Posts: 1,389

Threads: 90

Joined: Aug 2007

Hey Gottfried,

did you notice that the sum over the logarithms of the eigenvalues of the Carleman matrix of exp converge (for increasing matrix size)?

Moreover also if you take the n-th power of the logarithms, for any n.

This woudl be a direct consequence of the matrix power method (for non-integer iteration of exp) converging to an analytic function.

Posts: 767

Threads: 119

Joined: Aug 2007

08/28/2009, 08:43 PM
(This post was last modified: 08/28/2009, 08:53 PM by Gottfried.)
(08/28/2009, 11:27 AM)bo198214 Wrote: Hey Gottfried,

did you notice that the sum over the logarithms of the eigenvalues of the Carleman matrix of exp converge (for increasing matrix size)?

Moreover also if you take the n-th power of the logarithms, for any n.

This woudl be a direct consequence of the matrix power method (for non-integer iteration of exp) converging to an analytic function.

Hmm, for dim=8..24 I get them always near null at machine-precision (Pari/GP, 200digit or 800 digits internal prec). That means the product of the eigenvalues is near 1 no matter what dimension I select. There may be an error, however the procedure is simple. Here is the Pari/Gp code

Code:

`fmt(800,12)`

{for(dim=8,24,

B = VE(fS2F,dim)*VE(P,dim)~ ; \\ construct the Bell(transp. Carlemann)-matrix for exp(x)

tmpW = mateigen(B); \\ getting the eigenvectors in tmpW

tmpD=HadDiv(B*tmpW,tmpW)[1,]; \\ getting the eigenvalues in tmpD

\\ this is simpler than the "official"

\\ method: tmpD = diag(tmpW^-1 * B * tmpW)

sulog = sum(k=1,#tmpD,log(tmpD[k]));

print(dim," ",sulog);

)}

8 -3.255463966 E-808

9 -9.22381457 E-808

10 4.88319595 E-808

11 2.821402104 E-807

12 6.51092793 E-808

13 -8.35569084 E-807

14 -1.519216517 E-807

15 3.157800047 E-806

16 -3.393278608 E-805

17 -5.444003874 E-804

18 -4.808428794 E-804

19 2.106242431 E-801

20 -1.177930451 E-800

21 -2.356690583 E-799

22 -3.10781078241 E-798

23 -3.76652961891 E-797

24 -3.06512885635 E-797

Example eigenvalues for dim=24

[4.28673736924 E-11]

[0.00000000296568145370]

[0.0000000957784063100]

[0.00000191766745327]

[0.0000266643844037]

[0.000273351215354]

[0.00214050544928]

[0.0130807983589]

[0.0630992348914]

[0.240819894743]

[0.729062578542]

[1.00000000000]

[1.89149672765]

[5.08315258442]

[15.5889319581]

[54.8943446446]

[221.893591029]

[1035.09334661]

[5636.83816890]

[36538.7788311]

[290981.552989]

[3004492.63267]

[44636646.3387]

[1247092190.35]

Wouldn't say, this is exactly convergence with increasing dimension... ;-)

Gottfried

Gottfried Helms, Kassel

Posts: 1,389

Threads: 90

Joined: Aug 2007

Hm, I dont know whether it has a deeper meaning its just an observation

Posts: 767

Threads: 119

Joined: Aug 2007

(08/28/2009, 08:58 PM)bo198214 Wrote: Hm, I dont know whether it has a deeper meaning its just an observation

Hmm, this means, the determinant of this matrix is 1 for finite dimension - which extends then also for iterates/powers.

That the determinant is 1 for finite dimension results also from the determinants of its factors, the stirling- ind the bpascal-matrix. Both are triangular and have units on the diagonal, so

det(fS2F) = det(P) = 1 and det(B) = det(fS2F * P~) = det(fS2F)*det(P~) = 1 *1.

Gottfried Helms, Kassel

Posts: 767

Threads: 119

Joined: Aug 2007

(08/28/2009, 09:10 PM)Gottfried Wrote: (08/28/2009, 08:58 PM)bo198214 Wrote: Hm, I dont know whether it has a deeper meaning its just an observation

Hmm, this means, the determinant of this matrix is 1 for finite dimension - which extends then also for iterates/powers.

That the determinant is 1 for finite dimension results also from the determinants of its factors, the stirling- ind the bpascal-matrix. Both are triangular and have units on the diagonal, so

det(fS2F) = det(P) = 1 and det(B) = det(fS2F * P~) = det(fS2F)*det(P~) = 1 *1.

It is also interesting in contrast to the version, which has the powerseries developed at the first complex fixpoint for exp(x), x0 = 0.318131505205 + 1.33723570143*I .

With dim=64 I get -at least for the integer iterates -1,1,2 the expected values to 10 digits accuracy - however, the sum of the logarithms of the eigenvalues should be simply, but consequently,

log(x0^0) + log(x0^1) + log(x0^2) + ...

= log(x0)*(0 + 1 + 2 + 3 + ...) =???= log(x0) * zeta(-1) =-0.0265109587671 - 0.111436308453*I

....

well, better not to ride the horse to death...

Gottfried

Gottfried Helms, Kassel