void compute(int nlhs, mxArray *plhs[], int nrhs, const mxArray *prhs[], const int atria_preprocessing_given, METRIC dummy)
{
long bins = 32;
double start_dist = 0;
double maximal_search_radius = 0;
double scale_factor = 1;
long opt_flag = 0; // 0 => eucl.norm, 1 => max.norm, utm, 2 => eucl, silent 3 => max., silent
/* handle matrix I/O */
const long N = mxGetM(prhs[0]);
const long dim = mxGetN(prhs[0]);
const double* p = (double *)mxGetPr(prhs[0]);
const long R = mxGetM(prhs[1]);
const double* ref = (double *)mxGetPr(prhs[1]);
const double relative_range = (double) *((double *)mxGetPr(prhs[2]));
if (nrhs > 3) bins = (long) *((double *)mxGetPr(prhs[3]));
if (nrhs > 4) opt_flag = (long) *((double *)mxGetPr(prhs[4]));
if (N < 1) {
mexErrMsgTxt("Data set must consist of at least two points (row vectors)");
return;
}
if (dim < 1) {
mexErrMsgTxt("Data points must be at least of dimension one");
return;
}
if (relative_range <= 0) {
mexErrMsgTxt("Relative range must be greater zero");
return;
}
if ((mxGetN(prhs[1]) == 0) || (mxGetM(prhs[1]) == 0)) {
mexErrMsgTxt("Wrong reference indices given");
return;
}
if (bins < 2) {
mexErrMsgTxt("Number of bins should be two");
return;
}
if (R < 1) {
mexErrMsgTxt("At least one reference point must be given");
return;
}
if (mxGetN(prhs[1]) != dim) {
mexErrMsgTxt("Dimension of reference points and data set points is not the same");
return;
}
if ((opt_flag < 0) || (opt_flag > 3)) {
mexErrMsgTxt("Flag must be out of 0...3");
return;
}
point_set<METRIC> points(N,dim, p);
point_set<METRIC> ref_points(R, dim, ref);
ATRIA< point_set<METRIC> >* searcher = 0;
#ifdef MATLAB_MEX_FILE
if (atria_preprocessing_given) {
searcher = new ATRIA< point_set<METRIC> >(points, prhs[-1]); // this constructor used the data from the preprocessing
if (searcher->geterr()) {
delete searcher;
searcher = 0;
}
}
#endif
if (searcher == 0) {
searcher = new ATRIA< point_set<METRIC> >(points);
}
if (searcher->geterr()) {
mexErrMsgTxt("Error preparing searcher");
return;
}
if (mxGetM(prhs[2])*mxGetN(prhs[2]) == 2) {
start_dist = (double) ((double *)mxGetPr(prhs[2]))[0];
maximal_search_radius = (double) ((double *)mxGetPr(prhs[2]))[1];
if (start_dist <= 0) {
mexErrMsgTxt("Starting radius is zero or negativ");
return;
}
if (maximal_search_radius <= start_dist) {
mexErrMsgTxt("Maximal search radius must be greater than starting radius");
return;
}
}
else {
maximal_search_radius = relative_range * searcher->data_set_radius(); // compute the maximal search radius using information about attractor size
// try to determine an estimate for the minimum inter-point distance in the data set, but greater zero
double* const coord = new double[dim];
for (long n=0; n < 128; n++) {
const long actual = (long) (((double)rand() * (double) (N-2))/(double)RAND_MAX);
vector<neighbor> v;
for (long k=0; k < dim; k++) coord[k] = points.coordinate(actual,k);
searcher->search_k_neighbors(v, 1, coord, actual, actual); // search the nearest neighbor
if (v.size() > 0) {
if (start_dist == 0) start_dist = v[0].dist();
if (v[0].dist() > 0) {
if (v[0].dist() < start_dist) start_dist = v[0].dist();
}
}
}
if (start_dist <= 0) { // first try to search again for a minimum inter-point distance greater zero
for (long n=0; n < 512; n++) {
const long actual = (long) (((double)rand() * (double) (N-2))/(double)RAND_MAX);
vector<neighbor> v;
for (long k=0; k < dim; k++) coord[k] = points.coordinate(actual,k);
searcher->search_k_neighbors(v, 1, coord, actual, actual); // search the nearest neighbor
if (v.size() > 0) {
if (start_dist == 0) start_dist = v[0].dist();
if (v[0].dist() > 0) {
if (v[0].dist() < start_dist) start_dist = v[0].dist();
}
}
}
}
delete[] coord;
if (start_dist <= 0) { // give up if we cannot find an interpoint distance greater zero
mexErrMsgTxt("Cannot find an interpoint distance greater zero, maybe ill-conditioned data set given");
return;
}
if (maximal_search_radius <= start_dist) {
mexErrMsgTxt("Maximal search radius must be greater than starting radius");
return;
}
}
scale_factor = pow(maximal_search_radius/start_dist, 1.0/(bins-1));
if (be_verbose) {
mexPrintf("Number of reference points : %d\n", R);
mexPrintf("Upper bound for attractor size : %f\n", 2 * searcher->data_set_radius());
mexPrintf("Starting distance : %f\n", start_dist);
mexPrintf("Maximal search radius : %f\n", maximal_search_radius);
mexPrintf("Number of partitions used : %d\n\n", bins);
}
plhs[0] = mxCreateDoubleMatrix(bins, 1, mxREAL);
double* corrsums = (double *) mxGetPr(plhs[0]);
double* dists;
if (nlhs > 1) {
plhs[1] = mxCreateDoubleMatrix(bins, 1, mxREAL);
dists = (double *) mxGetPr(plhs[1]);
} else
dists = new double[bins];
double x = start_dist;
// initialize vectors
// corrsum[i] counts the number of points/distances (real) smaller than dists[i]
for (long bin=0; bin < bins; bin++) {
corrsums[bin] = 0;
dists[bin] = x;
x *= scale_factor;
}
long total_pairs = 0; // count the total number of points pairs that were at least theoretically tested
for (long n=0; n < R; n++) { /* iterate over all reference points */
vector<neighbor> v;
const long pairs = N;
if (pairs <= 0)
continue;
total_pairs += pairs;
searcher->search_range(v, maximal_search_radius, ref_points.point_begin(n), -1, -1);
if (v.size() > 0) {
for (vector<neighbor>::iterator i = v.begin(); i < v.end(); i++) { // v is unsorted (!!!)
const double dist = (*i).dist();
int mi, ni; // index of first and last (inclusive) element of array dists
if (dist >= dists[ni = bins-1]) // => we know that dist <= dists[ni]
continue;
if (dist < dists[mi = 0]) { // => we know that dists[mi] < dist
corrsums[0]++;
continue;
}
do
{
const int ki = (mi + ni) / 2; // "mi < ki < ni" if "mi + 1 < ni"
if (dist < dists[ki])
ni = ki; // "dists[mi] < dist <= dists[ni]" is still valid
else
mi = ki; // "dists[mi] < dist <= dists[ni]" is still valid
} while (ni > mi+1);
// now, "ni == mi+1" and "dists[mi] < dist <= dists[ni]" is true,
// so "dists[mi] < dist <= dists[mi+1]", so mi+1 is the desired bin where
// to count this distance
corrsums[ni]++;
}
}
}
if (total_pairs > 0) {
unsigned long sum = 0;
for (long bin=0; bin < bins; bin++) {
sum += (long) corrsums[bin];
corrsums[bin] = ((double) sum) / (double) total_pairs;
}
}
if (nlhs > 2) {
plhs[2] = mxCreateDoubleMatrix(1, 1, mxREAL);
*((double *) mxGetPr(plhs[2])) = (double) total_pairs;
}
if (!(nlhs > 1)) delete[] dists;
delete searcher;
}
void compute(int nlhs, mxArray *plhs[], int nrhs, const mxArray *prhs[], const int atria_preprocessing_given, METRIC dummy)
{
long i,j,n,k; /* loop variables */
/* handle matrix I/O */
const long N = mxGetM(prhs[0]);
const long dim = mxGetN(prhs[0]);
const double* p = (double *)mxGetPr(prhs[0]);
const long R = max(mxGetM(prhs[1]), mxGetN(prhs[1]));
const double* ref = (double *)mxGetPr(prhs[1]);
const double relative_range = (double) *((double *)mxGetPr(prhs[2]));
const long past = (long) *((double *)mxGetPr(prhs[3]));
if (N < 1) {
mexErrMsgTxt("Data set must consist of at least two points (row vectors)");
return;
}
if (dim < 2) {
mexErrMsgTxt("Data points must be at least of dimension two");
return;
}
if ((mxGetN(prhs[1]) == 0) || (mxGetM(prhs[1]) == 0)) {
mexErrMsgTxt("Wrong reference indices given");
return;
}
if (R < 1) {
mexErrMsgTxt("At least one reference index or point must be given");
return;
}
for (i=0; i < R; i++) {
if ((ref[i] < 1) || (ref[i]>N)) {
mexErrMsgTxt("Reference indices out of range");
return;
}
}
point_set<METRIC> points(N,dim, p);
ATRIA< point_set<METRIC> >* searcher = 0;
#ifdef MATLAB_MEX_FILE
if (atria_preprocessing_given) {
searcher = new ATRIA< point_set<METRIC> >(points, prhs[-1]); // this constructor used the data from the preprocessing
if (searcher->geterr()) {
delete searcher;
searcher = 0;
}
}
#endif
if (searcher == 0) {
searcher = new ATRIA< point_set<METRIC> >(points);
}
if (searcher->geterr()) {
mexErrMsgTxt("Error preparing searcher");
return;
}
double range;
if (relative_range > 0)
range = relative_range * searcher->data_set_radius(); // compute the maximal search radius using information about attractor size
else
range = fabs(relative_range); // if relative_range is negativ, use it's value (without sign) as maximal search radius
mexPrintf("Number of reference points : %d\n", R);
mexPrintf("Upper bound for attractor size : %f\n", 2 * searcher->data_set_radius());
mexPrintf("Maximal search radius : %f\n", range);
plhs[0] = mxCreateDoubleMatrix(1, 1, mxREAL);
double* out = (double *) mxGetPr(plhs[0]);
unsigned long counter = 0;
double sum = 0;
for (n=0; n < R; n++) { /* iterate over all reference points */
const long actual = (long) ref[n]-1; /* Matlab to C means indices change from 1 to 0, 2 to 1, 3 to 2 ...*/
if (actual > past) {
vector<neighbor> v;
searcher->search_range(v, range, points.point_begin(actual), actual-past, N); // don't search points from [actual-past .. N-1]
//overall_points += (actual-past); // count the total number of points pairs that were at least theoretically tested
if (v.size() > 0) {
vector<neighbor>::iterator i;
for (i = v.begin(); i < v.end(); i++) { // v is unsorted
const double dist = (*i).dist();
if (dist > 0) {
sum += log(dist/range);
counter++;
}
}
}
}
}
if (counter > 0) *out = -((double)counter)/sum;
else *out = 0;
delete searcher;
}
void compute(int nlhs, mxArray *plhs[], int nrhs, const mxArray *prhs[], const int atria_preprocessing_given, METRIC dummy)
{
long bins = 32;
double start_dist = 0;
double maximal_search_radius = 0;
double scale_factor = 1;
long opt_flag = 0; // 0 => eucl.norm, upper triangle matrix, 1 => max.norm, utm, 2 => eucl,full matrix, 3 => max.,full matrix
long Nref_min;
long Nref_max;
/* handle matrix I/O */
#ifdef C_STYLE_POINT_SET
const long N = mxGetN(prhs[0]);
const long dim = mxGetM(prhs[0]);
#else // this is the default
const long N = mxGetM(prhs[0]);
const long dim = mxGetN(prhs[0]);
#endif
const double* p = (double *)mxGetPr(prhs[0]);
const long Npairs = (long) *((double *)mxGetPr(prhs[1])); // number of pairs
if (mxGetM(prhs[1])*mxGetN(prhs[1]) == 3) {
Nref_min = (long) ((double *)mxGetPr(prhs[1]))[1];
Nref_max = (long) ((double *)mxGetPr(prhs[1]))[2];
} else {
Nref_min = 1;
Nref_max = N;
}
const double relative_range = (double) *((double *)mxGetPr(prhs[2]));
const long past = (long) *((double *)mxGetPr(prhs[3]));
if (nrhs > 4) bins = (long) *((double *)mxGetPr(prhs[4]));
if (nrhs > 5) opt_flag = (long) *((double *)mxGetPr(prhs[5]));
if (N < 1) {
mexErrMsgTxt("Data set must consist of at least two points (row vectors)");
return;
}
if (dim < 1) {
mexErrMsgTxt("Data points must be at least of dimension one");
return;
}
if (relative_range <= 0) {
mexErrMsgTxt("Relative range must be greater zero");
return;
}
if (bins < 2) {
mexErrMsgTxt("Number of bins should be two");
return;
}
if (Npairs < 1) {
mexErrMsgTxt("Number of pairs must be positive");
return;
}
if ((opt_flag < 0) || (opt_flag > 7)) {
mexErrMsgTxt("Flag must be out of 0..7");
return;
}
if (Nref_min < 1) Nref_min = 1;
if (Nref_max > N) Nref_max = N;
point_set<METRIC> points(N,dim, p);
ATRIA< point_set<METRIC> >* searcher = 0;
#ifdef MATLAB_MEX_FILE
if (atria_preprocessing_given) {
searcher = new ATRIA< point_set<METRIC> >(points, prhs[-1]); // this constructor used the data from the preprocessing
if (searcher->geterr()) {
delete searcher;
searcher = 0;
}
}
#endif
if (searcher == 0) {
searcher = new ATRIA< point_set<METRIC> >(points);
}
if (searcher->geterr()) {
mexErrMsgTxt("Error preparing searcher");
return;
}
if (mxGetM(prhs[2])*mxGetN(prhs[2]) == 2) {
start_dist = (double) ((double *)mxGetPr(prhs[2]))[0];
maximal_search_radius = (double) ((double *)mxGetPr(prhs[2]))[1];
if (start_dist <= 0) {
mexErrMsgTxt("Starting radius is zero or negativ");
return;
}
if (maximal_search_radius <= start_dist) {
mexErrMsgTxt("Maximal search radius must be greater than starting radius");
return;
}
}
else {
maximal_search_radius = relative_range * searcher->data_set_radius(); // compute the maximal search radius using information about attractor size
// try to determine an estimate for the minimum inter-point distance in the data set, but greater zero
for (long n=0; n < 128; n++) {
const long actual = (long) (((double)rand() * (double) (N-2))/(double)RAND_MAX);
vector<neighbor> v;
searcher->search_k_neighbors(v, 1, points.point_begin(actual), actual, actual); // search the nearest neighbor
if (v.size() > 0) {
if (start_dist == 0) start_dist = v[0].dist();
if (v[0].dist() > 0) {
if (v[0].dist() < start_dist) start_dist = v[0].dist();
}
}
}
if (start_dist <= 0) { // first try to search again for a minimum inter-point distance greater zero
for (long n=0; n < 512; n++) {
const long actual = (long) (((double)rand() * (double) (N-2))/(double)RAND_MAX);
vector<neighbor> v;
searcher->search_k_neighbors(v, 1, points.point_begin(actual), actual, actual); // search the nearest neighbor
if (v.size() > 0) {
if (start_dist == 0) start_dist = v[0].dist();
if (v[0].dist() > 0) {
if (v[0].dist() < start_dist) start_dist = v[0].dist();
}
}
}
}
if (start_dist <= 0) { // give up if we cannot find an interpoint distance greater zero
mexErrMsgTxt("Cannot find an interpoint distance greater zero, maybe ill-conditioned data set given");
return;
}
if (maximal_search_radius <= start_dist) {
mexErrMsgTxt("Maximal search radius must be greater than starting radius");
return;
}
}
scale_factor = pow(maximal_search_radius/start_dist, 1.0/(bins-1));
if (be_verbose) {
//mexPrintf("Number of reference points : %d\n", R);
mexPrintf("Number of data set points : %d\n", N);
mexPrintf("Number of pairs to find : %d\n", Npairs);
mexPrintf("Miniumum number of reference points : %d\n", Nref_min);
mexPrintf("Maximum number of reference points : %d\n", Nref_max);
mexPrintf("Upper bound for attractor size : %f\n", 2 * searcher->data_set_radius());
mexPrintf("Number of partitions used : %d\n", bins);
mexPrintf("Time window to exclude from search : %d\n", past);
mexPrintf("Minimal length scale : %f\n", start_dist);
mexPrintf("Starting at maximal length scale : %f\n", maximal_search_radius);
}
plhs[0] = mxCreateDoubleMatrix(bins, 1, mxREAL);
double* corrsums = (double *) mxGetPr(plhs[0]);
long* const ref = new long[N]; // needed to create random reference indices
long* const total_pairs = new long[bins]; // number of total pairs is not equal for all bins
long* const pairs_found = new long[bins]; // number of pairs found within distance dist
double* dists;
if (nlhs > 1) {
plhs[1] = mxCreateDoubleMatrix(bins, 1, mxREAL);
dists = (double *) mxGetPr(plhs[1]);
} else
dists = new double[bins];
double x = start_dist;
// initialize vectors
// pairs_found[i] counts the number of points/distances (real) smaller than dists[i]
for (long bin=0; bin < bins; bin++) {
pairs_found[bin] = 0;
total_pairs[bin] = 0;
dists[bin] = x;
x *= scale_factor;
}
for (long r=0; r < N; r++) ref[r] = r;
long R = 0; // number of reference points actually used
long bin = bins-1; // the current highest bin (and length scale) that needs to be filled over Npairs
while(((R < Nref_min) || (pairs_found[bin] < Npairs)) && (R < Nref_max)) {
vector<neighbor> v;
long first, last; // all points with indices i so that first <= i <= last are excluded(!) from search
long pairs = 0;
const long actual = random_permutation(ref, N, R++); // choose random index from 0...N-1, without reoccurences
if (opt_flag & (long)2) {
first = actual-past;
last = actual+past;
if (past >= 0)
pairs = lmax(0,first) + lmax(N-1-last,0);
else
pairs = N;
} else {
if (past >= 0)
first = actual-past;
else
first = actual;
last = N;
pairs = lmin(first,last); // don't search points from [actual-past .. N-1]
}
if (pairs <= 0) {
continue;
}
searcher->search_range(v, dists[bin], points.point_begin(actual), first, last);
if (v.size() > 0) {
for (vector<neighbor>::iterator i = v.begin(); i < v.end(); i++) { // v is unsorted (!!!)
const double d = (*i).dist();
for (long n = bin; n >= 0; n--) {
if (d > dists[n])
break;
pairs_found[n]++;
}
}
}
for (long n = 0; n <= bin; n++) {
total_pairs[n] += pairs;
}
// see if we can reduce length scale
int bins_changed = 0;
while((pairs_found[bin] >= Npairs) && (bin > 0) && (R >= Nref_min)) {
bin--;
bins_changed = 1;
}
if (be_verbose) {
if (bins_changed) {
mexPrintf("Reference points used so far : %d\n", R);
mexPrintf("Switching to length scale : %f\n", dists[bin]);
}
}
}
if (be_verbose) {
mexPrintf("Number of reference points used : %d\n", R);
}
for (long bin=0; bin < bins; bin++) {
corrsums[bin] = ((double) pairs_found[bin]) / ((double) total_pairs[bin]);
}
if (nlhs > 2) {
plhs[2] = mxCreateDoubleMatrix(bins, 1, mxREAL);
double* const y = mxGetPr(plhs[2]);
for (long b=0; b < bins; b++)
y[b] = (double) pairs_found[b];
}
if (nlhs > 3) {
plhs[3] = mxCreateDoubleMatrix(bins, 1, mxREAL);
double* const y = mxGetPr(plhs[3]);
for (long b=0; b < bins; b++)
y[b] = (double) total_pairs[b];
}
if (nlhs > 4) {
plhs[4] = mxCreateDoubleMatrix(R, 1, mxREAL);
double* const y = mxGetPr(plhs[4]);
for (long r=0; r < R; r++)
y[r] = (double) ref[r]+1; // C to Matlab means
}
delete[] total_pairs;
delete[] pairs_found;
delete[] ref;
if (!(nlhs > 1)) delete[] dists;
delete searcher;
}
void mexFunction(int nlhs, mxArray *plhs[], int nrhs, const mxArray *prhs[])
{
int atria_preprocessing_given = 0; // flag wheter preprocessing is already given on command line
ATRIA< point_set<euclidian_distance> >* searcher = 0;
// try to see if the first parameter given is an atria structure
// If this is true, the order of input parameters is shifted by one
// try to see if the first parameter given is an atria structure
// If this is true, the order of input parameters is shifted by one
if ((nrhs) && (mxIsStruct(prhs[0]))) {
atria_preprocessing_given = 1;
prhs++; // these two lines enable us to use the old argument parsing block without changing it
nrhs--;
}
/* check input args */
if (nrhs < 5)
{
mexErrMsgTxt("Largest lyapunov exponent : Data set of points (row vectors), reference indices, length of prediction, maximal number of neighbours and past must be given");
return;
}
/* handle matrix I/O */
const long N = mxGetM(prhs[0]);
const long dim = mxGetN(prhs[0]);
const double* p = (double *)mxGetPr(prhs[0]);
const long R = mxGetM(prhs[1]) * mxGetN(prhs[1]); // number of reference points
const double* ref = (double *)mxGetPr(prhs[1]);
const long length = (long) *((double *)mxGetPr(prhs[2])); // length of prediction
const long NNR = (long) *((double *)mxGetPr(prhs[3])); // number of nearest neighbors
const long past = (long) *((double *)mxGetPr(prhs[4])); // number of points to be excluded from search
if (fabs((double) N) - fabs((double)length) < 1) {
mexErrMsgTxt("Data set must consist of at least two points (row vectors)");
return;
}
if (dim < 1) {
mexErrMsgTxt("Data points must be at least of dimension one");
return;
}
if (length <= 0) {
mexErrMsgTxt("Prediction length must be greater zero");
return;
}
if ((mxGetN(prhs[1]) == 0) || (mxGetM(prhs[1]) == 0)) {
mexErrMsgTxt("Wrong reference indices given");
return;
}
if ((NNR < 1) || (NNR > N)) {
mexErrMsgTxt("Number of nearest neighbours must be within 1 and number of points in the data set");
return;
}
mexPrintf("Number of reference points : %d\n", R);
if (R < 1) {
mexErrMsgTxt("At least one reference index or point must be given");
return;
}
for (long i=0; i < R; i++) { // check reference indices
if ((ref[i] < 1) || (ref[i]>N-length)) {
mexErrMsgTxt("Reference indices out of range");
return;
}
}
#include "create_searcher.cpp"
plhs[0] = mxCreateDoubleMatrix(length+1, 1, mxREAL);
double* x = (double *) mxGetPr(plhs[0]);
const double logarithm_of_2 = log(2.0);
int issue_warning = 1;
long count = 0; // count the number of pairs with distance > 0
for (long step = 0; step <= length; step++) x[step] = 0;
for (long n=0; n < R; n++) { /* main loop, iterate over all reference points */
vector<neighbor>::iterator i;
vector<neighbor> v;
const long actual = (long) ref[n]-1; /* Matlab to C means indices change from 1 to 0, 2 to 1, 3 to 2 ...*/
// try to find at least NNR valid pairs (reference point - neighbor)
// a valid pair must fullfil several criteria :
// 1) both must have a known future of at least lenght steps, so that we
// don't run out of the data set size when computing the divergence
// 2) initial distance must be greater zero
// 3) inital indices must differ by at least "past" steps
for (long nnr=NNR; nnr < 4 * NNR; nnr++) {
vector<neighbor> v_test;
v.erase(v.begin(), v.end());
searcher->search_k_neighbors(v_test, nnr, points.point_begin(actual), actual-past, actual+past);
for (i = v_test.begin(); i < v_test.end(); i++) {
if (((*i).dist() > 0) && ((*i).index() < N-length)) {
v.push_back(*i);
}
}
if (v.size() >= NNR)
break;
}
for (i = v.begin(); i < v.end(); i++) { // v is the sorted vector of neighbors
const long nnindex = (*i).index();
const double nndist = (*i).dist();
const double log_nndist = log(nndist);
count++;
for (long step = 0; step <= length; step++) {
const double dist = points.distance(actual+step, nnindex+step);
if (dist > 0) {
x[step] += log(dist) - log_nndist;
}
else {
if (issue_warning) {
issue_warning = 0; // don't issue this warning again
mexPrintf("Largelyap warning : cannot take logarithm of distance zero \n");
}
}
}
}
}
if (count > 0) {
for (long step = 0; step <= length; step++) x[step] /= (count*logarithm_of_2);
}
delete searcher;
}
