/** * Copyright (C) 2012 - present by OpenGamma Inc. and the OpenGamma group of companies * * Please see distribution for license. */ package com.opengamma.analytics.math.minimization; import java.util.HashMap; import java.util.Map; import com.opengamma.analytics.math.matrix.DoubleMatrix1D; import com.opengamma.analytics.math.matrix.DoubleMatrix2D; import com.opengamma.util.ArgumentChecker; /** * For a set of N-1 "fit" parameters, produces N "model" parameters that sum to one */ public class SumToOne { private static final double TOL = 1e-9; private static final Map<Integer, int[][]> SETS = new HashMap<>(); private final int[][] _set; private final int _n; /** *For a set of N-1 "fit" parameters, produces N "model" parameters that sum to one * @param n The number of "model" parameters, N */ public SumToOne(final int n) { _set = getSet(n); _n = n; } /** * Transform from the N-1 "fit" parameters to the N "model" parameters * @param fitParms The N-1 "fit" parameters * @return The N "model" parameters */ public double[] transform(final double[] fitParms) { ArgumentChecker.isTrue(fitParms.length == _n - 1, "length of fitParms is {}, but must be {} ", fitParms.length, _n - 1); final double[] s2 = new double[_n - 1]; final double[] c2 = new double[_n - 1]; for (int j = 0; j < _n - 1; j++) { double temp = Math.sin(fitParms[j]); temp *= temp; s2[j] = temp; c2[j] = 1.0 - temp; } final double[] res = new double[_n]; for (int i = 0; i < _n; i++) { double prod = 1.0; for (int j = 0; j < _n - 1; j++) { if (_set[i][j] == 1) { prod *= s2[j]; } else if (_set[i][j] == -1) { prod *= c2[j]; } } res[i] = prod; } return res; } /** * Transform from the N-1 "fit" parameters to the N "model" parameters * @param fitParms The N-1 "fit" parameters * @return The N "model" parameters */ public DoubleMatrix1D transform(final DoubleMatrix1D fitParms) { return new DoubleMatrix1D(transform(fitParms.getData())); } /** * Inverse transform from the N "model" parameters to the N-1 "fit" parameters. Used mainly to find the start position of a optimisation routine * @param modelParms The N "model" parameters. <b>These must sum to one</b> * @return The N-1 "fit" parameters */ public double[] inverseTransform(final double[] modelParms) { ArgumentChecker.isTrue(modelParms.length == _n, "length of modelParms is {}, but must be {} ", modelParms.length, _n); final double[] res = new double[_n - 1]; final double[] cum = new double[_n + 1]; double sum = 0.0; for (int i = 0; i < _n; i++) { sum += modelParms[i]; cum[i + 1] = sum; } ArgumentChecker.isTrue(Math.abs(sum - 1.0) < TOL, "sum of elements is {}. Must be 1.0", sum); cal(cum, 1.0, 0, _n, 0, res); for (int i = 0; i < _n - 1; i++) { res[i] = Math.asin(Math.sqrt(res[i])); } return res; } /** * Inverse transform from the N "model" parameters to the N-1 "fit" parameters. Used mainly to find the start position of a optimisation routine * @param modelParms The N "model" parameters. <b>These must sum to one</b> * @return The N-1 "fit" parameters */ public DoubleMatrix1D inverseTransform(final DoubleMatrix1D modelParms) { return new DoubleMatrix1D(inverseTransform(modelParms.getData())); } /** * The N by N-1 Jacobian matrix between the N "model" parameters (that sum to one) and the N-1 "fit" parameters * @param fitParms The N-1 "fit" parameters * @return The N by N-1 Jacobian matrix */ public double[][] jacobian(final double[] fitParms) { ArgumentChecker.isTrue(fitParms.length == _n - 1, "length of fitParms is {}, but must be {} ", fitParms.length, _n - 1); final double[] sin = new double[_n - 1]; final double[] cos = new double[_n - 1]; for (int j = 0; j < _n - 1; j++) { sin[j] = Math.sin(fitParms[j]); cos[j] = Math.cos(fitParms[j]); } final double[] a = new double[_n]; for (int i = 0; i < _n; i++) { double prod = 1.0; for (int j = 0; j < _n - 1; j++) { if (_set[i][j] == 1) { prod *= sin[j]; } else if (_set[i][j] == -1) { prod *= cos[j]; } } a[i] = 2 * prod * prod; } final double[][] res = new double[_n][_n - 1]; for (int i = 0; i < _n; i++) { for (int j = 0; j < _n - 1; j++) { if (_set[i][j] == 1 && a[i] != 0.0) { res[i][j] = a[i] * cos[j] / sin[j]; } else if (_set[i][j] == -1 && a[i] != 0.0) { res[i][j] = -a[i] * sin[j] / cos[j]; } } } return res; } /** * The N by N-1 Jacobian matrix between the N "model" parameters (that sum to one) and the N-1 "fit" parameters * @param fitParms The N-1 "fit" parameters * @return The N by N-1 Jacobian matrix */ public DoubleMatrix2D jacobian(final DoubleMatrix1D fitParms) { return new DoubleMatrix2D(jacobian(fitParms.getData())); } private void cal(final double[] cum, final double factor, final int d, final int n, final int p1, final double[] res) { if (n == 1) { return; } final int n1 = n / 2; final int n2 = n - n1; final double s = (cum[p1 + n1] - cum[p1]) * factor; final double c = 1 - s; res[d] = s; cal(cum, factor / s, d + 1, n1, p1, res); cal(cum, factor / c, d + n1, n2, p1 + n1, res); } protected static int[][] getSet(final int n) { ArgumentChecker.isTrue(n > 1, "need n>1"); if (SETS.containsKey(n)) { return SETS.get(n); } int[][] res = new int[n][]; switch (n) { case 2: res[0] = new int[] {1 }; res[1] = new int[] {-1 }; break; case 3: res[0] = new int[] {1, 0 }; res[1] = new int[] {-1, 1 }; res[2] = new int[] {-1, -1 }; break; case 4: res[0] = new int[] {1, 1, 0 }; res[1] = new int[] {1, -1, 0 }; res[2] = new int[] {-1, 0, 1 }; res[3] = new int[] {-1, 0, -1 }; break; default: final int n1 = n / 2; final int n2 = n - n1; final int[][] set1 = getSet(n1); final int[][] set2 = (n1 == n2 ? set1 : getSet(n2)); res = new int[n][n - 1]; for (int i = 0; i < n1; i++) { res[i][0] = 1; System.arraycopy(set1[i], 0, res[i], 1, n1 - 1); } for (int i = 0; i < n2; i++) { res[i + n1][0] = -1; System.arraycopy(set2[i], 0, res[i + n1], n1, n2 - 1); } } SETS.put(n, res); return res; } }