## File: regress.c

package info (click to toggle)
chrony 1.21z-5
 123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628 /* $Header: /cvs/src/chrony/regress.c,v 1.32 2003/09/22 21:22:30 richard Exp$ ======================================================================= chronyd/chronyc - Programs for keeping computer clocks accurate. ********************************************************************** * Copyright (C) Richard P. Curnow 1997-2003 * * This program is free software; you can redistribute it and/or modify * it under the terms of version 2 of the GNU General Public License as * published by the Free Software Foundation. * * This program is distributed in the hope that it will be useful, but * WITHOUT ANY WARRANTY; without even the implied warranty of * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU * General Public License for more details. * * You should have received a copy of the GNU General Public License along * with this program; if not, write to the Free Software Foundation, Inc., * 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA * ********************************************************************** ======================================================================= Regression algorithms. */ #include #include #include #include #include "regress.h" #include "logging.h" #include "util.h" #define MAX_POINTS 128 void RGR_WeightedRegression (double *x, /* independent variable */ double *y, /* measured data */ double *w, /* weightings (large => data less reliable) */ int n, /* number of data points */ /* And now the results */ double *b0, /* estimated y axis intercept */ double *b1, /* estimated slope */ double *s2, /* estimated variance of data points */ double *sb0, /* estimated standard deviation of intercept */ double *sb1 /* estimated standard deviation of slope */ /* Could add correlation stuff later if required */ ) { double P, Q, U, V, W; double diff; double u, ui, aa; int i; if (n<3) { CROAK("Insufficient points"); } W = U = 0; for (i=0; i 0.0) && (resid[i] > 0.0))) { /* Nothing to do */ } else { nruns++; } } return nruns; } /* ================================================== */ /* Return a boolean indicating whether we had enough points for regression */ #define RESID_SIZE 1024 #define MIN_SAMPLES_FOR_REGRESS 3 int RGR_FindBestRegression (double *x, /* independent variable */ double *y, /* measured data */ double *w, /* weightings (large => data less reliable) */ int n, /* number of data points */ /* And now the results */ double *b0, /* estimated y axis intercept */ double *b1, /* estimated slope */ double *s2, /* estimated variance of data points */ double *sb0, /* estimated standard deviation of intercept */ double *sb1, /* estimated standard deviation of slope */ int *new_start, /* the new starting index to make the residuals pass the two tests */ int *n_runs, /* number of runs amongst the residuals */ int *dof /* degrees of freedom in statistics (needed to get confidence intervals later) */ ) { double P, Q, U, V, W; /* total */ double resid[RESID_SIZE]; double ss; double a, b, u, ui, aa; int start, nruns, npoints, npoints_left; int i; if (n < MIN_SAMPLES_FOR_REGRESS) { return 0; } start = 0; do { W = U = 0; for (i=start; i critical_runs10[npoints]) || (npoints_left < MIN_SAMPLES_FOR_REGRESS)) { break; } else { /* Try dropping one sample at a time until the runs test passes. */ ++start; } } while (1); /* Work out statistics from full dataset */ *b1 = b; *b0 = a; ss = 0.0; for (i=start; i 0 && !flags[u]) u--; if (flags[u]) u++; while (v < (n-1) && !flags[v]) v++; if (flags[v]) v--; do { if (v - u < 2) { if (x[v] < x[u]) { EXCH(x[v], x[u]); } flags[v] = flags[u] = 1; return x[index]; } else { pivind = (u + v) >> 1; EXCH(x[u], x[pivind]); piv = x[u]; /* New value */ l = u + 1; r = v; do { while (x[l] < piv) l++; while (x[r] > piv) r--; if (r <= l) break; EXCH(x[l], x[r]); l++; r--; } while (1); EXCH(x[u], x[r]); flags[r] = 1; /* Pivot now in correct place */ if (index == r) { return x[r]; } else if (index < r) { v = r - 1; } else if (index > r) { u = l; } else { CROAK("Impossible"); } } } while (1); } /* ================================================== */ #if 0 /* Not used, but this is how it can be done */ static double find_ordered_entry(double *x, int n, int index) { int flags[MAX_POINTS]; bzero(flags, n * sizeof(int)); return find_ordered_entry_with_flags(x, n, index, flags); } #endif /* ================================================== */ /* Find the median entry of an array x[] with n elements. */ static double find_median(double *x, int n) { int k; int flags[MAX_POINTS]; memset(flags, 0, n*sizeof(int)); k = n>>1; if (n&1) { return find_ordered_entry_with_flags(x, n, k, flags); } else { return 0.5 * (find_ordered_entry_with_flags(x, n, k, flags) + find_ordered_entry_with_flags(x, n, k-1, flags)); } } /* ================================================== */ /* This function evaluates the equation \sum_{i=0}^{n-1} x_i sign(y_i - a - b x_i) and chooses the value of a that minimises the absolute value of the result. (See pp703-704 of Numerical Recipes in C). */ static void eval_robust_residual (double *x, /* The independent points */ double *y, /* The dependent points */ int n, /* Number of points */ double b, /* Slope */ double *aa, /* Intercept giving smallest absolute value for the above equation */ double *rr /* Corresponding value of equation */ ) { int i; double a, res, del; double d[MAX_POINTS]; for (i=0; i 0.0) { res += x[i]; } else if (del < 0.0) { res -= x[i]; } } *aa = a; *rr = res; } /* ================================================== */ /* This routine performs a 'robust' regression, i.e. one which has low susceptibility to outliers amongst the data. If one thinks of a normal (least squares) linear regression in 2D being analogous to the arithmetic mean in 1D, this algorithm in 2D is roughly analogous to the median in 1D. This algorithm seems to work quite well until the number of outliers is approximately half the number of data points. The return value is a status indicating whether there were enough data points to run the routine or not. */ int RGR_FindBestRobustRegression (double *x, /* The independent axis points */ double *y, /* The dependent axis points (which may contain outliers). */ int n, /* The number of points */ double tol, /* The tolerance required in determining the value of b1 */ double *b0, /* The estimated Y-axis intercept */ double *b1, /* The estimated slope */ int *n_runs, /* The number of runs of residuals */ int *best_start /* The best starting index */ ) { int i; int start; int n_points; double a, b; double P, U, V, W, X; double resid, resids[MAX_POINTS]; double blo, bhi, bmid, rlo, rhi, rmid; double s2, sb, incr; double mx, dx, my, dy; int nruns = 0; if (n < 2) { return 0; } else if (n == 2) { /* Just a straight line fit (we need this for the manual mode) */ *b1 = (y[1] - y[0]) / (x[1] - x[0]); *b0 = y[0] - (*b1) * x[0]; *n_runs = 0; *best_start = 0; return 1; } /* else at least 3 points, apply normal algorithm */ start = 0; /* Loop to strip oldest points that cause the regression residuals to fail the number of runs test */ do { n_points = n - start; /* Use standard least squares regression to get starting estimate */ P = U = 0.0; for (i=start; i 0.0) { incr = 3.0 * sb; } else { incr = 3.0 * tol; } blo = b; bhi = b; do { blo -= incr; bhi += incr; /* We don't want 'a' yet */ eval_robust_residual(x + start, y + start, n_points, blo, &a, &rlo); eval_robust_residual(x + start, y + start, n_points, bhi, &a, &rhi); } while (rlo * rhi > 0.0); /* fn vals have same sign, i.e. root not in interval. */ /* OK, so the root for b lies in (blo, bhi). Start bisecting */ do { bmid = 0.5 * (blo + bhi); eval_robust_residual(x + start, y + start, n_points, bmid, &a, &rmid); if (rmid == 0.0) { break; } else if (rmid * rlo > 0.0) { blo = bmid; rlo = rmid; } else if (rmid * rhi > 0.0) { bhi = bmid; rhi = rmid; } else { CROAK("Impossible"); } } while ((bhi - blo) > tol); *b0 = a; *b1 = bmid; /* Number of runs test, but not if we're already down to the minimum number of points */ if (n_points == MIN_SAMPLES_FOR_REGRESS) { break; } for (i=start; i critical_runs10[n_points]) { break; } else { start++; } } while (1); *n_runs = nruns; *best_start = start; return 1; } /* ================================================== */