petz and Nándor Sieben for their help and support.

March 6, 2005

Abstract
Means are used in several applications from electronic engeneering to information theory, however there is no general theorem on how to extend a given M ( x , y )   mean function to multiple variable forms. In this article we would like to present a theorem, which gives one possible solution for this problem, for every M ( x , y )   mean function, acting on positive numbers and operators.

1 Introduction

The use of mean functions falls back to the early ages of mathematics, but they are used widely nowdays aswell. For example they have great importance in statistics, but also in volume visualization, in imaging surgery.
The study of electrical network connections implied the introduction of parallel sum of two positive semidefinite matrices in [2. Formerly, Anderson defined a matrix operation, called shorted operation to a subspace, for each positive semidefinite matrix. Anderson and Trapp in [3, have extended the theory of parallel addition and shorted operation to bounded linear positive operators on a Hilbert space and demonstrated its importance in operator theory. They have studied fundamental properties of these operations.
The axiomatic theory of means, for pair of positive operators, have been developed by Kubo and Ando in [7. This theory has found a number of applications in operator theory.
Several steps have been taken in the theory of means, but a general idea has not been laid down yet, on how to extend, or define an n   variable version of a given general M ( x , y )   mean function acting on numbers and operators. We would like to present a theory which gives a possible solution and a frame theory for further studies.
Definition 1.1. A two variable function M : R + × R + R +   according to [11, is called a mean function if
  • (i) M ( x , x ) = x   for every x R +   .
  • (ii) M ( x , y ) = M ( y , x )   for every x , y R +   .
  • (iii) If x < y   , then x < M ( x , y ) < y   .
  • (iv) If x < x   and y < y   , then M ( x , y ) < M ( x , y )   .
  • (v) M ( x , y )   is continuous.
The geometric mean x y   , the arithmetic mean ( x + y ) / 2   and the harmonic mean 2 / ( x 1 + y 1 )   are the most known examples but there are many other means aswell.
The above definition of means can be easily extended to positive ordered operators and matrices, by replacing the numbers with them.

2 Extending a mean to multiple variables

The definition of M ( x , y )   given in the first section, can be extended to multiple variable functions. The following definition is one possibility of extension.
Definition 2.1. An n   variable function M n : ( R + ) n R +   may be called a ( n   variable) mean function if
  • (i') M n ( x , . . . x ) = x   for every x R +   .
  • (ii') M n   is independent from the ordering of x 1 , x 2 , . . . x n   .
  • (iii') m i n ( x 1 , . . . x n ) M n ( x 1 , . . . x n ) m a x ( x 1 , . . . x n )   .
  • (iv') If ( i ) x i x i   , then M n ( x 1 , . . . x n ) M n ( x 1 , . . . x i , . . . x n )   .
  • (v') If ( i ) x i < x i   , then M n ( x 1 , . . . x n ) < M n ( x 1 , . . . x n )   .
  • (vi') M n ( x 1 , . . . x n )   is continuous.
Our goal is to algorithmically define an n   variable mean by using its less than n   variable forms. Firstly we will use the n 1   variable form of M   , to define the n   variable one, as an iteration's limit.
Definition 2.2. Let X = ( x 1 , . . . x n ) ( R + ) n   and M n 1   be an n 1   variable mean function. Let us consider the n 1   class variations of x n   . There are ( n n 1 ) = n   different variations. Let us define an iteration as
(2.1) x i 0 = x i i [ 1 , . . . n ] x i k + 1 = M n 1 ( V i n 1 ( X k ) ) (2.2)
where V i n 1 ( X k ) = V i n 1 ( x 1 k , . . . x n k )   is the i   th n 1   different variation of x n   ( n 1   are chosen from n   , which could be done in n   different ways).
Theorem 2.1. The x i k   sequences defined in definition  2.2 are convergent and their limits are the same, which could be defined as the M n   mean of the x n   numbers.
  • Proof. The iteration given in definition  2.2 has a contractive-like property by (iii'), which means that the sequence min X k   is monotonic increasing, max X k   is monotonic decreasing. Hence the limits lim k min X k   and lim k max X k   exist. From condition (iv'), one can see that the series' minimal and maximal elements are given by the following:
    (2.3) min X k + 1 = M n 1 ( V m i n n 1 ( X k ) ) max X k + 1 = M n 1 ( V m a x n 1 ( X k ) ) (2.4)
    where V m i n n 1 ( X k )   are the smallest n 1   numbers, V m a x n 1 ( X k )   are the largest n 1   numbers from the series X k   . By  2.3 and  2.4  min X k + 1   and max X k + 1   explicit dependence on min X k   and max X k   is given
    (2.5) lim k min X k + 1 = lim k M n 1 ( V m i n n 1 ( X k ) ) lim k max X k + 1 = lim k M n 1 ( V m a x n 1 ( X k ) ) (2.6)
    and by (vi') one can write
    (2.7) lim k min X k + 1 = M n 1 ( lim k V m i n n 1 ( X k ) ) lim k max X k + 1 = M n 1 ( lim k V m a x n 1 ( X k ) ) (2.8)
    which yields
    (2.9) lim k min X k + 1 = I = M n 1 ( I , . . . ) lim k max X k + 1 = J = M n 1 ( J , . . . ) (2.10)
    and this is only true when I = J   .
The above theorem and its proof yields the following remarks.
Remark 2.1. The iteratively defined mean function M n   , is invariant to the initial ordering of the V i n 1   variations.
Remark 2.2. The iteration  2.2 leaves the mean of the starting n numbers invariant through the sequence.
It is easy to verify the next two theorems, which have stressed importance in inequalities of means and operator means in our given context.
Corollary 2.2. Two different, iteratively defined mean function M n , 1   and M n , 2   , is in the same relation as their two variable forms ( M 2 , 1 < M 2 , 2   implies M n , 1 < M n , 2   ).
Corollary 2.3. Theorem  2.1 and the proof also works for ordered positive operators, acting on a H   Hilbert space, and for k × k   matrices.
We will show with some examples, that theorem  2.1 gives a sufficient definition of the n   variable mean.
Corollary 2.4. Theorem  2.1 applied on the n 1   variable arithmetic, geometric and harmonic mean, gives the corresponding n   variable mean.
  • Proof. According to the given serie's convergence in theorem  2.1 it is enough to prove that the minimum's or maximum's limit is the n   variable mean. We will prove it only for the arithmetic mean. Let us consider the numbers x 1 0 x n 0 R +   and the A n 1   arithmetic mean. By theorem 2.1, the sequence x 1 k   can explicitly be written and proven by induction with  2.2 ,  2.3 and  2.4 :
    x 1 k = { i = 1 n ( n 1 ) k 1 n x i 0 + x 1 0 ( n 1 ) k if k is even, i = 1 n ( n 1 ) k + 1 n x i 0 x n 0 ( n 1 ) k if k is odd, x n k = { i = 1 n ( n 1 ) k 1 n x i 0 + x n 0 ( n 1 ) k if k is even, i = 1 n ( n 1 ) k + 1 n x i 0 x 1 0 ( n 1 ) k if k is odd. (2.11)
    For the geometric mean the proof can be extended using the logarithmic function and its inverse for the limit:
    log ( x 1 x 2 . . . x n 1 n 1 ) = i = 1 n 1 log x i n 1 . (2.12)
    The proof for the harmonic mean can be given by inverses:
    ( n 1 x 1 1 + x 2 1 . . . x n 1 1 ) 1 = x 1 1 + x 2 1 . . . x n 1 1 n 1 . (2.13)
Our main idea of extending means to multiple variables is based on theorem  2.1 , which is theoretically enough but in practice is very insufficient. For example if we would like to compute n   numbers or matrices M n   mean, we should use the two variable main definition M 2   and extend the other ones from one to another.
In the next section we will prove that M n   can be directly extended from the corresponding M 2   .

3 Extending M n   directly from M 2  

Let us define the following iteration:
Definition 3.1. Let X = ( x 1 0 x n 0 ) ( R + ) n   and M = M 2   be a two variable mean function,
x i k + 1 = { M ( x 1 k , x 2 k ) if i = 1 , M ( x n 1 k , x n k ) if i = n , M ( x i 1 k , x i + 1 k ) else. (3.1)
Theorem 3.1. The iteration given in definition  3.1 for all n   is convergent and ( i ) lim k x i k = M n ( x 1 0 , . . . x n 0 )   where M n   is defined by theorem  2.1 .
  • Proof. Firstly we begin with proving the convergence. It is clear that the min x i k   is always the first element ( i = 1   ) and the max x i k   is always the last element ( i = n   ) of the series in definition  3.1 . Hence (iii) and definition  3.1 , min x i k   is increasing and max x i k   is decreasing. This yields:
    (3.2) lim k min x i k + 1 = lim k M ( x 1 k , x 2 k ) lim k max x i k + 1 = lim k M ( x n 1 k , x n k ) (3.3)
    and by (v):
    (3.4) lim k min x i k + 1 = M ( lim k x 1 k , lim k x 2 k ) lim k max x i k + 1 = M ( lim k x n 1 k , lim k x n k ) (3.5)
    which give
    (3.6) lim k min x i k + 1 = I = M ( I , lim k x 2 k ) , lim k max x i k + 1 = J = M ( lim k x n 1 k , J ) . (3.7)
    Considering the characteristics of the definition  3.1 , this can only be true when I = J   . Secondly we will prove the limit. For n = 3   the theorem is clear, because the two iterations, defined in theorems  2.1 and in  3.1 , are the same. Our next step is to prove for n + 1   , if it is true for n   .
    Let us consider the definition of M n + 1   in theorem  2.1 . Comparing the min x i k   (which is the first element i = 1   ) in theorem  3.1 , and min X k   (which equals M n ( V m i n n ( X k 1 ) )   by  2.3 ), we can see that min x i k min X k   , because of the inductional condition and the definition of M n   as a limit in theorem  2.1 . The same can be applied for max x i k   and max X k   which yields max x i k max X k   .
    Hence min x i k   and max x i k   are minoring and majoring, for every k   , min X k   and max X k   , but lim k min x i k = lim k max x i k   , so lim k min x i k = M n + 1 ( x 1 0 , . . . x n + 1 0 )   .
Furthermore there is special property in the iteration in definition  3.1 .
Definition 3.2. Let x i 0 R + i [ 1 , . . . n ]   and G   be a graph, with n verteces and edges given as that, there is one cycle in G   , which contains all verteces and edges (so it is at the same time a Hamiltonian and an Euler-cycle ). This implies that in G   , every vertex has two edges and all of them are bound together. Let us consider an optional one to one correspondence between x i 0   numbers and G   -s verteces. Taking every edge in G   as an M ( x j 0 , x l 0 )   (where M   is a mean function and x j 0   , x l 0   are assigned to the two ending points of the edge as previously given), we can define an iteration with an optional n   mappings,
x i k + 1 = M ( x j k , x l k ) i , j , l [ 1 , . . . n ] j l . (3.8)
Theorem 3.2. Every different iteration given in definition  3.2 , converge to the limit M n   mean function, and the iteration independently from the mapping x i k + 1 = M ( x j k , x l k )   converge on a higher or equally rate as the iteration given in definition  3.1 .
  • Proof. For n = 3   , it is easy to see that the theorem is true, because the iterations given in definitions  3.1 and  3.2 , are the same.
    Assume that the theorem is true for n   variable. Let us expand from an n   variable iteration defined in  3.2 with an optional mapping, to n + 1   variable.
    This can be done as replacing one edge with two edges and one vertex (mapped to a new number). Let us do this expansion in the following way. Take the first smallest n   numbers from n + 1   and set up on them the iteration given in definition  3.1 . From the inductional condition this iteration will have the slowest convergence rate, which means that its minimal and maximal elements will minor and major every other iteration in definition  3.2 . Let us replace the edge which gives the maximal element of the iteration given in definition  3.1 , with the two new edges and the remaining number (which is the greatest number out of the n + 1   ) as a vertex. This two edges with the corresponding two M 2 ( x , y )   will give the new iterations given for n + 1   numbers greatest two elements. This replacement cannot be done better in any other optional mapped iteration given in definition  3.2 aswell. But considering the inductional condition this yields that any n + 1   variable iteration given in definition  3.2 , cannot minor and major, with its maximal and minimal elements, the iteration given in definition  3.1 , hence theorem  3.2 is proven.
Corollary 3.3. The above theorems also work for ordered operators acting on a H   Hilbert space, and k × k   matrices aswell.
We will have to consider further examinations to define the above iterational definitions for inorderable matrices. In the next section we will study this problem.

4 Extending theorems  3.1 and  3.2 to unordered matrices and operators

The problem is with positive matrices and operators which satisfy A = B   and A B   . For the above matrices and operators, the function M ( A , B )   's and its arguments' relation is not explained and highly depend on the main characteristics of M ( A , B )   , so the given iterations in the above theorems must be specified.
Theorem 4.1. For any X 1 , . . . X n   positive operators or matrices the iteration given in definition  3.1 is convergent, as defined in theorem  3.1 .
  • Proof. If X 1 = X 2 = = X n   does not hold, than the iteration in definition  3.1 converges for all X i   , because after n   steps, the iteration will surely alter all of the X i k   -s (from one to another), so the iteration will converge.
    If X 1 = X 2 = = X n   does hold, we will have to define (according to [11) the following construction. Let a t   and a t   be monotone sequences as,
    a t , a t R + ( t ) a t 1 and a t 1 l i m t a t = 1 and l i m t a t = 1 . (4.1)
    Let X 1 = a t X 1   , X i = X i i [ 2 , . . . n ]   and X 1 = a t X 1   , X i = X i i [ 2 , . . . n ]   . Let us set the iteration given in definition  3.1 up on X i   , X i   and X i   . According to the first part of the proof, the ( X i ) k   and ( X i ) k   series are convergent for any t   , as given in theorem  3.1 . Considering the definition of sequences a t   and a t   in  4.1 , it is easy to verify by condition (iii), that for any t   , the ( X i ) k   series are minoring and ( X i ) k   series are majoring the series X i k   for any i   .
    Taking the limit k   , we get M n ( X 1 ( t ) , . . . X n ( t ) )   and M n ( X 1 ( t ) , . . . X n ( t ) )   .
    Hence M n ( X 1 ( t ) , . . . X n ( t ) )   and M n ( X 1 ( t ) , . . . X n ( t ) )   are Cauchy sequences in index t   and condition (vi'), they are convergent and
    l i m t M n ( X 1 ( t ) , . . . X n ( t ) ) = l i m t M n ( X 1 ( t ) , . . . X n ( t ) ) . (4.2)
    But ( X i ) k ( t )   and ( X i ) k ( t )   are minoring and majoring every X i k   for any i   and t   , so the limit l i m k X i k   exist and by  4.2 ,
    l i m t M n ( X 1 ( t ) , . . . X n ( t ) ) = l i m k X i k = l i m t M n ( X 1 ( t ) , . . . X n ( t ) ) (4.3)
    and theorem  4.1 is proven.
Corollary 4.2. Using the above proof, theorems  2.1 and  3.2 work for the unordered X i   -s.

5 Consequences

By the theorems given in our examinations generalize the extension of the two variable mean functions and gives a frame theory, which may be used in the future studies related to the extension of means to multiple variables.
An important outcome is, that these theorems are applying for operators and matrices and guarantee the existence of one possible extension.
It is known, that in several situations, there are more than one possible generalization of a mean. One example is the logarithmic mean,
L ( x , y ) = x y log x log y (5.1)
which has several extended forms according to [5, [8, [9, but our theorems may leave only one form valid. However with some means, it appears to be quite difficult to give the iterations limit in a closed form.
References

  1. T. Ando, C-K. Li and R. Mathias, Geometric means, Linear Algebra Appl.
  2. W. N. Jr. Anderson and R. J. Duffin, Series and parallel addition of matrices, J. Math. Anal. Appl. 26(1969), 576594.
  3. W. N. Jr. Anderson and G. E. Trapp, Shorted Operators II, Siam J. Appl. Math. 28(1975), 6071.
  4. R. Bhatia, Matrix Analysis, Springer-Verlag, New York, 1996.
  5. B. C. Carlson, The logarithmic mean, Amer. Math. Monthly 79, 615-618 (1972).
  6. F. Hiai and H. Kosaki, Means of Hilbert space operators, Lecture Notes In Maths. 1820, Springer, 2003.
  7. F. Kubo, T. Ando, Means of positive linear operators, Math. Ann. 246(1980), 205-224.
  8. S. Mustonen, Logarithmic mean for several arguments,
  9. E. Neuman, The weighted logarithmic mean, J. Math. Anal. Appl. 188(1994), 885-900.
  10. M. K. Vamanamurthy and M. Vuorinen, Inequalities for means. J. Math. Anal. Appl. 183(1994), 155-166.
  11. D. Petz, Means of positive numbers and operators, preprint (2004).