[ae3ce4e] | 1 | /* CylinderFit.c |
---|
| 2 | |
---|
| 3 | A simplified project designed to act as a template for your curve fitting function. |
---|
| 4 | The fitting function is a Cylinder form factor. No resolution effects are included (yet) |
---|
| 5 | */ |
---|
| 6 | |
---|
| 7 | #include "StandardHeaders.h" // Include ANSI headers, Mac headers |
---|
| 8 | #include "GaussWeights.h" |
---|
| 9 | #include "libCylinder.h" |
---|
| 10 | /* CylinderForm : calculates the form factor of a cylinder at the give x-value p->x |
---|
| 11 | |
---|
| 12 | Warning: |
---|
| 13 | The call to WaveData() below returns a pointer to the middle |
---|
| 14 | of an unlocked Macintosh handle. In the unlikely event that your |
---|
| 15 | calculations could cause memory to move, you should copy the coefficient |
---|
| 16 | values to local variables or an array before such operations. |
---|
| 17 | */ |
---|
| 18 | double |
---|
| 19 | CylinderForm(double dp[], double q) |
---|
| 20 | { |
---|
| 21 | int i; |
---|
| 22 | double Pi; |
---|
| 23 | double scale,radius,length,delrho,bkg,halfheight; //local variables of coefficient wave |
---|
| 24 | int nord=76; //order of integration |
---|
| 25 | double uplim,lolim; //upper and lower integration limits |
---|
| 26 | double summ,zi,yyy,answer,vcyl; //running tally of integration |
---|
| 27 | |
---|
| 28 | Pi = 4.0*atan(1.0); |
---|
| 29 | lolim = 0; |
---|
| 30 | uplim = Pi/2.0; |
---|
| 31 | |
---|
| 32 | summ = 0.0; //initialize intergral |
---|
| 33 | |
---|
| 34 | scale = dp[0]; //make local copies in case memory moves |
---|
| 35 | radius = dp[1]; |
---|
| 36 | length = dp[2]; |
---|
| 37 | delrho = dp[3]; |
---|
| 38 | bkg = dp[4]; |
---|
| 39 | halfheight = length/2.0; |
---|
| 40 | for(i=0;i<nord;i++) { |
---|
| 41 | zi = ( Gauss76Z[i]*(uplim-lolim) + uplim + lolim )/2.0; |
---|
| 42 | yyy = Gauss76Wt[i] * CylKernel(q, radius, halfheight, zi); |
---|
| 43 | summ += yyy; |
---|
| 44 | } |
---|
| 45 | |
---|
| 46 | answer = (uplim-lolim)/2.0*summ; |
---|
| 47 | // Multiply by contrast^2 |
---|
| 48 | answer *= delrho*delrho; |
---|
| 49 | //normalize by cylinder volume |
---|
| 50 | //NOTE that for this (Fournet) definition of the integral, one must MULTIPLY by Vcyl |
---|
| 51 | vcyl=Pi*radius*radius*length; |
---|
| 52 | answer *= vcyl; |
---|
| 53 | //convert to [cm-1] |
---|
| 54 | answer *= 1.0e8; |
---|
| 55 | //Scale |
---|
| 56 | answer *= scale; |
---|
| 57 | // add in the background |
---|
| 58 | answer += bkg; |
---|
| 59 | |
---|
| 60 | return answer; |
---|
| 61 | } |
---|
| 62 | |
---|
| 63 | /* EllipCyl76X : calculates the form factor of a elliptical cylinder at the given x-value p->x |
---|
| 64 | |
---|
| 65 | Uses 76 pt Gaussian quadrature for both integrals |
---|
| 66 | |
---|
| 67 | Warning: |
---|
| 68 | The call to WaveData() below returns a pointer to the middle |
---|
| 69 | of an unlocked Macintosh handle. In the unlikely event that your |
---|
| 70 | calculations could cause memory to move, you should copy the coefficient |
---|
| 71 | values to local variables or an array before such operations. |
---|
| 72 | */ |
---|
| 73 | double |
---|
| 74 | EllipCyl76(double dp[], double q) |
---|
| 75 | { |
---|
| 76 | int i,j; |
---|
| 77 | double Pi; |
---|
| 78 | double scale,ra,nu,length,delrho,bkg; //local variables of coefficient wave |
---|
| 79 | int nord=76; //order of integration |
---|
| 80 | double va,vb; //upper and lower integration limits |
---|
| 81 | double summ,zi,yyy,answer,vell; //running tally of integration |
---|
| 82 | double summj,vaj,vbj,zij,arg; //for the inner integration |
---|
| 83 | |
---|
| 84 | Pi = 4.0*atan(1.0); |
---|
| 85 | va = 0; |
---|
| 86 | vb = 1; //orintational average, outer integral |
---|
| 87 | vaj=0; |
---|
| 88 | vbj=Pi; //endpoints of inner integral |
---|
| 89 | |
---|
| 90 | summ = 0.0; //initialize intergral |
---|
| 91 | |
---|
| 92 | scale = dp[0]; //make local copies in case memory moves |
---|
| 93 | ra = dp[1]; |
---|
| 94 | nu = dp[2]; |
---|
| 95 | length = dp[3]; |
---|
| 96 | delrho = dp[4]; |
---|
| 97 | bkg = dp[5]; |
---|
| 98 | for(i=0;i<nord;i++) { |
---|
| 99 | //setup inner integral over the ellipsoidal cross-section |
---|
| 100 | summj=0; |
---|
| 101 | zi = ( Gauss76Z[i]*(vb-va) + va + vb )/2.0; //the "x" dummy |
---|
| 102 | arg = ra*sqrt(1-zi*zi); |
---|
| 103 | for(j=0;j<nord;j++) { |
---|
| 104 | //76 gauss points for the inner integral as well |
---|
| 105 | zij = ( Gauss76Z[j]*(vbj-vaj) + vaj + vbj )/2.0; //the "y" dummy |
---|
| 106 | yyy = Gauss76Wt[j] * EllipCylKernel(q,arg,nu,zij); |
---|
| 107 | summj += yyy; |
---|
| 108 | } |
---|
| 109 | //now calculate the value of the inner integral |
---|
| 110 | answer = (vbj-vaj)/2.0*summj; |
---|
| 111 | //divide integral by Pi |
---|
| 112 | answer /=Pi; |
---|
| 113 | |
---|
| 114 | //now calculate outer integral |
---|
| 115 | arg = q*length*zi/2; |
---|
| 116 | yyy = Gauss76Wt[i] * answer * sin(arg) * sin(arg) / arg / arg; |
---|
| 117 | summ += yyy; |
---|
| 118 | } |
---|
| 119 | answer = (vb-va)/2.0*summ; |
---|
| 120 | // Multiply by contrast^2 |
---|
| 121 | answer *= delrho*delrho; |
---|
| 122 | //normalize by cylinder volume |
---|
| 123 | //NOTE that for this (Fournet) definition of the integral, one must MULTIPLY by Vcyl |
---|
| 124 | vell = Pi*ra*(nu*ra)*length; |
---|
| 125 | answer *= vell; |
---|
| 126 | //convert to [cm-1] |
---|
| 127 | answer *= 1.0e8; |
---|
| 128 | //Scale |
---|
| 129 | answer *= scale; |
---|
| 130 | // add in the background |
---|
| 131 | answer += bkg; |
---|
| 132 | |
---|
| 133 | return answer; |
---|
| 134 | } |
---|
| 135 | |
---|
| 136 | /* EllipCyl20X : calculates the form factor of a elliptical cylinder at the given x-value p->x |
---|
| 137 | |
---|
| 138 | Uses 76 pt Gaussian quadrature for orientational integral |
---|
| 139 | Uses 20 pt quadrature for the inner integral over the elliptical cross-section |
---|
| 140 | |
---|
| 141 | Warning: |
---|
| 142 | The call to WaveData() below returns a pointer to the middle |
---|
| 143 | of an unlocked Macintosh handle. In the unlikely event that your |
---|
| 144 | calculations could cause memory to move, you should copy the coefficient |
---|
| 145 | values to local variables or an array before such operations. |
---|
| 146 | */ |
---|
| 147 | double |
---|
| 148 | EllipCyl20(double dp[], double q) |
---|
| 149 | { |
---|
| 150 | int i,j; |
---|
| 151 | double Pi; |
---|
| 152 | double scale,ra,nu,length,delrho,bkg; //local variables of coefficient wave |
---|
| 153 | int nordi=76; //order of integration |
---|
| 154 | int nordj=20; |
---|
| 155 | double va,vb; //upper and lower integration limits |
---|
| 156 | double summ,zi,yyy,answer,vell; //running tally of integration |
---|
| 157 | double summj,vaj,vbj,zij,arg; //for the inner integration |
---|
| 158 | |
---|
| 159 | Pi = 4.0*atan(1.0); |
---|
| 160 | va = 0; |
---|
| 161 | vb = 1; //orintational average, outer integral |
---|
| 162 | vaj=0; |
---|
| 163 | vbj=Pi; //endpoints of inner integral |
---|
| 164 | |
---|
| 165 | summ = 0.0; //initialize intergral |
---|
| 166 | |
---|
| 167 | scale = dp[0]; //make local copies in case memory moves |
---|
| 168 | ra = dp[1]; |
---|
| 169 | nu = dp[2]; |
---|
| 170 | length = dp[3]; |
---|
| 171 | delrho = dp[4]; |
---|
| 172 | bkg = dp[5]; |
---|
| 173 | for(i=0;i<nordi;i++) { |
---|
| 174 | //setup inner integral over the ellipsoidal cross-section |
---|
| 175 | summj=0; |
---|
| 176 | zi = ( Gauss76Z[i]*(vb-va) + va + vb )/2.0; //the "x" dummy |
---|
| 177 | arg = ra*sqrt(1-zi*zi); |
---|
| 178 | for(j=0;j<nordj;j++) { |
---|
| 179 | //20 gauss points for the inner integral |
---|
| 180 | zij = ( Gauss20Z[j]*(vbj-vaj) + vaj + vbj )/2.0; //the "y" dummy |
---|
| 181 | yyy = Gauss20Wt[j] * EllipCylKernel(q,arg,nu,zij); |
---|
| 182 | summj += yyy; |
---|
| 183 | } |
---|
| 184 | //now calculate the value of the inner integral |
---|
| 185 | answer = (vbj-vaj)/2.0*summj; |
---|
| 186 | //divide integral by Pi |
---|
| 187 | answer /=Pi; |
---|
| 188 | |
---|
| 189 | //now calculate outer integral |
---|
| 190 | arg = q*length*zi/2; |
---|
| 191 | yyy = Gauss76Wt[i] * answer * sin(arg) * sin(arg) / arg / arg; |
---|
| 192 | summ += yyy; |
---|
| 193 | } |
---|
| 194 | |
---|
| 195 | answer = (vb-va)/2.0*summ; |
---|
| 196 | // Multiply by contrast^2 |
---|
| 197 | answer *= delrho*delrho; |
---|
| 198 | //normalize by cylinder volume |
---|
| 199 | //NOTE that for this (Fournet) definition of the integral, one must MULTIPLY by Vcyl |
---|
| 200 | vell = Pi*ra*(nu*ra)*length; |
---|
| 201 | answer *= vell; |
---|
| 202 | //convert to [cm-1] |
---|
| 203 | answer *= 1.0e8; |
---|
| 204 | //Scale |
---|
| 205 | answer *= scale; |
---|
| 206 | // add in the background |
---|
| 207 | answer += bkg; |
---|
| 208 | |
---|
| 209 | return answer; |
---|
| 210 | } |
---|
| 211 | |
---|
| 212 | /* TriaxialEllipsoidX : calculates the form factor of a Triaxial Ellipsoid at the given x-value p->x |
---|
| 213 | |
---|
| 214 | Uses 76 pt Gaussian quadrature for both integrals |
---|
| 215 | |
---|
| 216 | Warning: |
---|
| 217 | The call to WaveData() below returns a pointer to the middle |
---|
| 218 | of an unlocked Macintosh handle. In the unlikely event that your |
---|
| 219 | calculations could cause memory to move, you should copy the coefficient |
---|
| 220 | values to local variables or an array before such operations. |
---|
| 221 | */ |
---|
| 222 | double |
---|
| 223 | TriaxialEllipsoid(double dp[], double q) |
---|
| 224 | { |
---|
| 225 | int i,j; |
---|
| 226 | double Pi; |
---|
| 227 | double scale,aa,bb,cc,delrho,bkg; //local variables of coefficient wave |
---|
| 228 | int nordi=76; //order of integration |
---|
| 229 | int nordj=76; |
---|
| 230 | double va,vb; //upper and lower integration limits |
---|
| 231 | double summ,zi,yyy,answer; //running tally of integration |
---|
| 232 | double summj,vaj,vbj,zij; //for the inner integration |
---|
| 233 | |
---|
| 234 | Pi = 4.0*atan(1.0); |
---|
| 235 | va = 0; |
---|
| 236 | vb = 1; //orintational average, outer integral |
---|
| 237 | vaj = 0; |
---|
| 238 | vbj = 1; //endpoints of inner integral |
---|
| 239 | |
---|
| 240 | summ = 0.0; //initialize intergral |
---|
| 241 | |
---|
| 242 | scale = dp[0]; //make local copies in case memory moves |
---|
| 243 | aa = dp[1]; |
---|
| 244 | bb = dp[2]; |
---|
| 245 | cc = dp[3]; |
---|
| 246 | delrho = dp[4]; |
---|
| 247 | bkg = dp[5]; |
---|
| 248 | for(i=0;i<nordi;i++) { |
---|
| 249 | //setup inner integral over the ellipsoidal cross-section |
---|
| 250 | summj=0; |
---|
| 251 | zi = ( Gauss76Z[i]*(vb-va) + va + vb )/2.0; //the "x" dummy |
---|
| 252 | for(j=0;j<nordj;j++) { |
---|
| 253 | //20 gauss points for the inner integral |
---|
| 254 | zij = ( Gauss76Z[j]*(vbj-vaj) + vaj + vbj )/2.0; //the "y" dummy |
---|
| 255 | yyy = Gauss76Wt[j] * TriaxialKernel(q,aa,bb,cc,zi,zij); |
---|
| 256 | summj += yyy; |
---|
| 257 | } |
---|
| 258 | //now calculate the value of the inner integral |
---|
| 259 | answer = (vbj-vaj)/2.0*summj; |
---|
| 260 | |
---|
| 261 | //now calculate outer integral |
---|
| 262 | yyy = Gauss76Wt[i] * answer; |
---|
| 263 | summ += yyy; |
---|
| 264 | } //final scaling is done at the end of the function, after the NT_FP64 case |
---|
| 265 | |
---|
| 266 | answer = (vb-va)/2.0*summ; |
---|
| 267 | // Multiply by contrast^2 |
---|
| 268 | answer *= delrho*delrho; |
---|
| 269 | //normalize by ellipsoid volume |
---|
| 270 | answer *= 4*Pi/3*aa*bb*cc; |
---|
| 271 | //convert to [cm-1] |
---|
| 272 | answer *= 1.0e8; |
---|
| 273 | //Scale |
---|
| 274 | answer *= scale; |
---|
| 275 | // add in the background |
---|
| 276 | answer += bkg; |
---|
| 277 | |
---|
| 278 | return answer; |
---|
| 279 | } |
---|
| 280 | |
---|
| 281 | /* ParallelepipedX : calculates the form factor of a Parallelepiped (a rectangular solid) |
---|
| 282 | at the given x-value p->x |
---|
| 283 | |
---|
| 284 | Uses 76 pt Gaussian quadrature for both integrals |
---|
| 285 | |
---|
| 286 | Warning: |
---|
| 287 | The call to WaveData() below returns a pointer to the middle |
---|
| 288 | of an unlocked Macintosh handle. In the unlikely event that your |
---|
| 289 | calculations could cause memory to move, you should copy the coefficient |
---|
| 290 | values to local variables or an array before such operations. |
---|
| 291 | */ |
---|
| 292 | double |
---|
| 293 | Parallelepiped(double dp[], double q) |
---|
| 294 | { |
---|
| 295 | int i,j; |
---|
| 296 | double scale,aa,bb,cc,delrho,bkg; //local variables of coefficient wave |
---|
| 297 | int nordi=76; //order of integration |
---|
| 298 | int nordj=76; |
---|
| 299 | double va,vb; //upper and lower integration limits |
---|
| 300 | double summ,yyy,answer; //running tally of integration |
---|
| 301 | double summj,vaj,vbj; //for the inner integration |
---|
| 302 | double mu,mudum,arg,sigma,uu,vol; |
---|
| 303 | |
---|
| 304 | |
---|
| 305 | // Pi = 4.0*atan(1.0); |
---|
| 306 | va = 0; |
---|
| 307 | vb = 1; //orintational average, outer integral |
---|
| 308 | vaj = 0; |
---|
| 309 | vbj = 1; //endpoints of inner integral |
---|
| 310 | |
---|
| 311 | summ = 0.0; //initialize intergral |
---|
| 312 | |
---|
| 313 | scale = dp[0]; //make local copies in case memory moves |
---|
| 314 | aa = dp[1]; |
---|
| 315 | bb = dp[2]; |
---|
| 316 | cc = dp[3]; |
---|
| 317 | delrho = dp[4]; |
---|
| 318 | bkg = dp[5]; |
---|
| 319 | |
---|
| 320 | mu = q*bb; |
---|
| 321 | vol = aa*bb*cc; |
---|
| 322 | // normalize all WRT bb |
---|
| 323 | aa = aa/bb; |
---|
| 324 | cc = cc/bb; |
---|
| 325 | |
---|
| 326 | for(i=0;i<nordi;i++) { |
---|
| 327 | //setup inner integral over the ellipsoidal cross-section |
---|
| 328 | summj=0; |
---|
| 329 | sigma = ( Gauss76Z[i]*(vb-va) + va + vb )/2.0; //the outer dummy |
---|
| 330 | |
---|
| 331 | for(j=0;j<nordj;j++) { |
---|
| 332 | //76 gauss points for the inner integral |
---|
| 333 | uu = ( Gauss76Z[j]*(vbj-vaj) + vaj + vbj )/2.0; //the inner dummy |
---|
| 334 | mudum = mu*sqrt(1-sigma*sigma); |
---|
| 335 | yyy = Gauss76Wt[j] * PPKernel(aa,mudum,uu); |
---|
| 336 | summj += yyy; |
---|
| 337 | } |
---|
| 338 | //now calculate the value of the inner integral |
---|
| 339 | answer = (vbj-vaj)/2.0*summj; |
---|
| 340 | |
---|
| 341 | arg = mu*cc*sigma/2; |
---|
| 342 | if ( arg == 0 ) { |
---|
| 343 | answer *= 1; |
---|
| 344 | } else { |
---|
| 345 | answer *= sin(arg)*sin(arg)/arg/arg; |
---|
| 346 | } |
---|
| 347 | |
---|
| 348 | //now sum up the outer integral |
---|
| 349 | yyy = Gauss76Wt[i] * answer; |
---|
| 350 | summ += yyy; |
---|
| 351 | } //final scaling is done at the end of the function, after the NT_FP64 case |
---|
| 352 | |
---|
| 353 | answer = (vb-va)/2.0*summ; |
---|
| 354 | // Multiply by contrast^2 |
---|
| 355 | answer *= delrho*delrho; |
---|
| 356 | //normalize by volume |
---|
| 357 | answer *= vol; |
---|
| 358 | //convert to [cm-1] |
---|
| 359 | answer *= 1.0e8; |
---|
| 360 | //Scale |
---|
| 361 | answer *= scale; |
---|
| 362 | // add in the background |
---|
| 363 | answer += bkg; |
---|
| 364 | |
---|
| 365 | return answer; |
---|
| 366 | } |
---|
| 367 | |
---|
| 368 | /* HollowCylinderX : calculates the form factor of a Hollow Cylinder |
---|
| 369 | at the given x-value p->x |
---|
| 370 | |
---|
| 371 | Uses 76 pt Gaussian quadrature for the single integral |
---|
| 372 | |
---|
| 373 | Warning: |
---|
| 374 | The call to WaveData() below returns a pointer to the middle |
---|
| 375 | of an unlocked Macintosh handle. In the unlikely event that your |
---|
| 376 | calculations could cause memory to move, you should copy the coefficient |
---|
| 377 | values to local variables or an array before such operations. |
---|
| 378 | */ |
---|
| 379 | double |
---|
| 380 | HollowCylinder(double dp[], double q) |
---|
| 381 | { |
---|
| 382 | int i; |
---|
| 383 | double scale,rcore,rshell,length,delrho,bkg; //local variables of coefficient wave |
---|
| 384 | int nord=76; //order of integration |
---|
| 385 | double va,vb,zi; //upper and lower integration limits |
---|
| 386 | double summ,answer,pi; //running tally of integration |
---|
| 387 | |
---|
| 388 | pi = 4.0*atan(1.0); |
---|
| 389 | va = 0; |
---|
| 390 | vb = 1; //limits of numerical integral |
---|
| 391 | |
---|
| 392 | summ = 0.0; //initialize intergral |
---|
| 393 | |
---|
| 394 | scale = dp[0]; //make local copies in case memory moves |
---|
| 395 | rcore = dp[1]; |
---|
| 396 | rshell = dp[2]; |
---|
| 397 | length = dp[3]; |
---|
| 398 | delrho = dp[4]; |
---|
| 399 | bkg = dp[5]; |
---|
| 400 | |
---|
| 401 | for(i=0;i<nord;i++) { |
---|
| 402 | zi = ( Gauss76Z[i]*(vb-va) + va + vb )/2.0; |
---|
| 403 | summ += Gauss76Wt[i] * HolCylKernel(q, rcore, rshell, length, zi); |
---|
| 404 | } |
---|
| 405 | |
---|
| 406 | answer = (vb-va)/2.0*summ; |
---|
| 407 | // Multiply by contrast^2 |
---|
| 408 | answer *= delrho*delrho; |
---|
| 409 | //normalize by volume |
---|
| 410 | answer *= pi*(rshell*rshell-rcore*rcore)*length; |
---|
| 411 | //convert to [cm-1] |
---|
| 412 | answer *= 1.0e8; |
---|
| 413 | //Scale |
---|
| 414 | answer *= scale; |
---|
| 415 | // add in the background |
---|
| 416 | answer += bkg; |
---|
| 417 | |
---|
| 418 | return answer; |
---|
| 419 | } |
---|
| 420 | |
---|
| 421 | /* EllipsoidFormX : calculates the form factor of an ellipsoid of revolution with semiaxes a:a:nua |
---|
| 422 | at the given x-value p->x |
---|
| 423 | |
---|
| 424 | Uses 76 pt Gaussian quadrature for the single integral |
---|
| 425 | |
---|
| 426 | Warning: |
---|
| 427 | The call to WaveData() below returns a pointer to the middle |
---|
| 428 | of an unlocked Macintosh handle. In the unlikely event that your |
---|
| 429 | calculations could cause memory to move, you should copy the coefficient |
---|
| 430 | values to local variables or an array before such operations. |
---|
| 431 | */ |
---|
| 432 | double |
---|
| 433 | EllipsoidForm(double dp[], double q) |
---|
| 434 | { |
---|
| 435 | int i; |
---|
| 436 | double scale,a,nua,delrho,bkg; //local variables of coefficient wave |
---|
| 437 | int nord=76; //order of integration |
---|
| 438 | double va,vb,zi; //upper and lower integration limits |
---|
| 439 | double summ,answer,pi; //running tally of integration |
---|
| 440 | |
---|
| 441 | pi = 4.0*atan(1.0); |
---|
| 442 | va = 0; |
---|
| 443 | vb = 1; //limits of numerical integral |
---|
| 444 | |
---|
| 445 | summ = 0.0; //initialize intergral |
---|
| 446 | |
---|
| 447 | scale = dp[0]; //make local copies in case memory moves |
---|
| 448 | nua = dp[1]; |
---|
| 449 | a = dp[2]; |
---|
| 450 | delrho = dp[3]; |
---|
| 451 | bkg = dp[4]; |
---|
| 452 | |
---|
| 453 | for(i=0;i<nord;i++) { |
---|
| 454 | zi = ( Gauss76Z[i]*(vb-va) + va + vb )/2.0; |
---|
| 455 | summ += Gauss76Wt[i] * EllipsoidKernel(q, a, nua, zi); |
---|
| 456 | } |
---|
| 457 | |
---|
| 458 | answer = (vb-va)/2.0*summ; |
---|
| 459 | // Multiply by contrast^2 |
---|
| 460 | answer *= delrho*delrho; |
---|
| 461 | //normalize by volume |
---|
| 462 | answer *= 4*pi/3*a*a*nua; |
---|
| 463 | //convert to [cm-1] |
---|
| 464 | answer *= 1.0e8; |
---|
| 465 | //Scale |
---|
| 466 | answer *= scale; |
---|
| 467 | // add in the background |
---|
| 468 | answer += bkg; |
---|
| 469 | |
---|
| 470 | return answer; |
---|
| 471 | } |
---|
| 472 | |
---|
| 473 | |
---|
| 474 | /* Cyl_PolyRadiusX : calculates the form factor of a cylinder at the given x-value p->x |
---|
| 475 | the cylinder has a polydisperse cross section |
---|
| 476 | |
---|
| 477 | */ |
---|
| 478 | double |
---|
| 479 | Cyl_PolyRadius(double dp[], double q) |
---|
| 480 | { |
---|
| 481 | int i; |
---|
| 482 | double scale,radius,length,pd,delrho,bkg; //local variables of coefficient wave |
---|
| 483 | int nord=20; //order of integration |
---|
| 484 | double uplim,lolim; //upper and lower integration limits |
---|
| 485 | double summ,zi,yyy,answer,Vpoly; //running tally of integration |
---|
| 486 | double range,zz,Pi; |
---|
| 487 | |
---|
| 488 | Pi = 4.0*atan(1.0); |
---|
| 489 | range = 3.4; |
---|
| 490 | |
---|
| 491 | summ = 0.0; //initialize intergral |
---|
| 492 | |
---|
| 493 | scale = dp[0]; //make local copies in case memory moves |
---|
| 494 | radius = dp[1]; |
---|
| 495 | length = dp[2]; |
---|
| 496 | pd = dp[3]; |
---|
| 497 | delrho = dp[4]; |
---|
| 498 | bkg = dp[5]; |
---|
| 499 | |
---|
| 500 | zz = (1.0/pd)*(1.0/pd) - 1.0; |
---|
| 501 | |
---|
| 502 | lolim = radius*(1.0-range*pd); //set the upper/lower limits to cover the distribution |
---|
| 503 | if(lolim<0) { |
---|
| 504 | lolim = 0; |
---|
| 505 | } |
---|
| 506 | if(pd>0.3) { |
---|
| 507 | range = 3.4 + (pd-0.3)*18.0; |
---|
| 508 | } |
---|
| 509 | uplim = radius*(1.0+range*pd); |
---|
| 510 | |
---|
| 511 | for(i=0;i<nord;i++) { |
---|
| 512 | zi = ( Gauss20Z[i]*(uplim-lolim) + uplim + lolim )/2.0; |
---|
| 513 | yyy = Gauss20Wt[i] * Cyl_PolyRadKernel(q, radius, length, zz, delrho, zi); |
---|
| 514 | summ += yyy; |
---|
| 515 | } |
---|
| 516 | |
---|
| 517 | answer = (uplim-lolim)/2.0*summ; |
---|
| 518 | //normalize by average cylinder volume |
---|
| 519 | //NOTE that for this (Fournet) definition of the integral, one must MULTIPLY by Vcyl |
---|
| 520 | Vpoly=Pi*radius*radius*length*(zz+2.0)/(zz+1.0); |
---|
| 521 | answer /= Vpoly; |
---|
| 522 | //convert to [cm-1] |
---|
| 523 | answer *= 1.0e8; |
---|
| 524 | //Scale |
---|
| 525 | answer *= scale; |
---|
| 526 | // add in the background |
---|
| 527 | answer += bkg; |
---|
| 528 | |
---|
| 529 | return answer; |
---|
| 530 | } |
---|
| 531 | |
---|
| 532 | /* Cyl_PolyLengthX : calculates the form factor of a cylinder at the given x-value p->x |
---|
| 533 | the cylinder has a polydisperse Length |
---|
| 534 | |
---|
| 535 | */ |
---|
| 536 | double |
---|
| 537 | Cyl_PolyLength(double dp[], double q) |
---|
| 538 | { |
---|
| 539 | int i; |
---|
| 540 | double scale,radius,length,pd,delrho,bkg; //local variables of coefficient wave |
---|
| 541 | int nord=20; //order of integration |
---|
| 542 | double uplim,lolim; //upper and lower integration limits |
---|
| 543 | double summ,zi,yyy,answer,Vpoly; //running tally of integration |
---|
| 544 | double range,zz,Pi; |
---|
| 545 | |
---|
| 546 | |
---|
| 547 | Pi = 4.0*atan(1.0); |
---|
| 548 | range = 3.4; |
---|
| 549 | |
---|
| 550 | summ = 0.0; //initialize intergral |
---|
| 551 | |
---|
| 552 | scale = dp[0]; //make local copies in case memory moves |
---|
| 553 | radius = dp[1]; |
---|
| 554 | length = dp[2]; |
---|
| 555 | pd = dp[3]; |
---|
| 556 | delrho = dp[4]; |
---|
| 557 | bkg = dp[5]; |
---|
| 558 | |
---|
| 559 | zz = (1.0/pd)*(1.0/pd) - 1.0; |
---|
| 560 | |
---|
| 561 | lolim = length*(1.0-range*pd); //set the upper/lower limits to cover the distribution |
---|
| 562 | if(lolim<0) { |
---|
| 563 | lolim = 0; |
---|
| 564 | } |
---|
| 565 | if(pd>0.3) { |
---|
| 566 | range = 3.4 + (pd-0.3)*18.0; |
---|
| 567 | } |
---|
| 568 | uplim = length*(1.0+range*pd); |
---|
| 569 | |
---|
| 570 | for(i=0;i<nord;i++) { |
---|
| 571 | zi = ( Gauss20Z[i]*(uplim-lolim) + uplim + lolim )/2.0; |
---|
| 572 | yyy = Gauss20Wt[i] * Cyl_PolyLenKernel(q, radius, length, zz, delrho, zi); |
---|
| 573 | summ += yyy; |
---|
| 574 | } |
---|
| 575 | |
---|
| 576 | answer = (uplim-lolim)/2.0*summ; |
---|
| 577 | //normalize by average cylinder volume (first moment) |
---|
| 578 | //NOTE that for this (Fournet) definition of the integral, one must MULTIPLY by Vcyl |
---|
| 579 | Vpoly=Pi*radius*radius*length; |
---|
| 580 | answer /= Vpoly; |
---|
| 581 | //convert to [cm-1] |
---|
| 582 | answer *= 1.0e8; |
---|
| 583 | //Scale |
---|
| 584 | answer *= scale; |
---|
| 585 | // add in the background |
---|
| 586 | answer += bkg; |
---|
| 587 | |
---|
| 588 | return answer; |
---|
| 589 | } |
---|
| 590 | |
---|
| 591 | /* CoreShellCylinderX : calculates the form factor of a cylinder at the given x-value p->x |
---|
| 592 | the cylinder has a core-shell structure |
---|
| 593 | |
---|
| 594 | */ |
---|
| 595 | double |
---|
| 596 | CoreShellCylinder(double dp[], double q) |
---|
| 597 | { |
---|
| 598 | int i; |
---|
| 599 | double scale,rcore,length,bkg; //local variables of coefficient wave |
---|
| 600 | double thick,rhoc,rhos,rhosolv; |
---|
| 601 | int nord=76; //order of integration |
---|
| 602 | double uplim,lolim,halfheight; //upper and lower integration limits |
---|
| 603 | double summ,zi,yyy,answer,Vcyl; //running tally of integration |
---|
| 604 | double Pi; |
---|
| 605 | |
---|
| 606 | Pi = 4.0*atan(1.0); |
---|
| 607 | |
---|
| 608 | lolim = 0.0; |
---|
| 609 | uplim = Pi/2.0; |
---|
| 610 | |
---|
| 611 | summ = 0.0; //initialize intergral |
---|
| 612 | |
---|
| 613 | scale = dp[0]; //make local copies in case memory moves |
---|
| 614 | rcore = dp[1]; |
---|
| 615 | thick = dp[2]; |
---|
| 616 | length = dp[3]; |
---|
| 617 | rhoc = dp[4]; |
---|
| 618 | rhos = dp[5]; |
---|
| 619 | rhosolv = dp[6]; |
---|
| 620 | bkg = dp[7]; |
---|
| 621 | |
---|
| 622 | halfheight = length/2.0; |
---|
| 623 | |
---|
| 624 | for(i=0;i<nord;i++) { |
---|
| 625 | zi = ( Gauss76Z[i]*(uplim-lolim) + uplim + lolim )/2.0; |
---|
| 626 | yyy = Gauss76Wt[i] * CoreShellCylKernel(q, rcore, thick, rhoc,rhos,rhosolv, halfheight, zi); |
---|
| 627 | summ += yyy; |
---|
| 628 | } |
---|
| 629 | |
---|
| 630 | answer = (uplim-lolim)/2.0*summ; |
---|
| 631 | // length is the total core length |
---|
| 632 | Vcyl=Pi*(rcore+thick)*(rcore+thick)*(length+2.0*thick); |
---|
| 633 | answer /= Vcyl; |
---|
| 634 | //convert to [cm-1] |
---|
| 635 | answer *= 1.0e8; |
---|
| 636 | //Scale |
---|
| 637 | answer *= scale; |
---|
| 638 | // add in the background |
---|
| 639 | answer += bkg; |
---|
| 640 | |
---|
| 641 | return answer; |
---|
| 642 | } |
---|
| 643 | |
---|
| 644 | |
---|
| 645 | /* PolyCoShCylinderX : calculates the form factor of a core-shell cylinder at the given x-value p->x |
---|
| 646 | the cylinder has a polydisperse CORE radius |
---|
| 647 | |
---|
| 648 | */ |
---|
| 649 | double |
---|
| 650 | PolyCoShCylinder(double dp[], double q) |
---|
| 651 | { |
---|
| 652 | int i; |
---|
| 653 | double scale,radius,length,sigma,bkg; //local variables of coefficient wave |
---|
| 654 | double rad,radthick,facthick,rhoc,rhos,rhosolv; |
---|
| 655 | int nord=20; //order of integration |
---|
| 656 | double uplim,lolim; //upper and lower integration limits |
---|
| 657 | double summ,yyy,answer,Vpoly; //running tally of integration |
---|
| 658 | double Pi,AR,Rsqrsumm,Rsqryyy,Rsqr; |
---|
| 659 | |
---|
| 660 | Pi = 4.0*atan(1.0); |
---|
| 661 | |
---|
| 662 | summ = 0.0; //initialize intergral |
---|
| 663 | Rsqrsumm = 0.0; |
---|
| 664 | |
---|
| 665 | scale = dp[0]; |
---|
| 666 | radius = dp[1]; |
---|
| 667 | sigma = dp[2]; //sigma is the standard mean deviation |
---|
| 668 | length = dp[3]; |
---|
| 669 | radthick = dp[4]; |
---|
| 670 | facthick= dp[5]; |
---|
| 671 | rhoc = dp[6]; |
---|
| 672 | rhos = dp[7]; |
---|
| 673 | rhosolv = dp[8]; |
---|
| 674 | bkg = dp[9]; |
---|
| 675 | |
---|
| 676 | lolim = exp(log(radius)-(4.*sigma)); |
---|
| 677 | if (lolim<0) { |
---|
| 678 | lolim=0; //to avoid numerical error when va<0 (-ve r value) |
---|
| 679 | } |
---|
| 680 | uplim = exp(log(radius)+(4.*sigma)); |
---|
| 681 | |
---|
| 682 | for(i=0;i<nord;i++) { |
---|
| 683 | rad = ( Gauss20Z[i]*(uplim-lolim) + uplim + lolim )/2.0; |
---|
| 684 | AR=(1.0/(rad*sigma*sqrt(2.0*Pi)))*exp(-(0.5*((log(radius/rad))/sigma)*((log(radius/rad))/sigma))); |
---|
| 685 | yyy = AR* Gauss20Wt[i] * CSCylIntegration(q,rad,radthick,facthick,rhoc,rhos,rhosolv,length); |
---|
| 686 | Rsqryyy= Gauss20Wt[i] * AR * (rad+radthick)*(rad+radthick); //SRK normalize to total dimensions |
---|
| 687 | summ += yyy; |
---|
| 688 | Rsqrsumm += Rsqryyy; |
---|
| 689 | } |
---|
| 690 | |
---|
| 691 | answer = (uplim-lolim)/2.0*summ; |
---|
| 692 | Rsqr = (uplim-lolim)/2.0*Rsqrsumm; |
---|
| 693 | //normalize by average cylinder volume |
---|
| 694 | Vpoly = Pi*Rsqr*(length+2*facthick); |
---|
| 695 | answer /= Vpoly; |
---|
| 696 | //convert to [cm-1] |
---|
| 697 | answer *= 1.0e8; |
---|
| 698 | //Scale |
---|
| 699 | answer *= scale; |
---|
| 700 | // add in the background |
---|
| 701 | answer += bkg; |
---|
| 702 | |
---|
| 703 | return answer; |
---|
| 704 | } |
---|
| 705 | |
---|
| 706 | /* OblateFormX : calculates the form factor of a core-shell Oblate ellipsoid at the given x-value p->x |
---|
| 707 | the ellipsoid has a core-shell structure |
---|
| 708 | |
---|
| 709 | */ |
---|
| 710 | double |
---|
| 711 | OblateForm(double dp[], double q) |
---|
| 712 | { |
---|
| 713 | int i; |
---|
| 714 | double scale,crmaj,crmin,trmaj,trmin,delpc,delps,bkg; |
---|
| 715 | int nord=76; //order of integration |
---|
| 716 | double uplim,lolim; //upper and lower integration limits |
---|
| 717 | double summ,zi,yyy,answer,oblatevol; //running tally of integration |
---|
| 718 | double Pi; |
---|
| 719 | |
---|
| 720 | Pi = 4.0*atan(1.0); |
---|
| 721 | |
---|
| 722 | lolim = 0.0; |
---|
| 723 | uplim = 1.0; |
---|
| 724 | |
---|
| 725 | summ = 0.0; //initialize intergral |
---|
| 726 | |
---|
| 727 | |
---|
| 728 | scale = dp[0]; //make local copies in case memory moves |
---|
| 729 | crmaj = dp[1]; |
---|
| 730 | crmin = dp[2]; |
---|
| 731 | trmaj = dp[3]; |
---|
| 732 | trmin = dp[4]; |
---|
| 733 | delpc = dp[5]; |
---|
| 734 | delps = dp[6]; |
---|
| 735 | bkg = dp[7]; |
---|
| 736 | |
---|
| 737 | for(i=0;i<nord;i++) { |
---|
| 738 | zi = ( Gauss76Z[i]*(uplim-lolim) + uplim + lolim )/2.0; |
---|
| 739 | yyy = Gauss76Wt[i] * gfn4(zi,crmaj,crmin,trmaj,trmin,delpc,delps,q); |
---|
| 740 | summ += yyy; |
---|
| 741 | } |
---|
| 742 | |
---|
| 743 | answer = (uplim-lolim)/2.0*summ; |
---|
| 744 | // normalize by particle volume |
---|
| 745 | oblatevol = 4*Pi/3*trmaj*trmaj*trmin; |
---|
| 746 | answer /= oblatevol; |
---|
| 747 | |
---|
| 748 | //convert to [cm-1] |
---|
| 749 | answer *= 1.0e8; |
---|
| 750 | //Scale |
---|
| 751 | answer *= scale; |
---|
| 752 | // add in the background |
---|
| 753 | answer += bkg; |
---|
| 754 | |
---|
| 755 | return answer; |
---|
| 756 | } |
---|
| 757 | |
---|
| 758 | /* ProlateFormX : calculates the form factor of a core-shell Prolate ellipsoid at the given x-value p->x |
---|
| 759 | the ellipsoid has a core-shell structure |
---|
| 760 | |
---|
| 761 | */ |
---|
| 762 | double |
---|
| 763 | ProlateForm(double dp[], double q) |
---|
| 764 | { |
---|
| 765 | int i; |
---|
| 766 | double scale,crmaj,crmin,trmaj,trmin,delpc,delps,bkg; |
---|
| 767 | int nord=76; //order of integration |
---|
| 768 | double uplim,lolim; //upper and lower integration limits |
---|
| 769 | double summ,zi,yyy,answer,prolatevol; //running tally of integration |
---|
| 770 | double Pi; |
---|
| 771 | |
---|
| 772 | Pi = 4.0*atan(1.0); |
---|
| 773 | |
---|
| 774 | lolim = 0.0; |
---|
| 775 | uplim = 1.0; |
---|
| 776 | |
---|
| 777 | summ = 0.0; //initialize intergral |
---|
| 778 | |
---|
| 779 | scale = dp[0]; //make local copies in case memory moves |
---|
| 780 | crmaj = dp[1]; |
---|
| 781 | crmin = dp[2]; |
---|
| 782 | trmaj = dp[3]; |
---|
| 783 | trmin = dp[4]; |
---|
| 784 | delpc = dp[5]; |
---|
| 785 | delps = dp[6]; |
---|
| 786 | bkg = dp[7]; |
---|
| 787 | |
---|
| 788 | for(i=0;i<nord;i++) { |
---|
| 789 | zi = ( Gauss76Z[i]*(uplim-lolim) + uplim + lolim )/2.0; |
---|
| 790 | yyy = Gauss76Wt[i] * gfn2(zi,crmaj,crmin,trmaj,trmin,delpc,delps,q); |
---|
| 791 | summ += yyy; |
---|
| 792 | } |
---|
| 793 | |
---|
| 794 | answer = (uplim-lolim)/2.0*summ; |
---|
| 795 | // normalize by particle volume |
---|
| 796 | prolatevol = 4*Pi/3*trmaj*trmin*trmin; |
---|
| 797 | answer /= prolatevol; |
---|
| 798 | |
---|
| 799 | //convert to [cm-1] |
---|
| 800 | answer *= 1.0e8; |
---|
| 801 | //Scale |
---|
| 802 | answer *= scale; |
---|
| 803 | // add in the background |
---|
| 804 | answer += bkg; |
---|
| 805 | |
---|
| 806 | return answer; |
---|
| 807 | } |
---|
| 808 | |
---|
| 809 | |
---|
| 810 | /* StackedDiscsX : calculates the form factor of a stacked "tactoid" of core shell disks |
---|
| 811 | like clay platelets that are not exfoliated |
---|
| 812 | |
---|
| 813 | */ |
---|
| 814 | double |
---|
| 815 | StackedDiscs(double dp[], double q) |
---|
| 816 | { |
---|
| 817 | int i; |
---|
| 818 | double scale,length,bkg,rcore,thick,rhoc,rhol,rhosolv,N,gsd; //local variables of coefficient wave |
---|
| 819 | double va,vb,vcyl,summ,yyy,zi,halfheight,d,answer; |
---|
| 820 | int nord=76; //order of integration |
---|
| 821 | double Pi; |
---|
| 822 | |
---|
| 823 | |
---|
| 824 | Pi = 4.0*atan(1.0); |
---|
| 825 | |
---|
| 826 | va = 0.0; |
---|
| 827 | vb = Pi/2.0; |
---|
| 828 | |
---|
| 829 | summ = 0.0; //initialize intergral |
---|
| 830 | |
---|
| 831 | scale = dp[0]; |
---|
| 832 | rcore = dp[1]; |
---|
| 833 | length = dp[2]; |
---|
| 834 | thick = dp[3]; |
---|
| 835 | rhoc = dp[4]; |
---|
| 836 | rhol = dp[5]; |
---|
| 837 | rhosolv = dp[6]; |
---|
| 838 | N = dp[7]; |
---|
| 839 | gsd = dp[8]; |
---|
| 840 | bkg = dp[9]; |
---|
| 841 | |
---|
| 842 | d=2.0*thick+length; |
---|
| 843 | halfheight = length/2.0; |
---|
| 844 | |
---|
| 845 | for(i=0;i<nord;i++) { |
---|
| 846 | zi = ( Gauss76Z[i]*(vb-va) + vb + va )/2.0; |
---|
| 847 | yyy = Gauss76Wt[i] * Stackdisc_kern(q, rcore, rhoc,rhol,rhosolv, halfheight,thick,zi,gsd,d,N); |
---|
| 848 | summ += yyy; |
---|
| 849 | } |
---|
| 850 | |
---|
| 851 | answer = (vb-va)/2.0*summ; |
---|
| 852 | // length is the total core length |
---|
| 853 | vcyl=Pi*rcore*rcore*(2.0*thick+length)*N; |
---|
| 854 | answer /= vcyl; |
---|
| 855 | //Convert to [cm-1] |
---|
| 856 | answer *= 1.0e8; |
---|
| 857 | //Scale |
---|
| 858 | answer *= scale; |
---|
| 859 | // add in the background |
---|
| 860 | answer += bkg; |
---|
| 861 | |
---|
| 862 | return answer; |
---|
| 863 | } |
---|
| 864 | |
---|
| 865 | |
---|
| 866 | /* LamellarFFX : calculates the form factor of a lamellar structure - no S(q) effects included |
---|
| 867 | |
---|
| 868 | */ |
---|
| 869 | double |
---|
| 870 | LamellarFF(double dp[], double q) |
---|
| 871 | { |
---|
| 872 | double scale,del,sig,contr,bkg; //local variables of coefficient wave |
---|
| 873 | double inten, qval,Pq; |
---|
| 874 | double Pi; |
---|
| 875 | |
---|
| 876 | |
---|
| 877 | Pi = 4.0*atan(1.0); |
---|
| 878 | scale = dp[0]; |
---|
| 879 | del = dp[1]; |
---|
| 880 | sig = dp[2]*del; |
---|
| 881 | contr = dp[3]; |
---|
| 882 | bkg = dp[4]; |
---|
| 883 | |
---|
| 884 | |
---|
| 885 | Pq = 2.0*contr*contr/qval/qval*(1.0-cos(qval*del)*exp(-0.5*qval*qval*sig*sig)); |
---|
| 886 | |
---|
| 887 | inten = 2.0*Pi*scale*Pq/(qval*qval); //this is now dimensionless... |
---|
| 888 | |
---|
| 889 | inten /= del; //normalize by the thickness (in A) |
---|
| 890 | |
---|
| 891 | inten *= 1.0e8; // 1/A to 1/cm |
---|
| 892 | |
---|
| 893 | return(inten+bkg); |
---|
| 894 | } |
---|
| 895 | |
---|
| 896 | /* LamellarPSX : calculates the form factor of a lamellar structure - with S(q) effects included |
---|
| 897 | ------- |
---|
| 898 | ------- resolution effects ARE included, but only a CONSTANT default value, not the real q-dependent resolution!! |
---|
| 899 | |
---|
| 900 | */ |
---|
| 901 | double |
---|
| 902 | LamellarPS(double dp[], double q) |
---|
| 903 | { |
---|
| 904 | double scale,dd,del,sig,contr,NN,Cp,bkg; //local variables of coefficient wave |
---|
| 905 | double inten, qval,Pq,Sq,alpha,temp,t1,t2,t3,dQ; |
---|
| 906 | double Pi,Euler,dQDefault,fii; |
---|
| 907 | int ii,NNint; |
---|
| 908 | |
---|
| 909 | Euler = 0.5772156649; // Euler's constant |
---|
| 910 | dQDefault = 0.0025; //[=] 1/A, q-resolution, default value |
---|
| 911 | dQ = dQDefault; |
---|
| 912 | |
---|
| 913 | Pi = 4.0*atan(1.0); |
---|
| 914 | qval = q; |
---|
| 915 | |
---|
| 916 | scale = dp[0]; |
---|
| 917 | dd = dp[1]; |
---|
| 918 | del = dp[2]; |
---|
| 919 | sig = dp[3]*del; |
---|
| 920 | contr = dp[4]; |
---|
| 921 | NN = trunc(dp[5]); //be sure that NN is an integer |
---|
| 922 | Cp = dp[6]; |
---|
| 923 | bkg = dp[7]; |
---|
| 924 | |
---|
| 925 | Pq = 2.0*contr*contr/qval/qval*(1.0-cos(qval*del)*exp(-0.5*qval*qval*sig*sig)); |
---|
| 926 | |
---|
| 927 | NNint = (int)NN; //cast to an integer for the loop |
---|
| 928 | ii=0; |
---|
| 929 | Sq = 0.0; |
---|
| 930 | for(ii=1;ii<(NNint-1);ii+=1) { |
---|
| 931 | |
---|
| 932 | fii = (double)ii; //do I really need to do this? |
---|
| 933 | |
---|
| 934 | temp = 0.0; |
---|
| 935 | alpha = Cp/4.0/Pi/Pi*(log(Pi*ii) + Euler); |
---|
| 936 | t1 = 2.0*dQ*dQ*dd*dd*alpha; |
---|
| 937 | t2 = 2.0*qval*qval*dd*dd*alpha; |
---|
| 938 | t3 = dQ*dQ*dd*dd*ii*ii; |
---|
| 939 | |
---|
| 940 | temp = 1.0-ii/NN; |
---|
| 941 | temp *= cos(dd*qval*ii/(1.0+t1)); |
---|
| 942 | temp *= exp(-1.0*(t2 + t3)/(2.0*(1.0+t1)) ); |
---|
| 943 | temp /= sqrt(1.0+t1); |
---|
| 944 | |
---|
| 945 | Sq += temp; |
---|
| 946 | } |
---|
| 947 | |
---|
| 948 | Sq *= 2.0; |
---|
| 949 | Sq += 1.0; |
---|
| 950 | |
---|
| 951 | inten = 2.0*Pi*scale*Pq*Sq/(dd*qval*qval); |
---|
| 952 | |
---|
| 953 | inten *= 1.0e8; // 1/A to 1/cm |
---|
| 954 | |
---|
| 955 | return(inten+bkg); |
---|
| 956 | } |
---|
| 957 | |
---|
| 958 | |
---|
| 959 | /* LamellarPS_HGX : calculates the form factor of a lamellar structure - with S(q) effects included |
---|
| 960 | ------- |
---|
| 961 | ------- resolution effects ARE included, but only a CONSTANT default value, not the real q-dependent resolution!! |
---|
| 962 | |
---|
| 963 | */ |
---|
| 964 | double |
---|
| 965 | LamellarPS_HG(double dp[], double q) |
---|
| 966 | { |
---|
| 967 | double scale,dd,delT,delH,SLD_T,SLD_H,SLD_S,NN,Cp,bkg; //local variables of coefficient wave |
---|
| 968 | double inten,qval,Pq,Sq,alpha,temp,t1,t2,t3,dQ,drh,drt; |
---|
| 969 | double Pi,Euler,dQDefault,fii; |
---|
| 970 | int ii,NNint; |
---|
| 971 | |
---|
| 972 | |
---|
| 973 | Euler = 0.5772156649; // Euler's constant |
---|
| 974 | dQDefault = 0.0025; //[=] 1/A, q-resolution, default value |
---|
| 975 | dQ = dQDefault; |
---|
| 976 | |
---|
| 977 | Pi = 4.0*atan(1.0); |
---|
| 978 | qval= q; |
---|
| 979 | |
---|
| 980 | scale = dp[0]; |
---|
| 981 | dd = dp[1]; |
---|
| 982 | delT = dp[2]; |
---|
| 983 | delH = dp[3]; |
---|
| 984 | SLD_T = dp[4]; |
---|
| 985 | SLD_H = dp[5]; |
---|
| 986 | SLD_S = dp[6]; |
---|
| 987 | NN = trunc(dp[7]); //be sure that NN is an integer |
---|
| 988 | Cp = dp[8]; |
---|
| 989 | bkg = dp[9]; |
---|
| 990 | |
---|
| 991 | |
---|
| 992 | drh = SLD_H - SLD_S; |
---|
| 993 | drt = SLD_T - SLD_S; //correction 13FEB06 by L.Porcar |
---|
| 994 | |
---|
| 995 | Pq = drh*(sin(qval*(delH+delT))-sin(qval*delT)) + drt*sin(qval*delT); |
---|
| 996 | Pq *= Pq; |
---|
| 997 | Pq *= 4.0/(qval*qval); |
---|
| 998 | |
---|
| 999 | NNint = (int)NN; //cast to an integer for the loop |
---|
| 1000 | ii=0; |
---|
| 1001 | Sq = 0.0; |
---|
| 1002 | for(ii=1;ii<(NNint-1);ii+=1) { |
---|
| 1003 | |
---|
| 1004 | fii = (double)ii; //do I really need to do this? |
---|
| 1005 | |
---|
| 1006 | temp = 0.0; |
---|
| 1007 | alpha = Cp/4.0/Pi/Pi*(log(Pi*ii) + Euler); |
---|
| 1008 | t1 = 2.0*dQ*dQ*dd*dd*alpha; |
---|
| 1009 | t2 = 2.0*qval*qval*dd*dd*alpha; |
---|
| 1010 | t3 = dQ*dQ*dd*dd*ii*ii; |
---|
| 1011 | |
---|
| 1012 | temp = 1.0-ii/NN; |
---|
| 1013 | temp *= cos(dd*qval*ii/(1.0+t1)); |
---|
| 1014 | temp *= exp(-1.0*(t2 + t3)/(2.0*(1.0+t1)) ); |
---|
| 1015 | temp /= sqrt(1.0+t1); |
---|
| 1016 | |
---|
| 1017 | Sq += temp; |
---|
| 1018 | } |
---|
| 1019 | |
---|
| 1020 | Sq *= 2.0; |
---|
| 1021 | Sq += 1.0; |
---|
| 1022 | |
---|
| 1023 | inten = 2.0*Pi*scale*Pq*Sq/(dd*qval*qval); |
---|
| 1024 | |
---|
| 1025 | inten *= 1.0e8; // 1/A to 1/cm |
---|
| 1026 | |
---|
| 1027 | return(inten+bkg); |
---|
| 1028 | } |
---|
| 1029 | |
---|
| 1030 | /* LamellarFF_HGX : calculates the form factor of a lamellar structure - no S(q) effects included |
---|
| 1031 | but extra SLD for head groups is included |
---|
| 1032 | |
---|
| 1033 | */ |
---|
| 1034 | double |
---|
| 1035 | LamellarFF_HG(double dp[], double q) |
---|
| 1036 | { |
---|
| 1037 | double scale,delT,delH,slds,sldh,sldt,bkg; //local variables of coefficient wave |
---|
| 1038 | double inten, qval,Pq,drh,drt; |
---|
| 1039 | double Pi; |
---|
| 1040 | |
---|
| 1041 | |
---|
| 1042 | Pi = 4.0*atan(1.0); |
---|
| 1043 | qval= q; |
---|
| 1044 | scale = dp[0]; |
---|
| 1045 | delT = dp[1]; |
---|
| 1046 | delH = dp[2]; |
---|
| 1047 | sldt = dp[3]; |
---|
| 1048 | sldh = dp[4]; |
---|
| 1049 | slds = dp[5]; |
---|
| 1050 | bkg = dp[6]; |
---|
| 1051 | |
---|
| 1052 | |
---|
| 1053 | drh = sldh - slds; |
---|
| 1054 | drt = sldt - slds; //correction 13FEB06 by L.Porcar |
---|
| 1055 | |
---|
| 1056 | Pq = drh*(sin(qval*(delH+delT))-sin(qval*delT)) + drt*sin(qval*delT); |
---|
| 1057 | Pq *= Pq; |
---|
| 1058 | Pq *= 4.0/(qval*qval); |
---|
| 1059 | |
---|
| 1060 | inten = 2.0*Pi*scale*Pq/(qval*qval); //dimensionless... |
---|
| 1061 | |
---|
| 1062 | inten /= 2.0*(delT+delH); //normalize by the bilayer thickness |
---|
| 1063 | |
---|
| 1064 | inten *= 1.0e8; // 1/A to 1/cm |
---|
| 1065 | |
---|
| 1066 | return(inten+bkg); |
---|
| 1067 | } |
---|
| 1068 | |
---|
| 1069 | /* FlexExclVolCylX : calculates the form factor of a flexible cylinder with a circular cross section |
---|
| 1070 | -- incorporates Wei-Ren Chen's fixes - 2006 |
---|
| 1071 | |
---|
| 1072 | */ |
---|
| 1073 | double |
---|
| 1074 | FlexExclVolCyl(double dp[], double q) |
---|
| 1075 | { |
---|
| 1076 | double scale,L,B,bkg,rad,qr,cont; |
---|
| 1077 | double Pi,flex,crossSect,answer; |
---|
| 1078 | |
---|
| 1079 | |
---|
| 1080 | Pi = 4.0*atan(1.0); |
---|
| 1081 | |
---|
| 1082 | scale = dp[0]; //make local copies in case memory moves |
---|
| 1083 | L = dp[1]; |
---|
| 1084 | B = dp[2]; |
---|
| 1085 | rad = dp[3]; |
---|
| 1086 | cont = dp[4]; |
---|
| 1087 | bkg = dp[5]; |
---|
| 1088 | |
---|
| 1089 | |
---|
| 1090 | qr = q*rad; |
---|
| 1091 | |
---|
| 1092 | flex = Sk_WR(q,L,B); |
---|
| 1093 | |
---|
| 1094 | crossSect = (2.0*NR_BessJ1(qr)/qr)*(2.0*NR_BessJ1(qr)/qr); |
---|
| 1095 | flex *= crossSect; |
---|
| 1096 | flex *= Pi*rad*rad*L; |
---|
| 1097 | flex *= cont*cont; |
---|
| 1098 | flex *= 1.0e8; |
---|
| 1099 | answer = scale*flex + bkg; |
---|
| 1100 | |
---|
| 1101 | return answer; |
---|
| 1102 | } |
---|
| 1103 | |
---|
| 1104 | /* FlexCyl_EllipX : calculates the form factor of a flexible cylinder with an elliptical cross section |
---|
| 1105 | -- incorporates Wei-Ren Chen's fixes - 2006 |
---|
| 1106 | |
---|
| 1107 | */ |
---|
| 1108 | double |
---|
| 1109 | FlexCyl_Ellip(double dp[], double q) |
---|
| 1110 | { |
---|
| 1111 | double scale,L,B,bkg,rad,qr,cont,ellRatio; |
---|
| 1112 | double Pi,flex,crossSect,answer; |
---|
| 1113 | |
---|
| 1114 | |
---|
| 1115 | Pi = 4.0*atan(1.0); |
---|
| 1116 | scale = dp[0]; //make local copies in case memory moves |
---|
| 1117 | L = dp[1]; |
---|
| 1118 | B = dp[2]; |
---|
| 1119 | rad = dp[3]; |
---|
| 1120 | ellRatio = dp[4]; |
---|
| 1121 | cont = dp[5]; |
---|
| 1122 | bkg = dp[6]; |
---|
| 1123 | |
---|
| 1124 | qr = q*rad; |
---|
| 1125 | |
---|
| 1126 | flex = Sk_WR(q,L,B); |
---|
| 1127 | |
---|
| 1128 | crossSect = EllipticalCross_fn(q,rad,(rad*ellRatio)); |
---|
| 1129 | flex *= crossSect; |
---|
| 1130 | flex *= Pi*rad*rad*ellRatio*L; |
---|
| 1131 | flex *= cont*cont; |
---|
| 1132 | flex *= 1.0e8; |
---|
| 1133 | answer = scale*flex + bkg; |
---|
| 1134 | |
---|
| 1135 | return answer; |
---|
| 1136 | } |
---|
| 1137 | |
---|
| 1138 | double |
---|
| 1139 | EllipticalCross_fn(double qq, double a, double b) |
---|
| 1140 | { |
---|
| 1141 | double uplim,lolim,Pi,summ,arg,zi,yyy,answer; |
---|
| 1142 | int i,nord=76; |
---|
| 1143 | |
---|
| 1144 | Pi = 4.0*atan(1.0); |
---|
| 1145 | lolim=0.0; |
---|
| 1146 | uplim=Pi/2.0; |
---|
| 1147 | summ=0.0; |
---|
| 1148 | |
---|
| 1149 | for(i=0;i<nord;i++) { |
---|
| 1150 | zi = ( Gauss76Z[i]*(uplim-lolim) + uplim + lolim )/2.0; |
---|
| 1151 | arg = qq*sqrt(a*a*sin(zi)*sin(zi)+b*b*cos(zi)*cos(zi)); |
---|
| 1152 | yyy = pow((2.0 * NR_BessJ1(arg) / arg),2); |
---|
| 1153 | yyy *= Gauss76Wt[i]; |
---|
| 1154 | summ += yyy; |
---|
| 1155 | } |
---|
| 1156 | answer = (uplim-lolim)/2.0*summ; |
---|
| 1157 | answer *= 2.0/Pi; |
---|
| 1158 | return(answer); |
---|
| 1159 | |
---|
| 1160 | } |
---|
| 1161 | /* FlexCyl_PolyLenX : calculates the form factor of a flecible cylinder at the given x-value p->x |
---|
| 1162 | the cylinder has a polydisperse Length |
---|
| 1163 | |
---|
| 1164 | */ |
---|
| 1165 | double |
---|
| 1166 | FlexCyl_PolyLen(double dp[], double q) |
---|
| 1167 | { |
---|
| 1168 | int i; |
---|
| 1169 | double scale,radius,length,pd,delrho,bkg,lb; //local variables of coefficient wave |
---|
| 1170 | int nord=20; //order of integration |
---|
| 1171 | double uplim,lolim; //upper and lower integration limits |
---|
| 1172 | double summ,zi,yyy,answer,Vpoly; //running tally of integration |
---|
| 1173 | double range,zz,Pi; |
---|
| 1174 | |
---|
| 1175 | Pi = 4.0*atan(1.0); |
---|
| 1176 | range = 3.4; |
---|
| 1177 | |
---|
| 1178 | summ = 0.0; //initialize intergral |
---|
| 1179 | scale = dp[0]; //make local copies in case memory moves |
---|
| 1180 | length = dp[1]; //radius |
---|
| 1181 | pd = dp[2]; // average length |
---|
| 1182 | lb = dp[3]; |
---|
| 1183 | radius = dp[4]; |
---|
| 1184 | delrho = dp[5]; |
---|
| 1185 | bkg = dp[6]; |
---|
| 1186 | |
---|
| 1187 | zz = (1.0/pd)*(1.0/pd) - 1.0; |
---|
| 1188 | |
---|
| 1189 | lolim = length*(1.0-range*pd); //set the upper/lower limits to cover the distribution |
---|
| 1190 | if(lolim<0) { |
---|
| 1191 | lolim = 0; |
---|
| 1192 | } |
---|
| 1193 | if(pd>0.3) { |
---|
| 1194 | range = 3.4 + (pd-0.3)*18.0; |
---|
| 1195 | } |
---|
| 1196 | uplim = length*(1.0+range*pd); |
---|
| 1197 | |
---|
| 1198 | for(i=0;i<nord;i++) { |
---|
| 1199 | zi = ( Gauss20Z[i]*(uplim-lolim) + uplim + lolim )/2.0; |
---|
| 1200 | yyy = Gauss20Wt[i] * FlePolyLen_kernel(q,radius,length,lb,zz,delrho,zi); |
---|
| 1201 | summ += yyy; |
---|
| 1202 | } |
---|
| 1203 | |
---|
| 1204 | answer = (uplim-lolim)/2.0*summ; |
---|
| 1205 | //normalize by average cylinder volume (first moment), using the average length |
---|
| 1206 | Vpoly=Pi*radius*radius*length; |
---|
| 1207 | answer /= Vpoly; |
---|
| 1208 | |
---|
| 1209 | answer *=delrho*delrho; |
---|
| 1210 | |
---|
| 1211 | //convert to [cm-1] |
---|
| 1212 | answer *= 1.0e8; |
---|
| 1213 | //Scale |
---|
| 1214 | answer *= scale; |
---|
| 1215 | // add in the background |
---|
| 1216 | answer += bkg; |
---|
| 1217 | |
---|
| 1218 | return answer; |
---|
| 1219 | } |
---|
| 1220 | |
---|
| 1221 | /* FlexCyl_PolyLenX : calculates the form factor of a flexible cylinder at the given x-value p->x |
---|
| 1222 | the cylinder has a polydisperse cross sectional radius |
---|
| 1223 | |
---|
| 1224 | */ |
---|
| 1225 | double |
---|
| 1226 | FlexCyl_PolyRad(double dp[], double q) |
---|
| 1227 | { |
---|
| 1228 | int i; |
---|
| 1229 | double scale,radius,length,pd,delrho,bkg,lb; //local variables of coefficient wave |
---|
| 1230 | int nord=76; //order of integration |
---|
| 1231 | double uplim,lolim; //upper and lower integration limits |
---|
| 1232 | double summ,zi,yyy,answer,Vpoly; //running tally of integration |
---|
| 1233 | double range,zz,Pi; |
---|
| 1234 | |
---|
| 1235 | |
---|
| 1236 | Pi = 4.0*atan(1.0); |
---|
| 1237 | range = 3.4; |
---|
| 1238 | |
---|
| 1239 | summ = 0.0; //initialize intergral |
---|
| 1240 | |
---|
| 1241 | scale = dp[0]; //make local copies in case memory moves |
---|
| 1242 | length = dp[1]; //radius |
---|
| 1243 | lb = dp[2]; // average length |
---|
| 1244 | radius = dp[3]; |
---|
| 1245 | pd = dp[4]; |
---|
| 1246 | delrho = dp[5]; |
---|
| 1247 | bkg = dp[6]; |
---|
| 1248 | |
---|
| 1249 | zz = (1.0/pd)*(1.0/pd) - 1.0; |
---|
| 1250 | |
---|
| 1251 | lolim = radius*(1.0-range*pd); //set the upper/lower limits to cover the distribution |
---|
| 1252 | if(lolim<0) { |
---|
| 1253 | lolim = 0; |
---|
| 1254 | } |
---|
| 1255 | if(pd>0.3) { |
---|
| 1256 | range = 3.4 + (pd-0.3)*18.0; |
---|
| 1257 | } |
---|
| 1258 | uplim = radius*(1.0+range*pd); |
---|
| 1259 | |
---|
| 1260 | for(i=0;i<nord;i++) { |
---|
| 1261 | //zi = ( Gauss20Z[i]*(uplim-lolim) + uplim + lolim )/2.0; |
---|
| 1262 | //yyy = Gauss20Wt[i] * FlePolyRad_kernel(q,radius,length,lb,zz,delrho,zi); |
---|
| 1263 | zi = ( Gauss76Z[i]*(uplim-lolim) + uplim + lolim )/2.0; |
---|
| 1264 | yyy = Gauss76Wt[i] * FlePolyRad_kernel(q,radius,length,lb,zz,delrho,zi); |
---|
| 1265 | summ += yyy; |
---|
| 1266 | } |
---|
| 1267 | |
---|
| 1268 | answer = (uplim-lolim)/2.0*summ; |
---|
| 1269 | //normalize by average cylinder volume (second moment), using the average radius |
---|
| 1270 | Vpoly = Pi*radius*radius*length*(zz+2.0)/(zz+1.0); |
---|
| 1271 | answer /= Vpoly; |
---|
| 1272 | |
---|
| 1273 | answer *=delrho*delrho; |
---|
| 1274 | |
---|
| 1275 | //convert to [cm-1] |
---|
| 1276 | answer *= 1.0e8; |
---|
| 1277 | //Scale |
---|
| 1278 | answer *= scale; |
---|
| 1279 | // add in the background |
---|
| 1280 | answer += bkg; |
---|
| 1281 | |
---|
| 1282 | return answer; |
---|
| 1283 | } |
---|
| 1284 | |
---|
| 1285 | /////////functions for WRC implementation of flexible cylinders |
---|
| 1286 | static double |
---|
| 1287 | Sk_WR(double q, double L, double b) |
---|
| 1288 | { |
---|
| 1289 | // |
---|
| 1290 | double p1,p2,p1short,p2short,q0,qconnect; |
---|
| 1291 | double C,epsilon,ans,q0short,Sexvmodify,pi; |
---|
| 1292 | |
---|
| 1293 | pi = 4.0*atan(1.0); |
---|
| 1294 | |
---|
| 1295 | p1 = 4.12; |
---|
| 1296 | p2 = 4.42; |
---|
| 1297 | p1short = 5.36; |
---|
| 1298 | p2short = 5.62; |
---|
| 1299 | q0 = 3.1; |
---|
| 1300 | qconnect = q0/b; |
---|
| 1301 | // |
---|
| 1302 | q0short = fmax(1.9/sqrt(Rgsquareshort(q,L,b)),3.0); |
---|
| 1303 | |
---|
| 1304 | // |
---|
| 1305 | if(L/b > 10.0) { |
---|
| 1306 | C = 3.06/pow((L/b),0.44); |
---|
| 1307 | epsilon = 0.176; |
---|
| 1308 | } else { |
---|
| 1309 | C = 1.0; |
---|
| 1310 | epsilon = 0.170; |
---|
| 1311 | } |
---|
| 1312 | // |
---|
| 1313 | |
---|
| 1314 | if( L > 4*b ) { // Longer Chains |
---|
| 1315 | if (q*b <= 3.1) { //Modified by Yun on Oct. 15, |
---|
| 1316 | Sexvmodify = Sexvnew(q, L, b); |
---|
| 1317 | ans = Sexvmodify + C * (4.0/15.0 + 7.0/(15.0*u_WR(q,L,b)) - (11.0/15.0 + 7.0/(15.0*u_WR(q,L,b)))*exp(-u_WR(q,L,b)))*(b/L); |
---|
| 1318 | } else { //q(i)*b > 3.1 |
---|
| 1319 | ans = a1long(q, L, b, p1, p2, q0)/(pow((q*b),p1)) + a2long(q, L, b, p1, p2, q0)/(pow((q*b),p2)) + pi/(q*L); |
---|
| 1320 | } |
---|
| 1321 | } else { //L <= 4*b Shorter Chains |
---|
| 1322 | if (q*b <= fmax(1.9/sqrt(Rgsquareshort(q,L,b)),3.0) ) { |
---|
| 1323 | if (q*b<=0.01) { |
---|
| 1324 | ans = 1.0 - Rgsquareshort(q,L,b)*(q*q)/3.0; |
---|
| 1325 | } else { |
---|
| 1326 | ans = Sdebye1(q,L,b); |
---|
| 1327 | } |
---|
| 1328 | } else { //q*b > max(1.9/sqrt(Rgsquareshort(q(i),L,b)),3) |
---|
| 1329 | ans = a1short(q,L,b,p1short,p2short,q0short)/(pow((q*b),p1short)) + a2short(q,L,b,p1short,p2short,q0short)/(pow((q*b),p2short)) + pi/(q*L); |
---|
| 1330 | } |
---|
| 1331 | } |
---|
| 1332 | |
---|
| 1333 | return(ans); |
---|
| 1334 | //return(a2long(q, L, b, p1, p2, q0)); |
---|
| 1335 | } |
---|
| 1336 | |
---|
| 1337 | //WR named this w (too generic) |
---|
| 1338 | static double |
---|
| 1339 | w_WR(double x) |
---|
| 1340 | { |
---|
| 1341 | double yy; |
---|
| 1342 | yy = 0.5*(1 + tanh((x - 1.523)/0.1477)); |
---|
| 1343 | |
---|
| 1344 | return (yy); |
---|
| 1345 | } |
---|
| 1346 | |
---|
| 1347 | // |
---|
| 1348 | static double |
---|
| 1349 | u1(double q, double L, double b) |
---|
| 1350 | { |
---|
| 1351 | double yy; |
---|
| 1352 | |
---|
| 1353 | yy = Rgsquareshort(q,L,b)*q*q; |
---|
| 1354 | |
---|
| 1355 | return (yy); |
---|
| 1356 | } |
---|
| 1357 | |
---|
| 1358 | // was named u |
---|
| 1359 | static double |
---|
| 1360 | u_WR(double q, double L, double b) |
---|
| 1361 | { |
---|
| 1362 | double yy; |
---|
| 1363 | yy = Rgsquare(q,L,b)*q*q; |
---|
| 1364 | return (yy); |
---|
| 1365 | } |
---|
| 1366 | |
---|
| 1367 | |
---|
| 1368 | |
---|
| 1369 | // |
---|
| 1370 | static double |
---|
| 1371 | Rgsquarezero(double q, double L, double b) |
---|
| 1372 | { |
---|
| 1373 | double yy; |
---|
| 1374 | yy = (L*b/6.0) * (1.0 - 1.5*(b/L) + 1.5*pow((b/L),2) - 0.75*pow((b/L),3)*(1.0 - exp(-2.0*(L/b)))); |
---|
| 1375 | |
---|
| 1376 | return (yy); |
---|
| 1377 | } |
---|
| 1378 | |
---|
| 1379 | // |
---|
| 1380 | static double |
---|
| 1381 | Rgsquareshort(double q, double L, double b) |
---|
| 1382 | { |
---|
| 1383 | double yy; |
---|
| 1384 | yy = AlphaSquare(L/b) * Rgsquarezero(q,L,b); |
---|
| 1385 | |
---|
| 1386 | return (yy); |
---|
| 1387 | } |
---|
| 1388 | |
---|
| 1389 | // |
---|
| 1390 | static double |
---|
| 1391 | Rgsquare(double q, double L, double b) |
---|
| 1392 | { |
---|
| 1393 | double yy; |
---|
| 1394 | yy = AlphaSquare(L/b)*L*b/6.0; |
---|
| 1395 | |
---|
| 1396 | return (yy); |
---|
| 1397 | } |
---|
| 1398 | |
---|
| 1399 | // |
---|
| 1400 | static double |
---|
| 1401 | AlphaSquare(double x) |
---|
| 1402 | { |
---|
| 1403 | double yy; |
---|
| 1404 | yy = pow( (1.0 + (x/3.12)*(x/3.12) + (x/8.67)*(x/8.67)*(x/8.67)),(0.176/3.0) ); |
---|
| 1405 | |
---|
| 1406 | return (yy); |
---|
| 1407 | } |
---|
| 1408 | |
---|
| 1409 | // ?? funciton is not used - but should the log actually be log10??? |
---|
| 1410 | static double |
---|
| 1411 | miu(double x) |
---|
| 1412 | { |
---|
| 1413 | double yy; |
---|
| 1414 | yy = (1.0/8.0)*(9.0*x - 2.0 + 2.0*log(1.0 + x)/x)*exp(1.0/2.565*(1.0/x + (1.0 - 1.0/(x*x))*log(1.0 + x))); |
---|
| 1415 | |
---|
| 1416 | return (yy); |
---|
| 1417 | } |
---|
| 1418 | |
---|
| 1419 | // |
---|
| 1420 | static double |
---|
| 1421 | Sdebye(double q, double L, double b) |
---|
| 1422 | { |
---|
| 1423 | double yy; |
---|
| 1424 | yy = 2.0*(exp(-u_WR(q,L,b)) + u_WR(q,L,b) -1.0)/(pow((u_WR(q,L,b)),2)); |
---|
| 1425 | |
---|
| 1426 | return (yy); |
---|
| 1427 | } |
---|
| 1428 | |
---|
| 1429 | // |
---|
| 1430 | static double |
---|
| 1431 | Sdebye1(double q, double L, double b) |
---|
| 1432 | { |
---|
| 1433 | double yy; |
---|
| 1434 | yy = 2.0*(exp(-u1(q,L,b)) + u1(q,L,b) -1.0)/( pow((u1(q,L,b)),2.0) ); |
---|
| 1435 | |
---|
| 1436 | return (yy); |
---|
| 1437 | } |
---|
| 1438 | |
---|
| 1439 | // |
---|
| 1440 | static double |
---|
| 1441 | Sexv(double q, double L, double b) |
---|
| 1442 | { |
---|
| 1443 | double yy,C1,C2,C3,miu,Rg2; |
---|
| 1444 | C1=1.22; |
---|
| 1445 | C2=0.4288; |
---|
| 1446 | C3=-1.651; |
---|
| 1447 | miu = 0.585; |
---|
| 1448 | |
---|
| 1449 | Rg2 = Rgsquare(q,L,b); |
---|
| 1450 | |
---|
| 1451 | yy = (1.0 - w_WR(q*sqrt(Rg2)))*Sdebye(q,L,b) + w_WR(q*sqrt(Rg2))*(C1*pow((q*sqrt(Rg2)),(-1.0/miu)) + C2*pow((q*sqrt(Rg2)),(-2.0/miu)) + C3*pow((q*sqrt(Rg2)),(-3.0/miu))); |
---|
| 1452 | |
---|
| 1453 | return (yy); |
---|
| 1454 | } |
---|
| 1455 | |
---|
| 1456 | // this must be WR modified version |
---|
| 1457 | static double |
---|
| 1458 | Sexvnew(double q, double L, double b) |
---|
| 1459 | { |
---|
| 1460 | double yy,C1,C2,C3,miu; |
---|
| 1461 | double del=1.05,C_star2,Rg2; |
---|
| 1462 | |
---|
| 1463 | C1=1.22; |
---|
| 1464 | C2=0.4288; |
---|
| 1465 | C3=-1.651; |
---|
| 1466 | miu = 0.585; |
---|
| 1467 | |
---|
| 1468 | //calculating the derivative to decide on the corection (cutoff) term? |
---|
| 1469 | // I have modified this from WRs original code |
---|
| 1470 | |
---|
| 1471 | if( (Sexv(q*del,L,b)-Sexv(q,L,b))/(q*del - q) >= 0.0 ) { |
---|
| 1472 | C_star2 = 0.0; |
---|
| 1473 | } else { |
---|
| 1474 | C_star2 = 1.0; |
---|
| 1475 | } |
---|
| 1476 | |
---|
| 1477 | Rg2 = Rgsquare(q,L,b); |
---|
| 1478 | |
---|
| 1479 | yy = (1.0 - w_WR(q*sqrt(Rg2)))*Sdebye(q,L,b) + C_star2*w_WR(q*sqrt(Rg2))*(C1*pow((q*sqrt(Rg2)),(-1.0/miu)) + C2*pow((q*sqrt(Rg2)),(-2.0/miu)) + C3*pow((q*sqrt(Rg2)),(-3.0/miu))); |
---|
| 1480 | |
---|
| 1481 | return (yy); |
---|
| 1482 | } |
---|
| 1483 | |
---|
| 1484 | // these are the messy ones |
---|
| 1485 | static double |
---|
| 1486 | a2short(double q, double L, double b, double p1short, double p2short, double q0) |
---|
| 1487 | { |
---|
| 1488 | double yy,Rg2_sh; |
---|
| 1489 | double t1,E,Rg2_sh2,Et1,Emt1,q02,q0p; |
---|
| 1490 | double pi; |
---|
| 1491 | |
---|
| 1492 | E = 2.718281828459045091; |
---|
| 1493 | pi = 4.0*atan(1.0); |
---|
| 1494 | Rg2_sh = Rgsquareshort(q,L,b); |
---|
| 1495 | Rg2_sh2 = Rg2_sh*Rg2_sh; |
---|
| 1496 | t1 = ((q0*q0*Rg2_sh)/(b*b)); |
---|
| 1497 | Et1 = pow(E,t1); |
---|
| 1498 | Emt1 =pow(E,-t1); |
---|
| 1499 | q02 = q0*q0; |
---|
| 1500 | q0p = pow(q0,(-4.0 + p2short) ); |
---|
| 1501 | |
---|
| 1502 | //E is the number e |
---|
| 1503 | yy = ((-(1.0/(L*((p1short - p2short))*Rg2_sh2)*((b*Emt1*q0p*((8.0*b*b*b*L - 8.0*b*b*b*Et1*L - 2.0*b*b*b*L*p1short + 2.0*b*b*b*Et1*L*p1short + 4.0*b*L*q02*Rg2_sh + 4.0*b*Et1*L*q02*Rg2_sh - 2.0*b*Et1*L*p1short*q02*Rg2_sh - Et1*pi*q02*q0*Rg2_sh2 + Et1*p1short*pi*q02*q0*Rg2_sh2))))))); |
---|
| 1504 | |
---|
| 1505 | return (yy); |
---|
| 1506 | } |
---|
| 1507 | |
---|
| 1508 | // |
---|
| 1509 | static double |
---|
| 1510 | a1short(double q, double L, double b, double p1short, double p2short, double q0) |
---|
| 1511 | { |
---|
| 1512 | double yy,Rg2_sh; |
---|
| 1513 | double t1,E,Rg2_sh2,Et1,Emt1,q02,q0p,b3; |
---|
| 1514 | double pi; |
---|
| 1515 | |
---|
| 1516 | E = 2.718281828459045091; |
---|
| 1517 | pi = 4.0*atan(1.0); |
---|
| 1518 | Rg2_sh = Rgsquareshort(q,L,b); |
---|
| 1519 | Rg2_sh2 = Rg2_sh*Rg2_sh; |
---|
| 1520 | b3 = b*b*b; |
---|
| 1521 | t1 = ((q0*q0*Rg2_sh)/(b*b)); |
---|
| 1522 | Et1 = pow(E,t1); |
---|
| 1523 | Emt1 =pow(E,-t1); |
---|
| 1524 | q02 = q0*q0; |
---|
| 1525 | q0p = pow(q0,(-4.0 + p1short) ); |
---|
| 1526 | |
---|
| 1527 | yy = ((1.0/(L*((p1short - p2short))*Rg2_sh2)*((b*Emt1*q0p*((8.0*b3*L - 8.0*b3*Et1*L - 2.0*b3*L*p2short + 2.0*b3*Et1*L*p2short + 4.0*b*L*q02*Rg2_sh + 4.0*b*Et1*L*q02*Rg2_sh - 2.0*b*Et1*L*p2short*q02*Rg2_sh - Et1*pi*q02*q0*Rg2_sh2 + Et1*p2short*pi*q02*q0*Rg2_sh2)))))); |
---|
| 1528 | |
---|
| 1529 | return(yy); |
---|
| 1530 | } |
---|
| 1531 | |
---|
| 1532 | // this one will be lots of trouble |
---|
| 1533 | static double |
---|
| 1534 | a2long(double q, double L, double b, double p1, double p2, double q0) |
---|
| 1535 | { |
---|
| 1536 | double yy,C1,C2,C3,C4,C5,miu,C,Rg2; |
---|
| 1537 | double t1,t2,t3,t4,t5,t6,t7,t8,t9,t10,pi; |
---|
| 1538 | double E,b2,b3,b4,q02,q03,q04,q05,Rg22; |
---|
| 1539 | |
---|
| 1540 | pi = 4.0*atan(1.0); |
---|
| 1541 | E = 2.718281828459045091; |
---|
| 1542 | if( L/b > 10.0) { |
---|
| 1543 | C = 3.06/pow((L/b),0.44); |
---|
| 1544 | } else { |
---|
| 1545 | C = 1.0; |
---|
| 1546 | } |
---|
| 1547 | |
---|
| 1548 | C1 = 1.22; |
---|
| 1549 | C2 = 0.4288; |
---|
| 1550 | C3 = -1.651; |
---|
| 1551 | C4 = 1.523; |
---|
| 1552 | C5 = 0.1477; |
---|
| 1553 | miu = 0.585; |
---|
| 1554 | |
---|
| 1555 | Rg2 = Rgsquare(q,L,b); |
---|
| 1556 | Rg22 = Rg2*Rg2; |
---|
| 1557 | b2 = b*b; |
---|
| 1558 | b3 = b*b*b; |
---|
| 1559 | b4 = b3*b; |
---|
| 1560 | q02 = q0*q0; |
---|
| 1561 | q03 = q0*q0*q0; |
---|
| 1562 | q04 = q03*q0; |
---|
| 1563 | q05 = q04*q0; |
---|
| 1564 | |
---|
| 1565 | t1 = (1.0/(b* p1*pow(q0,((-1.0) - p1 - p2)) - b*p2*pow(q0,((-1.0) - p1 - p2)) )); |
---|
| 1566 | |
---|
| 1567 | t2 = (b*C*(((-1.0*((14.0*b3)/(15.0*q03*Rg2))) + (14*b3*pow(E,(-((q02*Rg2)/b2))))/(15*q03*Rg2) + (2*pow(E,(-((q02*Rg2)/b2)))*q0*((11.0/15.0 + (7*b2)/(15*q02*Rg2)))*Rg2)/b)))/L; |
---|
| 1568 | |
---|
| 1569 | t3 = (sqrt(Rg2)*((C3*pow((((sqrt(Rg2)*q0)/b)),((-3)/miu)) + C2*pow((((sqrt(Rg2)*q0)/b)),((-2)/miu)) + C1*pow((((sqrt(Rg2)*q0)/b)),((-1.0)/miu))))*pow(sech_WR(((-C4) + (sqrt(Rg2)*q0)/b)/C5),2))/(2*C5); |
---|
| 1570 | |
---|
| 1571 | t4 = (b4*sqrt(Rg2)*(((-1) + pow(E,(-((q02*Rg2)/b2))) + (q02*Rg2)/b2))*pow(sech_WR(((-C4) + (sqrt(Rg2)*q0)/b)/C5),2))/(C5*q04*Rg22); |
---|
| 1572 | |
---|
| 1573 | t5 = (2*b4*(((2*q0*Rg2)/b - (2*pow(E,(-((q02*Rg2)/b2)))*q0*Rg2)/b))*((1 + 1.0/2.0*(((-1) - tanh(((-C4) + (sqrt(Rg2)*q0)/b)/C5))))))/(q04*Rg22); |
---|
| 1574 | |
---|
| 1575 | t6 = (8*b4*b*(((-1) + pow(E,(-((q02*Rg2)/b2))) + (q02*Rg2)/b2))*((1 + 1.0/2.0*(((-1) - tanh(((-C4) + (sqrt(Rg2)*q0)/b)/C5))))))/(q05*Rg22); |
---|
| 1576 | |
---|
| 1577 | t7 = (((-((3*C3*sqrt(Rg2)*pow((((sqrt(Rg2)*q0)/b)),((-1) - 3/miu)))/miu)) - (2*C2*sqrt(Rg2)*pow((((sqrt(Rg2)*q0)/b)),((-1) - 2/miu)))/miu - (C1*sqrt(Rg2)*pow((((sqrt(Rg2)*q0)/b)),((-1) - 1/miu)))/miu)); |
---|
| 1578 | |
---|
| 1579 | t8 = ((1 + tanh(((-C4) + (sqrt(Rg2)*q0)/b)/C5))); |
---|
| 1580 | |
---|
| 1581 | t9 = (b*C*((4.0/15.0 - pow(E,(-((q02*Rg2)/b2)))*((11.0/15.0 + (7*b2)/(15*q02*Rg2))) + (7*b2)/(15*q02*Rg2))))/L; |
---|
| 1582 | |
---|
| 1583 | t10 = (2*b4*(((-1) + pow(E,(-((q02*Rg2)/b2))) + (q02*Rg2)/b2))*((1 + 1.0/2.0*(((-1) - tanh(((-C4) + (sqrt(Rg2)*q0)/b)/C5))))))/(q04*Rg22); |
---|
| 1584 | |
---|
| 1585 | |
---|
| 1586 | yy = ((-1*(t1* ((-pow(q0,-p1)*(((b2*pi)/(L*q02) + t2 + t3 - t4 + t5 - t6 + 1.0/2.0*t7*t8)) - b*p1*pow(q0,((-1) - p1))*(((-((b*pi)/(L*q0))) + t9 + t10 + 1.0/2.0*((C3*pow((((sqrt(Rg2)*q0)/b)),((-3)/miu)) + C2*pow((((sqrt(Rg2)*q0)/b)),((-2)/miu)) + C1*pow((((sqrt(Rg2)*q0)/b)),((-1)/miu))))*((1 + tanh(((-C4) + (sqrt(Rg2)*q0)/b)/C5)))))))))); |
---|
| 1587 | |
---|
| 1588 | return (yy); |
---|
| 1589 | } |
---|
| 1590 | |
---|
| 1591 | //need to define this on my own |
---|
| 1592 | static double |
---|
| 1593 | sech_WR(double x) |
---|
| 1594 | { |
---|
| 1595 | return(1/cosh(x)); |
---|
| 1596 | } |
---|
| 1597 | |
---|
| 1598 | // |
---|
| 1599 | static double |
---|
| 1600 | a1long(double q, double L, double b, double p1, double p2, double q0) |
---|
| 1601 | { |
---|
| 1602 | double yy,C,C1,C2,C3,C4,C5,miu,Rg2; |
---|
| 1603 | double t1,t2,t3,t4,t5,t6,t7,t8,t9,t10,t11,t12,t13,t14,t15; |
---|
| 1604 | double E,pi; |
---|
| 1605 | double b2,b3,b4,q02,q03,q04,q05,Rg22; |
---|
| 1606 | |
---|
| 1607 | pi = 4.0*atan(1.0); |
---|
| 1608 | E = 2.718281828459045091; |
---|
| 1609 | |
---|
| 1610 | if( L/b > 10.0) { |
---|
| 1611 | C = 3.06/pow((L/b),0.44); |
---|
| 1612 | } else { |
---|
| 1613 | C = 1.0; |
---|
| 1614 | } |
---|
| 1615 | |
---|
| 1616 | C1 = 1.22; |
---|
| 1617 | C2 = 0.4288; |
---|
| 1618 | C3 = -1.651; |
---|
| 1619 | C4 = 1.523; |
---|
| 1620 | C5 = 0.1477; |
---|
| 1621 | miu = 0.585; |
---|
| 1622 | |
---|
| 1623 | Rg2 = Rgsquare(q,L,b); |
---|
| 1624 | Rg22 = Rg2*Rg2; |
---|
| 1625 | b2 = b*b; |
---|
| 1626 | b3 = b*b*b; |
---|
| 1627 | b4 = b3*b; |
---|
| 1628 | q02 = q0*q0; |
---|
| 1629 | q03 = q0*q0*q0; |
---|
| 1630 | q04 = q03*q0; |
---|
| 1631 | q05 = q04*q0; |
---|
| 1632 | |
---|
| 1633 | t1 = (b*C*((4.0/15.0 - pow(E,(-((q02*Rg2)/b2)))*((11.0/15.0 + (7.0*b2)/(15.0*q02*Rg2))) + (7.0*b2)/(15.0*q02*Rg2)))); |
---|
| 1634 | |
---|
| 1635 | t2 = (2.0*b4*(((-1.0) + pow(E,(-((q02*Rg2)/b2))) + (q02*Rg2)/b2))*((1.0 + 1.0/2.0*(((-1.0) - tanh(((-C4) + (sqrt(Rg2)*q0)/b)/C5)))))); |
---|
| 1636 | |
---|
| 1637 | t3 = ((C3*pow((((sqrt(Rg2)*q0)/b)),((-3.0)/miu)) + C2*pow((((sqrt(Rg2)*q0)/b)),((-2.0)/miu)) + C1*pow((((sqrt(Rg2)*q0)/b)),((-1.0)/miu)))); |
---|
| 1638 | |
---|
| 1639 | t4 = ((1.0 + tanh(((-C4) + (sqrt(Rg2)*q0)/b)/C5))); |
---|
| 1640 | |
---|
| 1641 | t5 = (1.0/(b*p1*pow(q0,((-1.0) - p1 - p2)) - b*p2*pow(q0,((-1.0) - p1 - p2)))); |
---|
| 1642 | |
---|
| 1643 | t6 = (b*C*(((-((14.0*b3)/(15.0*q03*Rg2))) + (14.0*b3*pow(E,(-((q02*Rg2)/b2))))/(15.0*q03*Rg2) + (2.0*pow(E,(-((q02*Rg2)/b2)))*q0*((11.0/15.0 + (7.0*b2)/(15.0*q02*Rg2)))*Rg2)/b))); |
---|
| 1644 | |
---|
| 1645 | t7 = (sqrt(Rg2)*((C3*pow((((sqrt(Rg2)*q0)/b)),((-3.0)/miu)) + C2*pow((((sqrt(Rg2)*q0)/b)),((-2.0)/miu)) + C1*pow((((sqrt(Rg2)*q0)/b)),((-1.0)/miu))))*pow(sech_WR(((-C4) + (sqrt(Rg2)*q0)/b)/C5),2)); |
---|
| 1646 | |
---|
| 1647 | t8 = (b4*sqrt(Rg2)*(((-1.0) + pow(E,(-((q02*Rg2)/b2))) + (q02*Rg2)/b2))*pow(sech_WR(((-C4) + (sqrt(Rg2)*q0)/b)/C5),2)); |
---|
| 1648 | |
---|
| 1649 | t9 = (2.0*b4*(((2.0*q0*Rg2)/b - (2.0*pow(E,(-((q02*Rg2)/b2)))*q0*Rg2)/b))*((1.0 + 1.0/2.0*(((-1.0) - tanh(((-C4) + (sqrt(Rg2)*q0)/b)/C5)))))); |
---|
| 1650 | |
---|
| 1651 | t10 = (8.0*b4*b*(((-1.0) + pow(E,(-((q02*Rg2)/b2))) + (q02*Rg2)/b2))*((1.0 + 1.0/2.0*(((-1.0) - tanh(((-C4) + (sqrt(Rg2)*q0)/b)/C5)))))); |
---|
| 1652 | |
---|
| 1653 | t11 = (((-((3.0*C3*sqrt(Rg2)*pow((((sqrt(Rg2)*q0)/b)),((-1.0) - 3.0/miu)))/miu)) - (2.0*C2*sqrt(Rg2)*pow((((sqrt(Rg2)*q0)/b)),((-1.0) - 2.0/miu)))/miu - (C1*sqrt(Rg2)*pow((((sqrt(Rg2)*q0)/b)),((-1.0) - 1.0/miu)))/miu)); |
---|
| 1654 | |
---|
| 1655 | t12 = ((1.0 + tanh(((-C4) + (sqrt(Rg2)*q0)/b)/C5))); |
---|
| 1656 | |
---|
| 1657 | t13 = (b*C*((4.0/15.0 - pow(E,(-((q02*Rg2)/b2)))*((11.0/15.0 + (7.0*b2)/(15.0*q02* Rg2))) + (7.0*b2)/(15.0*q02*Rg2)))); |
---|
| 1658 | |
---|
| 1659 | t14 = (2.0*b4*(((-1.0) + pow(E,(-((q02*Rg2)/b2))) + (q02*Rg2)/b2))*((1.0 + 1.0/2.0*(((-1.0) - tanh(((-C4) + (sqrt(Rg2)*q0)/b)/C5)))))); |
---|
| 1660 | |
---|
| 1661 | t15 = ((C3*pow((((sqrt(Rg2)*q0)/b)),((-3.0)/miu)) + C2*pow((((sqrt(Rg2)*q0)/b)),((-2.0)/miu)) + C1*pow((((sqrt(Rg2)*q0)/b)),((-1.0)/miu)))); |
---|
| 1662 | |
---|
| 1663 | |
---|
| 1664 | yy = (pow(q0,p1)*(((-((b*pi)/(L*q0))) +t1/L +t2/(q04*Rg22) + 1.0/2.0*t3*t4)) + (t5*((pow(q0,(p1 - p2))*(((-pow(q0,(-p1)))*(((b2*pi)/(L*q02) +t6/L +t7/(2.0*C5) -t8/(C5*q04*Rg22) +t9/(q04*Rg22) -t10/(q05*Rg22) + 1.0/2.0*t11*t12)) - b*p1*pow(q0,((-1.0) - p1))*(((-((b*pi)/(L*q0))) +t13/L +t14/(q04*Rg22) + 1.0/2.0*t15*((1.0 + tanh(((-C4) + (sqrt(Rg2)*q0)/b)/C5))))))))))); |
---|
| 1665 | |
---|
| 1666 | return (yy); |
---|
| 1667 | } |
---|
| 1668 | |
---|
| 1669 | |
---|
| 1670 | |
---|
| 1671 | /////////////// |
---|
| 1672 | |
---|
| 1673 | // |
---|
| 1674 | // FUNCTION gfn2: CONTAINS F(Q,A,B,mu)**2 AS GIVEN |
---|
| 1675 | // BY (53) AND (56,57) IN CHEN AND |
---|
| 1676 | // KOTLARCHYK REFERENCE |
---|
| 1677 | // |
---|
| 1678 | // <PROLATE ELLIPSOIDS> |
---|
| 1679 | // |
---|
| 1680 | double |
---|
| 1681 | gfn2(double xx, double crmaj, double crmin, double trmaj, double trmin, double delpc, double delps, double qq) |
---|
| 1682 | { |
---|
| 1683 | // local variables |
---|
| 1684 | double aa,bb,u2,ut2,uq,ut,vc,vt,gfnc,gfnt,gfn2,pi43,gfn,Pi; |
---|
| 1685 | |
---|
| 1686 | Pi = 4.0*atan(1.0); |
---|
| 1687 | |
---|
| 1688 | pi43=4.0/3.0*Pi; |
---|
| 1689 | aa = crmaj; |
---|
| 1690 | bb = crmin; |
---|
| 1691 | u2 = (aa*aa*xx*xx + bb*bb*(1.0-xx*xx)); |
---|
| 1692 | ut2 = (trmaj*trmaj*xx*xx + trmin*trmin*(1.0-xx*xx)); |
---|
| 1693 | uq = sqrt(u2)*qq; |
---|
| 1694 | ut= sqrt(ut2)*qq; |
---|
| 1695 | vc = pi43*aa*bb*bb; |
---|
| 1696 | vt = pi43*trmaj*trmin*trmin; |
---|
| 1697 | gfnc = 3.0*(sin(uq)/uq/uq - cos(uq)/uq)/uq*vc*delpc; |
---|
| 1698 | gfnt = 3.0*(sin(ut)/ut/ut - cos(ut)/ut)/ut*vt*delps; |
---|
| 1699 | gfn = gfnc+gfnt; |
---|
| 1700 | gfn2 = gfn*gfn; |
---|
| 1701 | |
---|
| 1702 | return (gfn2); |
---|
| 1703 | } |
---|
| 1704 | |
---|
| 1705 | // |
---|
| 1706 | // FUNCTION gfn4: CONTAINS F(Q,A,B,MU)**2 AS GIVEN |
---|
| 1707 | // BY (53) & (58-59) IN CHEN AND |
---|
| 1708 | // KOTLARCHYK REFERENCE |
---|
| 1709 | // |
---|
| 1710 | // <OBLATE ELLIPSOID> |
---|
| 1711 | // function gfn4 for oblate ellipsoids |
---|
| 1712 | double |
---|
| 1713 | gfn4(double xx, double crmaj, double crmin, double trmaj, double trmin, double delpc, double delps, double qq) |
---|
| 1714 | { |
---|
| 1715 | // local variables |
---|
| 1716 | double aa,bb,u2,ut2,uq,ut,vc,vt,gfnc,gfnt,tgfn,gfn4,pi43,Pi; |
---|
| 1717 | |
---|
| 1718 | Pi = 4.0*atan(1.0); |
---|
| 1719 | pi43=4.0/3.0*Pi; |
---|
| 1720 | aa = crmaj; |
---|
| 1721 | bb = crmin; |
---|
| 1722 | u2 = (bb*bb*xx*xx + aa*aa*(1.0-xx*xx)); |
---|
| 1723 | ut2 = (trmin*trmin*xx*xx + trmaj*trmaj*(1.0-xx*xx)); |
---|
| 1724 | uq = sqrt(u2)*qq; |
---|
| 1725 | ut= sqrt(ut2)*qq; |
---|
| 1726 | vc = pi43*aa*aa*bb; |
---|
| 1727 | vt = pi43*trmaj*trmaj*trmin; |
---|
| 1728 | gfnc = 3.0*(sin(uq)/uq/uq - cos(uq)/uq)/uq*vc*delpc; |
---|
| 1729 | gfnt = 3.0*(sin(ut)/ut/ut - cos(ut)/ut)/ut*vt*delps; |
---|
| 1730 | tgfn = gfnc+gfnt; |
---|
| 1731 | gfn4 = tgfn*tgfn; |
---|
| 1732 | |
---|
| 1733 | return (gfn4); |
---|
| 1734 | } |
---|
| 1735 | |
---|
| 1736 | double |
---|
| 1737 | FlePolyLen_kernel(double q, double radius, double length, double lb, double zz, double delrho, double zi) |
---|
| 1738 | { |
---|
| 1739 | double Pq,vcyl,dl; |
---|
| 1740 | double Pi,qr; |
---|
| 1741 | |
---|
| 1742 | Pi = 4.0*atan(1.0); |
---|
| 1743 | qr = q*radius; |
---|
| 1744 | |
---|
| 1745 | Pq = Sk_WR(q,zi,lb); //does not have cross section term |
---|
| 1746 | Pq *= (2.0*NR_BessJ1(qr)/qr)*(2.0*NR_BessJ1(qr)/qr); |
---|
| 1747 | |
---|
| 1748 | vcyl=Pi*radius*radius*zi; |
---|
| 1749 | Pq *= vcyl*vcyl; |
---|
| 1750 | |
---|
| 1751 | dl = SchulzPoint_cpr(zi,length,zz); |
---|
| 1752 | return (Pq*dl); |
---|
| 1753 | |
---|
| 1754 | } |
---|
| 1755 | |
---|
| 1756 | double |
---|
| 1757 | FlePolyRad_kernel(double q, double ravg, double Lc, double Lb, double zz, double delrho, double zi) |
---|
| 1758 | { |
---|
| 1759 | double Pq,vcyl,dr; |
---|
| 1760 | double Pi,qr; |
---|
| 1761 | |
---|
| 1762 | Pi = 4.0*atan(1.0); |
---|
| 1763 | qr = q*zi; |
---|
| 1764 | |
---|
| 1765 | Pq = Sk_WR(q,Lc,Lb); //does not have cross section term |
---|
| 1766 | Pq *= (2.0*NR_BessJ1(qr)/qr)*(2.0*NR_BessJ1(qr)/qr); |
---|
| 1767 | |
---|
| 1768 | vcyl=Pi*zi*zi*Lc; |
---|
| 1769 | Pq *= vcyl*vcyl; |
---|
| 1770 | |
---|
| 1771 | dr = SchulzPoint_cpr(zi,ravg,zz); |
---|
| 1772 | return (Pq*dr); |
---|
| 1773 | |
---|
| 1774 | } |
---|
| 1775 | |
---|
| 1776 | double |
---|
| 1777 | CSCylIntegration(double qq, double rad, double radthick, double facthick, double rhoc, double rhos, double rhosolv, double length) |
---|
| 1778 | { |
---|
| 1779 | double answer,halfheight,Pi; |
---|
| 1780 | double lolim,uplim,summ,yyy,zi; |
---|
| 1781 | int nord,i; |
---|
| 1782 | |
---|
| 1783 | // set up the integration end points |
---|
| 1784 | Pi = 4.0*atan(1.0); |
---|
| 1785 | nord = 76; |
---|
| 1786 | lolim = 0; |
---|
| 1787 | uplim = Pi/2; |
---|
| 1788 | halfheight = length/2.0; |
---|
| 1789 | |
---|
| 1790 | summ = 0.0; // initialize integral |
---|
| 1791 | i=0; |
---|
| 1792 | for(i=0;i<nord;i++) { |
---|
| 1793 | zi = ( Gauss76Z[i]*(uplim-lolim) + uplim + lolim )/2.0; |
---|
| 1794 | yyy = Gauss76Wt[i] * CScyl(qq, rad, radthick, facthick, rhoc,rhos,rhosolv, halfheight, zi); |
---|
| 1795 | summ += yyy; |
---|
| 1796 | } |
---|
| 1797 | |
---|
| 1798 | // calculate value of integral to return |
---|
| 1799 | answer = (uplim-lolim)/2.0*summ; |
---|
| 1800 | return (answer); |
---|
| 1801 | } |
---|
| 1802 | |
---|
| 1803 | double |
---|
| 1804 | CScyl(double qq, double rad, double radthick, double facthick, double rhoc, double rhos, double rhosolv, double length, double dum) |
---|
| 1805 | { |
---|
| 1806 | // qq is the q-value for the calculation (1/A) |
---|
| 1807 | // radius is the core radius of the cylinder (A) |
---|
| 1808 | // radthick and facthick are the radial and face layer thicknesses |
---|
| 1809 | // rho(n) are the respective SLD's |
---|
| 1810 | // length is the *Half* CORE-LENGTH of the cylinder |
---|
| 1811 | // dum is the dummy variable for the integration (theta) |
---|
| 1812 | |
---|
| 1813 | double dr1,dr2,besarg1,besarg2,vol1,vol2,sinarg1,sinarg2,t1,t2,retval; |
---|
| 1814 | double Pi; |
---|
| 1815 | |
---|
| 1816 | Pi = 4.0*atan(1.0); |
---|
| 1817 | |
---|
| 1818 | dr1 = rhoc-rhos; |
---|
| 1819 | dr2 = rhos-rhosolv; |
---|
| 1820 | vol1 = Pi*rad*rad*(2.0*length); |
---|
| 1821 | vol2 = Pi*(rad+radthick)*(rad+radthick)*(2.0*length+2.0*facthick); |
---|
| 1822 | |
---|
| 1823 | besarg1 = qq*rad*sin(dum); |
---|
| 1824 | besarg2 = qq*(rad+radthick)*sin(dum); |
---|
| 1825 | sinarg1 = qq*length*cos(dum); |
---|
| 1826 | sinarg2 = qq*(length+facthick)*cos(dum); |
---|
| 1827 | |
---|
| 1828 | t1 = 2.0*vol1*dr1*sin(sinarg1)/sinarg1*NR_BessJ1(besarg1)/besarg1; |
---|
| 1829 | t2 = 2.0*vol2*dr2*sin(sinarg2)/sinarg2*NR_BessJ1(besarg2)/besarg2; |
---|
| 1830 | |
---|
| 1831 | retval = ((t1+t2)*(t1+t2))*sin(dum); |
---|
| 1832 | return (retval); |
---|
| 1833 | |
---|
| 1834 | } |
---|
| 1835 | |
---|
| 1836 | |
---|
| 1837 | double |
---|
| 1838 | CoreShellCylKernel(double qq, double rcore, double thick, double rhoc, double rhos, double rhosolv, double length, double dum) |
---|
| 1839 | { |
---|
| 1840 | |
---|
| 1841 | double dr1,dr2,besarg1,besarg2,vol1,vol2,sinarg1,sinarg2,t1,t2,retval; |
---|
| 1842 | double Pi; |
---|
| 1843 | |
---|
| 1844 | Pi = 4.0*atan(1.0); |
---|
| 1845 | |
---|
| 1846 | dr1 = rhoc-rhos; |
---|
| 1847 | dr2 = rhos-rhosolv; |
---|
| 1848 | vol1 = Pi*rcore*rcore*(2.0*length); |
---|
| 1849 | vol2 = Pi*(rcore+thick)*(rcore+thick)*(2.0*length+2.0*thick); |
---|
| 1850 | |
---|
| 1851 | besarg1 = qq*rcore*sin(dum); |
---|
| 1852 | besarg2 = qq*(rcore+thick)*sin(dum); |
---|
| 1853 | sinarg1 = qq*length*cos(dum); |
---|
| 1854 | sinarg2 = qq*(length+thick)*cos(dum); |
---|
| 1855 | |
---|
| 1856 | t1 = 2.0*vol1*dr1*sin(sinarg1)/sinarg1*NR_BessJ1(besarg1)/besarg1; |
---|
| 1857 | t2 = 2.0*vol2*dr2*sin(sinarg2)/sinarg2*NR_BessJ1(besarg2)/besarg2; |
---|
| 1858 | |
---|
| 1859 | retval = ((t1+t2)*(t1+t2))*sin(dum); |
---|
| 1860 | |
---|
| 1861 | return (retval); |
---|
| 1862 | } |
---|
| 1863 | |
---|
| 1864 | double |
---|
| 1865 | Cyl_PolyLenKernel(double q, double radius, double len_avg, double zz, double delrho, double dumLen) |
---|
| 1866 | { |
---|
| 1867 | |
---|
| 1868 | double halfheight,uplim,lolim,zi,summ,yyy,Pi; |
---|
| 1869 | double answer,dr,Vcyl; |
---|
| 1870 | int i,nord; |
---|
| 1871 | |
---|
| 1872 | Pi = 4.0*atan(1.0); |
---|
| 1873 | lolim = 0; |
---|
| 1874 | uplim = Pi/2.0; |
---|
| 1875 | halfheight = dumLen/2.0; |
---|
| 1876 | nord=20; |
---|
| 1877 | summ = 0.0; |
---|
| 1878 | |
---|
| 1879 | //do the cylinder orientational average |
---|
| 1880 | for(i=0;i<nord;i++) { |
---|
| 1881 | zi = ( Gauss20Z[i]*(uplim-lolim) + uplim + lolim )/2.0; |
---|
| 1882 | yyy = Gauss20Wt[i] * CylKernel(q, radius, halfheight, zi); |
---|
| 1883 | summ += yyy; |
---|
| 1884 | } |
---|
| 1885 | answer = (uplim-lolim)/2.0*summ; |
---|
| 1886 | // Multiply by contrast^2 |
---|
| 1887 | answer *= delrho*delrho; |
---|
| 1888 | // don't do the normal scaling to volume here |
---|
| 1889 | // instead, multiply by VCyl^2 to get rid of the normalization for this radius of cylinder |
---|
| 1890 | Vcyl = Pi*radius*radius*dumLen; |
---|
| 1891 | answer *= Vcyl*Vcyl; |
---|
| 1892 | |
---|
| 1893 | dr = SchulzPoint_cpr(dumLen,len_avg,zz); |
---|
| 1894 | return(dr*answer); |
---|
| 1895 | } |
---|
| 1896 | |
---|
| 1897 | |
---|
| 1898 | double |
---|
| 1899 | Stackdisc_kern(double qq, double rcore, double rhoc, double rhol, double rhosolv, double length, double thick, double dum, double gsd, double d, double N) |
---|
| 1900 | { |
---|
| 1901 | // qq is the q-value for the calculation (1/A) |
---|
| 1902 | // rcore is the core radius of the cylinder (A) |
---|
| 1903 | // rho(n) are the respective SLD's |
---|
| 1904 | // length is the *Half* CORE-LENGTH of the cylinder = L (A) |
---|
| 1905 | // dum is the dummy variable for the integration (x in Feigin's notation) |
---|
| 1906 | |
---|
| 1907 | //Local variables |
---|
| 1908 | double totald,dr1,dr2,besarg1,besarg2,area,sinarg1,sinarg2,t1,t2,retval,sqq,dexpt; |
---|
| 1909 | double Pi; |
---|
| 1910 | int kk; |
---|
| 1911 | |
---|
| 1912 | Pi = 4.0*atan(1.0); |
---|
| 1913 | |
---|
| 1914 | dr1 = rhoc-rhosolv; |
---|
| 1915 | dr2 = rhol-rhosolv; |
---|
| 1916 | area = Pi*rcore*rcore; |
---|
| 1917 | totald=2.0*(thick+length); |
---|
| 1918 | |
---|
| 1919 | besarg1 = qq*rcore*sin(dum); |
---|
| 1920 | besarg2 = qq*rcore*sin(dum); |
---|
| 1921 | |
---|
| 1922 | sinarg1 = qq*length*cos(dum); |
---|
| 1923 | sinarg2 = qq*(length+thick)*cos(dum); |
---|
| 1924 | |
---|
| 1925 | t1 = 2*area*(2*length)*dr1*(sin(sinarg1)/sinarg1)*(NR_BessJ1(besarg1)/besarg1); |
---|
| 1926 | t2 = 2*area*dr2*(totald*sin(sinarg2)/sinarg2-2*length*sin(sinarg1)/sinarg1)*(NR_BessJ1(besarg2)/besarg2); |
---|
| 1927 | |
---|
| 1928 | retval =((t1+t2)*(t1+t2))*sin(dum); |
---|
| 1929 | |
---|
| 1930 | // loop for the structure facture S(q) |
---|
| 1931 | sqq=0.0; |
---|
| 1932 | for(kk=1;kk<N;kk+=1) { |
---|
| 1933 | dexpt=qq*cos(dum)*qq*cos(dum)*d*d*gsd*gsd*kk/2.0; |
---|
| 1934 | sqq=sqq+(N-kk)*cos(qq*cos(dum)*d*kk)*exp(-1.*dexpt); |
---|
| 1935 | } |
---|
| 1936 | |
---|
| 1937 | // end of loop for S(q) |
---|
| 1938 | sqq=1.0+2.0*sqq/N; |
---|
| 1939 | retval *= sqq; |
---|
| 1940 | |
---|
| 1941 | return(retval); |
---|
| 1942 | } |
---|
| 1943 | |
---|
| 1944 | |
---|
| 1945 | double |
---|
| 1946 | Cyl_PolyRadKernel(double q, double radius, double length, double zz, double delrho, double dumRad) |
---|
| 1947 | { |
---|
| 1948 | |
---|
| 1949 | double halfheight,uplim,lolim,zi,summ,yyy,Pi; |
---|
| 1950 | double answer,dr,Vcyl; |
---|
| 1951 | int i,nord; |
---|
| 1952 | |
---|
| 1953 | Pi = 4.0*atan(1.0); |
---|
| 1954 | lolim = 0; |
---|
| 1955 | uplim = Pi/2.0; |
---|
| 1956 | halfheight = length/2.0; |
---|
| 1957 | // nord=20; |
---|
| 1958 | nord=76; |
---|
| 1959 | summ = 0.0; |
---|
| 1960 | |
---|
| 1961 | //do the cylinder orientational average |
---|
| 1962 | // for(i=0;i<nord;i++) { |
---|
| 1963 | // zi = ( Gauss20Z[i]*(uplim-lolim) + uplim + lolim )/2.0; |
---|
| 1964 | // yyy = Gauss20Wt[i] * CylKernel(q, dumRad, halfheight, zi); |
---|
| 1965 | // summ += yyy; |
---|
| 1966 | // } |
---|
| 1967 | for(i=0;i<nord;i++) { |
---|
| 1968 | zi = ( Gauss76Z[i]*(uplim-lolim) + uplim + lolim )/2.0; |
---|
| 1969 | yyy = Gauss76Wt[i] * CylKernel(q, dumRad, halfheight, zi); |
---|
| 1970 | summ += yyy; |
---|
| 1971 | } |
---|
| 1972 | answer = (uplim-lolim)/2.0*summ; |
---|
| 1973 | // Multiply by contrast^2 |
---|
| 1974 | answer *= delrho*delrho; |
---|
| 1975 | // don't do the normal scaling to volume here |
---|
| 1976 | // instead, multiply by VCyl^2 to get rid of the normalization for this radius of cylinder |
---|
| 1977 | Vcyl = Pi*dumRad*dumRad*length; |
---|
| 1978 | answer *= Vcyl*Vcyl; |
---|
| 1979 | |
---|
| 1980 | dr = SchulzPoint_cpr(dumRad,radius,zz); |
---|
| 1981 | return(dr*answer); |
---|
| 1982 | } |
---|
| 1983 | |
---|
| 1984 | double |
---|
| 1985 | SchulzPoint_cpr(double dumRad, double radius, double zz) |
---|
| 1986 | { |
---|
| 1987 | double dr; |
---|
| 1988 | |
---|
| 1989 | dr = zz*log(dumRad) - gammaln(zz+1.0) + (zz+1.0)*log((zz+1.0)/radius)-(dumRad/radius*(zz+1.0)); |
---|
| 1990 | return(exp(dr)); |
---|
| 1991 | } |
---|
| 1992 | |
---|
| 1993 | static double |
---|
| 1994 | gammaln(double xx) |
---|
| 1995 | { |
---|
| 1996 | double x,y,tmp,ser; |
---|
| 1997 | static double cof[6]={76.18009172947146,-86.50532032941677, |
---|
| 1998 | 24.01409824083091,-1.231739572450155, |
---|
| 1999 | 0.1208650973866179e-2,-0.5395239384953e-5}; |
---|
| 2000 | int j; |
---|
| 2001 | |
---|
| 2002 | y=x=xx; |
---|
| 2003 | tmp=x+5.5; |
---|
| 2004 | tmp -= (x+0.5)*log(tmp); |
---|
| 2005 | ser=1.000000000190015; |
---|
| 2006 | for (j=0;j<=5;j++) ser += cof[j]/++y; |
---|
| 2007 | return -tmp+log(2.5066282746310005*ser/x); |
---|
| 2008 | } |
---|
| 2009 | |
---|
| 2010 | |
---|
| 2011 | double |
---|
| 2012 | EllipsoidKernel(double qq, double a, double nua, double dum) |
---|
| 2013 | { |
---|
| 2014 | double arg,nu,retval; //local variables |
---|
| 2015 | |
---|
| 2016 | nu = nua/a; |
---|
| 2017 | arg = qq*a*sqrt(1+dum*dum*(nu*nu-1)); |
---|
| 2018 | |
---|
| 2019 | retval = (sin(arg)-arg*cos(arg))/(arg*arg*arg); |
---|
| 2020 | retval *= retval; |
---|
| 2021 | retval *= 9; |
---|
| 2022 | |
---|
| 2023 | return(retval); |
---|
| 2024 | }//Function EllipsoidKernel() |
---|
| 2025 | |
---|
| 2026 | double |
---|
| 2027 | HolCylKernel(double qq, double rcore, double rshell, double length, double dum) |
---|
| 2028 | { |
---|
| 2029 | double gamma,arg1,arg2,lam1,lam2,psi,sinarg,t2,retval; //local variables |
---|
| 2030 | |
---|
| 2031 | gamma = rcore/rshell; |
---|
| 2032 | arg1 = qq*rshell*sqrt(1-dum*dum); //1=shell (outer radius) |
---|
| 2033 | arg2 = qq*rcore*sqrt(1-dum*dum); //2=core (inner radius) |
---|
| 2034 | lam1 = 2*NR_BessJ1(arg1)/arg1; |
---|
| 2035 | lam2 = 2*NR_BessJ1(arg2)/arg2; |
---|
| 2036 | psi = 1/(1-gamma*gamma)*(lam1 - gamma*gamma*lam2); //SRK 10/19/00 |
---|
| 2037 | |
---|
| 2038 | sinarg = qq*length*dum/2; |
---|
| 2039 | t2 = sin(sinarg)/sinarg; |
---|
| 2040 | |
---|
| 2041 | retval = psi*psi*t2*t2; |
---|
| 2042 | |
---|
| 2043 | return(retval); |
---|
| 2044 | }//Function HolCylKernel() |
---|
| 2045 | |
---|
| 2046 | double |
---|
| 2047 | PPKernel(double aa, double mu, double uu) |
---|
| 2048 | { |
---|
| 2049 | // mu passed in is really mu*sqrt(1-sig^2) |
---|
| 2050 | double arg1,arg2,Pi,tmp1,tmp2; //local variables |
---|
| 2051 | |
---|
| 2052 | Pi = 4.0*atan(1.0); |
---|
| 2053 | |
---|
| 2054 | //handle arg=0 separately, as sin(t)/t -> 1 as t->0 |
---|
| 2055 | arg1 = (mu/2)*cos(Pi*uu/2); |
---|
| 2056 | arg2 = (mu*aa/2)*sin(Pi*uu/2); |
---|
| 2057 | if(arg1==0) { |
---|
| 2058 | tmp1 = 1; |
---|
| 2059 | } else { |
---|
| 2060 | tmp1 = sin(arg1)*sin(arg1)/arg1/arg1; |
---|
| 2061 | } |
---|
| 2062 | |
---|
| 2063 | if (arg2==0) { |
---|
| 2064 | tmp2 = 1; |
---|
| 2065 | } else { |
---|
| 2066 | tmp2 = sin(arg2)*sin(arg2)/arg2/arg2; |
---|
| 2067 | } |
---|
| 2068 | |
---|
| 2069 | return (tmp1*tmp2); |
---|
| 2070 | |
---|
| 2071 | }//Function PPKernel() |
---|
| 2072 | |
---|
| 2073 | |
---|
| 2074 | double |
---|
| 2075 | TriaxialKernel(double q, double aa, double bb, double cc, double dx, double dy) |
---|
| 2076 | { |
---|
| 2077 | |
---|
| 2078 | double arg,val,pi; //local variables |
---|
| 2079 | |
---|
| 2080 | pi = 4.0*atan(1.0); |
---|
| 2081 | |
---|
| 2082 | arg = aa*aa*cos(pi*dx/2)*cos(pi*dx/2); |
---|
| 2083 | arg += bb*bb*sin(pi*dx/2)*sin(pi*dx/2)*(1-dy*dy); |
---|
| 2084 | arg += cc*cc*dy*dy; |
---|
| 2085 | arg = q*sqrt(arg); |
---|
| 2086 | val = 9 * ( (sin(arg) - arg*cos(arg))/(arg*arg*arg) ) * ( (sin(arg) - arg*cos(arg))/(arg*arg*arg) ); |
---|
| 2087 | |
---|
| 2088 | return (val); |
---|
| 2089 | |
---|
| 2090 | }//Function TriaxialKernel() |
---|
| 2091 | |
---|
| 2092 | |
---|
| 2093 | double |
---|
| 2094 | CylKernel(double qq, double rr,double h, double theta) |
---|
| 2095 | { |
---|
| 2096 | |
---|
| 2097 | // qq is the q-value for the calculation (1/A) |
---|
| 2098 | // rr is the radius of the cylinder (A) |
---|
| 2099 | // h is the HALF-LENGTH of the cylinder = L/2 (A) |
---|
| 2100 | |
---|
| 2101 | double besarg,bj,retval,d1,t1,b1,t2,b2; //Local variables |
---|
| 2102 | |
---|
| 2103 | |
---|
| 2104 | besarg = qq*rr*sin(theta); |
---|
| 2105 | |
---|
| 2106 | bj =NR_BessJ1(besarg); |
---|
| 2107 | |
---|
| 2108 | //* Computing 2nd power */ |
---|
| 2109 | d1 = sin(qq * h * cos(theta)); |
---|
| 2110 | t1 = d1 * d1; |
---|
| 2111 | //* Computing 2nd power */ |
---|
| 2112 | d1 = bj; |
---|
| 2113 | t2 = d1 * d1 * 4.0 * sin(theta); |
---|
| 2114 | //* Computing 2nd power */ |
---|
| 2115 | d1 = qq * h * cos(theta); |
---|
| 2116 | b1 = d1 * d1; |
---|
| 2117 | //* Computing 2nd power */ |
---|
| 2118 | d1 = qq * rr * sin(theta); |
---|
| 2119 | b2 = d1 * d1; |
---|
| 2120 | retval = t1 * t2 / b1 / b2; |
---|
| 2121 | |
---|
| 2122 | return (retval); |
---|
| 2123 | |
---|
| 2124 | }//Function CylKernel() |
---|
| 2125 | |
---|
| 2126 | double |
---|
| 2127 | EllipCylKernel(double qq, double ra,double nu, double theta) |
---|
| 2128 | { |
---|
| 2129 | //this is the function LAMBDA1^2 in Feigin's notation |
---|
| 2130 | // qq is the q-value for the calculation (1/A) |
---|
| 2131 | // ra is the transformed radius"a" in Feigin's notation |
---|
| 2132 | // nu is the ratio (major radius)/(minor radius) of the Ellipsoid [=] --- |
---|
| 2133 | // theta is the dummy variable of the integration |
---|
| 2134 | |
---|
| 2135 | double retval,arg; //Local variables |
---|
| 2136 | |
---|
| 2137 | arg = qq*ra*sqrt((1+nu*nu)/2+(1-nu*nu)*cos(theta)/2); |
---|
| 2138 | |
---|
| 2139 | retval = 2*NR_BessJ1(arg)/arg; |
---|
| 2140 | |
---|
| 2141 | //square it |
---|
| 2142 | retval *= retval; |
---|
| 2143 | |
---|
| 2144 | return(retval); |
---|
| 2145 | |
---|
| 2146 | }//Function EllipCylKernel() |
---|
| 2147 | |
---|
| 2148 | double NR_BessJ1(double x) |
---|
| 2149 | { |
---|
| 2150 | double ax,z; |
---|
| 2151 | double xx,y,ans,ans1,ans2; |
---|
| 2152 | |
---|
| 2153 | if ((ax=fabs(x)) < 8.0) { |
---|
| 2154 | y=x*x; |
---|
| 2155 | ans1=x*(72362614232.0+y*(-7895059235.0+y*(242396853.1 |
---|
| 2156 | +y*(-2972611.439+y*(15704.48260+y*(-30.16036606)))))); |
---|
| 2157 | ans2=144725228442.0+y*(2300535178.0+y*(18583304.74 |
---|
| 2158 | +y*(99447.43394+y*(376.9991397+y*1.0)))); |
---|
| 2159 | ans=ans1/ans2; |
---|
| 2160 | } else { |
---|
| 2161 | z=8.0/ax; |
---|
| 2162 | y=z*z; |
---|
| 2163 | xx=ax-2.356194491; |
---|
| 2164 | ans1=1.0+y*(0.183105e-2+y*(-0.3516396496e-4 |
---|
| 2165 | +y*(0.2457520174e-5+y*(-0.240337019e-6)))); |
---|
| 2166 | ans2=0.04687499995+y*(-0.2002690873e-3 |
---|
| 2167 | +y*(0.8449199096e-5+y*(-0.88228987e-6 |
---|
| 2168 | +y*0.105787412e-6))); |
---|
| 2169 | ans=sqrt(0.636619772/ax)*(cos(xx)*ans1-z*sin(xx)*ans2); |
---|
| 2170 | if (x < 0.0) ans = -ans; |
---|
| 2171 | } |
---|
| 2172 | |
---|
| 2173 | return(ans); |
---|
| 2174 | } |
---|