drvec 0.1.0__py3-none-any.whl

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
drvec/diagnose.py ADDED
@@ -0,0 +1,439 @@
1
+ """The per-series residual diagnosis of the .out: diagnose.c (drvec-v0.12,
2
+ engines/drvec/src/diagnose.c, vendored from drvarma), transcribed line for line
3
+ -- the statistics, the ASCII plot, the histogram, the ACF and PACF with their
4
+ Ljung-Box, and the cross-correlation functions. Everything is printed, so the
5
+ arithmetic keeps the C's order (sums in increasing t, pow() for the moments)."""
6
+ from __future__ import annotations
7
+
8
+ import math
9
+
10
+ GUIONS = "-------------+-------------------------+-------------------------+--------------"
11
+
12
+
13
+ def round_local(num):
14
+ """nlatools.c:round_local: half away from zero."""
15
+ t1 = abs(num)
16
+ t2 = math.floor(t1)
17
+ if t1 - t2 >= 0.5:
18
+ t2 = math.ceil(t1)
19
+ i = int(t2)
20
+ return i if num >= 0.0 else -i
21
+
22
+
23
+ def obs_to_date(beg_per, beg_sub, obs_no, freq):
24
+ """diagnose.c:ObsToDate -> (per, sub)."""
25
+ if obs_no + beg_sub - 1 <= freq:
26
+ return beg_per, beg_sub + obs_no - 1
27
+ q, rem = divmod(obs_no - (freq - beg_sub + 1), freq)
28
+ if rem > 0:
29
+ return beg_per + q + 1, rem
30
+ return beg_per + q, freq
31
+
32
+
33
+ def mean(d):
34
+ s = 0.0
35
+ for v in d:
36
+ s += v
37
+ return s / len(d)
38
+
39
+
40
+ def stdev(d):
41
+ ave = mean(d)
42
+ s = 0.0
43
+ for v in d:
44
+ s += (v - ave) * (v - ave)
45
+ return math.sqrt(s / len(d))
46
+
47
+
48
+ def skew(d):
49
+ ave, sd = mean(d), stdev(d)
50
+ if sd < 1e-20:
51
+ return 0.0
52
+ s = 0.0
53
+ for v in d:
54
+ s += math.pow((v - ave) / sd, 3.0)
55
+ return s / len(d)
56
+
57
+
58
+ def kurt(d):
59
+ ave, sd = mean(d), stdev(d)
60
+ if sd < 1e-20:
61
+ return 0.0
62
+ s = 0.0
63
+ for v in d:
64
+ s += math.pow((v - ave) / sd, 4.0)
65
+ return s / len(d) - 3.0
66
+
67
+
68
+ def maxval(d):
69
+ im, mx = 0, d[0]
70
+ for i in range(1, len(d)):
71
+ if d[i] > mx:
72
+ mx, im = d[i], i
73
+ return im
74
+
75
+
76
+ def minval(d):
77
+ im, mn = 0, d[0]
78
+ for i in range(1, len(d)):
79
+ if d[i] < mn:
80
+ mn, im = d[i], i
81
+ return im
82
+
83
+
84
+ def acf(d, lags, mu, var):
85
+ n = len(d)
86
+ out = [0.0] * (lags + 1)
87
+ for j in range(1, lags + 1):
88
+ c = 0.0
89
+ for i in range(n - j):
90
+ c += (d[i] - mu) * (d[i + j] - mu)
91
+ out[j] = c / (n * var)
92
+ return out
93
+
94
+
95
+ def pacf(corr, lags):
96
+ Mt = [[0.0] * (lags + 1) for _ in range(lags + 1)]
97
+ Mt[1][1] = corr[1]
98
+ for i in range(2, lags + 1):
99
+ s1 = s2 = 0.0
100
+ for j in range(1, i):
101
+ s1 += Mt[i - 1][j] * corr[i - j]
102
+ s2 += Mt[i - 1][j] * corr[j]
103
+ Mt[i][i] = (corr[i] - s1) / (1.0 - s2)
104
+ for j in range(1, i):
105
+ Mt[i][j] = Mt[i - 1][j] - Mt[i][i] * Mt[i - 1][i - j]
106
+ return [0.0] + [Mt[i][i] for i in range(1, lags + 1)]
107
+
108
+
109
+ def chitest(corr, lags, n):
110
+ c = 0.0
111
+ for i in range(1, lags + 1):
112
+ c += corr[i] * corr[i] / (n - i)
113
+ return c * (n * (n + 2))
114
+
115
+
116
+ def chitest_c(corr, lags, n):
117
+ c = 0.0
118
+ for i in range(1, lags + 1):
119
+ c += corr[i] * corr[i] / (n - i + 1)
120
+ return c * (n * (n + 2))
121
+
122
+
123
+ def plot_cor(corr, lags, isacf, n, freq, npar):
124
+ marcas = (" -1 0 1 L-B Q DF"
125
+ if isacf else " -1 0 1")
126
+ o = "Autocorrelation function (acf " if isacf else "Partial autocorrelation function (pacf "
127
+ o += "bands = %5.3f):\n\n%s\n%s\n" % (2.0 / math.sqrt(n), marcas, GUIONS)
128
+ for i in range(1, lags + 1):
129
+ t = [" "] * 53
130
+ if freq != 1 and i % freq == 0:
131
+ o += "%4d %7.3f +" % (i, corr[i])
132
+ sym = "+"
133
+ t[51] = "+"
134
+ else:
135
+ o += "%4d %7.3f |" % (i, corr[i])
136
+ sym = "*"
137
+ t[51] = "|"
138
+ pos = corr[i] * 25.0
139
+ posi = abs(round_local(pos))
140
+ if pos <= 0.0:
141
+ for j in range(25 - posi, 26):
142
+ t[j] = sym
143
+ else:
144
+ for j in range(25, 26 + posi):
145
+ t[j] = sym
146
+ t[25] = "|"
147
+ posi = round_local(2.0 / math.sqrt(n) * 25.0)
148
+ if t[25 + posi] == " ":
149
+ t[25 + posi] = ":"
150
+ if t[25 - posi] == " ":
151
+ t[25 - posi] = ":"
152
+ o += "".join(t)
153
+ lb = "%6.2f %3d" % (chitest(corr, i, n), i - npar)
154
+ if freq != 1 and i % freq == 0 and isacf and i - npar >= 1:
155
+ o += lb
156
+ if i % freq != 0 and isacf and i == lags and i - npar >= 1:
157
+ o += lb
158
+ if freq == 1 and isacf and i == lags and i - npar >= 1:
159
+ o += lb
160
+ o += "\n"
161
+ return o + "%s\n%s\n\n" % (GUIONS, marcas)
162
+
163
+
164
+ def plot_ccf(corr, lags, n):
165
+ marcas = " -1 0 1"
166
+ o = "CCF BANDS 2.0/SQRT(N) = %2.5f:\n" % (2.0 / math.sqrt(n))
167
+ o += "\n%s\n%s\n" % (marcas, GUIONS)
168
+ for i in range(1, 2 * lags + 2):
169
+ t = [" "] * 80
170
+ o += "%4d %7.3f |" % (i - lags - 1, corr[i])
171
+ t[51] = "|"
172
+ pos = corr[i] * 25.0
173
+ posi = int(pos + (0.5 if pos >= 0 else -0.5))
174
+ if pos <= 0.0:
175
+ for j in range(25 - abs(posi), 26):
176
+ t[j] = "*"
177
+ else:
178
+ for j in range(25, 26 + posi):
179
+ t[j] = "*"
180
+ t[25] = "|"
181
+ posi = int(2.0 / math.sqrt(n) * 25.0 + 0.5)
182
+ if t[25 + posi] == " ":
183
+ t[25 + posi] = ":"
184
+ if t[25 - posi] == " ":
185
+ t[25 - posi] = ":"
186
+ o += "".join(t) + "\n"
187
+ return o + "%s\n%s\n" % (GUIONS, marcas)
188
+
189
+
190
+ def ccf(d1, d2, lags, m1, m2, sd1, sd2):
191
+ n = len(d1)
192
+ out = [0.0] * (lags + 2)
193
+ for j in range(1, lags + 2):
194
+ c = 0.0
195
+ for i in range(n - j + 1):
196
+ c += (d1[i] - m1) * (d2[i + j - 1] - m2)
197
+ out[j] = c / (n * sd1 * sd2)
198
+ return out
199
+
200
+
201
+ def stat_ser(d, freq, by, bt):
202
+ n = len(d)
203
+ ey, et = obs_to_date(by, bt, n, freq)
204
+ mu = mean(d)
205
+ tmp = stdev(d)
206
+ sk, ku = skew(d), kurt(d)
207
+ imax, imin = maxval(d), minval(d)
208
+ o = "Unconditional residuals (seasonal period: %d)\n" % freq
209
+ o += "%d observations: " % n
210
+ o += ("from %d/%d to %d/%d\n" % (bt, by, et, ey)) if freq > 1 else ("from %d to %d\n" % (by, ey))
211
+ o += "\n"
212
+ o += " Mean: %18.6f\n" % mu
213
+ o += "Standard error of mean: %18.6f\n" % (tmp / math.sqrt(n))
214
+ o += " Variance: %18.6f\n" % (tmp * tmp)
215
+ o += " Standard deviation: %18.6f\n" % tmp
216
+ o += " Skewness: %18.6f\n" % sk
217
+ o += " Kurtosis: %18.6f\n" % ku
218
+ My, Mt = obs_to_date(by, bt, imax + 1, freq)
219
+ my, mt = obs_to_date(by, bt, imin + 1, freq)
220
+ if freq > 1:
221
+ o += " Minimum: %18.6f at %2d/%d (observation %3d)\n" % (d[imin], mt, my, imin + 1)
222
+ o += " Maximum: %18.6f at %2d/%d (observation %3d)\n" % (d[imax], Mt, My, imax + 1)
223
+ else:
224
+ o += " Minimum: %18.6f at %d (observation %3d)\n" % (d[imin], my, imin + 1)
225
+ o += " Maximum: %18.6f at %d (observation %3d)\n" % (d[imax], My, imax + 1)
226
+ return o + "\n", mu, tmp * tmp, imax, imin
227
+
228
+
229
+ def plot_ser(d, freq, by, bt, mu, var, imax, imin):
230
+ guions = list(GUIONS)
231
+ marcas = list(" 0 ")
232
+ r1, r2, r3, r4 = d[imax], d[imin], mu, math.sqrt(var)
233
+ if not (r4 > 0.0):
234
+ return "(residual series is constant; plot skipped)\n\n"
235
+ absmax = abs((r1 - r3) / r4)
236
+ if abs((r2 - r3) / r4) > absmax:
237
+ absmax = abs((r2 - r3) / r4)
238
+ if absmax <= 2.0:
239
+ absmax = 3.0
240
+ if not (absmax <= 8.0):
241
+ return "Warning: at least one observation above 8 sigmas\n"
242
+ hor = 25.0 / absmax
243
+ b1, b2 = hor, 2.0 * hor
244
+ for i in range(1, 9):
245
+ if absmax >= i:
246
+ c = str(i)[0]
247
+ k = round_local(i * hor)
248
+ guions[39 - k] = "+"
249
+ marcas[39 - k] = c
250
+ marcas[39 - k - 1] = "-"
251
+ guions[39 + k] = "+"
252
+ marcas[39 + k] = c
253
+ marcas[39 + k - 1] = "+"
254
+ G, Mk = "".join(guions), "".join(marcas)
255
+ o = "Standardized time series plot (original values on right-side column):\n\n"
256
+ o += "%s\n%s\n" % (Mk, G)
257
+ for i in range(1, len(d) + 1):
258
+ o += "%4d" % i
259
+ Ap, As = obs_to_date(by, bt, i, freq)
260
+ o += ("%7d " % Ap) if freq == 1 else ("%3d/%4d" % (As, Ap))
261
+ t = [" "] * 55
262
+ if freq != 1 and As == freq:
263
+ t[1] = t[53] = "+"
264
+ else:
265
+ t[1] = t[53] = "|"
266
+ z = (d[i - 1] - r3) / r4
267
+ if abs(z) >= 2.0:
268
+ t[0] = t[54] = ">"
269
+ pos = z * hor
270
+ t[27 + round_local(pos)] = "*"
271
+ if t[27] == " ":
272
+ t[27] = "|"
273
+ for k in (round_local(b1), -round_local(b1), round_local(b2), -round_local(b2)):
274
+ if t[27 + k] == " ":
275
+ t[27 + k] = ":"
276
+ o += "".join(t) + "%13.10f\n" % d[i - 1]
277
+ o += "%s\n%s\n\n" % (G, Mk)
278
+ o += (" +------------------------------------------+\n"
279
+ " | Table of standardized values |\n"
280
+ " | greater than or equal to 2.0 |\n"
281
+ " +------------------------------------------+\n"
282
+ " | |\n"
283
+ " | Observation Date Standardized value |\n"
284
+ " | |\n")
285
+ for i in range(1, len(d) + 1):
286
+ z = (d[i - 1] - r3) / r4
287
+ if abs(z) >= 2.0:
288
+ Ap, As = obs_to_date(by, bt, i, freq)
289
+ o += " |" + "%7d" % i
290
+ o += ("%13d " % Ap) if freq == 1 else ("%9d/%4d" % (As, Ap))
291
+ o += "%13.2f" % z + " |\n"
292
+ o += " +------------------------------------------+\n\n"
293
+ return o
294
+
295
+
296
+ def hist_ser(d, mu, var, imax, imin):
297
+ NumFil, NumCol = 17, 64
298
+ r1, r2 = mu, math.sqrt(var)
299
+ if not (r2 > 0.0):
300
+ return "(residual series is constant; histogram skipped)\n\n"
301
+ xmax = abs((d[imax] - r1) / r2)
302
+ if abs((d[imin] - r1) / r2) > xmax:
303
+ xmax = abs((d[imin] - r1) / r2)
304
+ if not (xmax <= 8.0):
305
+ return "Warning: at least one observation above 8 sigmas\n"
306
+ xmax = 4.0 if xmax <= 4.0 else 8.0
307
+ if xmax == 4.0:
308
+ nphor, no, yes = 4, " ", "...."
309
+ base1 = " +---+---+---+---+---+---+---+---+---+---+---+---+---+---+---+---+"
310
+ base2 = " -4 -3 -2 -1 0 +1 +2 +3 +4"
311
+ else:
312
+ nphor, no, yes = 2, " ", ".."
313
+ base1 = " +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+"
314
+ base2 = " -8 -7 -6 -5 -4 -3 -2 -1 0 +1 +2 +3 +4 +5 +6 +7 +8"
315
+ brk = [0.0, -xmax + 0.5]
316
+ while brk[-1] < xmax:
317
+ brk.append(brk[-1] + 0.5)
318
+ NumCat = len(brk) - 1
319
+ freqs = [0] * (NumCat + 1)
320
+ a1 = a2 = 0
321
+ for v in d:
322
+ num = (v - r1) / r2
323
+ if num <= brk[1]:
324
+ freqs[1] += 1
325
+ else:
326
+ for j in range(2, NumCat + 1):
327
+ if brk[j - 1] < num <= brk[j]:
328
+ freqs[j] += 1
329
+ break
330
+ if abs(num) >= 2.0:
331
+ a2 += 1
332
+ if abs(num) >= 1.0:
333
+ a1 += 1
334
+ fmax = max(freqs[1:NumCat + 1])
335
+ opf = fmax / 16.0
336
+ shist = [""] * NumFil
337
+ aux = [""] * NumFil
338
+ for j in range(2, NumFil + 1):
339
+ row = ""
340
+ for i in range(1, NumCat + 1):
341
+ row += yes if freqs[i] > opf * (NumFil - j) else no
342
+ shist[j - 1] += row + "|"
343
+ chk = [0] * (NumCat + 1)
344
+ for j in range(2, NumFil + 1):
345
+ for i in range(1, NumCat + 1):
346
+ if freqs[i] > opf * (NumFil - j) and chk[i] == 0:
347
+ s1 = "%d" % freqs[i]
348
+ if nphor == 2:
349
+ if len(s1) == 1:
350
+ s1 += " "
351
+ else:
352
+ if len(s1) == 2:
353
+ s1 = " " + s1
354
+ elif len(s1) == 1:
355
+ s1 = " " + s1
356
+ elif len(s1) == 3:
357
+ s1 += " "
358
+ if len(s1) > nphor:
359
+ s1 = "*" * nphor
360
+ aux[j - 2] += s1
361
+ chk[i] = 1
362
+ else:
363
+ aux[j - 2] += no
364
+ # strcpy(shist[0], aux[0]); shist[0][NumCol-1] = '|'; shist[0][NumCol] = 0:
365
+ # a 3-character label in a 4-character cell leaves aux[0] short, and then
366
+ # the '|' lands past the string's terminator and is not printed
367
+ a0 = aux[0]
368
+ if len(a0) >= NumCol - 1:
369
+ shist[0] = a0[:NumCol - 1] + "|"
370
+ else:
371
+ shist[0] = a0
372
+ for j in range(2, NumFil):
373
+ row = list(shist[j - 1])
374
+ for k, ch in enumerate(aux[j - 1]):
375
+ if ch != " ":
376
+ row[k] = ch
377
+ shist[j - 1] = "".join(row)
378
+ o = "Standardized time series histogram:\n\n%s\n%s\n" % (base2, base1)
379
+ for i in range(NumFil):
380
+ o += " |%s\n" % shist[i]
381
+ o += "%s\n%s\n\n" % (base1, base2)
382
+ n = len(d)
383
+ o += "%16d values outside (-1,+1): %5.2f %% (31.74 %% expected)\n" % (a1, (a1 * 100.0) / n)
384
+ o += "%16d values outside (-2,+2): %5.2f %% ( 4.56 %% expected)\n" % (a2, (a2 * 100.0) / n)
385
+ return o + "\n"
386
+
387
+
388
+ def diagnose(a, names, freq, start, offset, npar_ser):
389
+ """diagnose (diagnose.c:372). a: n x m residuals (lists); names: the m
390
+ labels; start: (year, sub) of the RAW sample; offset: where the residuals
391
+ start in it (drvec passes nobs_raw - nobs)."""
392
+ n, m = len(a), len(a[0])
393
+ o = "\n\n" + "=" * 61 + "\n RESIDUAL DIAGNOSTICS \n" + "=" * 61 + "\n"
394
+ by, bt = obs_to_date(start[0], start[1], offset + 1, freq)
395
+ for i in range(m):
396
+ d = [a[t][i] for t in range(n)]
397
+ o += "\n--- Residual series a[%d] (%s) ---\n" % (i + 1, names[i])
398
+ t, mu, var, imax, imin = stat_ser(d, freq, by, bt)
399
+ o += t
400
+ o += plot_ser(d, freq, by, bt, mu, var, imax, imin)
401
+ o += hist_ser(d, mu, var, imax, imin)
402
+ lags = (n - freq // 2) if n < 3 * (freq + 1) else 3 * (freq + 2)
403
+ if lags > n - 2:
404
+ lags = n - 2
405
+ corr = acf(d, lags, mu, var)
406
+ o += plot_cor(corr, lags, True, n, freq, npar_ser[i])
407
+ corr = pacf(corr, lags)
408
+ o += plot_cor(corr, lags, False, n, freq, npar_ser[i])
409
+ if m > 1:
410
+ o += "\n--- Cross-correlation functions between residuals ---\n"
411
+ lags = n if n < 6 else 9
412
+ if lags > n - 2:
413
+ lags = n - 2
414
+ for i in range(m - 1):
415
+ r1 = [a[t][i] for t in range(n)]
416
+ m1 = mean(r1)
417
+ sd1 = stdev(r1)
418
+ for j in range(i + 1, m):
419
+ r2 = [a[t][j] for t in range(n)]
420
+ m2 = mean(r2)
421
+ sd2 = stdev(r2)
422
+ o += "\nCROSS CORRELATION a[%d] (%s) - a[%d] (%s)\n" % (j + 1, names[j], i + 1, names[i])
423
+ o += " %s --> %s IF k > 0\n" % (names[j], names[i])
424
+ o += " %s --> %s IF k < 0\n" % (names[i], names[j])
425
+ c1 = ccf(r1, r2, lags, m1, m2, sd1, sd2)
426
+ c2 = ccf(r2, r1, lags, m2, m1, sd2, sd1)
427
+ tot = [0.0] * (2 * lags + 2)
428
+ for h in range(1, lags + 2):
429
+ tot[h] = c1[lags + 2 - h]
430
+ for h in range(2, lags + 2):
431
+ tot[lags + h] = c2[h]
432
+ o += plot_ccf(tot, lags, n)
433
+ q2 = chitest_c(c2, lags + 1, n)
434
+ q1 = chitest_c(c1, lags + 1, n)
435
+ o += "Q(%2d) = %6.3f k >= 0 Q(%2d) = %6.3f k > 0\n" % (
436
+ lags + 1, q2, lags, q2 - (c2[1] * c2[1] * (n + 2)))
437
+ o += "Q(%2d) = %6.3f k <= 0 Q(%2d) = %6.3f k < 0\n" % (
438
+ lags + 1, q1, lags, q1 - (c1[1] * c1[1] * (n + 2)))
439
+ return o + "\n" + "=" * 61 + "\n"
drvec/fdhess.py ADDED
@@ -0,0 +1,130 @@
1
+ """-fdhess: standard errors from the finite-difference Hessian AT THE OPTIMUM.
2
+
3
+ C: `exact_hessian_se` (drvec.c:1910-2007) with its objective `fdh_obj` (1877),
4
+ and `fdhess` (qnewtopt.c:404), which is drvarma's `_qnewt.fdhess` line for line.
5
+
6
+ The BFGS covariance (stderr.py) is read off the optimiser's last Hessian
7
+ update and so depends on the trajectory; this one depends only on the point.
8
+ The objective is the concentrated likelihood normalised AT THE OPTIMUM,
9
+
10
+ g(x) = (f1(x) / f1(x*))^m (f2(x) / f2(x*)), g(x*) = 1,
11
+
12
+ the Hessian of g by Dennis and Schnabel's A5.6.2 with eta = macheps, and
13
+
14
+ cov = 2 g(x*) H^{-1} / n (the same formula as est's)
15
+
16
+ A point the cast or AS 311 refuses counts as 1e10 and as a rejection. Any
17
+ rejection means the optimum is ON the boundary of the admissible region (an MA
18
+ root on the unit circle, BUG-49): the BFGS standard errors are kept and the
19
+ .out says why (BUG-34). So does a Hessian that fails choldcp with every
20
+ evaluation admissible (the optimiser did not stop at a maximum).
21
+ """
22
+ from __future__ import annotations
23
+
24
+ import math
25
+ from dataclasses import dataclass
26
+
27
+ import numpy as np
28
+ from drvarma import _qnewt
29
+
30
+ from .cast import shootx
31
+ from .loglik import XITOL
32
+ from .nla import choldcp
33
+
34
+ REFUSED = 1.0e10
35
+
36
+
37
+ @dataclass
38
+ class FdHess:
39
+ ok: bool # the s.e. are the Hessian's
40
+ nrej: int # finite-difference evaluations refused (boundary)
41
+ nneg: int = 0 # non-positive eigenvalues of H (when not ok, nrej == 0)
42
+ ratio: float = 0.0 # -min/max eigenvalue
43
+ dev: np.ndarray = None
44
+ cov: np.ndarray = None
45
+
46
+ @property
47
+ def why(self):
48
+ """The reason written to the .out when the s.e. could not be used."""
49
+ return ("the optimum is on the boundary (finite-difference steps left the "
50
+ "admissible region)" if self.nrej > 0 else
51
+ "the Hessian at the optimum is not positive definite")
52
+
53
+ def warning(self, npar):
54
+ """The C's stderr warning (drvec.c:1975-1991)."""
55
+ if self.ok:
56
+ return None
57
+ if self.nrej > 0:
58
+ return ("WARNING: -fdhess: the optimum lies ON the boundary of the admissible\n"
59
+ " region -- %d of the finite-difference evaluations fell\n"
60
+ " outside it -- so the unconstrained Hessian is not defined\n"
61
+ " there and the BFGS standard errors are kept. See the roots\n"
62
+ " of the AR and MA operators reported above: a modulus at one\n"
63
+ " identifies the binding direction.\n" % self.nrej)
64
+ return ("WARNING: -fdhess: the Hessian at the reported optimum is not positive\n"
65
+ " definite: %d of %d eigenvalues are non-positive, the most\n"
66
+ " negative being %.3g times the largest positive one. Every\n"
67
+ " evaluation was admissible, so the optimiser did not stop at\n"
68
+ " a maximum. The BFGS standard errors are kept.\n"
69
+ % (self.nneg, npar, self.ratio))
70
+
71
+
72
+ def _f12(spec, rd, x, xitol):
73
+ mu, phi, theta, qq, w, ifc = shootx(spec, rd, x)
74
+ if ifc > 0:
75
+ return None
76
+ from drvarma._engine import elf_c
77
+ _lg, f1, f2, _a, ifa = elf_c(spec.M, rd.n, spec.p, spec.q, mu, phi, theta, qq, w,
78
+ 1.0, xitol, False)
79
+ if ifa > 0:
80
+ return None
81
+ return f1, f2
82
+
83
+
84
+ def exact_hessian_se(spec, rd, x, xitol=XITOL):
85
+ """-> FdHess. `rd` is the estimation sample (n = rd.n is the C's nobs)."""
86
+ npar = len(x)
87
+ m = spec.M
88
+ base = _f12(spec, rd, list(x), xitol)
89
+ if base is None:
90
+ return FdHess(ok=False, nrej=0)
91
+ n1, n2 = base
92
+ nrej = [0]
93
+
94
+ def obj(xk1):
95
+ v = _f12(spec, rd, xk1[1:].tolist(), xitol)
96
+ if v is None:
97
+ nrej[0] += 1
98
+ return REFUSED
99
+ return pow(v[0] / n1, m) * (v[1] / n2)
100
+
101
+ xk = np.zeros(npar + 1)
102
+ xk[1:] = x
103
+ f = obj(xk) # = 1 by construction
104
+ nrej[0] = 0
105
+ H = np.zeros((npar + 1, npar + 1))
106
+ _qnewt.fdhess(obj, npar, xk, f, _qnewt.MACHEPS, H)
107
+ Hs = H[1:, 1:].copy()
108
+ ev = np.linalg.eigvals(Hs).real # eigenqr, for the warning only
109
+ mx = max(0.0, float(ev.max())) if npar else 0.0
110
+ mn = min(0.0, float(ev.min())) if npar else 0.0
111
+ nneg = int((ev <= 0.0).sum())
112
+ ratio = -mn / mx if mx > 0.0 else 0.0
113
+ if nrej[0] > 0:
114
+ return FdHess(ok=False, nrej=nrej[0], nneg=nneg, ratio=ratio)
115
+ L, ifc = choldcp([[float(v) for v in row] for row in Hs])
116
+ if ifc > 0:
117
+ return FdHess(ok=False, nrej=0, nneg=nneg, ratio=ratio)
118
+ Lk = np.zeros((npar + 1, npar + 1))
119
+ Lk[1:, 1:] = np.array(L)
120
+ cov = np.zeros((npar, npar))
121
+ dev = np.zeros(npar)
122
+ for i in range(1, npar + 1):
123
+ e = np.zeros(npar + 1)
124
+ e[i] = 1.0
125
+ _qnewt.cholsol(Lk, npar, e)
126
+ for j in range(1, npar + 1):
127
+ cov[j - 1, i - 1] = (2.0 * f * e[j]) / rd.n
128
+ c = cov[i - 1, i - 1]
129
+ dev[i - 1] = math.sqrt(c if c > 0.0 else 0.0)
130
+ return FdHess(ok=True, nrej=0, nneg=nneg, ratio=ratio, dev=dev, cov=cov)