Research Article

Optimal Bandwidth Selection for Kernel Density Functionals Estimation

Table 2

Optimal bandwidth selection in estimation of with comparison to classical bandwidth selection methods.

Bandwidth selection for estimation of
Proposed method Classical bandwidth for KDE
100 100 100 100 100 100

Normal
 Mean .086 .067 .236 .173 .221 .163 .303 .271 .358 .298 .300 .270
 1st quantile .079 .063 .203 .156 .198 .150 .278 .256 .296 .256 .263 .244
 Median .086 .067 .236 .173 .229 .170 .304 .272 .380 .320 .302 .272
 3rd quantile .094 .071 .267 .190 .254 .185 .329 .287 .432 .361 .334 .296
Cauchy
 Mean .137 .104 .312 .221 .283 .200 .488 .422 .428 .320 .410 .337
 1st quantile .114 .092 .258 .194 .223 .164 .406 .373 .317 .238 .342 .298
 Median .136 .103 .304 .217 .285 .201 .483 .416 .431 .335 .405 .339
 3rd quantile .157 .115 .359 .244 .347 .236 .554 .468 .548 .414 .475 .387
Pareto
 Mean .191 .147 .276 .188 .145 .099 .673 .595 .201 .143 .372 .290
 1st quantile .142 .122 .216 .163 .107 .076 .501 .495 .140 .106 .296 .254
 Median .180 .142 .263 .185 .138 .093 .635 .574 .193 .128 .357 .285
 3rd quantile .219 .166 .318 .208 .173 .114 .775 .673 .245 .174 .430 .322
Mix-Normal I
 Mean .146 .112 .566 .369 .508 .342 .578 .507 .551 .417 .505 .420
 1st quantile .140 .109 .398 .289 .352 .254 .549 .490 .390 .332 .440 .378
 Median .147 .112 .481 .334 .454 .316 .577 .507 .549 .407 .500 .414
 3rd quantile .154 .116 .634 .391 .593 .382 .608 .523 .719 .500 .563 .458
Mix-Normal II
 Mean .103 .080 .442 .272 .442 .247 .521 .461 .489 .359 .465 .387
 1st quantile .094 .076 .310 .220 .258 .173 .491 .443 .341 .257 .412 .348
 Median .106 .081 .363 .246 .350 .226 .522 .458 .490 .353 .461 .382
 3rd quantile .114 .087 .468 .282 .470 .278 .555 .481 .637 .456 .514 .419

ROT bandwidth for estimation of given by (16); 2DPI bandwidth for estimation of with pilot bandwidth ; 3DPI bandwidth for estimation of with pilot bandwidth ; 4ROT bandwidth for density estimation proposed in Silverman [7]; 5LSCV bandwidth for density estimation proposed in Bowman [8]; 6DPI bandwidth for density estimation proposed in Sheather and Jones [9].