1 #ifndef STAN_MATH_REV_MAT_FUN_LOG_DETERMINANT_SPD_HPP
2 #define STAN_MATH_REV_MAT_FUN_LOG_DETERMINANT_SPD_HPP
3
4 #include <stan/math/prim/scal/err/domain_error.hpp>
5 #include <stan/math/prim/scal/err/check_finite.hpp>
6 #include <stan/math/prim/mat/err/check_square.hpp>
7 #include <stan/math/prim/mat/fun/Eigen.hpp>
8 #include <stan/math/rev/core.hpp>
9
10 namespace stan {
11 namespace math {
12
13 template <int R, int C>
log_determinant_spd(const Eigen::Matrix<var,R,C> & m)14 inline var log_determinant_spd(const Eigen::Matrix<var, R, C>& m) {
15 using Eigen::Matrix;
16
17 check_square("log_determinant_spd", "m", m);
18
19 Matrix<double, R, C> m_d(m.rows(), m.cols());
20 for (int i = 0; i < m.size(); ++i)
21 m_d(i) = m(i).val();
22
23 Eigen::LDLT<Matrix<double, R, C> > ldlt(m_d);
24 if (ldlt.info() != Eigen::Success) {
25 double y = 0;
26 domain_error("log_determinant_spd", "matrix argument", y,
27 "failed LDLT factorization");
28 }
29
30 // compute the inverse of A (needed for the derivative)
31 m_d.setIdentity(m.rows(), m.cols());
32 ldlt.solveInPlace(m_d);
33
34 if (ldlt.isNegative() || (ldlt.vectorD().array() <= 1e-16).any()) {
35 double y = 0;
36 domain_error("log_determinant_spd", "matrix argument", y,
37 "matrix is negative definite");
38 }
39
40 double val = ldlt.vectorD().array().log().sum();
41
42 check_finite("log_determinant_spd",
43 "log determininant of the matrix argument", val);
44
45 vari** operands
46 = ChainableStack::instance().memalloc_.alloc_array<vari*>(m.size());
47 for (int i = 0; i < m.size(); ++i)
48 operands[i] = m(i).vi_;
49
50 double* gradients
51 = ChainableStack::instance().memalloc_.alloc_array<double>(m.size());
52 for (int i = 0; i < m.size(); ++i)
53 gradients[i] = m_d(i);
54
55 return var(
56 new precomputed_gradients_vari(val, m.size(), operands, gradients));
57 }
58
59 } // namespace math
60
61 } // namespace stan
62 #endif
63