• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 namespace Eigen {
2 
3 namespace internal {
4 
5 template <typename Scalar>
lmpar(Matrix<Scalar,Dynamic,Dynamic> & r,const VectorXi & ipvt,const Matrix<Scalar,Dynamic,1> & diag,const Matrix<Scalar,Dynamic,1> & qtb,Scalar delta,Scalar & par,Matrix<Scalar,Dynamic,1> & x)6 void lmpar(
7         Matrix< Scalar, Dynamic, Dynamic > &r,
8         const VectorXi &ipvt,
9         const Matrix< Scalar, Dynamic, 1 >  &diag,
10         const Matrix< Scalar, Dynamic, 1 >  &qtb,
11         Scalar delta,
12         Scalar &par,
13         Matrix< Scalar, Dynamic, 1 >  &x)
14 {
15     using std::abs;
16     using std::sqrt;
17     typedef DenseIndex Index;
18 
19     /* Local variables */
20     Index i, j, l;
21     Scalar fp;
22     Scalar parc, parl;
23     Index iter;
24     Scalar temp, paru;
25     Scalar gnorm;
26     Scalar dxnorm;
27 
28 
29     /* Function Body */
30     const Scalar dwarf = (std::numeric_limits<Scalar>::min)();
31     const Index n = r.cols();
32     eigen_assert(n==diag.size());
33     eigen_assert(n==qtb.size());
34     eigen_assert(n==x.size());
35 
36     Matrix< Scalar, Dynamic, 1 >  wa1, wa2;
37 
38     /* compute and store in x the gauss-newton direction. if the */
39     /* jacobian is rank-deficient, obtain a least squares solution. */
40     Index nsing = n-1;
41     wa1 = qtb;
42     for (j = 0; j < n; ++j) {
43         if (r(j,j) == 0. && nsing == n-1)
44             nsing = j - 1;
45         if (nsing < n-1)
46             wa1[j] = 0.;
47     }
48     for (j = nsing; j>=0; --j) {
49         wa1[j] /= r(j,j);
50         temp = wa1[j];
51         for (i = 0; i < j ; ++i)
52             wa1[i] -= r(i,j) * temp;
53     }
54 
55     for (j = 0; j < n; ++j)
56         x[ipvt[j]] = wa1[j];
57 
58     /* initialize the iteration counter. */
59     /* evaluate the function at the origin, and test */
60     /* for acceptance of the gauss-newton direction. */
61     iter = 0;
62     wa2 = diag.cwiseProduct(x);
63     dxnorm = wa2.blueNorm();
64     fp = dxnorm - delta;
65     if (fp <= Scalar(0.1) * delta) {
66         par = 0;
67         return;
68     }
69 
70     /* if the jacobian is not rank deficient, the newton */
71     /* step provides a lower bound, parl, for the zero of */
72     /* the function. otherwise set this bound to zero. */
73     parl = 0.;
74     if (nsing >= n-1) {
75         for (j = 0; j < n; ++j) {
76             l = ipvt[j];
77             wa1[j] = diag[l] * (wa2[l] / dxnorm);
78         }
79         // it's actually a triangularView.solveInplace(), though in a weird
80         // way:
81         for (j = 0; j < n; ++j) {
82             Scalar sum = 0.;
83             for (i = 0; i < j; ++i)
84                 sum += r(i,j) * wa1[i];
85             wa1[j] = (wa1[j] - sum) / r(j,j);
86         }
87         temp = wa1.blueNorm();
88         parl = fp / delta / temp / temp;
89     }
90 
91     /* calculate an upper bound, paru, for the zero of the function. */
92     for (j = 0; j < n; ++j)
93         wa1[j] = r.col(j).head(j+1).dot(qtb.head(j+1)) / diag[ipvt[j]];
94 
95     gnorm = wa1.stableNorm();
96     paru = gnorm / delta;
97     if (paru == 0.)
98         paru = dwarf / (std::min)(delta,Scalar(0.1));
99 
100     /* if the input par lies outside of the interval (parl,paru), */
101     /* set par to the closer endpoint. */
102     par = (std::max)(par,parl);
103     par = (std::min)(par,paru);
104     if (par == 0.)
105         par = gnorm / dxnorm;
106 
107     /* beginning of an iteration. */
108     while (true) {
109         ++iter;
110 
111         /* evaluate the function at the current value of par. */
112         if (par == 0.)
113             par = (std::max)(dwarf,Scalar(.001) * paru); /* Computing MAX */
114         wa1 = sqrt(par)* diag;
115 
116         Matrix< Scalar, Dynamic, 1 > sdiag(n);
117         qrsolv<Scalar>(r, ipvt, wa1, qtb, x, sdiag);
118 
119         wa2 = diag.cwiseProduct(x);
120         dxnorm = wa2.blueNorm();
121         temp = fp;
122         fp = dxnorm - delta;
123 
124         /* if the function is small enough, accept the current value */
125         /* of par. also test for the exceptional cases where parl */
126         /* is zero or the number of iterations has reached 10. */
127         if (abs(fp) <= Scalar(0.1) * delta || (parl == 0. && fp <= temp && temp < 0.) || iter == 10)
128             break;
129 
130         /* compute the newton correction. */
131         for (j = 0; j < n; ++j) {
132             l = ipvt[j];
133             wa1[j] = diag[l] * (wa2[l] / dxnorm);
134         }
135         for (j = 0; j < n; ++j) {
136             wa1[j] /= sdiag[j];
137             temp = wa1[j];
138             for (i = j+1; i < n; ++i)
139                 wa1[i] -= r(i,j) * temp;
140         }
141         temp = wa1.blueNorm();
142         parc = fp / delta / temp / temp;
143 
144         /* depending on the sign of the function, update parl or paru. */
145         if (fp > 0.)
146             parl = (std::max)(parl,par);
147         if (fp < 0.)
148             paru = (std::min)(paru,par);
149 
150         /* compute an improved estimate for par. */
151         /* Computing MAX */
152         par = (std::max)(parl,par+parc);
153 
154         /* end of an iteration. */
155     }
156 
157     /* termination. */
158     if (iter == 0)
159         par = 0.;
160     return;
161 }
162 
163 template <typename Scalar>
lmpar2(const ColPivHouseholderQR<Matrix<Scalar,Dynamic,Dynamic>> & qr,const Matrix<Scalar,Dynamic,1> & diag,const Matrix<Scalar,Dynamic,1> & qtb,Scalar delta,Scalar & par,Matrix<Scalar,Dynamic,1> & x)164 void lmpar2(
165         const ColPivHouseholderQR<Matrix< Scalar, Dynamic, Dynamic> > &qr,
166         const Matrix< Scalar, Dynamic, 1 >  &diag,
167         const Matrix< Scalar, Dynamic, 1 >  &qtb,
168         Scalar delta,
169         Scalar &par,
170         Matrix< Scalar, Dynamic, 1 >  &x)
171 
172 {
173     using std::sqrt;
174     using std::abs;
175     typedef DenseIndex Index;
176 
177     /* Local variables */
178     Index j;
179     Scalar fp;
180     Scalar parc, parl;
181     Index iter;
182     Scalar temp, paru;
183     Scalar gnorm;
184     Scalar dxnorm;
185 
186 
187     /* Function Body */
188     const Scalar dwarf = (std::numeric_limits<Scalar>::min)();
189     const Index n = qr.matrixQR().cols();
190     eigen_assert(n==diag.size());
191     eigen_assert(n==qtb.size());
192 
193     Matrix< Scalar, Dynamic, 1 >  wa1, wa2;
194 
195     /* compute and store in x the gauss-newton direction. if the */
196     /* jacobian is rank-deficient, obtain a least squares solution. */
197 
198 //    const Index rank = qr.nonzeroPivots(); // exactly double(0.)
199     const Index rank = qr.rank(); // use a threshold
200     wa1 = qtb;
201     wa1.tail(n-rank).setZero();
202     qr.matrixQR().topLeftCorner(rank, rank).template triangularView<Upper>().solveInPlace(wa1.head(rank));
203 
204     x = qr.colsPermutation()*wa1;
205 
206     /* initialize the iteration counter. */
207     /* evaluate the function at the origin, and test */
208     /* for acceptance of the gauss-newton direction. */
209     iter = 0;
210     wa2 = diag.cwiseProduct(x);
211     dxnorm = wa2.blueNorm();
212     fp = dxnorm - delta;
213     if (fp <= Scalar(0.1) * delta) {
214         par = 0;
215         return;
216     }
217 
218     /* if the jacobian is not rank deficient, the newton */
219     /* step provides a lower bound, parl, for the zero of */
220     /* the function. otherwise set this bound to zero. */
221     parl = 0.;
222     if (rank==n) {
223         wa1 = qr.colsPermutation().inverse() *  diag.cwiseProduct(wa2)/dxnorm;
224         qr.matrixQR().topLeftCorner(n, n).transpose().template triangularView<Lower>().solveInPlace(wa1);
225         temp = wa1.blueNorm();
226         parl = fp / delta / temp / temp;
227     }
228 
229     /* calculate an upper bound, paru, for the zero of the function. */
230     for (j = 0; j < n; ++j)
231         wa1[j] = qr.matrixQR().col(j).head(j+1).dot(qtb.head(j+1)) / diag[qr.colsPermutation().indices()(j)];
232 
233     gnorm = wa1.stableNorm();
234     paru = gnorm / delta;
235     if (paru == 0.)
236         paru = dwarf / (std::min)(delta,Scalar(0.1));
237 
238     /* if the input par lies outside of the interval (parl,paru), */
239     /* set par to the closer endpoint. */
240     par = (std::max)(par,parl);
241     par = (std::min)(par,paru);
242     if (par == 0.)
243         par = gnorm / dxnorm;
244 
245     /* beginning of an iteration. */
246     Matrix< Scalar, Dynamic, Dynamic > s = qr.matrixQR();
247     while (true) {
248         ++iter;
249 
250         /* evaluate the function at the current value of par. */
251         if (par == 0.)
252             par = (std::max)(dwarf,Scalar(.001) * paru); /* Computing MAX */
253         wa1 = sqrt(par)* diag;
254 
255         Matrix< Scalar, Dynamic, 1 > sdiag(n);
256         qrsolv<Scalar>(s, qr.colsPermutation().indices(), wa1, qtb, x, sdiag);
257 
258         wa2 = diag.cwiseProduct(x);
259         dxnorm = wa2.blueNorm();
260         temp = fp;
261         fp = dxnorm - delta;
262 
263         /* if the function is small enough, accept the current value */
264         /* of par. also test for the exceptional cases where parl */
265         /* is zero or the number of iterations has reached 10. */
266         if (abs(fp) <= Scalar(0.1) * delta || (parl == 0. && fp <= temp && temp < 0.) || iter == 10)
267             break;
268 
269         /* compute the newton correction. */
270         wa1 = qr.colsPermutation().inverse() * diag.cwiseProduct(wa2/dxnorm);
271         // we could almost use this here, but the diagonal is outside qr, in sdiag[]
272         // qr.matrixQR().topLeftCorner(n, n).transpose().template triangularView<Lower>().solveInPlace(wa1);
273         for (j = 0; j < n; ++j) {
274             wa1[j] /= sdiag[j];
275             temp = wa1[j];
276             for (Index i = j+1; i < n; ++i)
277                 wa1[i] -= s(i,j) * temp;
278         }
279         temp = wa1.blueNorm();
280         parc = fp / delta / temp / temp;
281 
282         /* depending on the sign of the function, update parl or paru. */
283         if (fp > 0.)
284             parl = (std::max)(parl,par);
285         if (fp < 0.)
286             paru = (std::min)(paru,par);
287 
288         /* compute an improved estimate for par. */
289         par = (std::max)(parl,par+parc);
290     }
291     if (iter == 0)
292         par = 0.;
293     return;
294 }
295 
296 } // end namespace internal
297 
298 } // end namespace Eigen
299