Generalizations of the functional equation of the mean sun

HARALD FRIPERTINGER*
LUDWIG REICH


*Supported by the Fonds zur Förderung der wissenschaftlichen Forschung P14342-MAT.
Abstract

Two generalizations U(c + m)y(n) = U(c)y(m + n) ( A c, m, n  (- A) of the functional equation of the mean sun are studied, where (A, +) is an Abelian group, K is a field, n is a positive integer, and both y: A --> Kn and U: A --> GL(n, k) (or U: A --> Mn(k) in the second case) are unknown functions, which will be determined by the equation.

1 Introduction

Local solar time is measured by a sundial. When the center of the sun is on an observer’s meridian, the observer’s local solar time is zero hours (noon). Because the earth moves with varying speed in its orbit at different times of the year and because the plane of the earth’s equator is inclined to its orbital plane, the length of the solar day is different depending on the time of year. It is more convenient to define time in terms of the average of local solar time. Such time, called mean solar time, may be thought of as being measured relative to an imaginary sun (the mean sun) that lies in the earth’s equatorial plane and about which the earth orbits with constant speed. Every mean solar day is of the same length.1

In [14] it is shown that the mean sun satisfies the functional equation

M  (c + t,f)T y(s) = M (c,f)T y(s + t),      A s,t,c  (-  R - p/2 < f < p/2,
(1)
where y(s) is a vector of length 1, which is the direction from the center of the earth to the sun at the time s (one day corresponds to 2p) expressed in a geocentric coordinate system. As a basis of this system we can choose two orthogonal vectors in the equatorial plane and one vector along the axis of the earth. M(c, f) is the matrix
           (                                   )
              - sinc  -  sin fcos c  cosf cosc
M  (c, f) =     cosc    - sin f sin c  cosf sinc    .
                 0        cosf         sin f

Then M(c, f)y(s) is the direction from the earth to the sun expressed in a local coordinate system on the surface of the earth in the point of longitude c and latitude f.

In the present paper we investigate generalizations of equation (1) for fixed f. To be more precise, first we will solve the following functional equation

U (c + m)y(n) =  U(c)y(m +  n),     A c, m,n  (-  A,
(2)
where (A, +) is an Abelian group, K is a field, n is a positive integer, and both y: A --> Kn and U: A --> GL(n, K) are unknown functions, which will be determined by (2). In some situations we will additionally have to assume that A = K. Later on we will study the more general situation when we replace GL(n, k) by Mn(K), the set of all n × n matrices over K. The following types of questions can be asked in connection with (2):
  1. Determine all solutions (U, y) of (2).
  2. For given U determine all y, such that (U, y) is a solution of (2).
  3. For given y determine all U, such that (U, y) is a solution of (2).
  4. Find relations between U and y for a solution (U, y) of (2).

We will mainly deal with problems of the second and third kind.

In Theorem 6 we describe in an appropriate system of coordinates the structure of the space SU of all solutions of (2) for a given U: A --> GL(n, K). We also state in this theorem how such a mapping U necessarily looks if a nontrivial solution y (i.e. y/=0) exists. A similar description of U-invariant subspaces of SU is given in Theorem 8. We emphasize that by our result (and similarly by the following theorems) the problem of solving (2) can be reduced, at least to some extent, to the problem of finding all exponential functions U11: A --> GL(k, K) (cf. the representation of U in Theorem 6), i.e. non singular matrices U11(c) satisfying the equation

U11(c + m) =  U11(c)U11(m),      A c,m  (-  A.

Here we assume that these functions are known and we refer the reader to [3].

In Theorem 9 we construct to a given subspace S0 of Kn the set of all mappings U and correspondingly the space S of all functions y, such that (U, y) satisfies (2) and S0 is exactly the set of all initial values y(0) for y  (- S. It is clear that this yields together with Theorem 6 an implicit description of the set of all solutions (U, y) of (2) by varying the subspace S0 of Kn. However, the space S and the mapping U obtained in this way from S0 may have the property that S is a proper subset of SU . Therefore we also deal with the problem to characterize the situation when S = SU .

From a mathematical point of view it seems also interesting to study the functional equation (2) for mappings U: A --> Mn(K). This situation is more complicated both with respect to the technical details and the construction (description) of the solutions U or y or (U, y). In Theorem 20 we start from a given mapping U: A --> Mn(K) and describe completely in appropriate coordinates the set of all functions y: A --> Kn, such that (U, y) is a solution of (2). Again this theorem provides necessary conditions on U for the existence of nontrivial solutions y of (2). We also show in Theorem 21 how to construct all mappings U: A --> Mn(K) and corresponding spaces S of functions y: A --> Kn, such that {y(0) |y  (-  S} is a given subspace S0 of Kn and (U, y) is a solution of (2), hence giving an implicit description of the general solution of (2) by varying S0. However, we were not able to contribute to the problem when S = SU .

The main difficulties in this last part seem to arise from the fact that there can occur solutions y (to a given U) with y(0) = 0 but y/=0 (cf. Lemma 18).

2 Regular matrices U(c)

Here in this part we always assume that U is a mapping from the abelian group A to GL(n, K).

Lemma 1. Let B, C be matrices in GL(n, k). Then (U, y) is a solution of (2) if and only if (V, By) is a solution of (2), where V (c) = CU(c)B-1.

Proof.  The pair (U, y) is a solution of (2) if U(c+m)y(n) = U(c)y(m+n) for all c, m, n  (- A. Since B and C are regular matrices, this is equivalent to CU(c + m)B-1By(n) = CU(c)B-1By(m + n) for all c, m, n  (- A.       []

For C = U(0)-1 we get CU(0) = I n, the identity matrix. Hence, without loss of generality we will always assume that U(0) = In.

Lemma 2. If (U, y) is a solution of (2), then

y(m) = U (m)y(0),      A m  (-  A,
(3)
[U (c + m) - U (c)U (m)]y(0) = 0,     A c, m  (-  A.
(4)
Proof.  Since U(0) = In, we get (3) from (2) for c = n = 0. And we get (4) from (2) and (3) for n = 0.       []

It is also possible to reverse the statement of Lemma 2.

Lemma 3. Assume U(0) = In and let y be given by (3). If (U, y) satisfies (4), then (U, y) is a solution of (2).

For any mapping U: A --> GL(n, k) let

SU := {y |(U, y) is a solution of (2)} and  S0U := {y(0) |y  (-  SU }.

Some basic properties of these two sets are collected in the following

Lemma 4. Both SU and SU0 are K-linear spaces and h: SU0 --> SU , given by h(y0) := U(.)y0, is a vector space isomorphism.

Proof.  It is clear that SU and SU0 are linear spaces. Assume that y0  (- SU0, then there is some y  (- SU , such that y(0) = y0. Since (U, y) satisfies (3), the function h is well defined. It is surjective, since for any y  (- SU we have h(y(0)) = U(.)y(0) = y(.) according to (3). The mapping h is also injective, since from h(y10) = h(y20) we derive U(c)y10 = U(c)y20, which implies for c = 0 (and U(0) = In) that y10 = y20. Finally we have to prove that h is a linear mapping. Let y10, y20  (- SU0 and let a1, a2  (- K, then a1y10 + a2y20  (- SU0 and h(a1y10 + a2y20) = U(.)(a1y10 + a2y20) = a1U(.)y10 + a2U(.)y20 = a1h(y10) + a2h(y20).       []

In conclusion, both SU and SU0 are m-dimensional linear spaces for some 0 < m < n.

There are some more interesting properties of SU and SU0.

Lemma 5. Let U: A --> GL(n, k). Then:

  1. SU is U(c0)-invariant for all c0  (- A (i.e. if y  (- SU , then also U(c0)y  (- SU ).
  2. SU is invariant under translations (i.e. if y  (- SU , then also y(. + c0)  (- SU for all c0  (- A).
  3. SU0 is U(c0)-invariant for all c0  (- A.
  4. SU0 = {y(c) |c  (-  A, y  (-  SU }
Proof. 
  1. Assume that z(c) := U(c0)y(c). Then (U, z) satisfies (2), since U(c + m)z(n) = U(c + m)U(c0)y(n) = U(c + m)U(c0)U(n)y(0) = U(c + m)U(c0 + n)y(0) = U(c + m)y(c0 + n) = U(c)y(m + c0 + n) = U(c)U(0)y(c0 + m + n) = U(c)U(c0)y(m + n) = U(c)z(m + n) by (3), (4), (3), (2), special form of U(0) and (2).
  2. Let z(c) := y(c+c0), then (U, z) satisfies (2), since U(c+m)z(n) = U(c+m)y(n + c0) = U(c)y(m + n + c0) = U(c)z(m + n) by (2).
  3. If y0  (- SU0, there exists some y  (- S U , such that y0 = y(0). From the first item of this lemma we know that U(c0)y  (- SU , hence U(c0)y(0) = U(c0)y0  (- SU0.
  4. According to the definition of SU0 we know that SU0  (_ {y(c) | c  (-  A, y  (-  SU }. Let y  (- SU and assume that c0  (- A \ {0}, then it follows from the second item of this lemma that z(.) := y(. + c0)  (- SU and y(c0) = z(0)  (- SU0.       []

If {b ,...,b }
  1      m denotes a basis of SU0, then there exists a matrix B  (- GL(n, k), such that Bbi = ei, the i-th unit vector in Kn. Applying this matrix B as a coordinate transformation on Kn as in Lemma 1 we get that SUB-1 - 1 = <e 1, ..., em>, the m-dimensional linear space generated by the first m unit vectors in Kn. Thus without loss of generality we may assume that SU = <e1, ..., em>.

Theorem 6. Let U: A --> GL(n, K) be a mapping, such that SU0 = <e1, ..., em> and U(0) = In. Then U(c) can be partitioned as a block matrix of the form

        (                  )
U (c) =    U11(c)    U12(c)  ,
          (0)n-m,m   U22(c)

where U11(c)  (- GL(m, K), U22(c)  (- GL(n - m, K) and U12(c)  (- Mm,n-m(K). These matrices satisfy the boundary conditions U11(0) = Im, U22(0) = In-m and U12(0) = (0)m,n-m, the zero matrix. Moreover, U11 is an exponential function, i.e. U11(c + m) = U11(c)U11(m) for all c, m  (- A.

Each y  (- SU can be expressed as

        (           )
y(c) =    U11(c)y(0)  ,      A c  (-  A,
              0

and y (0)  (- Km.

Proof.  Let y  (- SU , then y(c)  (- SU0 = <e1, ..., em>, so y(c) = (      )
  y(c)
    0 and y (c)  (- Km. We partition U(c) as a block matrix
        (                )
U (c) =   U11(c)  U12(c)   ,
          U21(c)  U22(c)
(5)
such that U11(c) is an m × m-matrix. From (3) we deduce that
y(c)   =  U11(c)y(0) + U12(c)0  = U11(c)y(0)
    0  =  U21(c)y(0) + U22(c)0  = U21(c)y(0).
Since y (0) is an arbitrary element of Km, it is clear that U 21(c) = (0)n-m,m for all c  (- A. Because of the fact that U(c) is regular, both U11 and U22 are regular matrices as well. From U(0) = In the boundary conditions follow. Inserting into (4) the form of U and y just described we get
(                        ) (      )
  U11(c + m)   U12(c + m)    y(0)
   (0)n-m,m    U22(c + m)      0    =

(                  ) (                   ) (     )
   U11(c)    U12(c)     U11(m)    U12(m)     y(0)
  (0)n-m,m   U22(c)    (0)n-m,m   U22(m)      0    ,

which means that U11(c + m)y (0) = U11(c)U11(m)y (0) for all y (0)  (- Km and c, m  (- A, so that U11 is an exponential function.       []

We are also interested in subspaces of SU . First we present a generalization of Lemma 5.

Lemma 7. Let S be a subspace of SU . Then the following statements are equivalent:

  1. S is a U(c0)-invariant space for all c0  (- A.
  2. S is invariant under translations.
  3. S0 is U(c 0)-invariant for all c0  (- A, where S0 := {y(0) |y  (-  S}.
Proof.  In order to prove that 1 implies 2, we set z(c) := y(c + c0) for arbitrary c0  (- A. Since S is U(c0)-invariant, U(c0)y  (- S. It is enough to prove that z = U(c0)y, since then z  (- S. For c  (- A we get z(c) = y(c0 + c) = U(c0 + c)y(0) = U(c0)U(c)y(0) = U(c0)y(c), so z = U(c0)y by (3), (4) and (3).

To each y0  (- S0 there exists y  (- S, such that y0 = y(0). Under the assumption 2, the function z(c) := y(c + c0) belongs to S for any c0  (- A. So z(0)  (- S0 and z(0) = y(c0) = U(c0)y(0) = U(c0)y0 by (3). Thus we proved that 2 implies 3.

In order to close the cycle of implications take y  (- S. Then y(0)  (- S0. For arbitrary c0  (- A also U(c0)y(0) belongs to S0. Hence, there exists z  (- S such that z(0) = U(c0)y(0). Taking into account that S is a subspace of SU we can write z as z(c) = U(c)z(0) = U(c)U(c0)y(0) = U(c + c0)y(0) = U(c0)U(c)y(0) = U(c0)y(c) by (3), (4), (4) and (3). Thus U(c0)y  (- S.       []

A generalization of Theorem 6 is

Theorem 8. Let S be a k-dimensional U-invariant subspace of SU . Then there exist coordinates in Kn, such that S0 = <e 1, ..., ek>, and U(c) is a block matrix of the form

        (                  )
U (c) =    U11(c)   U12(c)   ,
          (0)n-k,k  U22(c)
(6)
where U11(c)  (- GL(k, K), U22(c)  (- GL(n - k, K) and U12(c)  (- Mk,n-k(K), such that U11(0) = Ik, U22(0) = In-k, U12(0) = (0)k,n-k. Moreover, U11 is an exponential function, and y  (- S if and only if
                   (         )
y(c) = U (c)y(0) =    U11(c)q
                        0

for q  (- Kk.

So far we described solutions (U, y) of (2) when the mapping U was given. Now we will assume that a linear subspace S0 of Kn is given and we describe all solutions (U, y) of (2), such that SU0 = S0. Let S0 be a k-dimensional U-invariant subspace of Kn, then without loss of generality S0 = <e 1, ..., ek>.

Theorem 9. Let S0 = <e 1, ..., ek> be a subspace of Kn, and let U 11(c)  (- GL(k, K), U22(c)  (- GL(n - k, K) and U12(c)  (- Mk,n-k(K), such that U11(0) = Ik, U22(0) = In-k, U12(0) = (0)k,n-k. Moreover U11 is assumed to be an exponential function. Then

     {           (         )          }
                   U11(c)q           k
S :=   y |y(c) =       0     , q  (-  K

is a U-invariant subspace of SU , where U is given by (6).

Proof. 
                (                        ) (         )    (                   )
                  U11(c + m)   U12(c + m)    U11(n)q        U11(c + m)U11(n)q
U (c + m)y(n)  =    (0)n- k,k    U22(c + m)        0     =            0           =

(                 )   (                   )
  U11(c + m + n)q       U11(c)U11(m  + n)q
         0          =            0           = U (c)y(m + n)

      []

When does S = SU hold?

Lemma 10. The two spaces S and SU coincide if and only if for all j  (- Kn-k \ {0} there exists (c0, m0)  (- A2, such that

(                                                   )
  [U11(c0)U12(m0) + U12(c0)U22(m0) - U12(c0 + m0)]j   /=  0.
           [U22(c0)U22(m0) - U22(c0 +  m0)]j
(7)
Proof.  From Lemma 2 and Lemma 3 we know that S is a subspace of SU different from SU if and only if there exists y0 / (- S0, such that [U(c + m) - U(c)U(m)]y0 = 0 for all c, m  (- A. In other words, S = SU if and only if for each y0 / (- S0 there exists (c 0, m0)  (- A2, such that [U(c0+m0)-U(c0)U(m0)]y0/=0. When writing y0 in the form (  )
  q
  j for q  (- Kk and j  (- Kn-k, then y0  (- Kn \ S0 if and only if j/=0. Together with (6) we get
                           (                                              )
                       0     [U11(c)U12(m) + U12(c)U22(m) -  U12(c + m)]j
[U (c + m) - U (c)U (m)]y  =           [U22(c)U22(m) - U22(c + m)]j           ,

which finishes the proof.       []

Now we are going to present several examples for the situation S = SU , i.e. by Lemma 10 examples, where condition (7) is satisfied. Here we always assume that A = K. First we will deal with the second line of condition (7). Secondly, if this condition is not satisfied by all j, then let V denote the set

{j  (-  Kn -k |[U22(c)U22(m) - U22(c + m)]j = 0,  A c,m  (-  A} .

Thus V is an r-dimensional subspace of Kn-k for 0 < r < n - k. In order to satisfy the requirements of Lemma 10 in this situation as well, the first line in (7) must be satisfied for j  (- V .

Now we describe some examples how to construct U22: K --> GL(s, K) for s < n - k, such that

 A j  (-  Ks \{0}  E c ,m  (-  K : [U  (c )U  (m ) - U  (c  + m )]j /= 0.
                 0  0          22  0   22  0      22  0    0
(8)
Case charK/=2:
Set U22(c) = cIs for all c  (- K \ {0} with c  (- K \ {0,1}. Then c2/=c. For c0 = m0 = 1 we get [U22(1)U22(1)-U22(1+1)]j = [c2I s -cIs]j = (c2-c)I sj/=0 for all j/=0.
Case charK = 2 and |K | > 2:
There exists c  (- K\{0}, such that c2/=1. Let U 22(c) = cIs for all c  (- K \ {0}, then for c0 = m0 = 1 we get [U22(1)U22(1) - U22(1 + 1)]j = [U22(1)2 - U 22(0)]j = [c2I s - Is]j = (c2 - 1)I sj/=0 for all j/=0.
Case |K | = 2:
If s = 1 each mapping U22: K --> GL(1, K) = {1} is a homomorphism, so (8) cannot be satisfied. If s > 1 there exist matrices M  (- GL(s, K) of order 2s-1. As a permutation of the vectors in Ks the cycle decomposition of M consists of one fixed point, the 0-vector, and a cycle of length 2s-1. (Actually, cf. [2] 3.5 Theorem, there are f(2s - 1)/s irreducible polynomials of degree s over K = GF (2), such that the companion matrix of these polynomials is of order 2s - 1.) If U 22(1) = M, then for c0 = m0 = 1 we get [U22(1)U22(1) - U22(1 + 1)]j = [M2 - I s]j/=0 for all j/=0, since 2 < 2s - 1.

Now we describe examples how to construct U12: K --> Mk,n-k(K), such that

 A j  (-  Kn -k\{0}  E c0,m0  (-  K : [U11(c0)U12(m0) + U12(c0)U22(m0) - U12(c0 + m0)]j /= 0.
(9)
Again we assume that A = K. Furthermore we assume that both U11 and U22 are exponential functions. Hence r = n - k. From the preceding considerations we already know that U11(c)  (- GL(k, K), U22(c)  (- GL(r, K), U11(0) = Ik, U22(0) = Ir and U12(0) = (0)k,r. Again we describe several different cases:
Case charK/=2:
If k > r, then assume that U12(1) = (0)k,r and
          (                   )
            - 1   0   ...   0
             0   - 1        0
             ...        ...   ...      (         )
U   (2) =                        =     - Ir    ,
  12         0    0   ...  - 1       (0)k-r,r
             0.   ...  ...   0.
             ..              ..
             0   ...  ...   0

where the upper part is -Ir and the lower part is a 0-matrix of the dimension (k - r) × r. Then for c0 = m0 = 1 we get that U11(1)U12(1) + U12(1)U22(1) - U12(1 + 1) = -U12(2) and it is obvious that -U12(2)j/=0 for all j  (- Kr \ {0}.

For k < r one possible way to proceed is indicated in

Lemma 11. If there are enough elements in K, to be more precise, if |K | > 2 |~   ~| 
  r
  k + 1, then it is always possible to find c0 and m0 satisfying (9).

Proof.  There exist uniquely determined integers q, s, such that r = kq + s and 0 < s < k. If q > 0 assume that c1  (- K \ {0}, then -c1  (- K \ {0,c1}. Let U12(±c1) be given by
U12(c1) = ( Ik  (0)k,r-k ) .U22(c1) and U12(- c1) = U11(- c1) .(Ik (0)k,r-k ).

If q > 1 and |K | is big enough, then there exists c2  (- K \{0, ±c1} and we assume that

U12(c2) = ((0)k,k   Ik  (0)k,r-2k ) .U22(c2) and
U  (- c ) = U  (- c ) .( (0)   I   (0)      ).
 12    2     11    2      k,k   k     k,r-2k

Going on like this we can find elements c1, ..., cq  (- K and matrices U12(±ci). If s > 0 and |K | is big enough, then there exists cq+1  (- K \ {0,± c1,...,± cq} and we assume that

            (   (0)s,qk      Is   )
U12(cq+1) =                       .U22(cq+1) and
              (0)k-s,qk   (0)k- s,s
                           (                  )
                              (0)s,qk      Is
U12(- cq+1) = U11(- cq+1) .  (0)k-s,qk  (0)k-s,s   .

When j  (- Kr \ {0}, then there exists 1 < i < r, such that j i/=0. Hence, there exists j  (- {1,2,...,q + 1}, such that (j - 1)k < i < jk. For c0 = cj and m0 = -cj we have U11(cj)U12(-cj) + U12(cj)U22(-cj) - U12(0) = 2U12(cj)U22(-cj). According to the choice of i and j it is clear that (9) is satisfied.       []

This is a very general result, but it is not the best result which is possible.

Example 12. Let K be the prime field of characteristic 3, k = 1 and r = 2. In this case |K | < 2 |~ r ~| 
 k + 1, but it is also possible to find U12, such that (9) is satisfied. For instance U given by

        (         )              (          )
          1   1  0                 1   0  0
U (1) =   0   1  1       U (2) =   0   1  2

          0   0  1                 0   0  1

satisfies (9).

Case charK = 2 and |K | > 2:
Assume first k > r. There exists c  (- K \ {0, 1} and then c + 1 / (- {0,1, c}. Let furthermore U12(1) = U12(c) = (0)k,r and
             (         )
U12(c + 1) =      Ir     ,
                (0)k- r,r

then for c0 = 1 and m0 = c we get U11(1)U12(c) + U12(1)U22(c) - U12(c + 1) = U12(c + 1) and (9) is satisfied. For k < r the following lemma holds:

Lemma 13. If there are enough elements in K, to be more precise, if |K | > 2 |~  r ~| 
  k + 2, then it is always possible to find c0 and m0 satisfying (9).

Proof.  There exist uniquely determined integers q, s, such that r = kq + s and 0 < s < k. Assume U12(1) = (0)k,r. For 1 < i < q there exists ci  (- K \ {0,1,cj,cj + 1 |j < i}, such that ci + 1 / (- {0, 1,cj,cj + 1,ci| j < i}. Let U12(ci) and U12(ci + 1) be given by
U12(ci) = (0)k,r and
U12(ci + 1) = ( (0)k,(i- 1)k Ik (0)k,r-ik ) .

If s > 0 and |K | is big enough, then there exists cq+1  (- K, such that cq+1, cq+1 + 1  (- K \ {0,1,cj,cj + 1 |j < q} and we assume that

U12(cq+1) = (0)k,r and

                (                  )
U12(cq+1 + 1) =    (0)s,qk      Is     .
                  (0)k-s,qk  (0)k-s,s

Given j  (- Kr \ {0} there exists 1 < i < r, such that j i/=0. Hence, there exists j  (- {1,2,...,q + 1}, such that (j - 1)k < i < jk. For c0 = cj and m0 = 1 we have U11(cj)U12(1) + U12(cj)U22(1) - U12(cj + 1) = -U12(cj + 1). According to the choice of i and j it is clear that (9) is satisfied.       []

Case |K | = 2:
In the situation r > k we can only give partial results. If c0 = 0 or m0 = 0, then it is impossible to satisfy (9). Since U11 and U22 are exponential functions, the orders of U11(1) and U22(1) are divisors of 2. If both U11(1) = Ik and U22(1) = Ir, then it is also impossible to satisfy (9). If r = 1, then U22(1) is the identity matrix I1. From the previous statements it is clear that necessarily k > 1 and U11(1) must be a matrix of order 2. If U is defined by
         (                     )
            1  .0   ....   0   1                    ( 0 )
            0   ..   ..       0                      ..
U11(1) =    ...       ...  ...  ...     and   U12(1) =   .   ,
            ..            ..                          0
            .              .  0                      1
            0   0    0    0   1

then (9) is satisfied. For r = 2 assume that U11(1) is given as above and

                                   (  1  0 )
         (      )
           1   1                      0.  0.
U22(1) =   0   1    and   U12(1) =    ..  ..   ,
                                      0  0
                                      0  1

then again (9) is satisfied. For k = r = 3 and for any choice of U11(1), U22(1)  (- GL(3, K) of order dividing 2 the computer did not find a matrix U12(1) in M3(K) such that (9) is satisfied. Other cases were not studied so far.

If k > r it is not possible to satisfy (9), since there is only one possible choice c0 = m0 = 1, which determines exactly one matrix U11(1)U12(1) + U12(1)U22(1). This matrix describes a homomorphism from Kk to Kr, which has a kernel of dimension > k - r > 0.

3 The general situation

In this part we generalize the functional equation (2) by assuming that U(c) is not necessarily a regular matrix, i.e. U: A --> Mn(K). Also in this situation Lemma 1 holds. When we define SU and SU0 as it was done earlier, then SU and SU0 are K-linear spaces (cf. Lemma 4). Again SU0 is an m-dimensional subspace of Kn for 0 < m < n, and S U is invariant under translations, and SU0 = {y(c) |c  (-  A, y  (-  SU } (cf. Lemma 5). Without loss of generality we can assume (as in the earlier case) that there exists a basis of Kn, such that SU0 = <e1, ..., em>.

Since U(0) need not be a regular matrix, we do not get the results of Lemma 2, and in general there is no isomorphism between SU and SU0.

For c = n = 0 or n = 0 we derive from (2)

Lemma 14. Let (U, y) be a solution of (2), then

U (0)y(m) = U (m)y(0),      A m  (-  A,
(10)
U(c)y(m) =  U(0)y(c + m),      A c, m  (-  A.
(11)

If U(c) is partitioned as in (5) and y(c) is written as (      )
  y(c)
    0 for y (c)  (- Km, then from (10) we get

(  U  (0)  U  (0)) ( y(m) )    ( U  (m)  U   (m)) ( y(0) )
    11      12              =     11       12              ,
   U21(0)  U22(0)      0         U21(m)  U22(m)       0

which leads to the system of equations

U11(0)y(m)  =   U11(m)y(0)      A m  (-  A

U21(0)y(m)  =   U21(m)y(0)     A m  (-  A.
(12)

Lemma 15. Let (U, y) be a solution of (2). Then there exists a system of coordinates of Kn, such that

       (   U  (c)   U   (c) )
U(c) =      11        12     ,      A c  (-  A,
         (0)n- m,m   U22(c)
(13)
where the m × m-matrix U11(0) is the block matrix of the form
         (                      )
              Ik       (0)k,m- k
U11(0) =   (0)m- k,k  (0)m -k,m -k
(14)
for some k < m.
Proof.  According to Lemma 1 choose matrices C  (- GL(n, K) and B'  (- GL(m, K), such that
       (                    )    (                  )
            B'      (0)m,n- m         V11(0)    V12(0)
CU (0)   (0)n-m,m    In- m     =   (0)n-m,m   V22(0)

and

        (                       )
             Ik       (0)k,m -k
V11(0) =    (0)m -k,k   (0)m-k,m-k   .

Without loss of generality assume that U = V . From the second line of (12) we deduce that 0 = 0y (m) = U21(m)y (0) for all m  (- A. Since y (0) can arbitrarily be chosen in Km, it is clear that U21(m) = (0)n-m,m for all m  (- A.       []

Since SU0 = <e1, ..., em>, there exist y1, ..., ym  (- SU , such that yj(0) = ej, the j-th unit vector in Kn, for 1 < j < m. Let S U ' := <y 1, ..., ym>, then SU ' is an m-dimensional subspace of SU . In order to prove this, it is only necessary to show that y1, ..., ym are linearly independent. Let a1, ..., am  (- K, such that  sum i = 1maiyi = 0, then also  sum i = 1maiyi(0) = 0, which implies  sum i = 1maiei = 0, so that a1 = ... = am = 0.

For y  (- SU ' there exist uniquely defined a 1, ..., am  (- K such that y =  sum i = 1maiyi. These ai can be read from y(0), since y(0) =  sum i = 1maiei.

Define the m × m-matrix Y (c) corresponding to the chosen y1, ..., ym by

Y (c) = (y1(c),...,ym(c)),
(15)
i.e. the j-th column of Y (c) is the vector y j(c)  (- Km. Then for y  (- S U ' we have
                          (     )
        sum m                   a1
y(c) =     aiyi(c) =  Y(c)    ...    =  Y(c)y(0).
       i=1                  a
                              m
(16)
Replacing y by yj in the first line of (12) we get for all m  (- A
U11(0)yj(m) =  U11(m)yj(0) = U11(m)ej,    j =  1,...,m.

These equations are collected to the matrix equation

U11(0)Y (m) = U11(m),      A m  (-  A.
(17)
The special form of U from Lemma 15 inserted into (11) yields for y = yj the equation
U11(c)yj(m) =  U11(0)yj(c + m),     A c, m  (-  A.

Again these equations can be collected for j = 1, ..., m and we derive

U11(c)Y (m) = U11(0)Y (c + m),     A c, m  (-  A.
(18)
Equations (17) and (18) together yield
U11(0)[Y (c + m)-  Y (c)Y (m)] = (0)m,m,      A c,m  (-  A.
(19)
According to the special form of U11(0) described in Lemma 15 we partition Y (c) as a block matrix
        (                )
Y (c) =   Y11(c)  Y12(c)   ,
          Y21(c)  Y22(c)

such that Y 11(c) is a k × k-matrix. We note that the “auxiliary” matrix function Y : A --> Mm(K), which will help us to describe the space SU of solutions y (for given U), is in general not uniquely determined. However, from (17), from the decomposition of U11(0) in Lemma 15 and the corresponding decomposition of Y (c) we see that Y 11(c) and Y 12(c) are uniquely determined by U11(c), namely

                       (       ) (                )   (                      )
U  (m) = U  (0)Y (m) =   Ik  0     Y11(m)  Y12(m)   =    Y11(m)      Y12(m)     .
 11        11             0   0     Y21(m)  Y22(m)       (0)m -k,k   (0)m -k,m -k
(20)
Then (19) can be rewritten as
(       )[(                        )   (                )(                )]
  Ik   0    Y11(c + m)   Y12(c + m)  -   Y11(c)  Y12(c)     Y11(m)   Y12(m)   = (0)
   0   0    Y21(c + m)   Y22(c + m)      Y21(c)  Y22(c)     Y21(m)   Y22(m)        m,m

and we end up with the system of equations

Y11(c + m)  =   Y11(c)Y11(m) + Y12(c)Y21(m)
                                                 A c, m  (-  A.
Y12(c + m)  =   Y11(c)Y12(m) + Y12(c)Y22(m)
(21)
From Y (0) = Im we deduce that Y 11(0) = Ik, Y 21(0) = (0)m-k,k, Y 12(0) = (0)k,m-k and Y 22(0) = Im-k. If m is replaced by m1 + m2 and taking into account that + is an associative composition we get from the first line of (21) that Y 11(c+(m1+m2)) = Y 11(c)Y 11(m1+m2)+Y 12(c)Y 21(m1+m2) = Y 11(c)[Y 11(m1)Y 11(m2)+Y 12(m1)Y 21(m2)]+Y 12(c)Y 21(m1+m2) is equal to Y 11((c + m1) + m2) = Y 11(c + m1)Y 11(m2) + Y 12(c + m1)Y 21(m2) = [Y 11(c)Y 11(m1) + Y 12(c)Y 21(m1)]Y 11(m2) + [Y 11(c)Y 12(m1) + Y 12(c)Y 22(m1)]Y 21(m2), which yields
Y12(c)[Y21(m1 + m2) - Y21(m1)Y11(m2) - Y22(m1)Y21(m2)] = 0,   A c, m1,m2  (-  A.
(22)
In the same way we can derive from the second line of (21) that
Y12(c)[Y22(m1 + m2) - Y21(m1)Y12(m2) - Y22(m1)Y22(m2)] = 0,   A c, m1,m2  (-  A.
(23)
Each Y 12(c) determines a homomorphism from Km-k to Kk. Let W :=  /~\ c (- A ker Y 12(c). Then W is an r-dimensional subspace of Km-k for 0 < r < m - k with basis {          }
  ^d1,..., ^dr. Moreover, there exists an m - k - r-dimensional subspace V of Km-k, such that Km-k = V  o+ W . Let {^c1,...,^cm-k-r} be a basis of V . We embed Km-k in a natural way into Km by placing k zeros in front of each vector, i.e.
     (   )                     (   )
ci :=   0    (-  Km   and  di :=   0    (-  Km.
       ^ci                        ^di

Then it is possible to find a matrix B''  (- GL(m - k, K), such that the coordinate transformation on Km induced by

      (                   )
 '         Ik     (0)k,m- k
B  :=    (0)m -k,k     B''      (-  GL(m, K)

satisfies

  '
B  ei = ei          1 < i < k
B'ci = ek+i         1 < i < m -  k-  r
B'di = em -r+i      1 < i < r.

Let B be the corresponding coordinate transformation on Kn

      (                    )
           B'     (0)m,n-m
B  :=    (0)n-m,m    In- m      (-  GL(n, K).

If U is decomposed as in (13) and (14), then also UB has this property.

Without loss of generality we assume that the basis of Kn was chosen in such a way that Lemma 15 is satisfied and that {ek+1,...,em- r} and {em -r+1,...,em} are a basis of V or W respectively. Then it is useful and important to partition Y (c) further as a 3 × 3 block matrix of the form

        (                         )
           Z11(c)  Z12(c)   Z13(c)
Y (c) =    Z21(c)  Z22(c)   Z23(c)   ,
           Z  (c)  Z  (c)   Z  (c)
            31       32      33

such that Z11(c) = Y 11(c)  (- Mk(K), Z22(c)  (- Mm-k-r(K) and Z33(c)  (- Mr(K). Hence

                                      ( Z  (c) )            ( Z  (c)  Z  (c) )
Y12(c) = ( Z12(c)  Z13(c) ), Y21(c) =     21      , Y22(c) =    22       23      .
                                        Z31(c)                Z32(c)  Z33(c)

Let x = (  )
  ~x
  ^x denote a vector in Km-k, where ~x  (- Km-k-r and x^  (- Kr. Then x belongs to W if and only if ~x = 0. Moreover Y 12(c)|W = 0 for all c  (- A, which means Z13(c) = (0)k,r for all c  (- A. From the definition of W it is clear that Z12(c)~x = 0 for all c  (- A is equivalent to ~x = 0.

The first line of (21) reads now as

Z11(c + m) = Z11(c)Z11(m)  + Z12(c)Z21(m)    A c,m  (-  A.

From the second line of (21) we derive

Z12(c + m) = Z11(c)Z12(m) +  Z12(c)Z22(m),    A c,m  (-  A

and

(0)   = Z   (c + m) =  Z  (c)(0)   + Z  (c)Z  (m),    A c,m  (-  A.
   k,r     13           11      k,r    12     23

Hence, each column of Z23(m) is 0  (- Km-k-r, so that Z 23(m) = (0)m-k-r,m-k-r for all m  (- A.

From (22) we deduce

Z12(c)[Z21(m1 + m2) - Z21(m1)Z11(m2) - Z22(m1)Z21(m2)] = (0)k,k,    A c,m1, m2  (-  A.

Let M denote the matrix between the two braces [ and ], then each column of M is 0  (- Km-k-r and consequently M = (0) m-k-r,k. Hence, we proved that

Z  (m +  m )-  Z  (m )Z  (m ) - Z  (m )Z  (m ) = (0)       ,   A m ,m   (-  A.
 21  1    2     21  1  11  2     22  1   21   2      m -k-r,k      1  2

The same way we deduce from (23) that

Z12(c)[Z22(m1 + m2) - Z21(m1)Z12(m2) - Z22(m1)Z22(m2)] = (0)k,m -k-r,   A c, m1,m2  (-  A

and correspondingly

Z22(m1 + m2) - Z21(m1)Z12(m2) -  Z22(m1)Z22(m2) = (0)m- k- r,m- k- r,   A m1,m2  (-  A.

This finishes the proof of

Theorem 16. There exists a coordinate system of Km, such that Y (c) is a solution of (19) if and only if Y (c) can be written as

        (                            )
          Z11(c)  Z12(c)     (0)k,r
Y(c) =    Z21(c)  Z22(c)   (0)m- k-r,r   ,
          Z  (c)  Z   (c)    Z  (c)
           31       32        33

where

(                 )
  Z11(c)   Z12(c)
  Z   (c)  Z  (c)
    21       22

is an exponential function, Z11(c)  (- Mk(K), Z22(c)  (- Mm-k-r(K), Z33(c)  (- Mr(K), satisfying the conditions Z11(0) = Ik, Z22(0) = Im-k-r, Z33(0) = Ir, Z12(0) = (0)k,m-k-r, Z21(0) = (0)m-k-r,k, Z31(0) = (0)r,k and Z32(0) = (0)r,m-k-r. For c/=0 the matrices Z31(c), Z32(c), Z33(c) can be arbitrarily chosen.

Next we describe the structure of SU in more details.

Lemma 17. For each y  (- SU with y(0)/=0 there exists a subspace SU ' of S U , such that y  (- SU '.

Proof.  Since y(0)/=0 also y (0)/=0. Hence there exist z20, ..., zm0  (- Kn, such that {y(0),z02,...,z0m} is a basis of SU0 = <e1, ..., em>. Consequently, there exist z2, ..., zm  (- SU , such that zj(0) = zj0 for j = 2, ..., m. Then y, z2, ..., zm are linearly independent, which implies that <y, z2, ..., zm> is an m-dimensional subspace of SU . Hence, there exist y1, ..., ym  (- <y, z2, ..., zm>, such that yj(0) = ej for j = 1, ..., m and y  (- <y, z2, ..., zm> = <y1, ..., ym> =: SU '.       []

Let N(SU ) denote the set {z  (-  SU | z(0) = 0}. Then N(SU ) is a subspace of SU . The appearance of this subspace N(SU ) of SU , which is in general not {0}, is one of the main differences to the case of mappings U: A --> GL(n, K). We will see that N(SU ) is closely related to the space W . This is described in

Lemma 18. A function z: A --> Kn belongs to N(S U ) if and only if

                     (     )
                        0
z(0) = 0 and  z(c) =   ^z(c)   for ^z(c)  (-  W.
(24)
Proof.  The function z belongs to N(SU ) if and only if z (0) = 0 and U11(c + m)z (n) = U11(c)z (m + n) for all c, m, n  (- A. Especially for c = n = 0 and because of the particular form of U11 given in (20) we get
(                      ) (   )    (                      ) (      )
   Y11(m)      Y12(m)      0   =     Y11(0)     Y12(0)       ~z(m)   =
  (0)m -k,k  (0)m -k,m -k    0        (0)m-k,k  (0)m-k,m-k     ^z(m)

(                       ) (     )
      Ik       (0)k,m -k      ~z(m)
   (0)m -k,k   (0)m-k,m-k     ^z(m)  ,    A m  (-  A,

so that ~z (m) = 0 for all m  (- A. For n = 0 we derive

(                        ) (   )    (                      ) (     )
  Y11(c + m)  Y12(c + m)     0        Y11(c)      Y12(c)        0
   (0)m- k,k   (0)m -k,m -k    0   =    (0)m -k,k   (0)m-k,m-k     ^z(m)  ,

so that Y 12(c)^z (m) = 0 for all c, m  (- A. This however implies that ^z (m)  (- W for all m  (- A.

Assuming conversely that z (0) = 0, ~z (c) = 0 and ^z (c)  (- W for all c  (- A, then it is obvious that z  (- N(SU ).       []

In conclusion we get the following result:

Lemma 19. Let SU ' be an m-dimensional subspace of S U (constructed as above), then SU = SU '  o+ N(S U ).

Proof.  It is clear that SU '  /~\ N(S U ) = {0} and that SU '  o+ N(S U ) is a subspace of SU . Assume that x is an element of SU , then there exists y  (- SU ', such that y(0) = x(0). Then z(c) := x(c) - y(c) belongs to N(SU ). Hence x(c) = z(c) + y(c), which finishes the proof.       []

We notice that in the decomposition SU = SU '  o+ N(S U ) the space SU ' is in general not uniquely determined, whereas N(SU ) is unique by definition. However, SU ' can be any m-dimensional subspace of SU , such that the space of initial values y(0) (for y  (- SU ') is already SU0.

As an immediate consequence we get

dim  S  = dim S'  + dim N (S  ) = m + dim N (S  ).
      U         U           U                  U

The following Theorem 20 yields together with Theorem 16 the structure of the space SU , of all solutions of (2), for a given function U: A --> Mn(K). These theorems also contain necessary conditions on U in order to admit a nontrivial solution y.

Theorem 20. Let U: A --> Mn(K) be given and assume that dim SU0 = m. Then there exist coordinates in Kn and solutions (U, y j) of (2) for j = 1, ..., m, such that yj(0) = ej. Moreover, U(c) can be written as in (13) and U11(c) satisfies (20). Y 11(c) and Y 22(c) are the blocks in the first row of the matrix Y (c) given by (15). This matrix is also a solution of (19) and each element y of SU can be expressed as y(c) = (             )
  y(c) + z(c)
       0 for y given by (16) with arbitrary y (0)  (- Km and z (c) given by (24).

We finish by Theorem 21, which provides a construction of all solutions (U, y) of (2) by starting from an arbitrary subspace of initial values y(0) of Kn. This choice then leads via the block matrix Y (c) satisfying (19) to a matrix valued function U and a space S of solutions corresponding to U. In this general situation we do not discuss the problem when S = SU .

Theorem 21. If Y (c) satisfies (19), U11(m) is given by (20), y is given by (16) for arbitrary y (0)  (- Km and z (c) given by (24), then (U, y) is a solution of (2) for

       (             )
y(c) =   y(c) + z(c)   .
              0

Proof.  From the special form of U(c) and y(c) it is clear that (2) is satisfied if and only if U11(c + m)[y (n) + z (n)] = U11(c)[y (m + n) + z (m + n)] for all c, m, n  (- A. This is equivalent to U11(c + m)[Y (n)y (0) + z (n)] = U11(c)[Y (m + n)y (0) + z (m + n)]. Due to the definition of z (c) this is equivalent to U11(0)[Y (c + m)Y (n) - Y (c)Y (m + n)]y (0) = 0 for all c, m, n  (- A. Since y (0) is an arbitrary element of Km we derive
U11(0)[Y (c + m)Y (n) - Y (c)Y(m + n)] = (0)m,m,      A c, m,n  (-  A.

We can rewrite U11(0)[Y (c + m)Y (n) - Y (c)Y (m + n)] as U11(0)[Y (c + m)Y (n) - Y (c + m + n) + Y (c + m + n) - Y (c)Y (m + n)] = U11(0)[Y (c + m)Y (n) - Y (c + m + n)] + U11(0)[Y (c + m + n) - Y (c)Y (m + n)], which is equal to (0)m,m since (19) holds.       []

In order to determine all solutions (U, y) of (2) we start with an arbitrary m-dimensional subspace S0 of Kn for some 0 < m < n. Let {b1,...,bm} be a basis of S0, then there exists a matrix B  (- GL(n, K), such that Bbi = ei for 1 < i < m. Hence BS0 = <e 1, ..., em>. For each solution Y (c) of (19) described in Theorem 16 let U11(c) be given by (20) and U(c) be given by (13) with arbitrary matrices U12(c) and U22(c). Then each element y of

      {        (            )                                                   }
T  :=   y(c) =    y(c) + z(c)   |y(0)  (-  Km,  y(c) = Y (c)y(0), z(c) given by (24)
                      0

is together with U a solution of (2). Due to this construction T 0, the space of initial values y(0) for y  (- T , is equal to <e1, ..., em>. According to Lemma 1 each pair (UB, B-1y) for y  (- T is a solution of (2) and {B -1y(0)| y  (-  T} = S0. Hence, by varying S0 over all subspaces of Kn we determine all solutions (U, y) of (2).

References

[1]   H. Fripertinger and J. Schwaiger. Some applications of functional equations in astronomy. Grazer Mathematische Berichte, 344 (2001), 1-6.

[2]   R. Lidl and H. Niederreiter. Finite Fields, volume 20 of Encyclopedia of Mathematics and its Applications. Addison-Wesley Publishing Company, London, Amsterdam, Don Mills - Ontario, Sydney, Tokyo, 1983. ISBN 0-201-13519-1.

[3]   M.A. McKiernan. The matrix equation a(xoy) = a(x)+a(x)a(y)+a(y). Aequationes Mathematicae, 15 (1977), 213-223.

[4]   J. Schwaiger. Some applications of functional equations in astronomy. Aequationes Mathematicae, 60 (2000), p. 185. In Report of the meeting, The Thirty-seventh International Symposium on Functional Equations, May 16-23, 1999, Huntington, WV.

HARALD FRIPERTINGER
LUDWIG REICH
Institut für Mathematik
Karl-Franzens-Universität Graz
Heinrichstr. 36/4
A-8010 Graz
Austria
harald.fripertinger@kfunigraz.ac.at
ludwig.reich@kfunigraz.ac.at