[357] | 1 | #include <math.h> |
---|
[262] | 2 | |
---|
[32] | 3 | #include <itpp/base/bessel.h> |
---|
[384] | 4 | #include "exp_family.h" |
---|
[13] | 5 | |
---|
[477] | 6 | namespace bdm { |
---|
[13] | 7 | |
---|
[32] | 8 | Uniform_RNG UniRNG; |
---|
| 9 | Normal_RNG NorRNG; |
---|
| 10 | Gamma_RNG GamRNG; |
---|
| 11 | |
---|
[13] | 12 | using std::cout; |
---|
| 13 | |
---|
[737] | 14 | /////////// |
---|
[487] | 15 | |
---|
[737] | 16 | void BMEF::bayes ( const vec &yt, const vec &cond ) { |
---|
[679] | 17 | this->bayes_weighted ( yt, cond, 1.0 ); |
---|
[477] | 18 | }; |
---|
[170] | 19 | |
---|
[737] | 20 | void egiw::set_parameters ( int dimx0, ldmat V0, double nu0 ) { |
---|
[629] | 21 | dimx = dimx0; |
---|
| 22 | nPsi = V0.rows() - dimx; |
---|
[737] | 23 | |
---|
[629] | 24 | V = V0; |
---|
[737] | 25 | if ( nu0 < 0 ) { |
---|
[629] | 26 | nu = 0.1 + nPsi + 2 * dimx + 2; // +2 assures finite expected value of R |
---|
| 27 | // terms before that are sufficient for finite normalization |
---|
| 28 | } else { |
---|
| 29 | nu = nu0; |
---|
| 30 | } |
---|
| 31 | } |
---|
| 32 | |
---|
[96] | 33 | vec egiw::sample() const { |
---|
[725] | 34 | mat M; |
---|
| 35 | chmat R; |
---|
[737] | 36 | sample_mat ( M, R ); |
---|
| 37 | |
---|
| 38 | return concat ( cvectorize ( M ), cvectorize ( R.to_mat() ) ); |
---|
[96] | 39 | } |
---|
| 40 | |
---|
[737] | 41 | mat egiw::sample_mat ( int n ) const { |
---|
[730] | 42 | // TODO - correct approach - convert to product of norm * Wishart |
---|
| 43 | mat M; |
---|
| 44 | ldmat Vz; |
---|
| 45 | ldmat Lam; |
---|
[737] | 46 | factorize ( M, Vz, Lam ); |
---|
| 47 | |
---|
| 48 | chmat ChLam ( Lam.to_mat() ); |
---|
[730] | 49 | chmat iChLam; |
---|
[737] | 50 | ChLam.inv ( iChLam ); |
---|
| 51 | |
---|
| 52 | eWishartCh Omega; //inverse Wishart, result is R, |
---|
| 53 | Omega.set_parameters ( iChLam, nu - 2*nPsi - dimx ); // 2*nPsi is there to match numercial simulations - check if analytically correct |
---|
[878] | 54 | Omega.validate(); |
---|
[737] | 55 | |
---|
[730] | 56 | mat OmChi; |
---|
[737] | 57 | mat Z ( M.rows(), M.cols() ); |
---|
[730] | 58 | |
---|
| 59 | mat Mi; |
---|
| 60 | mat RChiT; |
---|
[737] | 61 | mat tmp ( dimension(), n ); |
---|
[847] | 62 | M=M.T();// ugly hack == decide what to do with M. |
---|
[737] | 63 | for ( int i = 0; i < n; i++ ) { |
---|
| 64 | OmChi = Omega.sample_mat(); |
---|
| 65 | RChiT = inv ( OmChi ); |
---|
| 66 | Z = randn ( M.rows(), M.cols() ); |
---|
| 67 | Mi = M + RChiT * Z * inv ( Vz._L().T() * diag ( sqrt ( Vz._D() ) ) ); |
---|
| 68 | |
---|
| 69 | tmp.set_col ( i, concat ( cvectorize ( Mi ), cvectorize ( RChiT*RChiT.T() ) ) ); |
---|
[730] | 70 | } |
---|
| 71 | return tmp; |
---|
| 72 | } |
---|
| 73 | |
---|
[737] | 74 | void egiw::sample_mat ( mat &Mi, chmat &Ri ) const { |
---|
| 75 | |
---|
[725] | 76 | // TODO - correct approach - convert to product of norm * Wishart |
---|
| 77 | mat M; |
---|
| 78 | ldmat Vz; |
---|
| 79 | ldmat Lam; |
---|
[737] | 80 | factorize ( M, Vz, Lam ); |
---|
| 81 | |
---|
[725] | 82 | chmat Ch; |
---|
[737] | 83 | Ch.setCh ( Lam._L() *diag ( sqrt ( Lam._D() ) ) ); |
---|
[725] | 84 | chmat iCh; |
---|
[737] | 85 | Ch.inv ( iCh ); |
---|
| 86 | |
---|
| 87 | eWishartCh Omega; //inverse Wishart, result is R, |
---|
| 88 | Omega.set_parameters ( iCh, nu - 2*nPsi - dimx ); // 2*nPsi is there to match numercial simulations - check if analytically correct |
---|
[878] | 89 | Omega.validate(); |
---|
[737] | 90 | |
---|
[725] | 91 | chmat Omi; |
---|
[737] | 92 | Omi.setCh ( Omega.sample_mat() ); |
---|
| 93 | |
---|
[810] | 94 | if (M._datasize()>0){ |
---|
| 95 | mat Z = randn ( M.rows(), M.cols() ); |
---|
| 96 | Mi = M + Omi._Ch() * Z * inv ( Vz._L() * diag ( sqrt ( Vz._D() ) ) ); |
---|
| 97 | } |
---|
[737] | 98 | Omi.inv ( Ri ); |
---|
[725] | 99 | } |
---|
| 100 | |
---|
[211] | 101 | double egiw::evallog_nn ( const vec &val ) const { |
---|
[810] | 102 | bdm_assert_debug(val.length()==dimx*(nPsi+dimx),"Incorrect cond in egiw::evallog_nn" ); |
---|
[761] | 103 | |
---|
[477] | 104 | int vend = val.length() - 1; |
---|
[168] | 105 | |
---|
[477] | 106 | if ( dimx == 1 ) { //same as the following, just quicker. |
---|
[170] | 107 | double r = val ( vend ); //last entry! |
---|
[477] | 108 | if ( r < 0 ) return -inf; |
---|
| 109 | vec Psi ( nPsi + dimx ); |
---|
[168] | 110 | Psi ( 0 ) = -1.0; |
---|
[477] | 111 | Psi.set_subvector ( 1, val ( 0, vend - 1 ) ); // fill the rest |
---|
[168] | 112 | |
---|
[477] | 113 | double Vq = V.qform ( Psi ); |
---|
| 114 | return -0.5* ( nu*log ( r ) + Vq / r ); |
---|
| 115 | } else { |
---|
[810] | 116 | mat Tmp; |
---|
| 117 | if (nPsi>0){ |
---|
| 118 | mat Th = reshape ( val ( 0, nPsi * dimx - 1 ), nPsi, dimx ); |
---|
| 119 | Tmp = concat_vertical ( -eye ( dimx ), Th ); |
---|
| 120 | } else { |
---|
| 121 | Tmp = -eye(dimx); |
---|
| 122 | } |
---|
[477] | 123 | fsqmat R ( reshape ( val ( nPsi*dimx, vend ), dimx, dimx ) ); |
---|
| 124 | double ldetR = R.logdet(); |
---|
[826] | 125 | if ( !std::isfinite(ldetR) ) return -inf; |
---|
[270] | 126 | fsqmat iR ( dimx ); |
---|
[168] | 127 | R.inv ( iR ); |
---|
[170] | 128 | |
---|
[395] | 129 | return -0.5* ( nu*ldetR + trace ( iR.to_mat() *Tmp.T() *V.to_mat() *Tmp ) ); |
---|
[168] | 130 | } |
---|
[96] | 131 | } |
---|
| 132 | |
---|
[168] | 133 | double egiw::lognc() const { |
---|
[96] | 134 | const vec& D = V._D(); |
---|
| 135 | |
---|
[477] | 136 | double m = nu - nPsi - dimx - 1; |
---|
[168] | 137 | #define log2 0.693147180559945286226763983 |
---|
| 138 | #define logpi 1.144729885849400163877476189 |
---|
| 139 | #define log2pi 1.83787706640935 |
---|
[178] | 140 | #define Inf std::numeric_limits<double>::infinity() |
---|
[168] | 141 | |
---|
[810] | 142 | double nkG = 0.5 * dimx * ( -nPsi * log2pi + sum ( log ( D.mid ( dimx, nPsi ) ) ) ); |
---|
[168] | 143 | // temporary for lgamma in Wishart |
---|
[477] | 144 | double lg = 0; |
---|
| 145 | for ( int i = 0; i < dimx; i++ ) { |
---|
| 146 | lg += lgamma ( 0.5 * ( m - i ) ); |
---|
| 147 | } |
---|
[168] | 148 | |
---|
[477] | 149 | double nkW = 0.5 * ( m * sum ( log ( D ( 0, dimx - 1 ) ) ) ) \ |
---|
| 150 | - 0.5 * dimx * ( m * log2 + 0.5 * ( dimx - 1 ) * log2pi ) - lg; |
---|
[168] | 151 | |
---|
[637] | 152 | // bdm_assert_debug ( ( ( -nkG - nkW ) > -Inf ) && ( ( -nkG - nkW ) < Inf ), "ARX improper" ); |
---|
[737] | 153 | if ( -nkG - nkW == Inf ) { |
---|
| 154 | cout << "??" << endl; |
---|
[665] | 155 | } |
---|
[477] | 156 | return -nkG - nkW; |
---|
[96] | 157 | } |
---|
| 158 | |
---|
[330] | 159 | vec egiw::est_theta() const { |
---|
[477] | 160 | if ( dimx == 1 ) { |
---|
[330] | 161 | const mat &L = V._L(); |
---|
| 162 | int end = L.rows() - 1; |
---|
| 163 | |
---|
[477] | 164 | mat iLsub = ltuinv ( L ( dimx, end, dimx, end ) ); |
---|
[330] | 165 | |
---|
[477] | 166 | vec L0 = L.get_col ( 0 ); |
---|
[330] | 167 | |
---|
[477] | 168 | return iLsub * L0 ( 1, end ); |
---|
| 169 | } else { |
---|
[565] | 170 | bdm_error ( "ERROR: est_theta() not implemented for dimx>1" ); |
---|
| 171 | return vec(); |
---|
[330] | 172 | } |
---|
| 173 | } |
---|
| 174 | |
---|
[737] | 175 | void egiw::factorize ( mat &M, ldmat &Vz, ldmat &Lam ) const { |
---|
[725] | 176 | const mat &L = V._L(); |
---|
| 177 | const vec &D = V._D(); |
---|
| 178 | int end = L.rows() - 1; |
---|
[737] | 179 | |
---|
[810] | 180 | Lam = ldmat ( L ( 0, dimx - 1, 0, dimx - 1 ), D ( 0, dimx - 1 ) ); //exp val of R |
---|
[737] | 181 | |
---|
[810] | 182 | if (dimx<=end){ |
---|
| 183 | Vz = ldmat ( L ( dimx, end, dimx, end ), D ( dimx, end ) ); |
---|
| 184 | mat iLsub = ltuinv ( Vz._L() ); |
---|
| 185 | // set mean value |
---|
| 186 | mat Lpsi = L ( dimx, end, 0, dimx - 1 ); |
---|
| 187 | M = iLsub * Lpsi; |
---|
| 188 | } |
---|
| 189 | /* if ( 0 ) { // test with Peterka |
---|
[737] | 190 | mat VF = V.to_mat(); |
---|
| 191 | mat Vf = VF ( 0, dimx - 1, 0, dimx - 1 ); |
---|
| 192 | mat Vzf = VF ( dimx, end, 0, dimx - 1 ); |
---|
| 193 | mat VZ = VF ( dimx, end, dimx, end ); |
---|
| 194 | |
---|
| 195 | mat Lam2 = Vf - Vzf.T() * inv ( VZ ) * Vzf; |
---|
[810] | 196 | }*/ |
---|
[725] | 197 | } |
---|
| 198 | |
---|
[330] | 199 | ldmat egiw::est_theta_cov() const { |
---|
| 200 | if ( dimx == 1 ) { |
---|
| 201 | const mat &L = V._L(); |
---|
| 202 | const vec &D = V._D(); |
---|
| 203 | int end = D.length() - 1; |
---|
| 204 | |
---|
[477] | 205 | mat Lsub = L ( 1, end, 1, end ); |
---|
[725] | 206 | // mat Dsub = diag ( D ( 1, end ) ); |
---|
[330] | 207 | |
---|
[737] | 208 | ldmat LD ( inv ( Lsub ).T(), 1.0 / D ( 1, end ) ); |
---|
[725] | 209 | return LD; |
---|
[330] | 210 | |
---|
[477] | 211 | } else { |
---|
[565] | 212 | bdm_error ( "ERROR: est_theta_cov() not implemented for dimx>1" ); |
---|
| 213 | return ldmat(); |
---|
[330] | 214 | } |
---|
| 215 | |
---|
| 216 | } |
---|
| 217 | |
---|
[96] | 218 | vec egiw::mean() const { |
---|
[168] | 219 | |
---|
[477] | 220 | if ( dimx == 1 ) { |
---|
| 221 | const vec &D = V._D(); |
---|
| 222 | int end = D.length() - 1; |
---|
[170] | 223 | |
---|
[270] | 224 | vec m ( dim ); |
---|
[330] | 225 | m.set_subvector ( 0, est_theta() ); |
---|
[477] | 226 | m ( end ) = D ( 0 ) / ( nu - nPsi - 2 * dimx - 2 ); |
---|
[170] | 227 | return m; |
---|
[477] | 228 | } else { |
---|
[170] | 229 | mat M; |
---|
| 230 | mat R; |
---|
[477] | 231 | mean_mat ( M, R ); |
---|
[737] | 232 | return concat ( cvectorize ( M ), cvectorize ( R ) ); |
---|
[168] | 233 | } |
---|
[170] | 234 | |
---|
| 235 | } |
---|
[262] | 236 | |
---|
| 237 | vec egiw::variance() const { |
---|
[629] | 238 | int l = V.rows(); |
---|
| 239 | // cut out rest of lower-right part of V |
---|
| 240 | // invert it |
---|
[977] | 241 | ldmat itmp; |
---|
[810] | 242 | if (dimx<l){ |
---|
| 243 | const ldmat tmp ( V, linspace ( dimx, l - 1 ) ); |
---|
| 244 | tmp.inv ( itmp ); |
---|
[977] | 245 | } |
---|
[629] | 246 | // following Wikipedia notation |
---|
[737] | 247 | // m=nu-nPsi-dimx-1, p=dimx |
---|
| 248 | double mp1p = nu - nPsi - 2 * dimx; // m-p+1 |
---|
| 249 | double mp1m = mp1p - 2; // m-p-1 |
---|
[977] | 250 | |
---|
[477] | 251 | if ( dimx == 1 ) { |
---|
[629] | 252 | double cove = V._D() ( 0 ) / mp1m ; |
---|
[977] | 253 | |
---|
[477] | 254 | vec var ( l ); |
---|
| 255 | var.set_subvector ( 0, diag ( itmp.to_mat() ) *cove ); |
---|
[737] | 256 | var ( l - 1 ) = cove * cove / ( mp1m - 2 ); |
---|
[262] | 257 | return var; |
---|
[477] | 258 | } else { |
---|
[737] | 259 | ldmat Vll ( V, linspace ( 0, dimx - 1 ) ); // top-left part of V |
---|
| 260 | mat Y = Vll.to_mat(); |
---|
| 261 | mat varY ( Y.rows(), Y.cols() ); |
---|
[977] | 262 | |
---|
[737] | 263 | double denom = ( mp1p - 1 ) * mp1m * mp1m * ( mp1m - 2 ); // (m-p)(m-p-1)^2(m-p-3) |
---|
[977] | 264 | |
---|
[737] | 265 | int i, j; |
---|
| 266 | for ( i = 0; i < Y.rows(); i++ ) { |
---|
| 267 | for ( j = 0; j < Y.cols(); j++ ) { |
---|
| 268 | varY ( i, j ) = ( mp1p * Y ( i, j ) * Y ( i, j ) + mp1m * Y ( i, i ) * Y ( j, j ) ) / denom; |
---|
[629] | 269 | } |
---|
| 270 | } |
---|
[737] | 271 | vec mean_dR = diag ( Y ) / mp1m; // corresponds to cove |
---|
| 272 | vec var_th = diag ( itmp.to_mat() ); |
---|
| 273 | vec var_Th ( mean_dR.length() *var_th.length() ); |
---|
[629] | 274 | // diagonal of diag(mean_dR) \kron diag(var_th) |
---|
[737] | 275 | for ( int i = 0; i < mean_dR.length(); i++ ) { |
---|
| 276 | var_Th.set_subvector ( i*var_th.length(), var_th*mean_dR ( i ) ); |
---|
[629] | 277 | } |
---|
[977] | 278 | |
---|
[737] | 279 | return concat ( var_Th, cvectorize ( varY ) ); |
---|
[262] | 280 | } |
---|
| 281 | } |
---|
| 282 | |
---|
[225] | 283 | void egiw::mean_mat ( mat &M, mat&R ) const { |
---|
[477] | 284 | const mat &L = V._L(); |
---|
| 285 | const vec &D = V._D(); |
---|
| 286 | int end = L.rows() - 1; |
---|
[225] | 287 | |
---|
[477] | 288 | ldmat ldR ( L ( 0, dimx - 1, 0, dimx - 1 ), D ( 0, dimx - 1 ) / ( nu - nPsi - 2*dimx - 2 ) ); //exp val of R |
---|
[225] | 289 | |
---|
[170] | 290 | // set mean value |
---|
[810] | 291 | if (dimx<=end){ |
---|
| 292 | mat iLsub = ltuinv ( L ( dimx, end, dimx, end ) ); |
---|
| 293 | mat Lpsi = L ( dimx, end, 0, dimx - 1 ); |
---|
| 294 | M = iLsub * Lpsi; |
---|
| 295 | } |
---|
[477] | 296 | R = ldR.to_mat() ; |
---|
[96] | 297 | } |
---|
| 298 | |
---|
[739] | 299 | void egiw::log_register ( bdm::logger& L, const string& prefix ) { |
---|
[970] | 300 | epdf::log_register ( L, prefix ); |
---|
| 301 | if ( log_level[logvartheta] ) { |
---|
| 302 | int th_dim = dim - dimx*dimx; // dimension - dimension of cov |
---|
| 303 | L.add_vector( log_level, logvartheta, RV ( th_dim ), prefix ); |
---|
[739] | 304 | } |
---|
| 305 | } |
---|
| 306 | |
---|
| 307 | void egiw::log_write() const { |
---|
[970] | 308 | epdf::log_write(); |
---|
| 309 | if ( log_level[logvartheta] ) { |
---|
[739] | 310 | mat M; |
---|
| 311 | ldmat Lam; |
---|
| 312 | ldmat Vz; |
---|
| 313 | factorize ( M, Vz, Lam ); |
---|
[970] | 314 | if( log_level[logvartheta] ) |
---|
| 315 | log_level.store( logvartheta, cvectorize ( est_theta_cov().to_mat() ) ); |
---|
[739] | 316 | } |
---|
| 317 | } |
---|
| 318 | |
---|
[956] | 319 | void egiw::from_setting ( const Setting &set ) { |
---|
| 320 | epdf::from_setting ( set ); |
---|
| 321 | UI::get ( dimx, set, "dimx", UI::compulsory ); |
---|
| 322 | if ( !UI::get ( nu, set, "nu", UI::optional ) ) { |
---|
| 323 | nu = -1; |
---|
| 324 | } |
---|
| 325 | mat V; |
---|
| 326 | if ( !UI::get ( V, set, "V", UI::optional ) ) { |
---|
| 327 | vec dV; |
---|
| 328 | UI::get ( dV, set, "dV", UI::compulsory ); |
---|
| 329 | set_parameters ( dimx, ldmat ( dV ), nu ); |
---|
| 330 | |
---|
| 331 | } else { |
---|
| 332 | set_parameters ( dimx, V, nu ); |
---|
| 333 | } |
---|
| 334 | } |
---|
| 335 | |
---|
| 336 | void egiw::to_setting ( Setting& set ) const { |
---|
| 337 | epdf::to_setting ( set ); |
---|
| 338 | UI::save ( dimx, set, "dimx" ); |
---|
| 339 | UI::save ( V.to_mat(), set, "V" ); |
---|
| 340 | UI::save ( nu, set, "nu" ); |
---|
| 341 | }; |
---|
| 342 | |
---|
| 343 | void egiw::validate() { |
---|
| 344 | eEF::validate(); |
---|
| 345 | dim = dimx * ( dimx + nPsi ); |
---|
| 346 | |
---|
| 347 | // check sizes, rvs etc. |
---|
| 348 | // also check if RV are meaningful!!! |
---|
| 349 | // meaningful = rv for theta and rv for r are split! |
---|
| 350 | } |
---|
[739] | 351 | void multiBM::bayes ( const vec &yt, const vec &cond ) { |
---|
| 352 | if ( frg < 1.0 ) { |
---|
| 353 | beta *= frg; |
---|
| 354 | last_lognc = est.lognc(); |
---|
| 355 | } |
---|
| 356 | beta += yt; |
---|
| 357 | if ( evalll ) { |
---|
| 358 | ll = est.lognc() - last_lognc; |
---|
| 359 | } |
---|
| 360 | } |
---|
| 361 | |
---|
| 362 | double multiBM::logpred ( const vec &yt ) const { |
---|
| 363 | eDirich pred ( est ); |
---|
| 364 | vec &beta = pred._beta(); |
---|
| 365 | |
---|
| 366 | double lll; |
---|
| 367 | if ( frg < 1.0 ) { |
---|
| 368 | beta *= frg; |
---|
| 369 | lll = pred.lognc(); |
---|
| 370 | } else if ( evalll ) { |
---|
| 371 | lll = last_lognc; |
---|
| 372 | } else { |
---|
| 373 | lll = pred.lognc(); |
---|
| 374 | } |
---|
| 375 | |
---|
| 376 | beta += yt; |
---|
| 377 | return pred.lognc() - lll; |
---|
| 378 | } |
---|
| 379 | void multiBM::flatten ( const BMEF* B ) { |
---|
| 380 | const multiBM* E = dynamic_cast<const multiBM*> ( B ); |
---|
| 381 | // sum(beta) should be equal to sum(B.beta) |
---|
| 382 | const vec &Eb = E->beta;//const_cast<multiBM*> ( E )->_beta(); |
---|
| 383 | beta *= ( sum ( Eb ) / sum ( beta ) ); |
---|
| 384 | if ( evalll ) { |
---|
| 385 | last_lognc = est.lognc(); |
---|
| 386 | } |
---|
| 387 | } |
---|
| 388 | |
---|
[32] | 389 | vec egamma::sample() const { |
---|
[477] | 390 | vec smp ( dim ); |
---|
[32] | 391 | int i; |
---|
| 392 | |
---|
[477] | 393 | for ( i = 0; i < dim; i++ ) { |
---|
| 394 | if ( beta ( i ) > std::numeric_limits<double>::epsilon() ) { |
---|
| 395 | GamRNG.setup ( alpha ( i ), beta ( i ) ); |
---|
| 396 | } else { |
---|
| 397 | GamRNG.setup ( alpha ( i ), std::numeric_limits<double>::epsilon() ); |
---|
[225] | 398 | } |
---|
[168] | 399 | #pragma omp critical |
---|
[32] | 400 | smp ( i ) = GamRNG(); |
---|
| 401 | } |
---|
| 402 | |
---|
| 403 | return smp; |
---|
| 404 | } |
---|
| 405 | |
---|
[102] | 406 | // mat egamma::sample ( int N ) const { |
---|
| 407 | // mat Smp ( rv.count(),N ); |
---|
| 408 | // int i,j; |
---|
[168] | 409 | // |
---|
[102] | 410 | // for ( i=0; i<rv.count(); i++ ) { |
---|
| 411 | // GamRNG.setup ( alpha ( i ),beta ( i ) ); |
---|
[168] | 412 | // |
---|
[102] | 413 | // for ( j=0; j<N; j++ ) { |
---|
| 414 | // Smp ( i,j ) = GamRNG(); |
---|
| 415 | // } |
---|
| 416 | // } |
---|
[168] | 417 | // |
---|
[102] | 418 | // return Smp; |
---|
| 419 | // } |
---|
[32] | 420 | |
---|
[211] | 421 | double egamma::evallog ( const vec &val ) const { |
---|
[96] | 422 | double res = 0.0; //the rest will be added |
---|
| 423 | int i; |
---|
| 424 | |
---|
[477] | 425 | if ( any ( val <= 0. ) ) return -inf; |
---|
| 426 | if ( any ( beta <= 0. ) ) return -inf; |
---|
| 427 | for ( i = 0; i < dim; i++ ) { |
---|
| 428 | res += ( alpha ( i ) - 1 ) * std::log ( val ( i ) ) - beta ( i ) * val ( i ); |
---|
[96] | 429 | } |
---|
[477] | 430 | double tmp = res - lognc();; |
---|
[565] | 431 | bdm_assert_debug ( std::isfinite ( tmp ), "Infinite value" ); |
---|
[214] | 432 | return tmp; |
---|
[96] | 433 | } |
---|
| 434 | |
---|
| 435 | double egamma::lognc() const { |
---|
[32] | 436 | double res = 0.0; //will be added |
---|
| 437 | int i; |
---|
| 438 | |
---|
[477] | 439 | for ( i = 0; i < dim; i++ ) { |
---|
| 440 | res += lgamma ( alpha ( i ) ) - alpha ( i ) * std::log ( beta ( i ) ) ; |
---|
[32] | 441 | } |
---|
| 442 | |
---|
| 443 | return res; |
---|
| 444 | } |
---|
| 445 | |
---|
[956] | 446 | void egamma::from_setting ( const Setting &set ) { |
---|
| 447 | epdf::from_setting ( set ); // reads rv |
---|
| 448 | UI::get ( alpha, set, "alpha", UI::compulsory ); |
---|
| 449 | UI::get ( beta, set, "beta", UI::compulsory ); |
---|
| 450 | } |
---|
| 451 | |
---|
| 452 | void egamma::to_setting ( Setting &set ) const |
---|
| 453 | { |
---|
| 454 | epdf::to_setting( set ); |
---|
| 455 | UI::save( alpha, set, "alpha" ); |
---|
| 456 | UI::save( beta, set, "beta" ); |
---|
| 457 | } |
---|
| 458 | |
---|
| 459 | |
---|
| 460 | void egamma::validate() { |
---|
| 461 | eEF::validate(); |
---|
| 462 | bdm_assert ( alpha.length() == beta.length(), "parameters do not match" ); |
---|
| 463 | dim = alpha.length(); |
---|
| 464 | } |
---|
| 465 | |
---|
[477] | 466 | void mgamma::set_parameters ( double k0, const vec &beta0 ) { |
---|
[461] | 467 | k = k0; |
---|
[487] | 468 | iepdf.set_parameters ( k * ones ( beta0.length() ), beta0 ); |
---|
[461] | 469 | } |
---|
[32] | 470 | |
---|
[956] | 471 | |
---|
| 472 | void mgamma::from_setting ( const Setting &set ) { |
---|
| 473 | pdf::from_setting ( set ); // reads rv and rvc |
---|
| 474 | vec betatmp; // ugly but necessary |
---|
| 475 | UI::get ( betatmp, set, "beta", UI::compulsory ); |
---|
| 476 | UI::get ( k, set, "k", UI::compulsory ); |
---|
| 477 | set_parameters ( k, betatmp ); |
---|
| 478 | } |
---|
| 479 | |
---|
| 480 | void mgamma::to_setting (Setting &set) const { |
---|
| 481 | pdf::to_setting(set); |
---|
| 482 | UI::save( _beta, set, "beta"); |
---|
| 483 | UI::save( k, set, "k"); |
---|
| 484 | |
---|
| 485 | } |
---|
| 486 | |
---|
| 487 | void mgamma::validate() { |
---|
| 488 | pdf_internal<egamma>::validate(); |
---|
| 489 | |
---|
| 490 | dim = _beta.length(); |
---|
| 491 | dimc = _beta.length(); |
---|
| 492 | } |
---|
| 493 | |
---|
[887] | 494 | void eEmp::resample ( RESAMPLING_METHOD method ) { |
---|
| 495 | ivec ind = zeros_i ( n ); |
---|
| 496 | bdm::resample(w,ind,method); |
---|
| 497 | // copy the internals according to ind |
---|
| 498 | for (int i = 0; i < n; i++ ) { |
---|
| 499 | if ( ind ( i ) != i ) { |
---|
| 500 | samples ( i ) = samples ( ind ( i ) ); |
---|
| 501 | } |
---|
| 502 | w ( i ) = 1.0 / n; |
---|
| 503 | } |
---|
| 504 | } |
---|
| 505 | |
---|
| 506 | void resample ( const vec &w, ivec &ind, RESAMPLING_METHOD method ) { |
---|
| 507 | int n = w.length(); |
---|
[637] | 508 | ind = zeros_i ( n ); |
---|
[32] | 509 | ivec N_babies = zeros_i ( n ); |
---|
| 510 | vec cumDist = cumsum ( w ); |
---|
| 511 | vec u ( n ); |
---|
[477] | 512 | int i, j, parent; |
---|
[32] | 513 | double u0; |
---|
[887] | 514 | |
---|
[32] | 515 | switch ( method ) { |
---|
[887] | 516 | case MULTINOMIAL: |
---|
| 517 | u ( n - 1 ) = pow ( UniRNG.sample(), 1.0 / n ); |
---|
| 518 | |
---|
| 519 | for ( i = n - 2; i >= 0; i-- ) { |
---|
| 520 | u ( i ) = u ( i + 1 ) * pow ( UniRNG.sample(), 1.0 / ( i + 1 ) ); |
---|
| 521 | } |
---|
| 522 | |
---|
| 523 | break; |
---|
| 524 | |
---|
| 525 | case STRATIFIED: |
---|
| 526 | |
---|
| 527 | for ( i = 0; i < n; i++ ) { |
---|
| 528 | u ( i ) = ( i + UniRNG.sample() ) / n; |
---|
| 529 | } |
---|
| 530 | |
---|
| 531 | break; |
---|
| 532 | |
---|
| 533 | case SYSTEMATIC: |
---|
| 534 | u0 = UniRNG.sample(); |
---|
| 535 | |
---|
| 536 | for ( i = 0; i < n; i++ ) { |
---|
| 537 | u ( i ) = ( i + u0 ) / n; |
---|
| 538 | } |
---|
| 539 | |
---|
| 540 | break; |
---|
| 541 | |
---|
| 542 | default: |
---|
| 543 | bdm_error ( "PF::resample(): Unknown resampling method" ); |
---|
[32] | 544 | } |
---|
[887] | 545 | |
---|
[32] | 546 | // U is now full |
---|
| 547 | j = 0; |
---|
[887] | 548 | |
---|
[477] | 549 | for ( i = 0; i < n; i++ ) { |
---|
[32] | 550 | while ( u ( i ) > cumDist ( j ) ) j++; |
---|
[887] | 551 | |
---|
[32] | 552 | N_babies ( j ) ++; |
---|
| 553 | } |
---|
| 554 | // We have assigned new babies for each Particle |
---|
| 555 | // Now, we fill the resulting index such that: |
---|
| 556 | // * particles with at least one baby should not move * |
---|
| 557 | // This assures that reassignment can be done inplace; |
---|
[887] | 558 | |
---|
[32] | 559 | // find the first parent; |
---|
[477] | 560 | parent = 0; |
---|
| 561 | while ( N_babies ( parent ) == 0 ) parent++; |
---|
[887] | 562 | |
---|
[32] | 563 | // Build index |
---|
[477] | 564 | for ( i = 0; i < n; i++ ) { |
---|
[32] | 565 | if ( N_babies ( i ) > 0 ) { |
---|
| 566 | ind ( i ) = i; |
---|
| 567 | N_babies ( i ) --; //this index was now replicated; |
---|
[477] | 568 | } else { |
---|
[32] | 569 | // test if the parent has been fully replicated |
---|
| 570 | // if yes, find the next one |
---|
[477] | 571 | while ( ( N_babies ( parent ) == 0 ) || ( N_babies ( parent ) == 1 && parent > i ) ) parent++; |
---|
[887] | 572 | |
---|
[32] | 573 | // Replicate parent |
---|
| 574 | ind ( i ) = parent; |
---|
[887] | 575 | |
---|
[32] | 576 | N_babies ( parent ) --; //this index was now replicated; |
---|
| 577 | } |
---|
[887] | 578 | |
---|
[32] | 579 | } |
---|
| 580 | } |
---|
| 581 | |
---|
[488] | 582 | void eEmp::set_statistics ( const vec &w0, const epdf &epdf0 ) { |
---|
| 583 | dim = epdf0.dimension(); |
---|
[477] | 584 | w = w0; |
---|
| 585 | w /= sum ( w0 );//renormalize |
---|
| 586 | n = w.length(); |
---|
[32] | 587 | samples.set_size ( n ); |
---|
| 588 | |
---|
[477] | 589 | for ( int i = 0; i < n; i++ ) { |
---|
[488] | 590 | samples ( i ) = epdf0.sample(); |
---|
[477] | 591 | } |
---|
[32] | 592 | } |
---|
[178] | 593 | |
---|
[225] | 594 | void eEmp::set_samples ( const epdf* epdf0 ) { |
---|
[477] | 595 | w = 1; |
---|
| 596 | w /= sum ( w );//renormalize |
---|
[178] | 597 | |
---|
[477] | 598 | for ( int i = 0; i < n; i++ ) { |
---|
| 599 | samples ( i ) = epdf0->sample(); |
---|
| 600 | } |
---|
[178] | 601 | } |
---|
| 602 | |
---|
[477] | 603 | void migamma_ref::from_setting ( const Setting &set ) { |
---|
[956] | 604 | migamma::from_setting(set); |
---|
[357] | 605 | vec ref; |
---|
[957] | 606 | double k,l; |
---|
| 607 | |
---|
[477] | 608 | UI::get ( ref, set, "ref" , UI::compulsory ); |
---|
[957] | 609 | UI::get( k, set, "k", UI::compulsory ); |
---|
| 610 | UI::get( l, set, "l", UI::compulsory ); |
---|
| 611 | set_parameters ( k, ref, l ); |
---|
[357] | 612 | } |
---|
[957] | 613 | |
---|
[956] | 614 | void migamma_ref::to_setting (Setting &set) const { |
---|
| 615 | migamma::to_setting(set); |
---|
| 616 | UI::save ( pow ( refl, 1/(1.0 - l) ), set, "ref"); |
---|
| 617 | UI::save(l,set,"l"); |
---|
| 618 | UI::save(k,set,"k"); |
---|
| 619 | } |
---|
[957] | 620 | |
---|
[477] | 621 | void mlognorm::from_setting ( const Setting &set ) { |
---|
[956] | 622 | pdf_internal<elognorm>::from_setting(set); |
---|
[477] | 623 | vec mu0; |
---|
[957] | 624 | double k; |
---|
[477] | 625 | UI::get ( mu0, set, "mu0", UI::compulsory ); |
---|
[957] | 626 | UI::get( k, set, "k", UI::compulsory ); |
---|
| 627 | set_parameters ( mu0.length(), k ); |
---|
[477] | 628 | condition ( mu0 ); |
---|
[357] | 629 | } |
---|
| 630 | |
---|
[956] | 631 | void mlognorm::to_setting (Setting &set) const { |
---|
| 632 | pdf_internal<elognorm>::to_setting(set); |
---|
| 633 | UI::save ( exp(mu + sig2), set, "mu0"); |
---|
[957] | 634 | |
---|
| 635 | // inversion of sig2 = 0.5 * log ( k * k + 1 ); |
---|
| 636 | double k = sqrt( exp( 2 * sig2 ) - 1 ); |
---|
| 637 | UI::save(k,set,"k"); |
---|
[956] | 638 | } |
---|
| 639 | |
---|
| 640 | |
---|
[739] | 641 | void mlstudent::condition ( const vec &cond ) { |
---|
| 642 | if ( cond.length() > 0 ) { |
---|
| 643 | iepdf._mu() = A * cond + mu_const; |
---|
| 644 | } else { |
---|
| 645 | iepdf._mu() = mu_const; |
---|
| 646 | } |
---|
| 647 | double zeta; |
---|
| 648 | //ugly hack! |
---|
| 649 | if ( ( cond.length() + 1 ) == Lambda.rows() ) { |
---|
| 650 | zeta = Lambda.invqform ( concat ( cond, vec_1 ( 1.0 ) ) ); |
---|
| 651 | } else { |
---|
| 652 | zeta = Lambda.invqform ( cond ); |
---|
| 653 | } |
---|
| 654 | _R = Re; |
---|
| 655 | _R *= ( 1 + zeta );// / ( nu ); << nu is in Re!!!!!! |
---|
| 656 | } |
---|
| 657 | |
---|
| 658 | void eEmp::qbounds ( vec &lb, vec &ub, double perc ) const { |
---|
| 659 | // lb in inf so than it will be pushed below; |
---|
| 660 | lb.set_size ( dim ); |
---|
| 661 | ub.set_size ( dim ); |
---|
| 662 | lb = std::numeric_limits<double>::infinity(); |
---|
| 663 | ub = -std::numeric_limits<double>::infinity(); |
---|
| 664 | int j; |
---|
| 665 | for ( int i = 0; i < n; i++ ) { |
---|
| 666 | for ( j = 0; j < dim; j++ ) { |
---|
| 667 | if ( samples ( i ) ( j ) < lb ( j ) ) { |
---|
| 668 | lb ( j ) = samples ( i ) ( j ); |
---|
| 669 | } |
---|
| 670 | if ( samples ( i ) ( j ) > ub ( j ) ) { |
---|
| 671 | ub ( j ) = samples ( i ) ( j ); |
---|
| 672 | } |
---|
| 673 | } |
---|
| 674 | } |
---|
| 675 | } |
---|
| 676 | |
---|
[956] | 677 | void eEmp::to_setting ( Setting &set ) const { |
---|
| 678 | epdf::to_setting( set ); |
---|
| 679 | UI::save ( samples, set, "samples" ); |
---|
| 680 | UI::save ( w, set, "w" ); |
---|
| 681 | } |
---|
[739] | 682 | |
---|
[956] | 683 | void eEmp::from_setting ( const Setting &set ) { |
---|
| 684 | epdf::from_setting( set ); |
---|
| 685 | |
---|
| 686 | UI::get( samples, set, "samples", UI::compulsory ); |
---|
| 687 | UI::get ( w, set, "w", UI::compulsory ); |
---|
| 688 | } |
---|
| 689 | |
---|
| 690 | void eEmp::validate (){ |
---|
| 691 | epdf::validate(); |
---|
| 692 | bdm_assert (samples.length()==w.length(),"samples and weigths are of different lengths"); |
---|
| 693 | n = w.length(); |
---|
| 694 | if (n>0) |
---|
| 695 | pdf::dim = samples ( 0 ).length(); |
---|
| 696 | } |
---|
| 697 | |
---|
| 698 | void eDirich::from_setting ( const Setting &set ) { |
---|
| 699 | epdf::from_setting ( set ); |
---|
| 700 | UI::get ( beta, set, "beta", UI::compulsory ); |
---|
| 701 | } |
---|
| 702 | void eDirich::validate() { |
---|
| 703 | //check rv |
---|
| 704 | eEF::validate(); |
---|
| 705 | dim = beta.length(); |
---|
| 706 | } |
---|
| 707 | |
---|
| 708 | void eDirich::to_setting ( Setting &set ) const |
---|
| 709 | { |
---|
| 710 | eEF::to_setting( set ); |
---|
| 711 | UI::save( beta, set, "beta" ); |
---|
| 712 | } |
---|
| 713 | |
---|
| 714 | void euni::from_setting ( const Setting &set ) { |
---|
| 715 | epdf::from_setting ( set ); // reads rv and rvc |
---|
| 716 | |
---|
| 717 | UI::get ( high, set, "high", UI::compulsory ); |
---|
| 718 | UI::get ( low, set, "low", UI::compulsory ); |
---|
| 719 | set_parameters ( low, high ); |
---|
| 720 | |
---|
| 721 | } |
---|
| 722 | |
---|
| 723 | void euni::to_setting (Setting &set) const { |
---|
| 724 | epdf::to_setting ( set ); |
---|
| 725 | UI::save ( high, set, "high" ); |
---|
| 726 | UI::save ( low, set, "low" ); |
---|
| 727 | } |
---|
| 728 | |
---|
| 729 | void euni::validate() { |
---|
| 730 | epdf::validate(); |
---|
| 731 | bdm_assert ( high.length() == low.length(), "Incompatible high and low vectors" ); |
---|
| 732 | dim = high.length(); |
---|
| 733 | bdm_assert ( min ( distance ) > 0.0, "bad support" ); |
---|
| 734 | } |
---|
| 735 | |
---|
| 736 | void mgdirac::from_setting(const Setting& set){ |
---|
| 737 | pdf::from_setting(set); |
---|
| 738 | g=UI::build<fnc>(set,"g",UI::compulsory); |
---|
| 739 | validate(); |
---|
| 740 | } |
---|
| 741 | void mgdirac::to_setting(Setting &set) const{ |
---|
| 742 | pdf::to_setting(set); |
---|
| 743 | UI::save(g.get(), set, "g"); |
---|
| 744 | } |
---|
| 745 | void mgdirac::validate() { |
---|
| 746 | pdf::validate(); |
---|
| 747 | dim = g->dimension(); |
---|
| 748 | dimc = g->dimensionc(); |
---|
| 749 | } |
---|
| 750 | |
---|
| 751 | void mDirich::from_setting ( const Setting &set ) { |
---|
| 752 | pdf::from_setting ( set ); // reads rv and rvc |
---|
| 753 | if ( _rv()._dsize() > 0 ) { |
---|
| 754 | rvc = _rv().copy_t ( -1 ); |
---|
| 755 | } |
---|
| 756 | vec beta0; |
---|
| 757 | if ( !UI::get ( beta0, set, "beta0", UI::optional ) ) { |
---|
| 758 | beta0 = ones ( _rv()._dsize() ); |
---|
| 759 | } |
---|
| 760 | if ( !UI::get ( betac, set, "betac", UI::optional ) ) { |
---|
| 761 | betac = 0.1 * ones ( _rv()._dsize() ); |
---|
| 762 | } |
---|
| 763 | _beta = beta0; |
---|
| 764 | |
---|
| 765 | UI::get ( k, set, "k", UI::compulsory ); |
---|
| 766 | } |
---|
| 767 | |
---|
| 768 | void mDirich::to_setting (Setting &set) const { |
---|
| 769 | pdf::to_setting(set); |
---|
| 770 | UI::save( _beta, set, "beta0"); |
---|
| 771 | UI::save( betac, set, "betac"); |
---|
| 772 | UI::save ( k, set, "k" ); |
---|
| 773 | } |
---|
| 774 | |
---|
| 775 | |
---|
| 776 | void mDirich::validate() { |
---|
| 777 | pdf_internal<eDirich>::validate(); |
---|
| 778 | bdm_assert ( _beta.length() == betac.length(), "beta0 and betac are not compatible" ); |
---|
| 779 | if ( _rv()._dsize() > 0 ) { |
---|
| 780 | bdm_assert ( ( _rv()._dsize() == dimension() ) , "Size of rv does not match with beta" ); |
---|
| 781 | } |
---|
| 782 | dimc = _beta.length(); |
---|
| 783 | } |
---|
| 784 | |
---|
| 785 | }; |
---|