1 namespace Eigen {
2 
3 namespace internal {
4 
5 template <typename Scalar>
dogleg(const Matrix<Scalar,Dynamic,Dynamic> & qrfac,const Matrix<Scalar,Dynamic,1> & diag,const Matrix<Scalar,Dynamic,1> & qtb,Scalar delta,Matrix<Scalar,Dynamic,1> & x)6 void dogleg(
7         const Matrix< Scalar, Dynamic, Dynamic >  &qrfac,
8         const Matrix< Scalar, Dynamic, 1 >  &diag,
9         const Matrix< Scalar, Dynamic, 1 >  &qtb,
10         Scalar delta,
11         Matrix< Scalar, Dynamic, 1 >  &x)
12 {
13     using std::abs;
14     using std::sqrt;
15 
16     typedef DenseIndex Index;
17 
18     /* Local variables */
19     Index i, j;
20     Scalar sum, temp, alpha, bnorm;
21     Scalar gnorm, qnorm;
22     Scalar sgnorm;
23 
24     /* Function Body */
25     const Scalar epsmch = NumTraits<Scalar>::epsilon();
26     const Index n = qrfac.cols();
27     eigen_assert(n==qtb.size());
28     eigen_assert(n==x.size());
29     eigen_assert(n==diag.size());
30     Matrix< Scalar, Dynamic, 1 >  wa1(n), wa2(n);
31 
32     /* first, calculate the gauss-newton direction. */
33     for (j = n-1; j >=0; --j) {
34         temp = qrfac(j,j);
35         if (temp == 0.) {
36             temp = epsmch * qrfac.col(j).head(j+1).maxCoeff();
37             if (temp == 0.)
38                 temp = epsmch;
39         }
40         if (j==n-1)
41             x[j] = qtb[j] / temp;
42         else
43             x[j] = (qtb[j] - qrfac.row(j).tail(n-j-1).dot(x.tail(n-j-1))) / temp;
44     }
45 
46     /* test whether the gauss-newton direction is acceptable. */
47     qnorm = diag.cwiseProduct(x).stableNorm();
48     if (qnorm <= delta)
49         return;
50 
51     // TODO : this path is not tested by Eigen unit tests
52 
53     /* the gauss-newton direction is not acceptable. */
54     /* next, calculate the scaled gradient direction. */
55 
56     wa1.fill(0.);
57     for (j = 0; j < n; ++j) {
58         wa1.tail(n-j) += qrfac.row(j).tail(n-j) * qtb[j];
59         wa1[j] /= diag[j];
60     }
61 
62     /* calculate the norm of the scaled gradient and test for */
63     /* the special case in which the scaled gradient is zero. */
64     gnorm = wa1.stableNorm();
65     sgnorm = 0.;
66     alpha = delta / qnorm;
67     if (gnorm == 0.)
68         goto algo_end;
69 
70     /* calculate the point along the scaled gradient */
71     /* at which the quadratic is minimized. */
72     wa1.array() /= (diag*gnorm).array();
73     // TODO : once unit tests cover this part,:
74     // wa2 = qrfac.template triangularView<Upper>() * wa1;
75     for (j = 0; j < n; ++j) {
76         sum = 0.;
77         for (i = j; i < n; ++i) {
78             sum += qrfac(j,i) * wa1[i];
79         }
80         wa2[j] = sum;
81     }
82     temp = wa2.stableNorm();
83     sgnorm = gnorm / temp / temp;
84 
85     /* test whether the scaled gradient direction is acceptable. */
86     alpha = 0.;
87     if (sgnorm >= delta)
88         goto algo_end;
89 
90     /* the scaled gradient direction is not acceptable. */
91     /* finally, calculate the point along the dogleg */
92     /* at which the quadratic is minimized. */
93     bnorm = qtb.stableNorm();
94     temp = bnorm / gnorm * (bnorm / qnorm) * (sgnorm / delta);
95     temp = temp - delta / qnorm * numext::abs2(sgnorm / delta) + sqrt(numext::abs2(temp - delta / qnorm) + (1.-numext::abs2(delta / qnorm)) * (1.-numext::abs2(sgnorm / delta)));
96     alpha = delta / qnorm * (1. - numext::abs2(sgnorm / delta)) / temp;
97 algo_end:
98 
99     /* form appropriate convex combination of the gauss-newton */
100     /* direction and the scaled gradient direction. */
101     temp = (1.-alpha) * (std::min)(sgnorm,delta);
102     x = temp * wa1 + alpha * x;
103 }
104 
105 } // end namespace internal
106 
107 } // end namespace Eigen
108