1dgenARMA11


Program to generate an ARMA(1,1) time series, for simulation studies.
Results are written to stdout.

Usage: 1dgenARMA11 [options]

Options:
========
 -num N  }  These equivalent options specify the length of the time
 -len N  }  series vector to generate.

 -nvec M  = The number of time series vectors to generate;
            if this option is not given, defaults to 1.

 -a a     = Specify ARMA(1,1) parameters 'a'.
 -b b     = Specify ARMA(1,1) parameter 'b' directly.
 -lam lam = Specify ARMA(1,1) parameter 'b' indirectly.

 -sig ss  = Set standard deviation of results [default=1].
 -norm    = Normalize time series so sum of squares is 1.
 -seed dd = Set random number seed.

  * The correlation coefficient r(k) of noise samples k units apart in time,
     for k >= 1, is given by r(k) = lam * a^(k-1)
     where                   lam  = (b+a)(1+a*b)/(1+2*a*b+b*b)
     (N.B.: lam=a when b=0 -- AR(1) noise has r(k)=a^k for k >= 0)
     (N.B.: lam=b when a=0 -- MA(1) noise has r(k)=b for k=1, r(k)=0 for k>1)
  * lam can be bigger or smaller than a, depending on the sign of b:
     b > 0 means lam > a;  b < 0 means lam < a.
  * What I call (a,b) here is sometimes called (p,q) in the ARMA literature.
  * For a noise model which is the sum of AR(1) and white noise, 0 < lam < a
     (i.e., a > 0  and  -a < b < 0 ).

 -CORcut cc = The exact ARMA(1,1) correlation matrix (for a != 0)
              has no non-zero entries.  The calculations in this
              program set correlations below a cutoff to zero.
              The default cutoff is 0.00010, but can be altered with
              this option.  The usual reason to use this option is
              to test the sensitivity of the results to the cutoff.

-----------------------------
A restricted ARMA(3,1) model:
-----------------------------
Skip the '-a', '-b', and '-lam' options, and use a model with 3 roots

 -arma31 a r theta vrat

 where the roots are z = a, z = r*exp(I*theta), z = r*exp(-I*theta)
 and vrat = s^2/(s^2+w^2) [so 0 < vrat < 1], where s = variance
 of the pure AR(3) component and w = variance of extra white noise
 added to the AR(3) process -- this is the 'restricted' ARMA(3,1).

 If the data has given TR, and you want a frequency of f Hz, in
 the noise model, then theta = 2 * PI * TR * f. If theta > PI,
 then you are modeling noise beyond the Nyquist frequency and
 the gods (and this program) won't be happy.

  # csh syntax for 'set' variable assignment commands
  set nt = 500
  set tr = 1
  set df = `ccalc "1/($nt*$tr)"`
  set f1 = 0.10
  set t1 = `ccalc "2*PI*$tr*$f1"`
  1dgenARMA11 -nvec 500 -len $nt -arma31 0.8 0.9 $t1 0.9 -CORcut 0.0001 \
       | 1dfft -nodetrend stdin: > qqq.1D
  3dTstat -mean -prefix stdout: qqq.1D \
       | 1dplot -stdin -num 201 -dt $df -xlabel 'frequency' -ylabel '|FFT|'
---------------------------------------------------------------------------
A similar option is now available for a restricted ARMA(5,1) model:
 -arma51 a r1 theta1 r2 theta2 vrat
where now the roots are
 z = a  z = r1*exp(I*theta1)  z = r1*exp(-I*theta1)
        z = r2*exp(I*theta2)  z = r2*exp(-I*theta2)
This model allows the simulation of two separate frequencies in the 'noise'.
---------------------------------------------------------------------------

Author: RWCox [for his own demented and deranged purposes]

Examples:
  1dgenARMA11 -num 200 -a .8 -lam 0.7 | 1dplot -stdin
  1dgenARMA11 -num 2000 -a .8 -lam 0.7 | 1dfft -nodetrend stdin: stdout: | 1dplot -stdin