[Search for users] [Overall Top Noters] [List of all Conferences] [Download this site]

Conference rusure::math

Title:Mathematics at DEC
Moderator:RUSURE::EDP
Created:Mon Feb 03 1986
Last Modified:Fri Jun 06 1997
Last Successful Update:Fri Jun 06 1997
Number of topics:2083
Total number of notes:14613

879.0. "Average series" by CLT::GILBERT () Thu May 26 1988 12:38

    Given a list of N numbers, each successive number in the series is
    the average of the preceding N numbers.  In terms of the initial
    N numbers, what is the limit of this series?
T.RTitleUserPersonal
Name
DateLines
879.1HPSTEK::XIAThu May 26 1988 15:277
    Re. 0
    One way to do it is to solve the characteristic equation:
    Nx^N - x^(N-1) -x^(N-2) - x^(N-3) - ... - 1 = 0.  I suspect that
    there are not way around it.  This equation might be solved, or
    might not by radicals (Though I have a feeling that it can, but Gosh, 
    I don't want to compute the Galois group!!)
    Eugene
879.2a hintCLT::GILBERTThu May 26 1988 18:192
    Hint:  Let f(x , x , ..., x ) be the limit.
                  1   2        N
879.3HPSTEK::XIAThu May 26 1988 20:0911
    Re .2
    There is a classical trick one can do when one computes the limit
    of a series when the series is defined recursively.  For example,
    If one has a recursive definition x = (x   )^2 then one can let
                                       N    N-1
    y be the limit (if the limit exists).  Then at limit we have 
    y = y^2 and simply solve it.  However, this trick does not work
    in the nontrivial linear recursions.  I am not sure if this trick
    is what you are hinting, if not please inform me.
    Thanks in advance.
    Eugene
879.4CHOVAX::YOUNGDumb, Expensive, Dumb ... (Pick Two)Mon May 30 1988 02:1210
    I may be a little brain-dead Peter, but your description seems
    confusing / ambiguous to me.
    
    I keep coming up with:
    
    	F(1) = k, F(2) = k, ... F(n) = k   ...  ?!?
    
    Could you provide a *leetle beety* example please?
    
    --  Barry
879.5needs proof, thoughVINO::JMUNZERTue May 31 1988 12:3118
    It seems to be:
    
    		sum from 1 to N of (j * x )
    					 j
    	limit = ---------------------------
    		  sum from 1 to N of (j)
    
    For instance, if N = 3, and you start with 0, 3, 6, you get
                         
    	0, 3, 6, 3, 4, 4.33, 3.78, 4.04, 4.05, 3.95, 4.01, 4.01, 3.99, etc.,
    
    which is 
    
    	1*0 + 2*3 + 3*6
    	---------------  =  4
    	       6
    
    John	
879.6there AREN'T N preceding numbersVIDEO::OSMANtype video::user$7:[osman]eric.vt240Tue May 31 1988 18:0436
>    Given a list of N numbers, each successive number in the series is
>    the average of the preceding N numbers.  In terms of the initial
>    N numbers, what is the limit of this series?


    Sorry to be a stickler, but as stated, this doesn't make sense.
    
    I interpret "given a list of N numbers" to mean that we have
    a finite list for some specific positive integer N.
    
    So, suppose N is 5.  Then your statement reads:
    
Given a list of 5 numbers, each successive number in the series is
the average of the preceding 5 numbers.  In terms of the initial
5 numbers, what is the limit of this series?
    
    o.k. "each successive" means consider the first then the second,
    then the third, then the fourth, then the fifth.
    
    So, we want the first number in the series to be the average of
    the preceding 5 numbers.  But there aren't 5 preceding numbers :-(
    
    Maybe you mean a CIRCULAR list ?  Or even just a SET ?
    
    Perhaps you mean this:
    
    	Given a set of N numbers with each being the average of
    	the rest, what is the limit of the values as N approaches
    	infinity.
    
    Well, all the numbers might be 7, but that's uninteresting.
    
    Please clarify...
    
    /Eric

879.7restatementZFC::DERAMOI am, therefore I'll think.Tue May 31 1988 21:4515
     What the problem means:
     
          Given N numbers a1, ..., aN.
     
          Define the infinite sequence x1, x2, ...
          by
         
               x1 = a1, ..., xN = aN,
               x_n = (x_(n-1) + ... x_(n-N)) / N    (n > N)
     
          Describe the behavior of the sequence x1, x2, ...
          (does it have limit points, does it converge, if
          so to what value as a function of a1, ..., aN, etc.)
     
     Dan
879.8Backward extensionTOOK::APPELLOFAm I position independent yet?Fri Jun 03 1988 11:136
    re .6
    
    The definition in .7 should clarify things for you.
    If you're worried about numbers preceding the given list,
    you can always recurse backwards.  That direction doesn't seem
    to have a limit, though.
879.9Proof (most of it, anyway)CLT::GILBERTFri Jun 03 1988 12:1557
    Let f(x , x , ..., x ) be the limit (with no proof that the limit exists).
           1   2        N
    We have the identity:

	f(x , x , ..., x ) = f(x , x , ..., x , (x + ... + x )/N)
	   1   2        N       2   3        N    1         N

    Now we claim (sans proof) that f is a linear function of its arguments:

	                      N
	f(x , x , ..., x ) = Sum c x
	   1   2        N    i=1  i i

    Substituting this into the above identity yields:

	 N          N             N
	Sum c x  = Sum c   x  +  Sum c x / N
	i=1  i i   i=2  i-1 i    i=1  N i

    Since this is an identity in the x , we have the N equations:
				      i
	c  =  c  / N
	 1     N

	c  =  c    + c  / N  = c    + c  ,  2 <= i <= N
	 i     i-1    N         i-1    1

    So that

	c  = i c  ,  1 <= i <= N
	 i      1

    And so
	                      N
	f(x , x , ..., x ) = Sum i c x
	   1   2        N    i=1    1 i

    We can determine c  from the identity: f(a, a, ..., a) = a, viz:
		      1

	                   N                N
	f(a, a, ..., a) = Sum i c a  = c a Sum i = c a C(N,2) = a
			  i=1    1      1  i=1      1
	
	c  = 1 / C(N,2) = 2 / (N (N+1))
	 1

    Finally,
	                        N
	f(x , x , ..., x ) = ( Sum i x ) / C(N,2)
	   1   2        N      i=1    i

    (BTW, John Munzer had the right answer in .5)


    What if each next term in the series is the *negative* of the average
    of the preceding N terms?
879.11sketch of complete proofHERON::BUCHANANnihilistic technofetishistSun Jun 05 1988 11:2584
	The technique that works best for this kind of problem is matrix 
algebra.   Define an N-dimensional vector,

u(n) = ( x(n) x(n+1) ... x(n+N-1) )'

where the ' denotes that I'm taking the transpose: i.e. it's a column
matrix not a row matrix.

	I can define easily a matrix A such that

u(n) = A.u(n-1) for all n

	It will look like this:

( 0  1         	     )
(    0  1            )
.       .  .         .
.          .  .      .
.	      .  .   .
(                0  1)
(1/N 1/N  .  .  . 1/N)

Now:

u(n) = A^(n-1).u(1)

	So the limiting behaviour of u(n) (and hence of x(n)) is determined by
the eigenvalues of A, i.e. the roots of the characteristic equation:

N.x^N -x^(N-1) -x^(N-2) ... -x -1

	However we don't need to solve this completely.   An inspection tells
us that:

(1) 1 is a root of this equation
(2) (by looking at the derivative) 1 is not a multiple root
(3) all other eigenvalues have modulus less than 1.

	This is great news, because it means that only the contribution to u(n)
from the eigenvector with eigenvalue 1 is important as n goes to infinity.   
The other components vanish.   And now we know there's a limit.

(The Frobenius-Perron Theorem is a generalization of this notion)

	The eigenvector is some multiple, c1, of (1  1  ... 1)'.   It remains
to find c1.   Let E be the matrix of eigenvectors written as column vectors,
with the distinguished eigenvector in the first column.

u(1) = E.c,

for some hitherto unknown vector c, of which the first value is c1, the value
that we're interested in.   We could invert E to get the answer, but that would
involve knowing what the other eigenvectors are, and we only know one.   We want
the first row of the inverse of E.

	The way round is this.   The matrix equation defining the eigenvectors
of A is:

A.E = E.D

where D is a (probably diagonal but let's not get into that) matrix of 
eigenvalues.   If we transpose the inverse of this equation, we get:

A~.E~ = E~.D~

where ~ denotes transpose of inverse.   Remember we wanted the first row of the
inverse of E?   Well that's the first column of E~, call it e, which just 
happens to be an eigenvector of A~!   Further the eigenvalues of A~ are the 
reciprocals of the eigenvalues of A, so e has eigenvalue 1.

	Note that all the other eigenvalues will have magnitude greater than
one, and so the assertion in .8, suggesting the sequence when extended to the 
left diverges is correct.

A~ is easy to compute, remembering how u(n) and u(n-1) are related, and e can
found from:

(A~-I).e = 0

	This is the pattern (1 2 ... N) that when normalized gives what John 
Munzer spotted.

Andrew Buchanan

879.12EasierHERON::BUCHANANnihilistic technofetishistSun Jun 05 1988 11:4419
Re: .9, if x(n) is defined to be the *negative* of the average of the preceding
N terms, then the limit is 0.

	Take the same approach as before.   The matrix A is identical except in
the Nth row, where all the terms are -1/N instead of 1/N.   The characteristic
equation is now:

N.t^N + t^(N-1) + ... + t + 1

	Once again, no eigenvalue can exceed 1 in modulus.   (To see this,
view the characteristic polynomial as the sum of N terms of the form t^N + t^i
if t is greater than one in modulus, then there is no way that the t^i term
can compensate for the t^N to "drag" the poly back to the origin.)

	But 1 is no longer a root of this poly.   And no eigenvalue can have
modulus 1, if N>1.   So all eigenvalues are smaller than one in modulus and 
hence u(n) vanishes in the limit.

Andrew Buchanan
879.13oh well, too lateZFC::DERAMOI am, therefore I&#039;ll think.Sun Jun 05 1988 23:5247
     Re .11, good job, that covers existence and uniqueness
     of the limit and its actual value.
     
     Essentially the same proof can be given without using
     matrices.  If you assume a solution of the form
     
                                               n
                   n-th term = constant times z
     
     and plug this in to solve for z, you get the same
     polynomial,
     
            N    N-1
          Nz  - z    - ... - 1 = 0
     
     The roots are z1 = 1, and z2, ..., zN have magnitude less
     than one (the polynomial is the same as in .11 so the roots
     are the same :-) and so the solution is a linear combination
     of the form [assuming the roots are all distinct]
     
                          n        n             n
          n-th term = c z    + c z   + ... + c z
                       1 1      2 2           N N
     
     With z1 = 1, and the other terms --> 0 as n --> oo,
     the limit exists and equals c1.  Solving for it in terms
     of the initial N terms gives the same result.
     
     Linear recurrences can be solved like this (assume the
     form of the solution, plug it in, and find the roots
     of the polynomial), like in .11 (give the answer explicitly
     in matrix form and explicitly compute it), by generating
     functions, by ....  You go through the same steps each
     way but just talk to yourself differently as you do it. :-)

     Dan
     
     p.s.  For a multiple root, say a root z of the polynomial
     has multiplicity k, its contribution to the form for the
     n-th term becomes 
     
               n       n      2 n            k k
            c z  + c nz  + c n z  + ... + c n z
             1      2       3              k
     
     But for z of magnitude less than one, all of these terms
     still go to zero as n increases without bound.