1 | |
---|
2 | /* Re Kolmogorov statistics, here is Birnbaum and Tingey's formula for the |
---|
3 | distribution of D+, the maximum of all positive deviations between a |
---|
4 | theoretical distribution function P(x) and an empirical one Sn(x) |
---|
5 | from n samples. |
---|
6 | |
---|
7 | + |
---|
8 | D = sup [P(x) - S (x)] |
---|
9 | n -inf < x < inf n |
---|
10 | |
---|
11 | |
---|
12 | [n(1-e)] |
---|
13 | + - v-1 n-v |
---|
14 | Pr{D > e} = > C e (e + v/n) (1 - e - v/n) |
---|
15 | n - n v |
---|
16 | v=0 |
---|
17 | |
---|
18 | [n(1-e)] is the largest integer not exceeding n(1-e). |
---|
19 | nCv is the number of combinations of n things taken v at a time. */ |
---|
20 | |
---|
21 | |
---|
22 | #include "mconf.h" |
---|
23 | #ifdef ANSIPROT |
---|
24 | extern double pow ( double, double ); |
---|
25 | extern double floor ( double ); |
---|
26 | extern double lgam ( double ); |
---|
27 | extern double exp ( double ); |
---|
28 | extern double sqrt ( double ); |
---|
29 | extern double log ( double ); |
---|
30 | extern double fabs ( double ); |
---|
31 | double smirnov ( int, double ); |
---|
32 | double kolmogorov ( double ); |
---|
33 | #else |
---|
34 | double pow (), floor (), lgam (), exp (), sqrt (), log (), fabs (); |
---|
35 | double smirnov (), kolmogorov (); |
---|
36 | #endif |
---|
37 | extern double MAXLOG; |
---|
38 | |
---|
39 | /* Exact Smirnov statistic, for one-sided test. */ |
---|
40 | double |
---|
41 | smirnov (n, e) |
---|
42 | int n; |
---|
43 | double e; |
---|
44 | { |
---|
45 | int v, nn; |
---|
46 | double evn, omevn, p, t, c, lgamnp1; |
---|
47 | |
---|
48 | if (n <= 0 || e < 0.0 || e > 1.0) |
---|
49 | return (-1.0); |
---|
50 | nn = floor ((double) n * (1.0 - e)); |
---|
51 | p = 0.0; |
---|
52 | if (n < 1013) |
---|
53 | { |
---|
54 | c = 1.0; |
---|
55 | for (v = 0; v <= nn; v++) |
---|
56 | { |
---|
57 | evn = e + ((double) v) / n; |
---|
58 | p += c * pow (evn, (double) (v - 1)) |
---|
59 | * pow (1.0 - evn, (double) (n - v)); |
---|
60 | /* Next combinatorial term; worst case error = 4e-15. */ |
---|
61 | c *= ((double) (n - v)) / (v + 1); |
---|
62 | } |
---|
63 | } |
---|
64 | else |
---|
65 | { |
---|
66 | lgamnp1 = lgam ((double) (n + 1)); |
---|
67 | for (v = 0; v <= nn; v++) |
---|
68 | { |
---|
69 | evn = e + ((double) v) / n; |
---|
70 | omevn = 1.0 - evn; |
---|
71 | if (fabs (omevn) > 0.0) |
---|
72 | { |
---|
73 | t = lgamnp1 |
---|
74 | - lgam ((double) (v + 1)) |
---|
75 | - lgam ((double) (n - v + 1)) |
---|
76 | + (v - 1) * log (evn) |
---|
77 | + (n - v) * log (omevn); |
---|
78 | if (t > -MAXLOG) |
---|
79 | p += exp (t); |
---|
80 | } |
---|
81 | } |
---|
82 | } |
---|
83 | return (p * e); |
---|
84 | } |
---|
85 | |
---|
86 | |
---|
87 | /* Kolmogorov's limiting distribution of two-sided test, returns |
---|
88 | probability that sqrt(n) * max deviation > y, |
---|
89 | or that max deviation > y/sqrt(n). |
---|
90 | The approximation is useful for the tail of the distribution |
---|
91 | when n is large. */ |
---|
92 | double |
---|
93 | kolmogorov (y) |
---|
94 | double y; |
---|
95 | { |
---|
96 | double p, t, r, sign, x; |
---|
97 | |
---|
98 | x = -2.0 * y * y; |
---|
99 | sign = 1.0; |
---|
100 | p = 0.0; |
---|
101 | r = 1.0; |
---|
102 | do |
---|
103 | { |
---|
104 | t = exp (x * r * r); |
---|
105 | p += sign * t; |
---|
106 | if (t == 0.0) |
---|
107 | break; |
---|
108 | r += 1.0; |
---|
109 | sign = -sign; |
---|
110 | } |
---|
111 | while ((t / p) > 1.1e-16); |
---|
112 | return (p + p); |
---|
113 | } |
---|
114 | |
---|
115 | /* Functional inverse of Smirnov distribution |
---|
116 | finds e such that smirnov(n,e) = p. */ |
---|
117 | double |
---|
118 | smirnovi (n, p) |
---|
119 | int n; |
---|
120 | double p; |
---|
121 | { |
---|
122 | double e, t, dpde; |
---|
123 | |
---|
124 | if (p <= 0.0 || p > 1.0) |
---|
125 | { |
---|
126 | mtherr ("smirnovi", DOMAIN); |
---|
127 | return 0.0; |
---|
128 | } |
---|
129 | /* Start with approximation p = exp(-2 n e^2). */ |
---|
130 | e = sqrt (-log (p) / (2.0 * n)); |
---|
131 | do |
---|
132 | { |
---|
133 | /* Use approximate derivative in Newton iteration. */ |
---|
134 | t = -2.0 * n * e; |
---|
135 | dpde = 2.0 * t * exp (t * e); |
---|
136 | if (fabs (dpde) > 0.0) |
---|
137 | t = (p - smirnov (n, e)) / dpde; |
---|
138 | else |
---|
139 | { |
---|
140 | mtherr ("smirnovi", UNDERFLOW); |
---|
141 | return 0.0; |
---|
142 | } |
---|
143 | e = e + t; |
---|
144 | if (e >= 1.0 || e <= 0.0) |
---|
145 | { |
---|
146 | mtherr ("smirnovi", OVERFLOW); |
---|
147 | return 0.0; |
---|
148 | } |
---|
149 | } |
---|
150 | while (fabs (t / e) > 1e-10); |
---|
151 | return (e); |
---|
152 | } |
---|
153 | |
---|
154 | |
---|
155 | /* Functional inverse of Kolmogorov statistic for two-sided test. |
---|
156 | Finds y such that kolmogorov(y) = p. |
---|
157 | If e = smirnovi (n,p), then kolmogi(2 * p) / sqrt(n) should |
---|
158 | be close to e. */ |
---|
159 | double |
---|
160 | kolmogi (p) |
---|
161 | double p; |
---|
162 | { |
---|
163 | double y, t, dpdy; |
---|
164 | |
---|
165 | if (p <= 0.0 || p > 1.0) |
---|
166 | { |
---|
167 | mtherr ("kolmogi", DOMAIN); |
---|
168 | return 0.0; |
---|
169 | } |
---|
170 | /* Start with approximation p = 2 exp(-2 y^2). */ |
---|
171 | y = sqrt (-0.5 * log (0.5 * p)); |
---|
172 | do |
---|
173 | { |
---|
174 | /* Use approximate derivative in Newton iteration. */ |
---|
175 | t = -2.0 * y; |
---|
176 | dpdy = 4.0 * t * exp (t * y); |
---|
177 | if (fabs (dpdy) > 0.0) |
---|
178 | t = (p - kolmogorov (y)) / dpdy; |
---|
179 | else |
---|
180 | { |
---|
181 | mtherr ("kolmogi", UNDERFLOW); |
---|
182 | return 0.0; |
---|
183 | } |
---|
184 | y = y + t; |
---|
185 | } |
---|
186 | while (fabs (t / y) > 1e-10); |
---|
187 | return (y); |
---|
188 | } |
---|
189 | |
---|
190 | |
---|
191 | #ifdef SALONE |
---|
192 | /* Type in a number. */ |
---|
193 | void |
---|
194 | getnum (s, px) |
---|
195 | char *s; |
---|
196 | double *px; |
---|
197 | { |
---|
198 | char str[30]; |
---|
199 | |
---|
200 | printf (" %s (%.15e) ? ", s, *px); |
---|
201 | gets (str); |
---|
202 | if (str[0] == '\0' || str[0] == '\n') |
---|
203 | return; |
---|
204 | sscanf (str, "%lf", px); |
---|
205 | printf ("%.15e\n", *px); |
---|
206 | } |
---|
207 | |
---|
208 | /* Type in values, get answers. */ |
---|
209 | void |
---|
210 | main () |
---|
211 | { |
---|
212 | int n; |
---|
213 | double e, p, ps, pk, ek, y; |
---|
214 | |
---|
215 | n = 5; |
---|
216 | e = 0.0; |
---|
217 | p = 0.1; |
---|
218 | loop: |
---|
219 | ps = n; |
---|
220 | getnum ("n", &ps); |
---|
221 | n = ps; |
---|
222 | if (n <= 0) |
---|
223 | { |
---|
224 | printf ("? Operator error.\n"); |
---|
225 | goto loop; |
---|
226 | } |
---|
227 | /* |
---|
228 | getnum ("e", &e); |
---|
229 | ps = smirnov (n, e); |
---|
230 | y = sqrt ((double) n) * e; |
---|
231 | printf ("y = %.4e\n", y); |
---|
232 | pk = kolmogorov (y); |
---|
233 | printf ("Smirnov = %.15e, Kolmogorov/2 = %.15e\n", ps, pk / 2.0); |
---|
234 | */ |
---|
235 | getnum ("p", &p); |
---|
236 | e = smirnovi (n, p); |
---|
237 | printf ("Smirnov e = %.15e\n", e); |
---|
238 | y = kolmogi (2.0 * p); |
---|
239 | ek = y / sqrt ((double) n); |
---|
240 | printf ("Kolmogorov e = %.15e\n", ek); |
---|
241 | goto loop; |
---|
242 | } |
---|
243 | #endif |
---|