Mercurial > repos > vipints > rdiff
view rDiff/src/locfit/Source/libmut.c @ 1:08d3a6143873
Deleted selected files
| author | vipints |
|---|---|
| date | Tue, 08 Oct 2013 06:54:25 -0400 |
| parents | 0f80a5141704 |
| children |
line wrap: on
line source
/* * Copyright 1996-2006 Catherine Loader. */ #include "mex.h" /* * Copyright 1996-2006 Catherine Loader. */ #include <math.h> #include "mut.h" /* stirlerr(n) = log(n!) - log( sqrt(2*pi*n)*(n/e)^n ) */ #define S0 0.083333333333333333333 /* 1/12 */ #define S1 0.00277777777777777777778 /* 1/360 */ #define S2 0.00079365079365079365079365 /* 1/1260 */ #define S3 0.000595238095238095238095238 /* 1/1680 */ #define S4 0.0008417508417508417508417508 /* 1/1188 */ /* error for 0, 0.5, 1.0, 1.5, ..., 14.5, 15.0. */ static double sferr_halves[31] = { 0.0, /* n=0 - wrong, place holder only */ 0.1534264097200273452913848, /* 0.5 */ 0.0810614667953272582196702, /* 1.0 */ 0.0548141210519176538961390, /* 1.5 */ 0.0413406959554092940938221, /* 2.0 */ 0.03316287351993628748511048, /* 2.5 */ 0.02767792568499833914878929, /* 3.0 */ 0.02374616365629749597132920, /* 3.5 */ 0.02079067210376509311152277, /* 4.0 */ 0.01848845053267318523077934, /* 4.5 */ 0.01664469118982119216319487, /* 5.0 */ 0.01513497322191737887351255, /* 5.5 */ 0.01387612882307074799874573, /* 6.0 */ 0.01281046524292022692424986, /* 6.5 */ 0.01189670994589177009505572, /* 7.0 */ 0.01110455975820691732662991, /* 7.5 */ 0.010411265261972096497478567, /* 8.0 */ 0.009799416126158803298389475, /* 8.5 */ 0.009255462182712732917728637, /* 9.0 */ 0.008768700134139385462952823, /* 9.5 */ 0.008330563433362871256469318, /* 10.0 */ 0.007934114564314020547248100, /* 10.5 */ 0.007573675487951840794972024, /* 11.0 */ 0.007244554301320383179543912, /* 11.5 */ 0.006942840107209529865664152, /* 12.0 */ 0.006665247032707682442354394, /* 12.5 */ 0.006408994188004207068439631, /* 13.0 */ 0.006171712263039457647532867, /* 13.5 */ 0.005951370112758847735624416, /* 14.0 */ 0.005746216513010115682023589, /* 14.5 */ 0.005554733551962801371038690 /* 15.0 */ }; double stirlerr(n) double n; { double nn; if (n<15.0) { nn = 2.0*n; if (nn==(int)nn) return(sferr_halves[(int)nn]); return(mut_lgamma(n+1.0) - (n+0.5)*log((double)n)+n - HF_LG_PIx2); } nn = (double)n; nn = nn*nn; if (n>500) return((S0-S1/nn)/n); if (n>80) return((S0-(S1-S2/nn)/nn)/n); if (n>35) return((S0-(S1-(S2-S3/nn)/nn)/nn)/n); return((S0-(S1-(S2-(S3-S4/nn)/nn)/nn)/nn)/n); } double bd0(x,np) double x, np; { double ej, s, s1, v; int j; if (fabs(x-np)<0.1*(x+np)) { s = (x-np)*(x-np)/(x+np); v = (x-np)/(x+np); ej = 2*x*v; v = v*v; for (j=1; ;++j) { ej *= v; s1 = s+ej/((j<<1)+1); if (s1==s) return(s1); s = s1; } } return(x*log(x/np)+np-x); } /* Raw binomial probability calculation. (1) This has both p and q arguments, when one may be represented more accurately than the other (in particular, in df()). (2) This should NOT check that inputs x and n are integers. This should be done in the calling function, where necessary. (3) Does not check for 0<=p<=1 and 0<=q<=1 or NaN's. Do this in the calling function. */ double dbinom_raw(x,n,p,q,give_log) double x, n, p, q; int give_log; { double f, lc; if (p==0.0) return((x==0) ? D_1 : D_0); if (q==0.0) return((x==n) ? D_1 : D_0); if (x==0) { lc = (p<0.1) ? -bd0(n,n*q) - n*p : n*log(q); return( DEXP(lc) ); } if (x==n) { lc = (q<0.1) ? -bd0(n,n*p) - n*q : n*log(p); return( DEXP(lc) ); } if ((x<0) | (x>n)) return( D_0 ); lc = stirlerr(n) - stirlerr(x) - stirlerr(n-x) - bd0(x,n*p) - bd0(n-x,n*q); f = (PIx2*x*(n-x))/n; return( FEXP(f,lc) ); } double dbinom(x,n,p,give_log) int x, n; double p; int give_log; { if ((p<0) | (p>1) | (n<0)) return(INVALID_PARAMS); if (x<0) return( D_0 ); return( dbinom_raw((double)x,(double)n,p,1-p,give_log) ); } /* Poisson probability lb^x exp(-lb) / x!. I don't check that x is an integer, since other functions that call dpois_raw() (i.e. dgamma) may use a fractional x argument. */ double dpois_raw(x,lambda,give_log) int give_log; double x, lambda; { if (lambda==0) return( (x==0) ? D_1 : D_0 ); if (x==0) return( DEXP(-lambda) ); if (x<0) return( D_0 ); return(FEXP( PIx2*x, -stirlerr(x)-bd0(x,lambda) )); } double dpois(x,lambda,give_log) int x, give_log; double lambda; { if (lambda<0) return(INVALID_PARAMS); if (x<0) return( D_0 ); return( dpois_raw((double)x,lambda,give_log) ); } double dbeta(x,a,b,give_log) double x, a, b; int give_log; { double f, p; if ((a<=0) | (b<=0)) return(INVALID_PARAMS); if ((x<=0) | (x>=1)) return(D_0); if (a<1) { if (b<1) /* a<1, b<1 */ { f = a*b/((a+b)*x*(1-x)); p = dbinom_raw(a,a+b,x,1-x,give_log); } else /* a<1, b>=1 */ { f = a/x; p = dbinom_raw(a,a+b-1,x,1-x,give_log); } } else { if (b<1) /* a>=1, b<1 */ { f = b/(1-x); p = dbinom_raw(a-1,a+b-1,x,1-x,give_log); } else /* a>=1, b>=1 */ { f = a+b-1; p = dbinom_raw(a-1,(a-1)+(b-1),x,1-x,give_log); } } return( (give_log) ? p + log(f) : p*f ); } /* * To evaluate the F density, write it as a Binomial probability * with p = x*m/(n+x*m). For m>=2, use the simplest conversion. * For m<2, (m-2)/2<0 so the conversion will not work, and we must use * a second conversion. Note the division by p; this seems unavoidable * for m < 2, since the F density has a singularity as x (or p) -> 0. */ double df(x,m,n,give_log) double x, m, n; int give_log; { double p, q, f, dens; if ((m<=0) | (n<=0)) return(INVALID_PARAMS); if (x <= 0.0) return(D_0); f = 1.0/(n+x*m); q = n*f; p = x*m*f; if (m>=2) { f = m*q/2; dens = dbinom_raw((m-2)/2.0, (m+n-2)/2.0, p, q, give_log); } else { f = m*m*q / (2*p*(m+n)); dens = dbinom_raw(m/2.0, (m+n)/2.0, p, q, give_log); } return((give_log) ? log(f)+dens : f*dens); } /* * Gamma density, * lb^r x^{r-1} exp(-lb*x) * p(x;r,lb) = ----------------------- * (r-1)! * * If USE_SCALE is defined below, the lb argument will be interpreted * as a scale parameter (i.e. replace lb by 1/lb above). Otherwise, * it is interpreted as a rate parameter, as above. */ /* #define USE_SCALE */ double dgamma(x,r,lambda,give_log) int give_log; double x, r, lambda; { double pr; if ((r<=0) | (lambda<0)) return(INVALID_PARAMS); if (x<=0.0) return( D_0 ); #ifdef USE_SCALE lambda = 1.0/lambda; #endif if (r<1) { pr = dpois_raw(r,lambda*x,give_log); return( (give_log) ? pr + log(r/x) : pr*r/x ); } pr = dpois_raw(r-1.0,lambda*x,give_log); return( (give_log) ? pr + log(lambda) : lambda*pr); } double dchisq(x, df, give_log) double x, df; int give_log; { return(dgamma(x, df/2.0, 0.5 ,give_log)); /* #ifdef USE_SCALE 2.0 #else 0.5 #endif ,give_log)); */ } /* * Given a sequence of r successes and b failures, we sample n (\le b+r) * items without replacement. The hypergeometric probability is the * probability of x successes: * * dbinom(x,r,p) * dbinom(n-x,b,p) * p(x;r,b,n) = --------------------------------- * dbinom(n,r+b,p) * * for any p. For numerical stability, we take p=n/(r+b); with this choice, * the denominator is not exponentially small. */ double dhyper(x,r,b,n,give_log) int x, r, b, n, give_log; { double p, q, p1, p2, p3; if ((r<0) | (b<0) | (n<0) | (n>r+b)) return( INVALID_PARAMS ); if (x<0) return(D_0); if (n==0) return((x==0) ? D_1 : D_0); p = ((double)n)/((double)(r+b)); q = ((double)(r+b-n))/((double)(r+b)); p1 = dbinom_raw((double)x,(double)r,p,q,give_log); p2 = dbinom_raw((double)(n-x),(double)b,p,q,give_log); p3 = dbinom_raw((double)n,(double)(r+b),p,q,give_log); return( (give_log) ? p1 + p2 - p3 : p1*p2/p3 ); } /* probability of x failures before the nth success. */ double dnbinom(x,n,p,give_log) double n, p; int x, give_log; { double prob, f; if ((p<0) | (p>1) | (n<=0)) return(INVALID_PARAMS); if (x<0) return( D_0 ); prob = dbinom_raw(n,x+n,p,1-p,give_log); f = n/(n+x); return((give_log) ? log(f) + prob : f*prob); } double dt(x, df, give_log) double x, df; int give_log; { double t, u, f; if (df<=0.0) return(INVALID_PARAMS); /* exp(t) = Gamma((df+1)/2) /{ sqrt(df/2) * Gamma(df/2) } = sqrt(df/2) / ((df+1)/2) * Gamma((df+3)/2) / Gamma((df+2)/2). This form leads to a computation that should be stable for all values of df, including df -> 0 and df -> infinity. */ t = -bd0(df/2.0,(df+1)/2.0) + stirlerr((df+1)/2.0) - stirlerr(df/2.0); if (x*x>df) u = log( 1+ x*x/df ) * df/2; else u = -bd0(df/2.0,(df+x*x)/2.0) + x*x/2.0; f = PIx2*(1+x*x/df); return( FEXP(f,t-u) ); } /* * Copyright 1996-2006 Catherine Loader. */ /* * Provides mut_erf() and mut_erfc() functions. Also mut_pnorm(). * Had too many problems with erf()'s built into math libraries * (when they existed). Hence the need to write my own... * * Algorithm from Walter Kr\"{a}mer, Frithjof Blomquist. * "Algorithms with Guaranteed Error Bounds for the Error Function * and Complementary Error Function" * Preprint 2000/2, Bergische Universt\"{a}t GH Wuppertal * http://www.math.uni-wuppertal.de/wrswt/preprints/prep_00_2.pdf * * Coded by Catherine Loader, September 2006. */ #include "mut.h" double erf1(double x) /* erf; 0 < x < 0.65) */ { double p[5] = {1.12837916709551256e0, /* 2/sqrt(pi) */ 1.35894887627277916e-1, 4.03259488531795274e-2, 1.20339380863079457e-3, 6.49254556481904354e-5}; double q[5] = {1.00000000000000000e0, 4.53767041780002545e-1, 8.69936222615385890e-2, 8.49717371168693357e-3, 3.64915280629351082e-4}; double x2, p4, q4; x2 = x*x; p4 = p[0] + p[1]*x2 + p[2]*x2*x2 + p[3]*x2*x2*x2 + p[4]*x2*x2*x2*x2; q4 = q[0] + q[1]*x2 + q[2]*x2*x2 + q[3]*x2*x2*x2 + q[4]*x2*x2*x2*x2; return(x*p4/q4); } double erf2(double x) /* erfc; 0.65 <= x < 2.2 */ { double p[6] = {9.99999992049799098e-1, 1.33154163936765307e0, 8.78115804155881782e-1, 3.31899559578213215e-1, 7.14193832506776067e-2, 7.06940843763253131e-3}; double q[7] = {1.00000000000000000e0, 2.45992070144245533e0, 2.65383972869775752e0, 1.61876655543871376e0, 5.94651311286481502e-1, 1.26579413030177940e-1, 1.25304936549413393e-2}; double x2, p5, q6; x2 = x*x; p5 = p[0] + p[1]*x + p[2]*x2 + p[3]*x2*x + p[4]*x2*x2 + p[5]*x2*x2*x; q6 = q[0] + q[1]*x + q[2]*x2 + q[3]*x2*x + q[4]*x2*x2 + q[5]*x2*x2*x + q[6]*x2*x2*x2; return( exp(-x2)*p5/q6 ); } double erf3(double x) /* erfc; 2.2 < x <= 6 */ { double p[6] = {9.99921140009714409e-1, 1.62356584489366647e0, 1.26739901455873222e0, 5.81528574177741135e-1, 1.57289620742838702e-1, 2.25716982919217555e-2}; double q[7] = {1.00000000000000000e0, 2.75143870676376208e0, 3.37367334657284535e0, 2.38574194785344389e0, 1.05074004614827206e0, 2.78788439273628983e-1, 4.00072964526861362e-2}; double x2, p5, q6; x2 = x*x; p5 = p[0] + p[1]*x + p[2]*x2 + p[3]*x2*x + p[4]*x2*x2 + p[5]*x2*x2*x; q6 = q[0] + q[1]*x + q[2]*x2 + q[3]*x2*x + q[4]*x2*x2 + q[5]*x2*x2*x + q[6]*x2*x2*x2; return( exp(-x2)*p5/q6 ); } double erf4(double x) /* erfc; x > 6.0 */ { double p[5] = {5.64189583547756078e-1, 8.80253746105525775e0, 3.84683103716117320e1, 4.77209965874436377e1, 8.08040729052301677e0}; double q[5] = {1.00000000000000000e0, 1.61020914205869003e1, 7.54843505665954743e1, 1.12123870801026015e2, 3.73997570145040850e1}; double x2, p4, q4; if (x>26.5432) return(0.0); x2 = 1.0/(x*x); p4 = p[0] + p[1]*x2 + p[2]*x2*x2 + p[3]*x2*x2*x2 + p[4]*x2*x2*x2*x2; q4 = q[0] + q[1]*x2 + q[2]*x2*x2 + q[3]*x2*x2*x2 + q[4]*x2*x2*x2*x2; return(exp(-x*x)*p4/(x*q4)); } double mut_erfc(double x) { if (x<0.0) return(2.0-mut_erfc(-x)); if (x==0.0) return(1.0); if (x<0.65) return(1.0-erf1(x)); if (x<2.2) return(erf2(x)); if (x<6.0) return(erf3(x)); return(erf4(x)); } double mut_erf(double x) { if (x<0.0) return(-mut_erf(-x)); if (x==0.0) return(0.0); if (x<0.65) return(erf1(x)); if (x<2.2) return(1.0-erf2(x)); if (x<6.0) return(1.0-erf3(x)); return(1.0-erf4(x)); } double mut_pnorm(double x) { if (x<0.0) return(mut_erfc(-x/SQRT2)/2); return((1.0+mut_erf(x/SQRT2))/2); } /* * Copyright 1996-2006 Catherine Loader. */ #include "mut.h" static double lookup_gamma[21] = { 0.0, /* place filler */ 0.572364942924699971, /* log(G(0.5)) = log(sqrt(pi)) */ 0.000000000000000000, /* log(G(1)) = log(0!) */ -0.120782237635245301, /* log(G(3/2)) = log(sqrt(pi)/2)) */ 0.000000000000000000, /* log(G(2)) = log(1!) */ 0.284682870472919181, /* log(G(5/2)) = log(3sqrt(pi)/4) */ 0.693147180559945286, /* log(G(3)) = log(2!) */ 1.200973602347074287, /* etc */ 1.791759469228054957, 2.453736570842442344, 3.178053830347945752, 3.957813967618716511, 4.787491742782045812, 5.662562059857141783, 6.579251212010101213, 7.534364236758732680, 8.525161361065414667, 9.549267257300996903, 10.604602902745250859, 11.689333420797268559, 12.801827480081469091 }; /* * coefs are B(2n)/(2n(2n-1)) 2n(2n-1) = * 2n B(2n) 2n(2n-1) coef * 2 1/6 2 1/12 * 4 -1/30 12 -1/360 * 6 1/42 30 1/1260 * 8 -1/30 56 -1/1680 * 10 5/66 90 1/1188 * 12 -691/2730 132 691/360360 */ double mut_lgamma(double x) { double f, z, x2, se; int ix; /* lookup table for common values. */ ix = (int)(2*x); if (((ix>=1) & (ix<=20)) && (ix==2*x)) return(lookup_gamma[ix]); f = 1.0; while (x <= 15) { f *= x; x += 1.0; } x2 = 1.0/(x*x); z = (x-0.5)*log(x) - x + HF_LG_PIx2; se = (13860 - x2*(462 - x2*(132 - x2*(99 - 140*x2))))/(166320*x); return(z + se - log(f)); } double mut_lgammai(int i) /* log(Gamma(i/2)) for integer i */ { if (i>20) return(mut_lgamma(i/2.0)); return(lookup_gamma[i]); } /* * Copyright 1996-2006 Catherine Loader. */ /* * A is a n*p matrix, find the cholesky decomposition * of the first p rows. In most applications, will want n=p. * * chol_dec(A,n,p) computes the decomoposition R'R=A. * (note that R is stored in the input A). * chol_solve(A,v,n,p) computes (R'R)^{-1}v * chol_hsolve(A,v,n,p) computes (R')^{-1}v * chol_isolve(A,v,n,p) computes (R)^{-1}v * chol_qf(A,v,n,p) computes ||(R')^{-1}v||^2. * chol_mult(A,v,n,p) computes (R'R)v * * The solve functions assume that A is already decomposed. * chol_solve(A,v,n,p) is equivalent to applying chol_hsolve() * and chol_isolve() in sequence. */ #include <math.h> #include "mut.h" void chol_dec(A,n,p) double *A; int n, p; { int i, j, k; for (j=0; j<p; j++) { k = n*j+j; for (i=0; i<j; i++) A[k] -= A[n*j+i]*A[n*j+i]; if (A[k]<=0) { for (i=j; i<p; i++) A[n*i+j] = 0.0; } else { A[k] = sqrt(A[k]); for (i=j+1; i<p; i++) { for (k=0; k<j; k++) A[n*i+j] -= A[n*i+k]*A[n*j+k]; A[n*i+j] /= A[n*j+j]; } } } for (j=0; j<p; j++) for (i=j+1; i<p; i++) A[n*j+i] = 0.0; } int chol_solve(A,v,n,p) double *A, *v; int n, p; { int i, j; for (i=0; i<p; i++) { for (j=0; j<i; j++) v[i] -= A[i*n+j]*v[j]; v[i] /= A[i*n+i]; } for (i=p-1; i>=0; i--) { for (j=i+1; j<p; j++) v[i] -= A[j*n+i]*v[j]; v[i] /= A[i*n+i]; } return(p); } int chol_hsolve(A,v,n,p) double *A, *v; int n, p; { int i, j; for (i=0; i<p; i++) { for (j=0; j<i; j++) v[i] -= A[i*n+j]*v[j]; v[i] /= A[i*n+i]; } return(p); } int chol_isolve(A,v,n,p) double *A, *v; int n, p; { int i, j; for (i=p-1; i>=0; i--) { for (j=i+1; j<p; j++) v[i] -= A[j*n+i]*v[j]; v[i] /= A[i*n+i]; } return(p); } double chol_qf(A,v,n,p) double *A, *v; int n, p; { int i, j; double sum; sum = 0.0; for (i=0; i<p; i++) { for (j=0; j<i; j++) v[i] -= A[i*n+j]*v[j]; v[i] /= A[i*n+i]; sum += v[i]*v[i]; } return(sum); } int chol_mult(A,v,n,p) double *A, *v; int n, p; { int i, j; double sum; for (i=0; i<p; i++) { sum = 0; for (j=i; j<p; j++) sum += A[j*n+i]*v[j]; v[i] = sum; } for (i=p-1; i>=0; i--) { sum = 0; for (j=0; j<=i; j++) sum += A[i*n+j]*v[j]; v[i] = sum; } return(1); } /* * Copyright 1996-2006 Catherine Loader. */ #include <stdio.h> #include <math.h> #include "mut.h" #define E_MAXIT 20 #define E_TOL 1.0e-10 #define SQR(x) ((x)*(x)) double e_tol(D,p) double *D; int p; { double mx; int i; if (E_TOL <= 0.0) return(0.0); mx = D[0]; for (i=1; i<p; i++) if (D[i*(p+1)]>mx) mx = D[i*(p+1)]; return(E_TOL*mx); } void eig_dec(X,P,d) double *X, *P; int d; { int i, j, k, iter, ms; double c, s, r, u, v; for (i=0; i<d; i++) for (j=0; j<d; j++) P[i*d+j] = (i==j); for (iter=0; iter<E_MAXIT; iter++) { ms = 0; for (i=0; i<d; i++) for (j=i+1; j<d; j++) if (SQR(X[i*d+j]) > 1.0e-15*fabs(X[i*d+i]*X[j*d+j])) { c = (X[j*d+j]-X[i*d+i])/2; s = -X[i*d+j]; r = sqrt(c*c+s*s); c /= r; s = sqrt((1-c)/2)*(2*(s>0)-1); c = sqrt((1+c)/2); for (k=0; k<d; k++) { u = X[i*d+k]; v = X[j*d+k]; X[i*d+k] = u*c+v*s; X[j*d+k] = v*c-u*s; } for (k=0; k<d; k++) { u = X[k*d+i]; v = X[k*d+j]; X[k*d+i] = u*c+v*s; X[k*d+j] = v*c-u*s; } X[i*d+j] = X[j*d+i] = 0.0; for (k=0; k<d; k++) { u = P[k*d+i]; v = P[k*d+j]; P[k*d+i] = u*c+v*s; P[k*d+j] = v*c-u*s; } ms = 1; } if (ms==0) return; } mut_printf("eig_dec not converged\n"); } int eig_solve(J,x) jacobian *J; double *x; { int d, i, j, rank; double *D, *P, *Q, *w; double tol; D = J->Z; P = Q = J->Q; d = J->p; w = J->wk; tol = e_tol(D,d); rank = 0; for (i=0; i<d; i++) { w[i] = 0.0; for (j=0; j<d; j++) w[i] += P[j*d+i]*x[j]; } for (i=0; i<d; i++) if (D[i*d+i]>tol) { w[i] /= D[i*(d+1)]; rank++; } for (i=0; i<d; i++) { x[i] = 0.0; for (j=0; j<d; j++) x[i] += Q[i*d+j]*w[j]; } return(rank); } int eig_hsolve(J,v) jacobian *J; double *v; { int i, j, p, rank; double *D, *Q, *w; double tol; D = J->Z; Q = J->Q; p = J->p; w = J->wk; tol = e_tol(D,p); rank = 0; for (i=0; i<p; i++) { w[i] = 0.0; for (j=0; j<p; j++) w[i] += Q[j*p+i]*v[j]; } for (i=0; i<p; i++) { if (D[i*p+i]>tol) { v[i] = w[i]/sqrt(D[i*(p+1)]); rank++; } else v[i] = 0.0; } return(rank); } int eig_isolve(J,v) jacobian *J; double *v; { int i, j, p, rank; double *D, *Q, *w; double tol; D = J->Z; Q = J->Q; p = J->p; w = J->wk; tol = e_tol(D,p); rank = 0; for (i=0; i<p; i++) { if (D[i*p+i]>tol) { v[i] = w[i]/sqrt(D[i*(p+1)]); rank++; } else v[i] = 0.0; } for (i=0; i<p; i++) { w[i] = 0.0; for (j=0; j<p; j++) w[i] += Q[i*p+j]*v[j]; } return(rank); } double eig_qf(J,v) jacobian *J; double *v; { int i, j, p; double sum, tol; p = J->p; sum = 0.0; tol = e_tol(J->Z,p); for (i=0; i<p; i++) if (J->Z[i*p+i]>tol) { J->wk[i] = 0.0; for (j=0; j<p; j++) J->wk[i] += J->Q[j*p+i]*v[j]; sum += J->wk[i]*J->wk[i]/J->Z[i*p+i]; } return(sum); } /* * Copyright 1996-2006 Catherine Loader. */ /* * Integrate a function f over a circle or disc. */ #include "mut.h" void setM(M,r,s,c,b) double *M, r, s, c; int b; { M[0] =-r*s; M[1] = r*c; M[2] = b*c; M[3] = b*s; M[4] =-r*c; M[5] = -s; M[6] = -s; M[7] = 0.0; M[8] =-r*s; M[9] = c; M[10]= c; M[11]= 0.0; } void integ_circ(f,r,orig,res,mint,b) int (*f)(), mint, b; double r, *orig, *res; { double y, x[2], theta, tres[MXRESULT], M[12], c, s; int i, j, nr; y = 0; for (i=0; i<mint; i++) { theta = 2*PI*(double)i/(double)mint; c = cos(theta); s = sin(theta); x[0] = orig[0]+r*c; x[1] = orig[1]+r*s; if (b!=0) { M[0] =-r*s; M[1] = r*c; M[2] = b*c; M[3] = b*s; M[4] =-r*c; M[5] = -s; M[6] = -s; M[7] = 0.0; M[8] =-r*s; M[9] = c; M[10]= c; M[11]= 0.0; } nr = f(x,2,tres,M); if (i==0) setzero(res,nr); for (j=0; j<nr; j++) res[j] += tres[j]; } y = 2 * PI * ((b==0)?r:1.0) / mint; for (j=0; j<nr; j++) res[j] *= y; } void integ_disc(f,fb,fl,res,resb,mg) int (*f)(), (*fb)(), *mg; double *fl, *res, *resb; { double x[2], y, r, tres[MXRESULT], *orig, rmin, rmax, theta, c, s, M[12]; int ct, ctb, i, j, k, nr, nrb, w; orig = &fl[2]; rmax = fl[1]; rmin = fl[0]; y = 0.0; ct = ctb = 0; for (j=0; j<mg[1]; j++) { theta = 2*PI*(double)j/(double)mg[1]; c = cos(theta); s = sin(theta); for (i= (rmin>0) ? 0 : 1; i<=mg[0]; i++) { r = rmin + (rmax-rmin)*i/mg[0]; w = (2+2*(i&1)-(i==0)-(i==mg[0])); x[0] = orig[0] + r*c; x[1] = orig[1] + r*s; nr = f(x,2,tres,NULL); if (ct==0) setzero(res,nr); for (k=0; k<nr; k++) res[k] += w*r*tres[k]; ct++; if (((i==0) | (i==mg[0])) && (fb!=NULL)) { setM(M,r,s,c,1-2*(i==0)); nrb = fb(x,2,tres,M); if (ctb==0) setzero(resb,nrb); ctb++; for (k=0; k<nrb; k++) resb[k] += tres[k]; } } } /* for (i= (rmin>0) ? 0 : 1; i<=mg[0]; i++) { r = rmin + (rmax-rmin)*i/mg[0]; w = (2+2*(i&1)-(i==0)-(i==mg[0])); for (j=0; j<mg[1]; j++) { theta = 2*PI*(double)j/(double)mg[1]; c = cos(theta); s = sin(theta); x[0] = orig[0] + r*c; x[1] = orig[1] + r*s; nr = f(x,2,tres,NULL); if (ct==0) setzero(res,nr); ct++; for (k=0; k<nr; k++) res[k] += w*r*tres[k]; if (((i==0) | (i==mg[0])) && (fb!=NULL)) { setM(M,r,s,c,1-2*(i==0)); nrb = fb(x,2,tres,M); if (ctb==0) setzero(resb,nrb); ctb++; for (k=0; k<nrb; k++) resb[k] += tres[k]; } } } */ for (j=0; j<nr; j++) res[j] *= 2*PI*(rmax-rmin)/(3*mg[0]*mg[1]); for (j=0; j<nrb; j++) resb[j] *= 2*PI/mg[1]; } /* * Copyright 1996-2006 Catherine Loader. */ /* * Multivariate integration of a vector-valued function * using Monte-Carlo method. * * uses drand48() random number generator. Does not seed. */ #include <stdlib.h> #include "mut.h" extern void setzero(); static double M[(1+MXIDIM)*MXIDIM*MXIDIM]; void monte(f,ll,ur,d,res,n) int (*f)(), d, n; double *ll, *ur, *res; { int i, j, nr; #ifdef WINDOWS mut_printf("Sorry, Monte-Carlo Integration not enabled.\n"); for (i=0; i<n; i++) res[i] = 0.0; #else double z, x[MXIDIM], tres[MXRESULT]; srand48(234L); for (i=0; i<n; i++) { for (j=0; j<d; j++) x[j] = ll[j] + (ur[j]-ll[j])*drand48(); nr = f(x,d,tres,NULL); if (i==0) setzero(res,nr); for (j=0; j<nr; j++) res[j] += tres[j]; } z = 1; for (i=0; i<d; i++) z *= (ur[i]-ll[i]); for (i=0; i<nr; i++) res[i] *= z/n; #endif } /* * Copyright 1996-2006 Catherine Loader. */ /* * Multivariate integration of a vector-valued function * using Simpson's rule. */ #include <math.h> #include "mut.h" extern void setzero(); static double M[(1+MXIDIM)*MXIDIM*MXIDIM]; /* third order corners */ void simp3(fd,x,d,resd,delta,wt,i0,i1,mg,ct,res2,index) int (*fd)(), d, wt, i0, i1, *mg, ct, *index; double *x, *resd, *delta, *res2; { int k, l, m, nrd; double zb; for (k=i1+1; k<d; k++) if ((index[k]==0) | (index[k]==mg[k])) { setzero(M,d*d); m = 0; zb = 1.0; for (l=0; l<d; l++) if ((l!=i0) & (l!=i1) & (l!=k)) { M[m*d+l] = 1.0; m++; zb *= delta[l]; } M[(d-3)*d+i0] = (index[i0]==0) ? -1 : 1; M[(d-2)*d+i1] = (index[i1]==0) ? -1 : 1; M[(d-1)*d+k] = (index[k]==0) ? -1 : 1; nrd = fd(x,d,res2,M); if ((ct==0) & (i0==0) & (i1==1) & (k==2)) setzero(resd,nrd); for (l=0; l<nrd; l++) resd[l] += wt*zb*res2[l]; } } /* second order corners */ void simp2(fc,fd,x,d,resc,resd,delta,wt,i0,mg,ct,res2,index) int (*fc)(), (*fd)(), d, wt, i0, *mg, ct, *index; double *x, *resc, *resd, *delta, *res2; { int j, k, l, nrc; double zb; for (j=i0+1; j<d; j++) if ((index[j]==0) | (index[j]==mg[j])) { setzero(M,d*d); l = 0; zb = 1; for (k=0; k<d; k++) if ((k!=i0) & (k!=j)) { M[l*d+k] = 1.0; l++; zb *= delta[k]; } M[(d-2)*d+i0] = (index[i0]==0) ? -1 : 1; M[(d-1)*d+j] = (index[j]==0) ? -1 : 1; nrc = fc(x,d,res2,M); if ((ct==0) & (i0==0) & (j==1)) setzero(resc,nrc); for (k=0; k<nrc; k++) resc[k] += wt*zb*res2[k]; if (fd!=NULL) simp3(fd,x,d,resd,delta,wt,i0,j,mg,ct,res2,index); } } /* first order boundary */ void simp1(fb,fc,fd,x,d,resb,resc,resd,delta,wt,mg,ct,res2,index) int (*fb)(), (*fc)(), (*fd)(), d, wt, *mg, ct, *index; double *x, *resb, *resc, *resd, *delta, *res2; { int i, j, k, nrb; double zb; for (i=0; i<d; i++) if ((index[i]==0) | (index[i]==mg[i])) { setzero(M,(1+d)*d*d); k = 0; for (j=0; j<d; j++) if (j!=i) { M[k*d+j] = 1; k++; } M[(d-1)*d+i] = (index[i]==0) ? -1 : 1; nrb = fb(x,d,res2,M); zb = 1; for (j=0; j<d; j++) if (i!=j) zb *= delta[j]; if ((ct==0) && (i==0)) for (j=0; j<nrb; j++) resb[j] = 0.0; for (j=0; j<nrb; j++) resb[j] += wt*zb*res2[j]; if (fc!=NULL) simp2(fc,fd,x,d,resc,resd,delta,wt,i,mg,ct,res2,index); } } void simpson4(f,fb,fc,fd,ll,ur,d,res,resb,resc,resd,mg,res2) int (*f)(), (*fb)(), (*fc)(), (*fd)(), d, *mg; double *ll, *ur, *res, *resb, *resc, *resd, *res2; { int ct, i, j, nr, wt, index[MXIDIM]; double x[MXIDIM], delta[MXIDIM], z; for (i=0; i<d; i++) { index[i] = 0; x[i] = ll[i]; if (mg[i]&1) mg[i]++; delta[i] = (ur[i]-ll[i])/(3*mg[i]); } ct = 0; while(1) { wt = 1; for (i=0; i<d; i++) wt *= (4-2*(index[i]%2==0)-(index[i]==0)-(index[i]==mg[i])); nr = f(x,d,res2,NULL); if (ct==0) setzero(res,nr); for (i=0; i<nr; i++) res[i] += wt*res2[i]; if (fb!=NULL) simp1(fb,fc,fd,x,d,resb,resc,resd,delta,wt,mg,ct,res2,index); /* compute next grid point */ for (i=0; i<d; i++) { index[i]++; if (index[i]>mg[i]) { index[i] = 0; x[i] = ll[i]; if (i==d-1) /* done */ { z = 1.0; for (j=0; j<d; j++) z *= delta[j]; for (j=0; j<nr; j++) res[j] *= z; return; } } else { x[i] = ll[i] + 3*delta[i]*index[i]; i = d; } } ct++; } } void simpsonm(f,ll,ur,d,res,mg,res2) int (*f)(), d, *mg; double *ll, *ur, *res, *res2; { simpson4(f,NULL,NULL,NULL,ll,ur,d,res,NULL,NULL,NULL,mg,res2); } double simpson(f,l0,l1,m) double (*f)(), l0, l1; int m; { double x, sum; int i; sum = 0; for (i=0; i<=m; i++) { x = ((m-i)*l0 + i*l1)/m; sum += (2+2*(i&1)-(i==0)-(i==m)) * f(x); } return( (l1-l0) * sum / (3*m) ); } /* * Copyright 1996-2006 Catherine Loader. */ #include "mut.h" static double *res, *resb, *orig, rmin, rmax; static int ct0; void sphM(M,r,u) double *M, r, *u; { double h, u1[3], u2[3]; /* set the orthogonal unit vectors. */ h = sqrt(u[0]*u[0]+u[1]*u[1]); if (h<=0) { u1[0] = u2[1] = 1.0; u1[1] = u1[2] = u2[0] = u2[2] = 0.0; } else { u1[0] = u[1]/h; u1[1] = -u[0]/h; u1[2] = 0.0; u2[0] = u[2]*u[0]/h; u2[1] = u[2]*u[1]/h; u2[2] = -h; } /* parameterize the sphere as r(cos(t)cos(v)u + sin(t)u1 + cos(t)sin(v)u2). * first layer of M is (dx/dt, dx/dv, dx/dr) at t=v=0. */ M[0] = r*u1[0]; M[1] = r*u1[1]; M[2] = r*u1[2]; M[3] = r*u2[0]; M[4] = r*u2[1]; M[5] = r*u2[2]; M[6] = u[0]; M[7] = u[1]; M[8] = u[2]; /* next layers are second derivative matrix of components of x(r,t,v). * d^2x/dt^2 = d^2x/dv^2 = -ru; d^2x/dtdv = 0; * d^2x/drdt = u1; d^2x/drdv = u2; d^2x/dr^2 = 0. */ M[9] = M[13] = -r*u[0]; M[11]= M[15] = u1[0]; M[14]= M[16] = u2[0]; M[10]= M[12] = M[17] = 0.0; M[18]= M[22] = -r*u[1]; M[20]= M[24] = u1[1]; M[23]= M[25] = u2[1]; M[19]= M[21] = M[26] = 0.0; M[27]= M[31] = -r*u[1]; M[29]= M[33] = u1[1]; M[32]= M[34] = u2[1]; M[28]= M[30] = M[35] = 0.0; } double ip3(a,b) double *a, *b; { return(a[0]*b[0] + a[1]*b[1] + a[2]*b[2]); } void rn3(a) double *a; { double s; s = sqrt(ip3(a,a)); a[0] /= s; a[1] /= s; a[2] /= s; } double sptarea(a,b,c) double *a, *b, *c; { double ea, eb, ec, yab, yac, ybc, sab, sac, sbc; double ab[3], ac[3], bc[3], x1[3], x2[3]; ab[0] = a[0]-b[0]; ab[1] = a[1]-b[1]; ab[2] = a[2]-b[2]; ac[0] = a[0]-c[0]; ac[1] = a[1]-c[1]; ac[2] = a[2]-c[2]; bc[0] = b[0]-c[0]; bc[1] = b[1]-c[1]; bc[2] = b[2]-c[2]; yab = ip3(ab,a); yac = ip3(ac,a); ybc = ip3(bc,b); x1[0] = ab[0] - yab*a[0]; x2[0] = ac[0] - yac*a[0]; x1[1] = ab[1] - yab*a[1]; x2[1] = ac[1] - yac*a[1]; x1[2] = ab[2] - yab*a[2]; x2[2] = ac[2] - yac*a[2]; sab = ip3(x1,x1); sac = ip3(x2,x2); ea = acos(ip3(x1,x2)/sqrt(sab*sac)); x1[0] = ab[0] + yab*b[0]; x2[0] = bc[0] - ybc*b[0]; x1[1] = ab[1] + yab*b[1]; x2[1] = bc[1] - ybc*b[1]; x1[2] = ab[2] + yab*b[2]; x2[2] = bc[2] - ybc*b[2]; sbc = ip3(x2,x2); eb = acos(ip3(x1,x2)/sqrt(sab*sbc)); x1[0] = ac[0] + yac*c[0]; x2[0] = bc[0] + ybc*c[0]; x1[1] = ac[1] + yac*c[1]; x2[1] = bc[1] + ybc*c[1]; x1[2] = ac[2] + yac*c[2]; x2[2] = bc[2] + ybc*c[2]; ec = acos(ip3(x1,x2)/sqrt(sac*sbc)); /* * Euler's formula is a+b+c-PI, except I've cheated... * a=ea, c=ec, b=PI-eb, which is more stable. */ return(ea+ec-eb); } void li(x,f,fb,mint,ar) double *x, ar; int (*f)(), (*fb)(), mint; { int i, j, nr, nrb, ct1, w; double u[3], r, M[36]; double sres[MXRESULT], tres[MXRESULT]; /* divide mint by 2, and force to even (Simpson's rule...) * to make comparable with rectangular interpretation of mint */ mint <<= 1; if (mint&1) mint++; ct1 = 0; for (i= (rmin==0) ? 1 : 0; i<=mint; i++) { r = rmin + (rmax-rmin)*i/mint; w = 2+2*(i&1)-(i==0)-(i==mint); u[0] = orig[0]+x[0]*r; u[1] = orig[1]+x[1]*r; u[2] = orig[2]+x[2]*r; nr = f(u,3,tres,NULL); if (ct1==0) setzero(sres,nr); for (j=0; j<nr; j++) sres[j] += w*r*r*tres[j]; ct1++; if ((fb!=NULL) && (i==mint)) /* boundary */ { sphM(M,rmax,x); nrb = fb(u,3,tres,M); if (ct0==0) for (j=0; j<nrb; j++) resb[j] = 0.0; for (j=0; j<nrb; j++) resb[j] += tres[j]*ar; } } if (ct0==0) for (j=0; j<nr; j++) res[j] = 0.0; ct0++; for (j=0; j<nr; j++) res[j] += sres[j] * ar * (rmax-rmin)/(3*mint); } void sphint(f,fb,a,b,c,lev,mint,cent) double *a, *b, *c; int (*f)(), (*fb)(), lev, mint, cent; { double x[3], ab[3], ac[3], bc[3], ar; int i; if (lev>1) { ab[0] = a[0]+b[0]; ab[1] = a[1]+b[1]; ab[2] = a[2]+b[2]; rn3(ab); ac[0] = a[0]+c[0]; ac[1] = a[1]+c[1]; ac[2] = a[2]+c[2]; rn3(ac); bc[0] = b[0]+c[0]; bc[1] = b[1]+c[1]; bc[2] = b[2]+c[2]; rn3(bc); lev >>= 1; if (cent==0) { sphint(f,fb,a,ab,ac,lev,mint,1); sphint(f,fb,ab,bc,ac,lev,mint,0); } else { sphint(f,fb,a,ab,ac,lev,mint,1); sphint(f,fb,b,ab,bc,lev,mint,1); sphint(f,fb,c,ac,bc,lev,mint,1); sphint(f,fb,ab,bc,ac,lev,mint,1); } return; } x[0] = a[0]+b[0]+c[0]; x[1] = a[1]+b[1]+c[1]; x[2] = a[2]+b[2]+c[2]; rn3(x); ar = sptarea(a,b,c); for (i=0; i<8; i++) { if (i>0) { x[0] = -x[0]; if (i%2 == 0) x[1] = -x[1]; if (i==4) x[2] = -x[2]; } switch(cent) { case 2: /* the reflection and its 120', 240' rotations */ ab[0] = x[0]; ab[1] = x[2]; ab[2] = x[1]; li(ab,f,fb,mint,ar); ab[0] = x[2]; ab[1] = x[1]; ab[2] = x[0]; li(ab,f,fb,mint,ar); ab[0] = x[1]; ab[1] = x[0]; ab[2] = x[2]; li(ab,f,fb,mint,ar); case 1: /* and the 120' and 240' rotations */ ab[0] = x[1]; ab[1] = x[2]; ab[2] = x[0]; li(ab,f,fb,mint,ar); ac[0] = x[2]; ac[1] = x[0]; ac[2] = x[1]; li(ac,f,fb,mint,ar); case 0: /* and the triangle itself. */ li( x,f,fb,mint,ar); } } } void integ_sphere(f,fb,fl,Res,Resb,mg) double *fl, *Res, *Resb; int (*f)(), (*fb)(), *mg; { double a[3], b[3], c[3]; a[0] = 1; a[1] = a[2] = 0; b[1] = 1; b[0] = b[2] = 0; c[2] = 1; c[0] = c[1] = 0; res = Res; resb=Resb; orig = &fl[2]; rmin = fl[0]; rmax = fl[1]; ct0 = 0; sphint(f,fb,a,b,c,mg[1],mg[0],0); } /* * Copyright 1996-2006 Catherine Loader. */ /* * solving symmetric equations using the jacobian structure. Currently, three * methods can be used: cholesky decomposition, eigenvalues, eigenvalues on * the correlation matrix. * * jacob_dec(J,meth) decompose the matrix, meth=JAC_CHOL, JAC_EIG, JAC_EIGD * jacob_solve(J,v) J^{-1}v * jacob_hsolve(J,v) (R')^{-1/2}v * jacob_isolve(J,v) (R)^{-1/2}v * jacob_qf(J,v) v' J^{-1} v * jacob_mult(J,v) (R'R) v (pres. CHOL only) * where for each decomposition, R'R=J, although the different decomp's will * produce different R's. * * To set up the J matrix: * first, allocate storage: jac_alloc(J,p,wk) * where p=dimension of matrix, wk is a numeric vector of length * jac_reqd(p) (or NULL, to allocate automatically). * now, copy the numeric values to J->Z (numeric vector with length p*p). * (or, just set J->Z to point to the data vector. But remember this * will be overwritten by the decomposition). * finally, set: * J->st=JAC_RAW; * J->p = p; * * now, call jac_dec(J,meth) (optional) and the solve functions as required. * */ #include "math.h" #include "mut.h" #define DEF_METH JAC_EIGD int jac_reqd(int p) { return(2*p*(p+1)); } double *jac_alloc(J,p,wk) jacobian *J; int p; double *wk; { if (wk==NULL) wk = (double *)calloc(2*p*(p+1),sizeof(double)); if ( wk == NULL ) { printf("Problem allocating memory for wk\n");fflush(stdout); } J->Z = wk; wk += p*p; J->Q = wk; wk += p*p; J->wk= wk; wk += p; J->dg= wk; wk += p; return(wk); } void jacob_dec(J, meth) jacobian *J; int meth; { int i, j, p; if (J->st != JAC_RAW) return; J->sm = J->st = meth; switch(meth) { case JAC_EIG: eig_dec(J->Z,J->Q,J->p); return; case JAC_EIGD: p = J->p; for (i=0; i<p; i++) J->dg[i] = (J->Z[i*(p+1)]<=0) ? 0.0 : 1/sqrt(J->Z[i*(p+1)]); for (i=0; i<p; i++) for (j=0; j<p; j++) J->Z[i*p+j] *= J->dg[i]*J->dg[j]; eig_dec(J->Z,J->Q,J->p); J->st = JAC_EIGD; return; case JAC_CHOL: chol_dec(J->Z,J->p,J->p); return; default: mut_printf("jacob_dec: unknown method %d",meth); } } int jacob_solve(J,v) /* (X^T W X)^{-1} v */ jacobian *J; double *v; { int i, rank; if (J->st == JAC_RAW) jacob_dec(J,DEF_METH); switch(J->st) { case JAC_EIG: return(eig_solve(J,v)); case JAC_EIGD: for (i=0; i<J->p; i++) v[i] *= J->dg[i]; rank = eig_solve(J,v); for (i=0; i<J->p; i++) v[i] *= J->dg[i]; return(rank); case JAC_CHOL: return(chol_solve(J->Z,v,J->p,J->p)); } mut_printf("jacob_solve: unknown method %d",J->st); return(0); } int jacob_hsolve(J,v) /* J^{-1/2} v */ jacobian *J; double *v; { int i; if (J->st == JAC_RAW) jacob_dec(J,DEF_METH); switch(J->st) { case JAC_EIG: return(eig_hsolve(J,v)); case JAC_EIGD: /* eigenvalues on corr matrix */ for (i=0; i<J->p; i++) v[i] *= J->dg[i]; return(eig_hsolve(J,v)); case JAC_CHOL: return(chol_hsolve(J->Z,v,J->p,J->p)); } mut_printf("jacob_hsolve: unknown method %d\n",J->st); return(0); } int jacob_isolve(J,v) /* J^{-1/2} v */ jacobian *J; double *v; { int i, r; if (J->st == JAC_RAW) jacob_dec(J,DEF_METH); switch(J->st) { case JAC_EIG: return(eig_isolve(J,v)); case JAC_EIGD: /* eigenvalues on corr matrix */ r = eig_isolve(J,v); for (i=0; i<J->p; i++) v[i] *= J->dg[i]; return(r); case JAC_CHOL: return(chol_isolve(J->Z,v,J->p,J->p)); } mut_printf("jacob_hsolve: unknown method %d\n",J->st); return(0); } double jacob_qf(J,v) /* vT J^{-1} v */ jacobian *J; double *v; { int i; if (J->st == JAC_RAW) jacob_dec(J,DEF_METH); switch (J->st) { case JAC_EIG: return(eig_qf(J,v)); case JAC_EIGD: for (i=0; i<J->p; i++) v[i] *= J->dg[i]; return(eig_qf(J,v)); case JAC_CHOL: return(chol_qf(J->Z,v,J->p,J->p)); default: mut_printf("jacob_qf: invalid method\n"); return(0.0); } } int jacob_mult(J,v) /* J v */ jacobian *J; double *v; { if (J->st == JAC_RAW) jacob_dec(J,DEF_METH); switch (J->st) { case JAC_CHOL: return(chol_mult(J->Z,v,J->p,J->p)); default: mut_printf("jacob_mult: invalid method\n"); return(0); } } /* * Copyright 1996-2006 Catherine Loader. */ /* * Routines for maximization of a one dimensional function f() * over an interval [xlo,xhi]. In all cases. the flag argument * controls the return: * flag='x', the maximizer xmax is returned. * otherwise, maximum f(xmax) is returned. * * max_grid(f,xlo,xhi,n,flag) * grid maximization of f() over [xlo,xhi] with n intervals. * * max_golden(f,xlo,xhi,n,tol,err,flag) * golden section maximization. * If n>2, an initial grid search is performed with n intervals * (this helps deal with local maxima). * convergence criterion is |x-xmax| < tol. * err is an error flag. * if flag='x', return value is xmax. * otherwise, return value is f(xmax). * * max_quad(f,xlo,xhi,n,tol,err,flag) * quadratic maximization. * * max_nr() * newton-raphson, handles multivariate case. * * TODO: additional error checking, non-convergence stop. */ #include <math.h> #include "mut.h" #define max_val(a,b) ((flag=='x') ? a : b) double max_grid(f,xlo,xhi,n,flag) double (*f)(), xlo, xhi; int n; char flag; { int i, mi; double x, y, mx, my; for (i=0; i<=n; i++) { x = xlo + (xhi-xlo)*i/n; y = f(x); if ((i==0) || (y>my)) { mx = x; my = y; mi = i; } } if (mi==0) return(max_val(xlo,my)); if (mi==n) return(max_val(xhi,my)); return(max_val(mx,my)); } double max_golden(f,xlo,xhi,n,tol,err,flag) double (*f)(), xhi, xlo, tol; int n, *err; char flag; { double dlt, x0, x1, x2, x3, y0, y1, y2, y3; *err = 0; if (n>2) { dlt = (xhi-xlo)/n; x0 = max_grid(f,xlo,xhi,n,'x'); if (xlo<x0) xlo = x0-dlt; if (xhi>x0) xhi = x0+dlt; } x0 = xlo; y0 = f(xlo); x3 = xhi; y3 = f(xhi); x1 = gold_rat*x0 + (1-gold_rat)*x3; y1 = f(x1); x2 = gold_rat*x3 + (1-gold_rat)*x0; y2 = f(x2); while (fabs(x3-x0)>tol) { if ((y1>=y0) && (y1>=y2)) { x3 = x2; y3 = y2; x2 = x1; y2 = y1; x1 = gold_rat*x0 + (1-gold_rat)*x3; y1 = f(x1); } else if ((y2>=y3) && (y2>=y1)) { x0 = x1; y0 = y1; x1 = x2; y1 = y2; x2 = gold_rat*x3 + (1-gold_rat)*x0; y2 = f(x2); } else { if (y3>y0) { x0 = x2; y0 = y2; } else { x3 = x1; y3 = y1; } x1 = gold_rat*x0 + (1-gold_rat)*x3; y1 = f(x1); x2 = gold_rat*x3 + (1-gold_rat)*x0; y2 = f(x2); } } if (y0>=y1) return(max_val(x0,y0)); if (y3>=y2) return(max_val(x3,y3)); return((y1>y2) ? max_val(x1,y1) : max_val(x2,y2)); } double max_quad(f,xlo,xhi,n,tol,err,flag) double (*f)(), xhi, xlo, tol; int n, *err; char flag; { double x0, x1, x2, xnew, y0, y1, y2, ynew, a, b; *err = 0; if (n>2) { x0 = max_grid(f,xlo,xhi,n,'x'); if (xlo<x0) xlo = x0-1.0/n; if (xhi>x0) xhi = x0+1.0/n; } x0 = xlo; y0 = f(x0); x2 = xhi; y2 = f(x2); x1 = (x0+x2)/2; y1 = f(x1); while (x2-x0>tol) { /* first, check (y0,y1,y2) is a peak. If not, * next interval is the halve with larger of (y0,y2). */ if ((y0>y1) | (y2>y1)) { if (y0>y2) { x2 = x1; y2 = y1; } else { x0 = x1; y0 = y1; } x1 = (x0+x2)/2; y1 = f(x1); } else /* peak */ { a = (y1-y0)*(x2-x1) + (y1-y2)*(x1-x0); b = ((y1-y0)*(x2-x1)*(x2+x1) + (y1-y2)*(x1-x0)*(x1+x0))/2; /* quadratic maximizer is b/a. But first check if a's too * small, since we may be close to constant. */ if ((a<=0) | (b<x0*a) | (b>x2*a)) { /* split the larger halve */ xnew = ((x2-x1) > (x1-x0)) ? (x1+x2)/2 : (x0+x1)/2; } else { xnew = b/a; if (10*xnew < (9*x0+x1)) xnew = (9*x0+x1)/10; if (10*xnew > (9*x2+x1)) xnew = (9*x2+x1)/10; if (fabs(xnew-x1) < 0.001*(x2-x0)) { if ((x2-x1) > (x1-x0)) xnew = (99*x1+x2)/100; else xnew = (99*x1+x0)/100; } } ynew = f(xnew); if (xnew>x1) { if (ynew >= y1) { x0 = x1; y0 = y1; x1 = xnew; y1 = ynew; } else { x2 = xnew; y2 = ynew; } } else { if (ynew >= y1) { x2 = x1; y2 = y1; x1 = xnew; y1 = ynew; } else { x0 = xnew; y0 = ynew; } } } } return(max_val(x1,y1)); } double max_nr(F, coef, old_coef, f1, delta, J, p, maxit, tol, err) double *coef, *old_coef, *f1, *delta, tol; int (*F)(), p, maxit, *err; jacobian *J; { double old_f, f, lambda; int i, j, fr; double nc, nd, cut; int rank; *err = NR_OK; J->p = p; fr = F(coef, &f, f1, J->Z); J->st = JAC_RAW; for (i=0; i<maxit; i++) { memcpy(old_coef,coef,p*sizeof(double)); old_f = f; rank = jacob_solve(J,f1); memcpy(delta,f1,p*sizeof(double)); if (rank==0) /* NR won't move! */ delta[0] = -f/f1[0]; lambda = 1.0; nc = innerprod(old_coef,old_coef,p); nd = innerprod(delta, delta, p); cut = sqrt(nc/nd); if (cut>1.0) cut = 1.0; cut *= 0.0001; do { for (j=0; j<p; j++) coef[j] = old_coef[j] + lambda*delta[j]; f = old_f - 1.0; fr = F(coef, &f, f1, J->Z); J->st = JAC_RAW; if (fr==NR_BREAK) return(old_f); lambda = (fr==NR_REDUCE) ? lambda/2 : lambda/10.0; } while ((lambda>cut) & (f <= old_f - 1.0e-3)); if (f < old_f - 1.0e-3) { *err = NR_NDIV; return(f); } if (fr==NR_REDUCE) return(f); if (fabs(f-old_f) < tol) return(f); } *err = NR_NCON; return(f); } /* * Copyright 1996-2006 Catherine Loader. */ #include <math.h> #include "mut.h" /* qr decomposition of X (n*p organized by column). * Take w for the ride, if not NULL. */ void qr(X,n,p,w) double *X, *w; int n, p; { int i, j, k, mi; double c, s, mx, nx, t; for (j=0; j<p; j++) { mi = j; mx = fabs(X[(n+1)*j]); nx = mx*mx; /* find the largest remaining element in j'th column, row mi. * flip that row with row j. */ for (i=j+1; i<n; i++) { nx += X[j*n+i]*X[j*n+i]; if (fabs(X[j*n+i])>mx) { mi = i; mx = fabs(X[j*n+i]); } } for (i=j; i<p; i++) { t = X[i*n+j]; X[i*n+j] = X[i*n+mi]; X[i*n+mi] = t; } if (w!=NULL) { t = w[j]; w[j] = w[mi]; w[mi] = t; } /* want the diag. element -ve, so we do the `good' Householder reflect. */ if (X[(n+1)*j]>0) { for (i=j; i<p; i++) X[i*n+j] = -X[i*n+j]; if (w!=NULL) w[j] = -w[j]; } nx = sqrt(nx); c = nx*(nx-X[(n+1)*j]); if (c!=0) { for (i=j+1; i<p; i++) { s = 0; for (k=j; k<n; k++) s += X[i*n+k]*X[j*n+k]; s = (s-nx*X[i*n+j])/c; for (k=j; k<n; k++) X[i*n+k] -= s*X[j*n+k]; X[i*n+j] += s*nx; } if (w != NULL) { s = 0; for (k=j; k<n; k++) s += w[k]*X[n*j+k]; s = (s-nx*w[j])/c; for (k=j; k<n; k++) w[k] -= s*X[n*j+k]; w[j] += s*nx; } X[j*n+j] = nx; } } } void qrinvx(R,x,n,p) double *R, *x; int n, p; { int i, j; for (i=p-1; i>=0; i--) { for (j=i+1; j<p; j++) x[i] -= R[j*n+i]*x[j]; x[i] /= R[i*n+i]; } } void qrtinvx(R,x,n,p) double *R, *x; int n, p; { int i, j; for (i=0; i<p; i++) { for (j=0; j<i; j++) x[i] -= R[i*n+j]*x[j]; x[i] /= R[i*n+i]; } } void qrsolv(R,x,n,p) double *R, *x; int n, p; { qrtinvx(R,x,n,p); qrinvx(R,x,n,p); } /* * Copyright 1996-2006 Catherine Loader. */ /* * solve f(x)=c by various methods, with varying stability etc... * xlo and xhi should be initial bounds for the solution. * convergence criterion is |f(x)-c| < tol. * * double solve_bisect(f,c,xmin,xmax,tol,bd_flag,err) * double solve_secant(f,c,xmin,xmax,tol,bd_flag,err) * Bisection and secant methods for solving of f(x)=c. * xmin and xmax are starting values and bound for solution. * tol = convergence criterion, |f(x)-c| < tol. * bd_flag = if (xmin,xmax) doesn't bound a solution, what action to take? * BDF_NONE returns error. * BDF_EXPRIGHT increases xmax. * BDF_EXPLEFT decreases xmin. * err = error flag. * The (xmin,xmax) bound is not formally necessary for the secant method. * But having such a bound vastly improves stability; the code performs * a bisection step whenever the iterations run outside the bounds. * * double solve_nr(f,f1,c,x0,tol,err) * Newton-Raphson solution of f(x)=c. * f1 = f'(x). * x0 = starting value. * tol = convergence criteria, |f(x)-c| < tol. * err = error flag. * No stability checks at present. * * double solve_fp(f,x0,tol) * fixed-point iteration to solve f(x)=x. * x0 = starting value. * tol = convergence criteria, stops when |f(x)-x| < tol. * Convergence requires |f'(x)|<1 in neighborhood of true solution; * f'(x) \approx 0 gives the fastest convergence. * No stability checks at present. * * TODO: additional error checking, non-convergence stop. */ #include <math.h> #include "mut.h" typedef struct { double xmin, xmax, x0, x1; double ymin, ymax, y0, y1; } solvest; int step_expand(f,c,sv,bd_flag) double (*f)(), c; solvest *sv; int bd_flag; { double x, y; if (sv->ymin*sv->ymax <= 0.0) return(0); if (bd_flag == BDF_NONE) { mut_printf("invalid bracket\n"); return(1); /* error */ } if (bd_flag == BDF_EXPRIGHT) { while (sv->ymin*sv->ymax > 0) { x = sv->xmax + 2*(sv->xmax-sv->xmin); y = f(x) - c; sv->xmin = sv->xmax; sv->xmax = x; sv->ymin = sv->ymax; sv->ymax = y; } return(0); } if (bd_flag == BDF_EXPLEFT) { while (sv->ymin*sv->ymax > 0) { x = sv->xmin - 2*(sv->xmax-sv->xmin); y = f(x) - c; sv->xmax = sv->xmin; sv->xmin = x; sv->ymax = sv->ymin; sv->ymin = y; } return(0); } mut_printf("step_expand: unknown bd_flag %d.\n",bd_flag); return(1); } int step_addin(sv,x,y) solvest *sv; double x, y; { sv->x1 = sv->x0; sv->x0 = x; sv->y1 = sv->y0; sv->y0 = y; if (y*sv->ymin > 0) { sv->xmin = x; sv->ymin = y; return(0); } if (y*sv->ymax > 0) { sv->xmax = x; sv->ymax = y; return(0); } if (y==0) { sv->xmin = sv->xmax = x; sv->ymin = sv->ymax = 0; return(0); } return(1); } int step_bisect(f,c,sv) double (*f)(), c; solvest *sv; { double x, y; x = sv->x0 = (sv->xmin + sv->xmax)/2; y = sv->y0 = f(x)-c; return(step_addin(sv,x,y)); } double solve_bisect(f,c,xmin,xmax,tol,bd_flag,err) double (*f)(), c, xmin, xmax, tol; int bd_flag, *err; { solvest sv; int z; *err = 0; sv.xmin = xmin; sv.ymin = f(xmin)-c; sv.xmax = xmax; sv.ymax = f(xmax)-c; *err = step_expand(f,c,&sv,bd_flag); if (*err>0) return(sv.xmin); while(1) /* infinite loop if f is discontinuous */ { z = step_bisect(f,c,&sv); if (z) { *err = 1; return(sv.x0); } if (fabs(sv.y0)<tol) return(sv.x0); } } int step_secant(f,c,sv) double (*f)(), c; solvest *sv; { double x, y; if (sv->y0==sv->y1) return(step_bisect(f,c,sv)); x = sv->x0 + (sv->x1-sv->x0)*sv->y0/(sv->y0-sv->y1); if ((x<=sv->xmin) | (x>=sv->xmax)) return(step_bisect(f,c,sv)); y = f(x)-c; return(step_addin(sv,x,y)); } double solve_secant(f,c,xmin,xmax,tol,bd_flag,err) double (*f)(), c, xmin, xmax, tol; int bd_flag, *err; { solvest sv; int z; *err = 0; sv.xmin = xmin; sv.ymin = f(xmin)-c; sv.xmax = xmax; sv.ymax = f(xmax)-c; *err = step_expand(f,c,&sv,bd_flag); if (*err>0) return(sv.xmin); sv.x0 = sv.xmin; sv.y0 = sv.ymin; sv.x1 = sv.xmax; sv.y1 = sv.ymax; while(1) /* infinite loop if f is discontinuous */ { z = step_secant(f,c,&sv); if (z) { *err = 1; return(sv.x0); } if (fabs(sv.y0)<tol) return(sv.x0); } } double solve_nr(f,f1,c,x0,tol,err) double (*f)(), (*f1)(), c, x0, tol; int *err; { double y; do { y = f(x0)-c; x0 -= y/f1(x0); } while (fabs(y)>tol); return(x0); } double solve_fp(f,x0,tol,maxit) double (*f)(), x0, tol; int maxit; { double x1; int i; for (i=0; i<maxit; i++) { x1 = f(x0); if (fabs(x1-x0)<tol) return(x1); x0 = x1; } return(x1); /* although it hasn't converged */ } /* * Copyright 1996-2006 Catherine Loader. */ #include "mut.h" void svd(x,p,q,d,mxit) /* svd of square matrix */ double *x, *p, *q; int d, mxit; { int i, j, k, iter, ms, zer; double r, u, v, cp, cm, sp, sm, c1, c2, s1, s2, mx; for (i=0; i<d; i++) for (j=0; j<d; j++) p[i*d+j] = q[i*d+j] = (i==j); for (iter=0; iter<mxit; iter++) { ms = 0; for (i=0; i<d; i++) for (j=i+1; j<d; j++) { s1 = fabs(x[i*d+j]); s2 = fabs(x[j*d+i]); mx = (s1>s2) ? s1 : s2; zer = 1; if (mx*mx>1.0e-15*fabs(x[i*d+i]*x[j*d+j])) { if (fabs(x[i*(d+1)])<fabs(x[j*(d+1)])) { for (k=0; k<d; k++) { u = x[i*d+k]; x[i*d+k] = x[j*d+k]; x[j*d+k] = u; u = p[k*d+i]; p[k*d+i] = p[k*d+j]; p[k*d+j] = u; } for (k=0; k<d; k++) { u = x[k*d+i]; x[k*d+i] = x[k*d+j]; x[k*d+j] = u; u = q[k*d+i]; q[k*d+i] = q[k*d+j]; q[k*d+j] = u; } } cp = x[i*(d+1)]+x[j*(d+1)]; sp = x[j*d+i]-x[i*d+j]; r = sqrt(cp*cp+sp*sp); if (r>0) { cp /= r; sp /= r; } else { cp = 1.0; zer = 0;} cm = x[i*(d+1)]-x[j*(d+1)]; sm = x[i*d+j]+x[j*d+i]; r = sqrt(cm*cm+sm*sm); if (r>0) { cm /= r; sm /= r; } else { cm = 1.0; zer = 0;} c1 = cm+cp; s1 = sm+sp; r = sqrt(c1*c1+s1*s1); if (r>0) { c1 /= r; s1 /= r; } else { c1 = 1.0; zer = 0;} if (fabs(s1)>ms) ms = fabs(s1); c2 = cm+cp; s2 = sp-sm; r = sqrt(c2*c2+s2*s2); if (r>0) { c2 /= r; s2 /= r; } else { c2 = 1.0; zer = 0;} for (k=0; k<d; k++) { u = x[i*d+k]; v = x[j*d+k]; x[i*d+k] = c1*u+s1*v; x[j*d+k] = c1*v-s1*u; u = p[k*d+i]; v = p[k*d+j]; p[k*d+i] = c1*u+s1*v; p[k*d+j] = c1*v-s1*u; } for (k=0; k<d; k++) { u = x[k*d+i]; v = x[k*d+j]; x[k*d+i] = c2*u-s2*v; x[k*d+j] = s2*u+c2*v; u = q[k*d+i]; v = q[k*d+j]; q[k*d+i] = c2*u-s2*v; q[k*d+j] = s2*u+c2*v; } if (zer) x[i*d+j] = x[j*d+i] = 0.0; ms = 1; } } if (ms==0) iter=mxit+10; } if (iter==mxit) mut_printf("Warning: svd not converged.\n"); for (i=0; i<d; i++) if (x[i*d+i]<0) { x[i*d+i] = -x[i*d+i]; for (j=0; j<d; j++) p[j*d+i] = -p[j*d+i]; } } int svdsolve(x,w,P,D,Q,d,tol) /* original X = PDQ^T; comp. QD^{-1}P^T x */ double *x, *w, *P, *D, *Q, tol; int d; { int i, j, rank; double mx; if (tol>0) { mx = D[0]; for (i=1; i<d; i++) if (D[i*(d+1)]>mx) mx = D[i*(d+1)]; tol *= mx; } rank = 0; for (i=0; i<d; i++) { w[i] = 0.0; for (j=0; j<d; j++) w[i] += P[j*d+i]*x[j]; } for (i=0; i<d; i++) if (D[i*d+i]>tol) { w[i] /= D[i*(d+1)]; rank++; } for (i=0; i<d; i++) { x[i] = 0.0; for (j=0; j<d; j++) x[i] += Q[i*d+j]*w[j]; } return(rank); } void hsvdsolve(x,w,P,D,Q,d,tol) /* original X = PDQ^T; comp. D^{-1/2}P^T x */ double *x, *w, *P, *D, *Q, tol; int d; { int i, j; double mx; if (tol>0) { mx = D[0]; for (i=1; i<d; i++) if (D[i*(d+1)]>mx) mx = D[i*(d+1)]; tol *= mx; } for (i=0; i<d; i++) { w[i] = 0.0; for (j=0; j<d; j++) w[i] += P[j*d+i]*x[j]; } for (i=0; i<d; i++) if (D[i*d+i]>tol) w[i] /= sqrt(D[i*(d+1)]); for (i=0; i<d; i++) x[i] = w[i]; } /* * Copyright 1996-2006 Catherine Loader. */ /* * Includes some miscellaneous vector functions: * setzero(v,p) sets all elements of v to 0. * unitvec(x,k,p) sets x to k'th unit vector e_k. * innerprod(v1,v2,p) inner product. * addouter(A,v1,v2,p,c) A <- A + c * v_1 v2^T * multmatscal(A,z,n) A <- A*z * matrixmultiply(A,B,C,m,n,p) C(m*p) <- A(m*n) * B(n*p) * transpose(x,m,n) inline transpose * m_trace(x,n) trace * vecsum(x,n) sum elements. */ #include "mut.h" void setzero(v,p) double *v; int p; { int i; for (i=0; i<p; i++) v[i] = 0.0; } void unitvec(x,k,p) double *x; int k, p; { setzero(x,p); x[k] = 1.0; } double innerprod(v1,v2,p) double *v1, *v2; int p; { int i; double s; s = 0; for (i=0; i<p; i++) s += v1[i]*v2[i]; return(s); } void addouter(A,v1,v2,p,c) double *A, *v1, *v2, c; int p; { int i, j; for (i=0; i<p; i++) for (j=0; j<p; j++) A[i*p+j] += c*v1[i]*v2[j]; } void multmatscal(A,z,n) double *A, z; int n; { int i; for (i=0; i<n; i++) A[i] *= z; } /* matrix multiply A (m*n) times B (n*p). * store in C (m*p). * all matrices stored by column. */ void matrixmultiply(A,B,C,m,n,p) double *A, *B, *C; int m, n, p; { int i, j, k, ij; for (i=0; i<m; i++) for (j=0; j<p; j++) { ij = j*m+i; C[ij] = 0.0; for (k=0; k<n; k++) C[ij] += A[k*m+i] * B[j*n+k]; } } /* * transpose() transposes an m*n matrix in place. * At input, the matrix has n rows, m columns and * x[0..n-1] is the is the first column. * At output, the matrix has m rows, n columns and * x[0..m-1] is the first column. */ void transpose(x,m,n) double *x; int m, n; { int t0, t, ti, tj; double z; for (t0=1; t0<m*n-2; t0++) { ti = t0%m; tj = t0/m; do { t = ti*n+tj; ti= t%m; tj= t/m; } while (t<t0); z = x[t]; x[t] = x[t0]; x[t0] = z; } } /* trace of an n*n square matrix. */ double m_trace(x,n) double *x; int n; { int i; double sum; sum = 0; for (i=0; i<n; i++) sum += x[i*(n+1)]; return(sum); } double vecsum(v,n) double *v; int n; { int i; double sum; sum = 0.0; for (i=0; i<n; i++) sum += v[i]; return(sum); } /* * Copyright 1996-2006 Catherine Loader. */ /* miscellaneous functions that may not be defined in the math libraries. The implementations are crude. mut_daws(x) -- dawson's function mut_exp(x) -- exp(x), but it won't overflow. where required, these must be #define'd in header files. also includes ptail(x) -- exp(x*x/2)*int_{-\infty}^x exp(-u^2/2)du for x < -6. logit(x) -- logistic function. expit(x) -- inverse of logit. factorial(n)-- factorial */ #include "mut.h" double mut_exp(x) double x; { if (x>700.0) return(1.014232054735004e+304); return(exp(x)); } double mut_daws(x) double x; { static double val[] = { 0, 0.24485619356002, 0.46034428261948, 0.62399959848185, 0.72477845900708, 0.76388186132749, 0.75213621001998, 0.70541701910853, 0.63998807456541, 0.56917098836654, 0.50187821196415, 0.44274283060424, 0.39316687916687, 0.35260646480842, 0.31964847250685, 0.29271122077502, 0.27039629581340, 0.25160207761769, 0.23551176224443, 0.22153505358518, 0.20924575719548, 0.19833146819662, 0.18855782729305, 0.17974461154688, 0.17175005072385 }; double h, f0, f1, f2, y, z, xx; int j, m; if (x<0) return(-mut_daws(-x)); if (x>6) { /* Tail series: 1/x + 1/x^3 + 1.3/x^5 + 1.3.5/x^7 + ... */ y = z = 1/x; j = 0; while (((f0=(2*j+1)/(x*x))<1) && (y>1.0e-10*z)) { y *= f0; z += y; j++; } return(z); } m = (int) (4*x); h = x-0.25*m; if (h>0.125) { m++; h = h-0.25; } xx = 0.25*m; f0 = val[m]; f1 = 1-xx*f0; z = f0+h*f1; y = h; j = 2; while (fabs(y)>z*1.0e-10) { f2 = -(j-1)*f0-xx*f1; y *= h/j; z += y*f2; f0 = f1; f1 = f2; j++; } return(z); } double ptail(x) /* exp(x*x/2)*int_{-\infty}^x exp(-u^2/2)du for x < -6 */ double x; { double y, z, f0; int j; y = z = -1.0/x; j = 0; while ((fabs(f0= -(2*j+1)/(x*x))<1) && (fabs(y)>1.0e-10*z)) { y *= f0; z += y; j++; } return(z); } double logit(x) double x; { return(log(x/(1-x))); } double expit(x) double x; { double u; if (x<0) { u = exp(x); return(u/(1+u)); } return(1/(1+exp(-x))); } int factorial(n) int n; { if (n<=1) return(1.0); return(n*factorial(n-1)); } /* * Copyright 1996-2006 Catherine Loader. */ /* * Constrained maximization of a bivariate function. * maxbvgrid(f,x,ll,ur,m0,m1) * maximizes over a grid of m0*m1 points. Returns the maximum, * and the maximizer through the array x. Usually this is a starter, * to choose between local maxima, followed by other routines to refine. * * maxbvstep(f,x,ymax,h,ll,ur,err) * essentially multivariate bisection. A 3x3 grid of points is * built around the starting value (x,ymax). This grid is moved * around (step size h[0] and h[1] in the two dimensions) until * the maximum is in the middle. Then, the step size is halved. * Usually, this will be called in a loop. * The error flag is set if the maximum can't be centered in a * reasonable number of steps. * * maxbv(f,x,h,ll,ur,m0,m1,tol) * combines the two previous functions. It begins with a grid search * (if m0>0 and m1>0), followed by refinement. Refines until both h * components are < tol. */ #include "mut.h" #define max(a,b) ((a)>(b) ? (a) : (b)) #define min(a,b) ((a)<(b) ? (a) : (b)) double maxbvgrid(f,x,ll,ur,m0,m1,con) double (*f)(), *x, *ll, *ur; int m0, m1, *con; { int i, j, im, jm; double y, ymax; im = -1; for (i=0; i<=m0; i++) { x[0] = ((m0-i)*ll[0] + i*ur[0])/m0; for (j=0; j<=m1; j++) { x[1] = ((m1-j)*ll[1] + j*ur[1])/m1; y = f(x); if ((im==-1) || (y>ymax)) { im = i; jm = j; ymax = y; } } } x[0] = ((m0-im)*ll[0] + im*ur[0])/m0; x[1] = ((m1-jm)*ll[1] + jm*ur[1])/m1; con[0] = (im==m0)-(im==0); con[1] = (jm==m1)-(jm==0); return(ymax); } double maxbvstep(f,x,ymax,h,ll,ur,err,con) double (*f)(), *x, ymax, *h, *ll, *ur; int *err, *con; { int i, j, ij, imax, steps, cts[2]; double newx, X[9][2], y[9]; imax =4; y[4] = ymax; for (i=(con[0]==-1)-1; i<2-(con[0]==1); i++) for (j=(con[1]==-1)-1; j<2-(con[1]==1); j++) { ij = 3*i+j+4; X[ij][0] = x[0]+i*h[0]; if (X[ij][0] < ll[0]+0.001*h[0]) X[ij][0] = ll[0]; if (X[ij][0] > ur[0]-0.001*h[0]) X[ij][0] = ur[0]; X[ij][1] = x[1]+j*h[1]; if (X[ij][1] < ll[1]+0.001*h[1]) X[ij][1] = ll[1]; if (X[ij][1] > ur[1]-0.001*h[1]) X[ij][1] = ur[1]; if (ij != 4) { y[ij] = f(X[ij]); if (y[ij]>ymax) { imax = ij; ymax = y[ij]; } } } steps = 0; cts[0] = cts[1] = 0; while ((steps<20) && (imax != 4)) { steps++; if ((con[0]>-1) && ((imax/3)==0)) /* shift right */ { cts[0]--; for (i=8; i>2; i--) { X[i][0] = X[i-3][0]; y[i] = y[i-3]; } imax = imax+3; if (X[imax][0]==ll[0]) con[0] = -1; else { newx = X[imax][0]-h[0]; if (newx < ll[0]+0.001*h[0]) newx = ll[0]; for (i=(con[1]==-1); i<3-(con[1]==1); i++) { X[i][0] = newx; y[i] = f(X[i]); if (y[i]>ymax) { ymax = y[i]; imax = i; } } con[0] = 0; } } if ((con[0]<1) && ((imax/3)==2)) /* shift left */ { cts[0]++; for (i=0; i<6; i++) { X[i][0] = X[i+3][0]; y[i] = y[i+3]; } imax = imax-3; if (X[imax][0]==ur[0]) con[0] = 1; else { newx = X[imax][0]+h[0]; if (newx > ur[0]-0.001*h[0]) newx = ur[0]; for (i=6+(con[1]==-1); i<9-(con[1]==1); i++) { X[i][0] = newx; y[i] = f(X[i]); if (y[i]>ymax) { ymax = y[i]; imax = i; } } con[0] = 0; } } if ((con[1]>-1) && ((imax%3)==0)) /* shift up */ { cts[1]--; for (i=9; i>0; i--) if (i%3 > 0) { X[i][1] = X[i-1][1]; y[i] = y[i-1]; } imax = imax+1; if (X[imax][1]==ll[1]) con[1] = -1; else { newx = X[imax][1]-h[1]; if (newx < ll[1]+0.001*h[1]) newx = ll[1]; for (i=3*(con[0]==-1); i<7-(con[0]==1); i=i+3) { X[i][1] = newx; y[i] = f(X[i]); if (y[i]>ymax) { ymax = y[i]; imax = i; } } con[1] = 0; } } if ((con[1]<1) && ((imax%3)==2)) /* shift down */ { cts[1]++; for (i=0; i<9; i++) if (i%3 < 2) { X[i][1] = X[i+1][1]; y[i] = y[i+1]; } imax = imax-1; if (X[imax][1]==ur[1]) con[1] = 1; else { newx = X[imax][1]+h[1]; if (newx > ur[1]-0.001*h[1]) newx = ur[1]; for (i=2+3*(con[0]==-1); i<9-(con[0]==1); i=i+3) { X[i][1] = newx; y[i] = f(X[i]); if (y[i]>ymax) { ymax = y[i]; imax = i; } } con[1] = 0; } } /* if we've taken 3 steps in one direction, try increasing the * corresponding h. */ if ((cts[0]==-2) | (cts[0]==2)) { h[0] = 2*h[0]; cts[0] = 0; } if ((cts[1]==-2) | (cts[1]==2)) { h[1] = 2*h[1]; cts[1] = 0; } } if (steps==40) *err = 1; else { h[0] /= 2.0; h[1] /= 2.0; *err = 0; } x[0] = X[imax][0]; x[1] = X[imax][1]; return(y[imax]); } #define BQMmaxp 5 int boxquadmin(J,b0,p,x0,ll,ur) jacobian *J; double *b0, *x0, *ll, *ur; int p; { double b[BQMmaxp], x[BQMmaxp], L[BQMmaxp*BQMmaxp], C[BQMmaxp*BQMmaxp], d[BQMmaxp]; double f, fmin; int i, imin, m, con[BQMmaxp], rlx; if (p>BQMmaxp) mut_printf("boxquadmin: maxp is 5.\n"); if (J->st != JAC_RAW) mut_printf("boxquadmin: must start with JAC_RAW.\n"); m = 0; setzero(L,p*p); setzero(x,p); memcpy(C,J->Z,p*p*sizeof(double)); for (i=0; i<p; i++) con[i] = 0; do { /* first, keep minimizing and add constraints, one at a time. */ do { matrixmultiply(C,x,b,p,p,1); for (i=0; i<p; i++) b[i] += b0[i]; conquadmin(J,b,p,L,d,m); /* if C matrix is not pd, don't even bother. * this relies on having used cholesky dec. */ if ((J->Z[0]==0.0) | (J->Z[3]==0.0)) return(1); fmin = 1.0; for (i=0; i<p; i++) if (con[i]==0) { f = 1.0; if (x0[i]+x[i]+b[i] < ll[i]) f = (ll[i]-x[i]-x0[i])/b[i]; if (x0[i]+x[i]+b[i] > ur[i]) f = (ur[i]-x[i]-x0[i])/b[i]; if (f<fmin) fmin = f; imin = i; } for (i=0; i<p; i++) x[i] += fmin*b[i]; if (fmin<1.0) { L[m*p+imin] = 1; m++; con[imin] = (b[imin]<0) ? -1 : 1; } } while ((fmin < 1.0) & (m<p)); /* now, can I relax any constraints? * compute slopes at current point. Can relax if: * slope is -ve on a lower boundary. * slope is +ve on an upper boundary. */ rlx = 0; if (m>0) { matrixmultiply(C,x,b,p,p,1); for (i=0; i<p; i++) b[i] += b0[i]; for (i=0; i<p; i++) { if ((con[i]==-1)&& (b[i]<0)) { con[i] = 0; rlx = 1; } if ((con[i]==1) && (b[i]>0)) { con[i] = 0; rlx = 1; } } if (rlx) /* reconstruct the constraint matrix */ { setzero(L,p*p); m = 0; for (i=0; i<p; i++) if (con[i] != 0) { L[m*p+i] = 1; m++; } } } } while (rlx); memcpy(b0,x,p*sizeof(double)); /* this is how far we should move from x0 */ return(0); } double maxquadstep(f,x,ymax,h,ll,ur,err,con) double (*f)(), *x, ymax, *h, *ll, *ur; int *err, *con; { jacobian J; double b[2], c[2], d, jwork[12]; double x0, x1, y0, y1, ym, h0, xl[2], xu[2], xi[2]; int i, m; *err = 0; /* first, can we relax any of the initial constraints? * if so, just do one step away from the boundary, and * return for restart. */ for (i=0; i<2; i++) if (con[i] != 0) { xi[0] = x[0]; xi[1] = x[1]; xi[i] = x[i]-con[i]*h[i]; y0 = f(xi); if (y0>ymax) { memcpy(x,xi,2*sizeof(double)); con[i] = 0; return(y0); } } /* now, all initial constraints remain active. */ m = 9; for (i=0; i<2; i++) if (con[i]==0) { m /= 3; xl[0] = x[0]; xl[1] = x[1]; xl[i] = max(x[i]-h[i],ll[i]); y0 = f(xl); x0 = xl[i]-x[i]; y0 -= ymax; xu[0] = x[0]; xu[1] = x[1]; xu[i] = min(x[i]+h[i],ur[i]); y1 = f(xu); x1 = xu[i]-x[i]; y1 -= ymax; if (x0*x1*(x1-x0)==0) { *err = 1; return(0.0); } b[i] = (x0*x0*y1-x1*x1*y0)/(x0*x1*(x0-x1)); c[i] = 2*(x0*y1-x1*y0)/(x0*x1*(x1-x0)); if (c[i] >= 0.0) { *err = 1; return(0.0); } xi[i] = (b[i]<0) ? xl[i] : xu[i]; } else { c[i] = -1.0; b[i] = 0.0; } /* enforce initial constraints */ if ((con[0]==0) && (con[1]==0)) { x0 = xi[0]-x[0]; x1 = xi[1]-x[1]; ym = f(xi) - ymax - b[0]*x0 - c[0]*x0*x0/2 - b[1]*x1 - c[1]*x1*x1/2; d = ym/(x0*x1); } else d = 0.0; /* now, maximize the quadratic. * y[4] + b0*x0 + b1*x1 + 0.5(c0*x0*x0 + c1*x1*x1 + 2*d*x0*x1) * -ve everything, to call quadmin. */ jac_alloc(&J,2,jwork); J.Z[0] = -c[0]; J.Z[1] = J.Z[2] = -d; J.Z[3] = -c[1]; J.st = JAC_RAW; J.p = 2; b[0] = -b[0]; b[1] = -b[1]; *err = boxquadmin(&J,b,2,x,ll,ur); if (*err) return(ymax); /* test to see if this step successfully increases... */ for (i=0; i<2; i++) { xi[i] = x[i]+b[i]; if (xi[i]<ll[i]+1e-8*h[i]) xi[i] = ll[i]; if (xi[i]>ur[i]-1e-8*h[i]) xi[i] = ur[i]; } y1 = f(xi); if (y1 < ymax) /* no increase */ { *err = 1; return(ymax); } /* wonderful. update x, h, with the restriction that h can't decrease * by a factor over 10, or increase by over 2. */ for (i=0; i<2; i++) { x[i] = xi[i]; if (x[i]==ll[i]) con[i] = -1; if (x[i]==ur[i]) con[i] = 1; h0 = fabs(b[i]); h0 = min(h0,2*h[i]); h0 = max(h0,h[i]/10); h[i] = min(h0,(ur[i]-ll[i])/2.0); } return(y1); } double maxbv(f,x,h,ll,ur,m0,m1,tol) double (*f)(), *x, *h, *ll, *ur, tol; int m0, m1; { double ymax; int err, con[2]; con[0] = con[1] = 0; if ((m0>0) & (m1>0)) { ymax = maxbvgrid(f,x,ll,ur,m0,m1,con); h[0] = (ur[0]-ll[0])/(2*m0); h[1] = (ur[1]-ll[1])/(2*m1); } else { x[0] = (ll[0]+ur[0])/2; x[1] = (ll[1]+ur[1])/2; h[0] = (ur[0]-ll[0])/2; h[1] = (ur[1]-ll[1])/2; ymax = f(x); } while ((h[0]>tol) | (h[1]>tol)) { ymax = maxbvstep(f,x,ymax,h,ll,ur,&err,con); if (err) mut_printf("maxbvstep failure\n"); } return(ymax); } double maxbvq(f,x,h,ll,ur,m0,m1,tol) double (*f)(), *x, *h, *ll, *ur, tol; int m0, m1; { double ymax; int err, con[2]; con[0] = con[1] = 0; if ((m0>0) & (m1>0)) { ymax = maxbvgrid(f,x,ll,ur,m0,m1,con); h[0] = (ur[0]-ll[0])/(2*m0); h[1] = (ur[1]-ll[1])/(2*m1); } else { x[0] = (ll[0]+ur[0])/2; x[1] = (ll[1]+ur[1])/2; h[0] = (ur[0]-ll[0])/2; h[1] = (ur[1]-ll[1])/2; ymax = f(x); } while ((h[0]>tol) | (h[1]>tol)) { /* first, try a quadratric step */ ymax = maxquadstep(f,x,ymax,h,ll,ur,&err,con); /* if the quadratic step fails, move the grid around */ if (err) { ymax = maxbvstep(f,x,ymax,h,ll,ur,&err,con); if (err) { mut_printf("maxbvstep failure\n"); return(ymax); } } } return(ymax); } /* * Copyright 1996-2006 Catherine Loader. */ #include "mut.h" prf mut_printf = (prf)printf; void mut_redirect(newprf) prf newprf; { mut_printf = newprf; } /* * Copyright 1996-2006 Catherine Loader. */ /* * function to find order of observations in an array. * * mut_order(x,ind,i0,i1) * x array to find order of. * ind integer array of indexes. * i0,i1 (integers) range to order. * * at output, ind[i0...i1] are permuted so that * x[ind[i0]] <= x[ind[i0+1]] <= ... <= x[ind[i1]]. * (with ties, output order of corresponding indices is arbitrary). * The array x is unchanged. * * Typically, if x has length n, then i0=0, i1=n-1 and * ind is (any permutation of) 0...n-1. */ #include "mut.h" double med3(x0,x1,x2) double x0, x1, x2; { if (x0<x1) { if (x2<x0) return(x0); if (x1<x2) return(x1); return(x2); } /* x1 < x0 */ if (x2<x1) return(x1); if (x0<x2) return(x0); return(x2); } void mut_order(x,ind,i0,i1) double *x; int *ind, i0, i1; { double piv; int i, l, r, z; if (i1<=i0) return; piv = med3(x[ind[i0]],x[ind[i1]],x[ind[(i0+i1)/2]]); l = i0; r = i0-1; /* at each stage, * x[i0..l-1] < piv * x[l..r] = piv * x[r+1..i-1] > piv * then, decide where to put x[i]. */ for (i=i0; i<=i1; i++) { if (x[ind[i]]==piv) { r++; z = ind[i]; ind[i] = ind[r]; ind[r] = z; } else if (x[ind[i]]<piv) { r++; z = ind[i]; ind[i] = ind[r]; ind[r] = ind[l]; ind[l] = z; l++; } } if (l>i0) mut_order(x,ind,i0,l-1); if (r<i1) mut_order(x,ind,r+1,i1); } /* * Copyright 1996-2006 Catherine Loader. */ #include "mut.h" #define LOG_2 0.6931471805599453094172321214581765680755 #define IBETA_LARGE 1.0e30 #define IBETA_SMALL 1.0e-30 #define IGAMMA_LARGE 1.0e30 #define DOUBLE_EP 2.2204460492503131E-16 double ibeta(x, a, b) double x, a, b; { int flipped = 0, i, k, count; double I = 0, temp, pn[6], ak, bk, next, prev, factor, val; if (x <= 0) return(0); if (x >= 1) return(1); /* use ibeta(x,a,b) = 1-ibeta(1-x,b,z) */ if ((a+b+1)*x > (a+1)) { flipped = 1; temp = a; a = b; b = temp; x = 1 - x; } pn[0] = 0.0; pn[2] = pn[3] = pn[1] = 1.0; count = 1; val = x/(1.0-x); bk = 1.0; next = 1.0; do { count++; k = count/2; prev = next; if (count%2 == 0) ak = -((a+k-1.0)*(b-k)*val)/((a+2.0*k-2.0)*(a+2.0*k-1.0)); else ak = ((a+b+k-1.0)*k*val)/((a+2.0*k)*(a+2.0*k-1.0)); pn[4] = bk*pn[2] + ak*pn[0]; pn[5] = bk*pn[3] + ak*pn[1]; next = pn[4] / pn[5]; for (i=0; i<=3; i++) pn[i] = pn[i+2]; if (fabs(pn[4]) >= IBETA_LARGE) for (i=0; i<=3; i++) pn[i] /= IBETA_LARGE; if (fabs(pn[4]) <= IBETA_SMALL) for (i=0; i<=3; i++) pn[i] /= IBETA_SMALL; } while (fabs(next-prev) > DOUBLE_EP*prev); /* factor = a*log(x) + (b-1)*log(1-x); factor -= mut_lgamma(a+1) + mut_lgamma(b) - mut_lgamma(a+b); */ factor = dbeta(x,a,b,1) + log(x/a); I = exp(factor) * next; return(flipped ? 1-I : I); } /* * Incomplete gamma function. * int_0^x u^{df-1} e^{-u} du / Gamma(df). */ double igamma(x, df) double x, df; { double factor, term, gintegral, pn[6], rn, ak, bk; int i, count, k; if (x <= 0.0) return(0.0); if (df < 1.0) return( dgamma(x,df+1.0,1.0,0) + igamma(x,df+1.0) ); factor = x * dgamma(x,df,1.0,0); /* factor = exp(df*log(x) - x - lgamma(df)); */ if (x > 1.0 && x >= df) { pn[0] = 0.0; pn[2] = pn[1] = 1.0; pn[3] = x; count = 1; rn = 1.0 / x; do { count++; k = count / 2; gintegral = rn; if (count%2 == 0) { bk = 1.0; ak = (double)k - df; } else { bk = x; ak = (double)k; } pn[4] = bk*pn[2] + ak*pn[0]; pn[5] = bk*pn[3] + ak*pn[1]; rn = pn[4] / pn[5]; for (i=0; i<4; i++) pn[i] = pn[i+2]; if (pn[4] > IGAMMA_LARGE) for (i=0; i<4; i++) pn[i] /= IGAMMA_LARGE; } while (fabs(gintegral-rn) > DOUBLE_EP*rn); gintegral = 1.0 - factor*rn; } else { /* For x<df, use the series * dpois(df,x)*( 1 + x/(df+1) + x^2/((df+1)(df+2)) + ... ) * This could be slow if df large and x/df is close to 1. */ gintegral = term = 1.0; rn = df; do { rn += 1.0; term *= x/rn; gintegral += term; } while (term > DOUBLE_EP*gintegral); gintegral *= factor/df; } return(gintegral); } double pf(q, df1, df2) double q, df1, df2; { return(ibeta(q*df1/(df2+q*df1), df1/2, df2/2)); } /* * Copyright 1996-2006 Catherine Loader. */ #include "mut.h" #include <string.h> /* quadmin: minimize the quadratic, * 2<x,b> + x^T A x. * x = -A^{-1} b. * * conquadmin: min. subject to L'x = d (m constraints) * x = -A^{-1}(b+Ly) (y = Lagrange multiplier) * y = -(L'A^{-1}L)^{-1} (L'A^{-1}b) * x = -A^{-1}b + A^{-1}L (L'A^{-1}L)^{-1} [(L'A^{-1})b + d] * (non-zero d to be added!!) * * qprogmin: min. subject to L'x >= 0. */ void quadmin(J,b,p) jacobian *J; double *b; int p; { int i; jacob_dec(J,JAC_CHOL); i = jacob_solve(J,b); if (i<p) mut_printf("quadmin singular %2d %2d\n",i,p); for (i=0; i<p; i++) b[i] = -b[i]; } /* project vector a (length n) onto * columns of X (n rows, m columns, organized by column). * store result in y. */ #define pmaxm 10 #define pmaxn 100 void project(a,X,y,n,m) double *a, *X, *y; int n, m; { double xta[pmaxm], R[pmaxn*pmaxm]; int i; if (n>pmaxn) mut_printf("project: n too large\n"); if (m>pmaxm) mut_printf("project: m too large\n"); for (i=0; i<m; i++) xta[i] = innerprod(a,&X[i*n],n); memcpy(R,X,m*n*sizeof(double)); qr(R,n,m,NULL); qrsolv(R,xta,n,m); matrixmultiply(X,xta,y,n,m,1); } void resproj(a,X,y,n,m) double *a, *X, *y; int n, m; { int i; project(a,X,y,n,m); for (i=0; i<n; i++) y[i] = a[i]-y[i]; } /* x = -A^{-1}b + A^{-1}L (L'A^{-1}L)^{-1} [(L'A^{-1})b + d] */ void conquadmin(J,b,n,L,d,m) jacobian *J; double *b, *L, *d; int m, n; { double bp[10], L0[100]; int i, j; if (n>10) mut_printf("conquadmin: max. n is 10.\n"); memcpy(L0,L,n*m*sizeof(double)); jacob_dec(J,JAC_CHOL); for (i=0; i<m; i++) jacob_hsolve(J,&L[i*n]); jacob_hsolve(J,b); resproj(b,L,bp,n,m); jacob_isolve(J,bp); for (i=0; i<n; i++) b[i] = -bp[i]; qr(L,n,m,NULL); qrsolv(L,d,n,m); for (i=0; i<n; i++) { bp[i] = 0; for (j=0; j<m; j++) bp[i] += L0[j*n+i]*d[j]; } jacob_solve(J,bp); for (i=0; i<n; i++) b[i] += bp[i]; } void qactivemin(J,b,n,L,d,m,ac) jacobian *J; double *b, *L, *d; int m, n, *ac; { int i, nac; double M[100], dd[10]; nac = 0; for (i=0; i<m; i++) if (ac[i]>0) { memcpy(&M[nac*n],&L[i*n],n*sizeof(double)); dd[nac] = d[i]; nac++; } conquadmin(J,b,n,M,dd,nac); } /* return 1 for full step; 0 if new constraint imposed. */ int movefrom(x0,x,n,L,d,m,ac) double *x0, *x, *L, *d; int n, m, *ac; { int i, imin; double c0, c1, lb, lmin; lmin = 1.0; for (i=0; i<m; i++) if (ac[i]==0) { c1 = innerprod(&L[i*n],x,n)-d[i]; if (c1<0.0) { c0 = innerprod(&L[i*n],x0,n)-d[i]; if (c0<0.0) { if (c1<c0) { lmin = 0.0; imin = 1; } } else { lb = c0/(c0-c1); if (lb<lmin) { lmin = lb; imin = i; } } } } for (i=0; i<n; i++) x0[i] = lmin*x[i]+(1-lmin)*x0[i]; if (lmin==1.0) return(1); ac[imin] = 1; return(0); } int qstep(J,b,x0,n,L,d,m,ac,deac) jacobian *J; double *b, *x0, *L, *d; int m, n, *ac, deac; { double x[10]; int i; if (m>10) mut_printf("qstep: too many constraints.\n"); if (deac) { for (i=0; i<m; i++) if (ac[i]==1) { ac[i] = 0; memcpy(x,b,n*sizeof(double)); qactivemin(J,x,n,L,d,m,ac); if (innerprod(&L[i*n],x,n)>d[i]) /* deactivate this constraint; should rem. */ i = m+10; else ac[i] = 1; } if (i==m) return(0); /* no deactivation possible */ } do { if (!deac) { memcpy(x,b,n*sizeof(double)); qactivemin(J,x,n,L,d,m,ac); } i = movefrom(x0,x,n,L,d,m,ac); deac = 0; } while (i==0); return(1); } /* * x0 is starting value; should satisfy constraints. * L is n*m constraint matrix. * ac is active constraint vector: * ac[i]=0, inactive. * ac[i]=1, active, but can be deactivated. * ac[i]=2, active, cannot be deactivated. */ void qprogmin(J,b,x0,n,L,d,m,ac) jacobian *J; double *b, *x0, *L, *d; int m, n, *ac; { int i; for (i=0; i<m; i++) if (ac[i]==0) { if (innerprod(&L[i*n],x0,n) < d[i]) ac[i] = 1; } jacob_dec(J,JAC_CHOL); qstep(J,b,x0,n,L,d,m,ac,0); while (qstep(J,b,x0,n,L,d,m,ac,1)); } void qpm(A,b,x0,n,L,d,m,ac) double *A, *b, *x0, *L, *d; int *n, *m, *ac; { jacobian J; double wk[1000]; jac_alloc(&J,*n,wk); memcpy(J.Z,A,(*n)*(*n)*sizeof(double)); J.p = *n; J.st = JAC_RAW; qprogmin(&J,b,x0,*n,L,d,*m,ac); }