# -*- org -*-

#+CATEGORY: statistics

* From: James Theiler <jt@lanl.gov>

To: John Lamb <J.D.Lamb@btinternet.com>

Cc: gsl-discuss@sources.redhat.com

Subject: Re: Collecting statistics for time dependent data?

Date: Thu, 9 Dec 2004 14:18:36 -0700 (MST)

On Thu, 9 Dec 2004, John Lamb wrote:

] Raimondo Giammanco wrote:

] > Hello,

] > 

] >  I was wondering if there is a way to compute "running" statistics with

] > gsl.

] > 

] Yes you can do it, but there's nothing in GSL that does it and its eay 

] enough that you don't need GSL. Something like (untested)

] 

] double update_mean( double* mean, int* n, double x ){

]    if( *n == 1 )

]      *mean = x;

]    else

]      *mean = (1 - (double)1 / *n ) * *mean + x / n;

] }

] 

] will work and you can derive a similar method for updating the variance 

] using the usual textbook formula.

] 

] var[x] = (1/n) sum x^2_i - mean(x)^2

] 

] I don't know if there is a method that avoids the rounding errors. I 

] don't know why so many textbooks repeat this formula without the 

] slightest warning that it can go so badly wrong.

] 

]

Stably updating mean and variance is remarkably nontrivial.  There was

a series of papers in Comm ACM that discussed the issue; the final one

(that I know of) refers back to the earlier ones, and it can be found

in D.H.D. West, Updating mean and variance estimates: an improved

method, Comm ACM 22:9, 532 (1979)  [* I see Luke Stras just sent this

reference! *].  I'll just copy out the pseudocode since the paper is

old enough that it might not be easy to find.  This, by the way, is

generalized for weighted data, so it assumes that you get a weight and

a data value (W_i and X_i) that you use to update the estimates XBAR

and S2:

    SUMW = W_1

    M = X_1

    T = 0

    For i=2,3,...,n 

    {

       Q = X_i - M

       TEMP = SUM + W_i    // typo: He meant SUMW

       R = Q*W_i/TEMP

       M = M + R

       T = T + R*SUMW*Q

       SUMW = TEMP

    }

    XBAR = M

    S2 = T*n/((n-1)*SUMW)

jt 

-- 

James Theiler            Space and Remote Sensing Sciences

MS-B244, ISR-2, LANL     Los Alamos National Laboratory

Los Alamos, NM 87545     http://nis-www.lanl.gov/~jt

* Look at STARPAC ftp://ftp.ucar.edu/starpac/ and Statlib

http://lib.stat.cmu.edu/ for more ideas

* Try using the Kahan summation formula to improve accuracy for the

NIST tests (see Brian for details, below is a sketch of the algorithm).

      sum = x(1)

      c = 0

      DO i = 2, 1000000, 1

         y = x(i) - c

         t = sum + y

         c = (t - sum) - y

         sum = t

      ENDDO

* Prevent incorrect use of unsorted data for quartile calculations

using a typedef for sorted data (?)

* Rejection of outliers

* Time series. Auto correlation, cross-correlation, smoothing (moving

average), detrending, various econometric things. Integrated

quantities (area under the curve). Interpolation of noisy data/fitting

-- maybe add that to the existing interpolation stuff.What about

missing data and gaps?

   There is a new GNU package called gretl which does econometrics

* Statistical tests (equal means, equal variance, etc).