Let be a compact linear (or more generally affine) operator
from a Banach space into itself. For each , the sequence of iterates , , 1, and its averages , 1, are either
bounded or approach infinity.
Let be a set and a map from into . For an , the sequence of iterations can be considered as a trajectory of a dynamical system where time is the discrete nonnegative integer: starting with the initial (time ) state , the state at time is . Suppose now that and is the transformation defined by an complex matrix . What can one say about the general behavior of trajectories of ? More generally, we consider affine maps on , that is, maps of the form , where is linear and is a constant vector, and we also allow to be infinite dimensional. Moreover, we study the behavior of the sequence of averages
Note that the method of averaging was used in [1] in approximating solutions of a system of linear equations. In case of linear operators, problems about the linear span of the iterates can be found in [2].
Recall that a square matrix is called nilpotent if for some nonnegative integer .
Throughout this paper, for nonnegative integers denotes the binomial coefficient
By convention, for and if .
If , the sum is considered as an empty sum and its value is 0.
Theorem 1. Let be an affine map defined by , where is an complex matrix and a constant vector in . Let be a norm on . Then for any vector , the sequence
is either bounded or .
Proof. By Jordan canonical decomposition theorem, for some subspaces with the following properties: (a) each is an invariant subspace of , that is, for all , and (b) there exists and a nilpotent matrix such that for all . Let be the algebraic projection of onto associated with the decomposition . Define a new norm on by
Let be a vector in . Let
For any vector , we have the following equalities:
The last equality follows from the commutative property of and since is invariant in each . Fix an and write , and . Let be the smallest nonnegative integer such that and the smallest nonnegative integer such that . Then for and (if or , the corresponding sum below is defined as 0), one has
where
Using the identity , we have, for , , and, for , is given recursively by
and (by subtracting from )
from which we get an alternate formula for :
Note that (12) is also valid for for is an empty sum. We will show that, for ,
where and are constant vectors independent of , and, for ,
where and are constant vectors independent of . More precisely, for ,
where for and for . Indeed substituting (12) into (8), we get
Now (13) follows by considering cases , or . Next consider the case . Substituting into (8), we get
Now (14) follows by considering cases , or . We now proceed to finish the proof of the theorem. Case 1 (). If one of these is nonzero, then we see from (13) that as ; otherwise is a constant vector.Case 2 ( and ). If one of these is nonzero, then we see from (13) that as ; otherwise the sequence , is bounded.Case 3 (). If one of these is nonzero, then we see from (14) that as ; otherwise is the constant vector .Case 4 (). We see from (13) that the sequence , converges to the constant vector . We conclude that for each , the sequence is either bounded or tends to infinity. If one of these sequences tends to infinity, then, since , the sequence , tends to infinity in norm and hence in norm , since the two norms are equivalent. Otherwise all sequences , are bounded for all , and from which it follows that , is bounded. This completes the proof.
The following example shows that in infinite-dimensional spaces, Theorem 1 is false.
Example 2. For nonnegative integers , let . Define for , and for . Define a linear operator such that . Then the sequence of iterates , contains subsequences that converge to 0 (e.g., ), subsequences that approach infinity (e.g., ), and infinitely many bounded nonconvergent subsequences (e.g., for ). Clearly is bounded but not compact.
Remark 3. For linear, Theorem 1 appeared in a 1997 unpublished article “On the behavior of the iterates of a matrix” of the author.
In the case that the mapping in Theorem 1 is linear, we can actually say more.
Theorem 4. Let be a linear map. Let be a norm on . Then for any vector , either or or for sufficiently large ’s, where are positive numbers with .
Proof. If in Theorem 1, then the constant vectors and in its proof are 0. It follows that in all cases where , is bounded and not convergent to 0, it is bounded away from 0. If is bounded away from 0 for some , then so is since . Otherwise converges to 0 for all and hence so does .
Example 5. The following simple example shows that linearity is needed in Theorem 4. Let be the map , where is nonzero and . Then and for all positive integers , showing that , is neither convergent to 0 nor bounded away from zero.
Definition 6. Let be a Banach space and a linear operator. one says that has property () if is a direct sum of two closed subspaces such that (1) each is invariant under , (2) is finite dimensional, and (3) there exists and a positive integer such that for all and all .
Note that by Gelfand’s spectral radius theorem, condition (12) above is equivalent to that , as an operator on , has spectral radius less than 1.
It is well known that every compact operator, or more generally, Riesz operator, has property (P); see, for example, [3].
Theorem 7. Let be a Banach space and a bounded operator having property (P). Let be a constant vector in , and let for . Then for any vector , either is bounded or . Moreover, if , then either or or for sufficiently large ’s, where .
Proof. Let , be as in Definition 6. Let , be the projections of onto , respectively. Define a norm on as
is equivalent to and commutes with . For any , write . Then for and one has
showing that , is bounded. By Theorem 1, , is either bounded or approaching infinity. Hence , is either bounded or approaching infinity. Since norms and are equivalent, the same is true for . If , then and as , and the last part of the theorem follows readily from Theorem 4.
Corollary 8. Let be a Banach space and a Riesz operator. Let be a constant vector in , and let for . Then for any vector , either is bounded or . Moreover, if , then either or or for sufficiently large ’s, where .
Corollary 9. Let be a Banach space and a compact operator. Let be a constant vector in , and let for . Then for any vector , either is bounded or . Moreover, if , then either or or for sufficiently large ’s, where .
Let us consider now the behavior of the sequence of averages of
We have that
As in the proof of Theorem 1, we may assume that and that are defined as in there. By replacing by , by 0, and hence by 0 and by in the proof of Theorem 1, we obtain the following:
for ,
where
for ,
By expanding , we have
where
Assume that . By subtracting from and using the geometric series formula, one gets
Using the relation , we see that
We will prove in the appendix that
for , where
Note that
so that approaches 1 as . Substituting the formulae we obtain thus far into , with , we get, for ,
where
and for and for .
If , then is equal to
We have
where the last equality follows from repeatedly applying the identity . From this it follows that
Therefore, for ,
Case 1 (). Suppose that . Then from (32) and (33), dominates all other coefficients. Since as and , we see that as . The same is true if since . So assume that . In the trivial case , we have for all . So assume that . Direct checking (see the appendix, item 1) shows that so that is a polynomial in . Then from (33), we have
where for fixed , are polynomials in . Since for any polynomial we see that if
as a polynomial in , is identically zero, then clearly from (32) we have , otherwise .
Case 2 (). Suppose that . Since as and , we see that as if , and it approaches 0 or is bounded if . The same is true if since . So assume that . In the trivial case , we have for all . So assume that . As in Case 1, we consider the polynomial (in )
so that
If degree of is one or less, then we see from above that is bounded and hence is bounded by (32). If degree of is two or more, then as and hence so is .
Case 3 (). We refer to (38). If , then and . If and , then . If , and for all , then . In all other cases .
The previous discussions yield the proof of the following.
Theorem 10. Let be an affine map defined by , where is an complex matrix and a constant vector in . Let be a norm on . For any vector , define
Then the sequence
is either bounded or .
If is linear, that is, if , we can say more.
Theorem 11. Let be linear map. Let be a norm on . For any vector , define
Then the sequence
is either (i) convergent to 0, or (ii) , or (iii) for sufficiently large ’s, where are positive numbers with .
Proof. If in Theorem 10, then , and by examining its proof we see that in all cases where is bounded and not convergent to 0, it is bounded away from 0. (In Case 2 of the proof, if is of degree one, then is bounded away from 0, and if is of degree 0, that is, a constant vector, then .).
The following example shows that Theorem 11 is false if the map is not linear.
Example 12. Let
and let . Define by . Let . Then
which does not converge to 0 and has a subsequence converging to 0.
Theorems 10 and 11 are also valid in Banach spaces. For proof, one only has to use these theorems and note that the average of the bounded (resp., null convergent) sequence , in the proof of Theorem 7 is also bounded (resp., null convergent). Thus we have the following theorem.
Theorem 13. Let be a Banach space. Let be an affine map defined by , where is an operator with property (P) and a constant vector in . Let be the norm on . For any vector , define
Then the sequence
is either bounded or . Moreover, if , then the sequence
is either (i) convergent to 0, or (ii) , or (iii) for sufficiently large ’s, where .
Recall that compact operators, or more generally, Riesz operators, have property (P). So the above theorem is valid for these operators.
Our last objective is to prove the following theorem. The result concerns in the case . It shows in particular that is identically 0 if and only if , and is the unique fixed point of , that is, , so that in case (i) in the proof of Theorem 10 we actually have for all , not just as .
Theorem 14. Let be a fixed complex number, . Let , and
be defined as previously. Let . The degree of is at most if and only if
is identically 0 if and only if
The proof will follow from the following discussions.
Consider . We have
and, for ,
The coefficient of in the sum above is obtained from the following calculations:
Our first objective is to write as a combination of . The proof for the following item 1 can be found in [4]. Let be a polynomial of degree . Let . Then
where
Moreover, for any integer with ,
Applying the above to the polynomial , we get
where
Moreover, for any ,
Let , be defined as above. Then
so that
Proof. Note that the degree of is and the coefficient of is . Applying item 1 to and using the identity , we get
Write . Note that is times in item 2. This change is necessary because there is an that was factored out of the summation sign in our expansion of .By item 3, the coefficient of in the expansion of is . Therefore the coefficient of in the expansion of is given by
To emphasize that depends on , we write as , so
We have that and for . We define for negative . We have the following recursive relations, which will not be used in the sequel:
We also define
so that
Then
where , and is the unsigned coefficient of in the expansion of the polynomial
that is, , and .In particular, we have that
Proof. For fixed ,
is a polynomial in with . Fix . We will show that the coefficient of in is 0. Now the coefficient is , where
Since
we have, for each , that
From the definition of are the corresponding terms (with the same coefficient) in . It follows that . For , we have that since , for . Thus the coefficient of is . This completes the proof.
Let be variables. Let be positive integers. Define to be the sum of
where are nonnegative integers such that and . Define . Define to be the sum of
where are nonnegative integers such that and . Define and . Then we have the following identity:
for any positive integers .
Proof. Consider a typical term in the expansion of the summand , sign disregarded. Let us write . We may assume that , otherwise the term belongs to a previous summand; this assumption also implies that does not belong to any previous (smaller ) summand. Let
Then the cardinality of must be greater than or equal to because of . Also the term appears in the expansion of the summand for exactly times. Next let us consider how many times appears in the expansion of the next summand . can result from multiplying a term in with a term in . Denote the exponent of in by , which is either 1 or 0. Denote by and the set of terms in and , respectively. Consider the following two sets:
Note that each corresponds to exactly one such that . Clearly has exactly elements, and these elements yield the same number of ’s which cancel out with the previous ones because of the sign change. Each consists of factors from , so has elements which yield the same number of ’s. If , that is, , then is empty and we are done since no further ’s will result. If not, we consider the next summand . Denote by and the set of terms in and , respectively. Consider the following two sets:
Note that it is , not , in the above definition of . Also note that for each , must be 0 or else is impossible. Then it is clear that and , as before. The ’s resulting from cancel out with the previous same number of ’s. The same statement about above applies to and the process continues. This process will continue for at most times since . This proves that the ’s occurring in are all canceled out.
We now finish the proof of Theorem 14. Let has degree of at most if and only if the coefficient of is 0. Since , this amounts to
which is equivalent to
So the assertion in Theorem 14 is true for . Suppose that the assertion is true for some ; this implies that coefficients of in are 0 for . We will prove that it is also true for . This will complete the proof by induction.By induction hypothesis, we have that
By applying to both sides times, recalling that , we get
for all . Substituting (88) into the formula for , we find
for all . Using , we get
If we write for the term
in and substitute (89) and (90), we find that, for ,
where for the sum is an empty sum and hence its value is 0, and
Since and are polynomials in of degree , the terms in contain only powers of power less than . And since the coefficients of powers of power greater than are 0 by induction hypothesis, we see that the coefficient of in is equal to the limit
Fix an , . Write for the number
The coefficient of the vector in for each , is, by (92),
which can be rewritten as
and which by item (5) is equal to
where is defined as in item (6), with . We also have that
where is defined as in item (6), with .The coefficient of the vector in is
Note that does not appear in for ; so the coefficient of in the sum
is the same as that in the sum
Therefore the coefficient of in the sum
which we will call , is times
(Note that so .)Now we will show that the polynomial part of is a constant, that is, the coefficients of , are zero. Fix . Since occurs in the sum only when , that is, , we see that the coefficient of in is plus
where we have used item (6) for the 0. Hence the coefficient of is 0.The constant term in is given by the coefficient of , which is plus
where we have used item (6) with and . Adding the term back in we conclude that the said coefficient is
which by item (5) is
On the other hand, by (93), the coefficient of in can be written as
Therefore the coefficient of , in is
which approaches
as since as . Since , , and only appears in , we see that the above sentence is also valid for . It then follows from (93) and (94) that the coefficient of in is
Therefore is of degree or less if and only if
The last part of the theorem corresponds to in the above equation.This completes the proof of Theorem 14.
Appendix
We prove by induction that
for , where
As we noted in the paper proper this is true for since . By item (1) below can be rewritten as
where
Taking the derivative of with respect to , we find
Then item (3) below proves that
This completes the induction since, as stated in the paper proper,
We have that
Proof.
We have that
For ,
Proof.
We have that
Let
and let . Then
Proof. We have that
It is easy to see that
Thus (A.13) is the sum of
The coefficient of in is
valid even when or since by convention for . This is equal to the coefficient of in (A.14) by item (2) above.
Acknowledgment
The author wishes to thank Professors Sing-Cheong Ong and Timothy Murphy for their discussions and Professors Roger Horn and Charles Johnson for their communications.
References
T. C. Lim, “Nonexpansive matrices with applications to solutions of linear systems by fixed point iterations,” Fixed Point Theory and Applications, vol. 2010, Article ID 821928, 13 pages, 2010.