1 // Arguments: Doubles, Doubles, Doubles
2 #include <stan/math/prim.hpp>
3 
4 using stan::math::square;
5 using stan::math::var;
6 using std::numeric_limits;
7 using std::vector;
8 
9 class AgradCdfLogNormal : public AgradCdfLogTest {
10  public:
valid_values(vector<vector<double>> & parameters,vector<double> & cdf_log)11   void valid_values(vector<vector<double> >& parameters,
12                     vector<double>& cdf_log) {
13     vector<double> param(3);
14 
15     param[0] = 0;  // y
16     param[1] = 0;  // mu
17     param[2] = 1;  // sigma
18     parameters.push_back(param);
19     cdf_log.push_back(-0.6931471805599452862268);  // expected cdf_log
20 
21     param[0] = 1;  // y
22     param[1] = 0;  // mu
23     param[2] = 1;  // sigma
24     parameters.push_back(param);
25     cdf_log.push_back(-0.1727537790234499326392);  // expected cdf_log
26 
27     param[0] = -2;   // y
28     param[1] = 0;    // mu
29     param[2] = 1.5;  // sigma
30     parameters.push_back(param);
31     cdf_log.push_back(-2.3945773661586438052);  // expected cdf_log
32 
33     param[0] = -3.5;  // y
34     param[1] = 1.9;   // mu
35     param[2] = 7.2;   // sigma
36     parameters.push_back(param);
37     cdf_log.push_back(-1.484448229919656192521);  // expected cdf_log
38 
39     param[0] = 7.5;  // y
40     param[1] = 0;    // mu
41     param[2] = 1;    // sigma
42     parameters.push_back(param);
43     cdf_log.push_back(-3.1908916729109475e-14);  // expected cdf_log
44   }
45 
invalid_values(vector<size_t> & index,vector<double> & value)46   void invalid_values(vector<size_t>& index, vector<double>& value) {
47     // y
48 
49     // mu
50     index.push_back(1U);
51     value.push_back(numeric_limits<double>::infinity());
52 
53     index.push_back(1U);
54     value.push_back(-numeric_limits<double>::infinity());
55 
56     // sigma
57     index.push_back(2U);
58     value.push_back(0.0);
59 
60     index.push_back(2U);
61     value.push_back(-1.0);
62 
63     index.push_back(2U);
64     value.push_back(-numeric_limits<double>::infinity());
65   }
66 
has_lower_bound()67   bool has_lower_bound() { return false; }
68 
has_upper_bound()69   bool has_upper_bound() { return false; }
70 
71   template <typename T_y, typename T_loc, typename T_scale, typename T3,
72             typename T4, typename T5>
cdf_log(const T_y & y,const T_loc & mu,const T_scale & sigma,const T3 &,const T4 &,const T5 &)73   stan::return_type_t<T_y, T_loc, T_scale> cdf_log(const T_y& y,
74                                                    const T_loc& mu,
75                                                    const T_scale& sigma,
76                                                    const T3&, const T4&,
77                                                    const T5&) {
78     return stan::math::normal_cdf_log(y, mu, sigma);
79   }
80 
81   template <typename T_y, typename T_loc, typename T_scale, typename T3,
82             typename T4, typename T5>
cdf_log_function(const T_y & y,const T_loc & mu,const T_scale & sigma,const T3 &,const T4 &,const T5 &)83   stan::return_type_t<T_y, T_loc, T_scale> cdf_log_function(
84       const T_y& y, const T_loc& mu, const T_scale& sigma, const T3&, const T4&,
85       const T5&) {
86     using stan::math::LOG_HALF;
87     using stan::math::SQRT_PI;
88     using stan::math::SQRT_TWO;
89     using std::exp;
90     using std::log;
91     using std::log1p;
92     using std::pow;
93 
94     stan::return_type_t<T_y, T_loc, T_scale> cdf_log(0.0);
95 
96     // define constants used in numerical approximation of log(CDF) for x<-20
97     stan::return_type_t<T_y, T_loc, T_scale> p[6]
98         = {-0.000658749161529837803157, -0.0160837851487422766278,
99            -0.125781726111229246204,    -0.360344899949804439429,
100            -0.305326634961232344035,    -0.0163153871373020978498};
101     stan::return_type_t<T_y, T_loc, T_scale> q[6]
102         = {0.00233520497626869185443, 0.0605183413124413191178,
103            0.527905102951428412248,   1.87295284992346047209,
104            2.56852019228982242072,    1.0};
105     stan::return_type_t<T_y, T_loc, T_scale> scaled_diff(0.0);
106     scaled_diff += (y - mu) / (sigma * SQRT_TWO);
107     stan::return_type_t<T_y, T_loc, T_scale> sigma2pi(0.0);
108     sigma2pi += sigma * SQRT_TWO;
109 
110     // use erfc() instead of erf() in order to retain precision since for x>0
111     // erfc()->0
112     if (scaled_diff > 0.0) {
113       // CDF(x) = 1/2 + 1/2erf(x) = 1 - 1/2erfc(x)
114       cdf_log += log1p(-0.5 * erfc(scaled_diff));
115     } else if (scaled_diff > -20.0) {
116       // CDF(x) = 1/2 - 1/2erf(-x) = 1/2erfc(-x)
117       cdf_log += log(erfc(-scaled_diff)) + LOG_HALF;
118     } else {
119       // check whether scaled_diff^10 term will overflow
120       if (10.0 * log(-scaled_diff) < log(std::numeric_limits<double>::max())) {
121         // entering territory where erfc(-x)~0
122         // need to use direct numerical approximation of cdf_log instead
123         // the following based on W. J. Cody, Math. Comp. 23(107):631-638 (1969)
124         // CDF(x) = 1/2erfc(-x)
125         stan::return_type_t<T_y, T_loc, T_scale> temp_p = 0.0;
126         stan::return_type_t<T_y, T_loc, T_scale> temp_q = 0.0;
127         stan::return_type_t<T_y, T_loc, T_scale> x2 = square(scaled_diff);
128         stan::return_type_t<T_y, T_loc, T_scale> pow_x = 0.0;
129         for (int j = 0; j < 6; j++) {
130           pow_x = pow(scaled_diff, -2 * j);
131           temp_p = temp_p - p[j] * pow_x;
132           temp_q = temp_q - q[j] * pow_x;
133         }
134         cdf_log += log(0.5) + log(1.0 / SQRT_PI + (temp_p / temp_q) / x2)
135                    - log(-scaled_diff) - x2;
136       } else {
137         cdf_log += stan::math::negative_infinity();
138       }
139     }
140     return cdf_log;
141   }
142 };
143