View Javadoc

1   /*
2    * Licensed to the Apache Software Foundation (ASF) under one or more
3    * contributor license agreements.  See the NOTICE file distributed with
4    * this work for additional information regarding copyright ownership.
5    * The ASF licenses this file to You under the Apache License, Version 2.0
6    * (the "License"); you may not use this file except in compliance with
7    * the License.  You may obtain a copy of the License at
8    *
9    *      http://www.apache.org/licenses/LICENSE-2.0
10   *
11   * Unless required by applicable law or agreed to in writing, software
12   * distributed under the License is distributed on an "AS IS" BASIS,
13   * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
14   * See the License for the specific language governing permissions and
15   * limitations under the License.
16   */
17  package org.apache.commons.math.analysis.solvers;
18  
19  import org.apache.commons.math.util.FastMath;
20  
21  /**
22   * This class implements the <a href="http://mathworld.wolfram.com/MullersMethod.html">
23   * Muller's Method</a> for root finding of real univariate functions. For
24   * reference, see <b>Elementary Numerical Analysis</b>, ISBN 0070124477,
25   * chapter 3.
26   * <p>
27   * Muller's method applies to both real and complex functions, but here we
28   * restrict ourselves to real functions.
29   * This class differs from {@link MullerSolver} in the way it avoids complex
30   * operations.</p>
31   * Muller's original method would have function evaluation at complex point.
32   * Since our f(x) is real, we have to find ways to avoid that. Bracketing
33   * condition is one way to go: by requiring bracketing in every iteration,
34   * the newly computed approximation is guaranteed to be real.</p>
35   * <p>
36   * Normally Muller's method converges quadratically in the vicinity of a
37   * zero, however it may be very slow in regions far away from zeros. For
38   * example, f(x) = exp(x) - 1, min = -50, max = 100. In such case we use
39   * bisection as a safety backup if it performs very poorly.</p>
40   * <p>
41   * The formulas here use divided differences directly.</p>
42   *
43   * @version $Id: MullerSolver.java 1183138 2011-10-13 22:21:04Z erans $
44   * @since 1.2
45   * @see MullerSolver2
46   */
47  public class MullerSolver extends AbstractUnivariateRealSolver {
48  
49      /** Default absolute accuracy. */
50      private static final double DEFAULT_ABSOLUTE_ACCURACY = 1e-6;
51  
52      /**
53       * Construct a solver with default accuracy (1e-6).
54       */
55      public MullerSolver() {
56          this(DEFAULT_ABSOLUTE_ACCURACY);
57      }
58      /**
59       * Construct a solver.
60       *
61       * @param absoluteAccuracy Absolute accuracy.
62       */
63      public MullerSolver(double absoluteAccuracy) {
64          super(absoluteAccuracy);
65      }
66      /**
67       * Construct a solver.
68       *
69       * @param relativeAccuracy Relative accuracy.
70       * @param absoluteAccuracy Absolute accuracy.
71       */
72      public MullerSolver(double relativeAccuracy,
73                          double absoluteAccuracy) {
74          super(relativeAccuracy, absoluteAccuracy);
75      }
76  
77      /**
78       * {@inheritDoc}
79       */
80      @Override
81      protected double doSolve() {
82          final double min = getMin();
83          final double max = getMax();
84          final double initial = getStartValue();
85  
86          final double functionValueAccuracy = getFunctionValueAccuracy();
87  
88          verifySequence(min, initial, max);
89  
90          // check for zeros before verifying bracketing
91          final double fMin = computeObjectiveValue(min);
92          if (FastMath.abs(fMin) < functionValueAccuracy) {
93              return min;
94          }
95          final double fMax = computeObjectiveValue(max);
96          if (FastMath.abs(fMax) < functionValueAccuracy) {
97              return max;
98          }
99          final double fInitial = computeObjectiveValue(initial);
100         if (FastMath.abs(fInitial) <  functionValueAccuracy) {
101             return initial;
102         }
103 
104         verifyBracketing(min, max);
105 
106         if (isBracketing(min, initial)) {
107             return solve(min, initial, fMin, fInitial);
108         } else {
109             return solve(initial, max, fInitial, fMax);
110         }
111     }
112 
113     /**
114      * Find a real root in the given interval.
115      *
116      * @param min Lower bound for the interval.
117      * @param max Upper bound for the interval.
118      * @param fMin function value at the lower bound.
119      * @param fMax function value at the upper bound.
120      * @return the point at which the function value is zero.
121      */
122     private double solve(double min, double max,
123                          double fMin, double fMax) {
124         final double relativeAccuracy = getRelativeAccuracy();
125         final double absoluteAccuracy = getAbsoluteAccuracy();
126         final double functionValueAccuracy = getFunctionValueAccuracy();
127 
128         // [x0, x2] is the bracketing interval in each iteration
129         // x1 is the last approximation and an interpolation point in (x0, x2)
130         // x is the new root approximation and new x1 for next round
131         // d01, d12, d012 are divided differences
132 
133         double x0 = min;
134         double y0 = fMin;
135         double x2 = max;
136         double y2 = fMax;
137         double x1 = 0.5 * (x0 + x2);
138         double y1 = computeObjectiveValue(x1);
139 
140         double oldx = Double.POSITIVE_INFINITY;
141         while (true) {
142             // Muller's method employs quadratic interpolation through
143             // x0, x1, x2 and x is the zero of the interpolating parabola.
144             // Due to bracketing condition, this parabola must have two
145             // real roots and we choose one in [x0, x2] to be x.
146             final double d01 = (y1 - y0) / (x1 - x0);
147             final double d12 = (y2 - y1) / (x2 - x1);
148             final double d012 = (d12 - d01) / (x2 - x0);
149             final double c1 = d01 + (x1 - x0) * d012;
150             final double delta = c1 * c1 - 4 * y1 * d012;
151             final double xplus = x1 + (-2.0 * y1) / (c1 + FastMath.sqrt(delta));
152             final double xminus = x1 + (-2.0 * y1) / (c1 - FastMath.sqrt(delta));
153             // xplus and xminus are two roots of parabola and at least
154             // one of them should lie in (x0, x2)
155             final double x = isSequence(x0, xplus, x2) ? xplus : xminus;
156             final double y = computeObjectiveValue(x);
157 
158             // check for convergence
159             final double tolerance = FastMath.max(relativeAccuracy * FastMath.abs(x), absoluteAccuracy);
160             if (FastMath.abs(x - oldx) <= tolerance ||
161                 FastMath.abs(y) <= functionValueAccuracy) {
162                 return x;
163             }
164 
165             // Bisect if convergence is too slow. Bisection would waste
166             // our calculation of x, hopefully it won't happen often.
167             // the real number equality test x == x1 is intentional and
168             // completes the proximity tests above it
169             boolean bisect = (x < x1 && (x1 - x0) > 0.95 * (x2 - x0)) ||
170                              (x > x1 && (x2 - x1) > 0.95 * (x2 - x0)) ||
171                              (x == x1);
172             // prepare the new bracketing interval for next iteration
173             if (!bisect) {
174                 x0 = x < x1 ? x0 : x1;
175                 y0 = x < x1 ? y0 : y1;
176                 x2 = x > x1 ? x2 : x1;
177                 y2 = x > x1 ? y2 : y1;
178                 x1 = x; y1 = y;
179                 oldx = x;
180             } else {
181                 double xm = 0.5 * (x0 + x2);
182                 double ym = computeObjectiveValue(xm);
183                 if (FastMath.signum(y0) + FastMath.signum(ym) == 0.0) {
184                     x2 = xm; y2 = ym;
185                 } else {
186                     x0 = xm; y0 = ym;
187                 }
188                 x1 = 0.5 * (x0 + x2);
189                 y1 = computeObjectiveValue(x1);
190                 oldx = Double.POSITIVE_INFINITY;
191             }
192         }
193     }
194 }