| 1 |
/* |
| 2 |
* // Copyright (c) Radzivon Bartoshyk 7/2025. All rights reserved. |
| 3 |
* // |
| 4 |
* // Redistribution and use in source and binary forms, with or without modification, |
| 5 |
* // are permitted provided that the following conditions are met: |
| 6 |
* // |
| 7 |
* // 1. Redistributions of source code must retain the above copyright notice, this |
| 8 |
* // list of conditions and the following disclaimer. |
| 9 |
* // |
| 10 |
* // 2. Redistributions in binary form must reproduce the above copyright notice, |
| 11 |
* // this list of conditions and the following disclaimer in the documentation |
| 12 |
* // and/or other materials provided with the distribution. |
| 13 |
* // |
| 14 |
* // 3. Neither the name of the copyright holder nor the names of its |
| 15 |
* // contributors may be used to endorse or promote products derived from |
| 16 |
* // this software without specific prior written permission. |
| 17 |
* // |
| 18 |
* // THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" |
| 19 |
* // AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE |
| 20 |
* // IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE |
| 21 |
* // DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE |
| 22 |
* // FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL |
| 23 |
* // DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR |
| 24 |
* // SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER |
| 25 |
* // CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, |
| 26 |
* // OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE |
| 27 |
* // OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. |
| 28 |
*/ |
| 29 |
use crate::bessel::j1_coeffs::{J1_ZEROS, J1_ZEROS_VALUE}; |
| 30 |
use crate::bessel::j1f_coeffs::J1F_COEFFS; |
| 31 |
use crate::bessel::trigo_bessel::sin_small; |
| 32 |
use crate::double_double::DoubleDouble; |
| 33 |
use crate::polyeval::{f_polyeval7, f_polyeval10, f_polyeval12, f_polyeval14}; |
| 34 |
use crate::rounding::CpuCeil; |
| 35 |
use crate::sincos_reduce::rem2pif_any; |
| 36 |
|
| 37 |
/// Bessel of the first kind of order 1 |
| 38 |
/// |
| 39 |
/// Max ULP 0.5 |
| 40 |
/// |
| 41 |
/// Note about accuracy: |
| 42 |
/// - Close to zero Bessel have tiny values such that testing against MPFR must be done exactly |
| 43 |
/// in the same precision, since any nearest representable number have ULP > 0.5. |
| 44 |
/// For example `J1(0.000000000000000000000000000000000000023509886)` in single precision |
| 45 |
/// have an error at least 0.72 ULP for any number with extended precision, |
| 46 |
/// that would be represented in f32. |
| 47 |
pub fn f_j1f(x: f32) -> f32 { |
| 48 |
let ux = x.to_bits().wrapping_shl(1); |
| 49 |
if ux >= 0xffu32 << 24 || ux == 0 { |
| 50 |
// |x| == 0, |x| == inf, |x| == NaN |
| 51 |
if ux == 0 { |
| 52 |
// |x| == 0 |
| 53 |
return x; |
| 54 |
} |
| 55 |
if x.is_infinite() { |
| 56 |
return 0.; |
| 57 |
} |
| 58 |
return x + f32::NAN; // x == NaN |
| 59 |
} |
| 60 |
|
| 61 |
let ax = x.to_bits() & 0x7fff_ffff; |
| 62 |
|
| 63 |
if ax < 0x429533c2u32 { |
| 64 |
// |x| <= 74.60109 |
| 65 |
if ax < 0x3e800000u32 { |
| 66 |
// |x| <= 0.25 |
| 67 |
if ax <= 0x34000000u32 { |
| 68 |
// |x| <= f32::EPSILON |
| 69 |
// taylor series for J1(x) ~ x/2 + O(x^3) |
| 70 |
return x * 0.5; |
| 71 |
} |
| 72 |
return poly_near_zero(x); |
| 73 |
} |
| 74 |
return small_argument_path(x); |
| 75 |
} |
| 76 |
|
| 77 |
// Exceptional cases: |
| 78 |
if ax == 0x6ef9be45 { |
| 79 |
return if x.is_sign_negative() { |
| 80 |
f32::from_bits(0x187d8a8f) |
| 81 |
} else { |
| 82 |
-f32::from_bits(0x187d8a8f) |
| 83 |
}; |
| 84 |
} else if ax == 0x7f0e5a38 { |
| 85 |
return if x.is_sign_negative() { |
| 86 |
-f32::from_bits(0x131f680b) |
| 87 |
} else { |
| 88 |
f32::from_bits(0x131f680b) |
| 89 |
}; |
| 90 |
} |
| 91 |
|
| 92 |
j1f_asympt(x) as f32 |
| 93 |
} |
| 94 |
|
| 95 |
#[inline] |
| 96 |
fn j1f_rsqrt(x: f64) -> f64 { |
| 97 |
(1. / x) * x.sqrt() |
| 98 |
} |
| 99 |
|
| 100 |
/* |
| 101 |
Evaluates: |
| 102 |
J1 = sqrt(2/(PI*x)) * beta(x) * cos(x - 3*PI/4 - alpha(x)) |
| 103 |
discarding 1*PI/2 using identities gives: |
| 104 |
J1 = sqrt(2/(PI*x)) * beta(x) * sin(x - PI/4 - alpha(x)) |
| 105 |
|
| 106 |
to avoid squashing small (-PI/4 - alpha(x)) into a large x actual expansion is: |
| 107 |
|
| 108 |
J1 = sqrt(2/(PI*x)) * beta(x) * sin((x mod 2*PI) - PI/4 - alpha(x)) |
| 109 |
*/ |
| 110 |
#[inline] |
| 111 |
fn j1f_asympt(x: f32) -> f64 { |
| 112 |
static SGN: [f64; 2] = [1., -1.]; |
| 113 |
let sign_scale = SGN[x.is_sign_negative() as usize]; |
| 114 |
let x = f32::from_bits(x.to_bits() & 0x7fff_ffff); |
| 115 |
|
| 116 |
let dx = x as f64; |
| 117 |
|
| 118 |
let alpha = j1f_asympt_alpha(dx); |
| 119 |
let beta = j1f_asympt_beta(dx); |
| 120 |
|
| 121 |
let angle = rem2pif_any(x); |
| 122 |
|
| 123 |
const SQRT_2_OVER_PI: f64 = f64::from_bits(0x3fe9884533d43651); |
| 124 |
const MPI_OVER_4: f64 = f64::from_bits(0xbfe921fb54442d18); |
| 125 |
|
| 126 |
let x0pi34 = MPI_OVER_4 - alpha; |
| 127 |
let r0 = angle + x0pi34; |
| 128 |
|
| 129 |
let m_sin = sin_small(r0); |
| 130 |
|
| 131 |
let z0 = beta * m_sin; |
| 132 |
let scale = SQRT_2_OVER_PI * j1f_rsqrt(dx); |
| 133 |
|
| 134 |
scale * z0 * sign_scale |
| 135 |
} |
| 136 |
|
| 137 |
/** |
| 138 |
Note expansion generation below: this is negative series expressed in Sage as positive, |
| 139 |
so before any real evaluation `x=1/x` should be applied. |
| 140 |
|
| 141 |
Generated by SageMath: |
| 142 |
```python |
| 143 |
def binomial_like(n, m): |
| 144 |
prod = QQ(1) |
| 145 |
z = QQ(4)*(n**2) |
| 146 |
for k in range(1,m + 1): |
| 147 |
prod *= (z - (2*k - 1)**2) |
| 148 |
return prod / (QQ(2)**(2*m) * (ZZ(m).factorial())) |
| 149 |
|
| 150 |
R = LaurentSeriesRing(RealField(300), 'x',default_prec=300) |
| 151 |
x = R.gen() |
| 152 |
|
| 153 |
def Pn_asymptotic(n, y, terms=10): |
| 154 |
# now y = 1/x |
| 155 |
return sum( (-1)**m * binomial_like(n, 2*m) / (QQ(2)**(2*m)) * y**(QQ(2)*m) for m in range(terms) ) |
| 156 |
|
| 157 |
def Qn_asymptotic(n, y, terms=10): |
| 158 |
return sum( (-1)**m * binomial_like(n, 2*m + 1) / (QQ(2)**(2*m + 1)) * y**(QQ(2)*m + 1) for m in range(terms) ) |
| 159 |
|
| 160 |
P = Pn_asymptotic(1, x, 50) |
| 161 |
Q = Qn_asymptotic(1, x, 50) |
| 162 |
|
| 163 |
R_series = (-Q/P) |
| 164 |
|
| 165 |
# alpha is atan(R_series) so we're doing Taylor series atan expansion on R_series |
| 166 |
|
| 167 |
arctan_series_Z = sum([QQ(-1)**k * x**(QQ(2)*k+1) / RealField(700)(RealField(700)(2)*k+1) for k in range(25)]) |
| 168 |
alpha_series = arctan_series_Z(R_series) |
| 169 |
|
| 170 |
# see the series |
| 171 |
print(alpha_series) |
| 172 |
``` |
| 173 |
|
| 174 |
See notes/bessel_asympt.ipynb for generation |
| 175 |
**/ |
| 176 |
#[inline] |
| 177 |
pub(crate) fn j1f_asympt_alpha(x: f64) -> f64 { |
| 178 |
const C: [u64; 12] = [ |
| 179 |
0xbfd8000000000000, |
| 180 |
0x3fc5000000000000, |
| 181 |
0xbfd7bccccccccccd, |
| 182 |
0x4002f486db6db6db, |
| 183 |
0xc03e9fbf40000000, |
| 184 |
0x4084997b55945d17, |
| 185 |
0xc0d4a914195269d9, |
| 186 |
0x412cd1b53816aec1, |
| 187 |
0xc18aa4095d419351, |
| 188 |
0x41ef809305f11b9d, |
| 189 |
0xc2572e6809ed618b, |
| 190 |
0x42c4c5b6057839f9, |
| 191 |
]; |
| 192 |
let recip = 1. / x; |
| 193 |
let x2 = recip * recip; |
| 194 |
let p = f_polyeval12( |
| 195 |
x2, |
| 196 |
f64::from_bits(C[0]), |
| 197 |
f64::from_bits(C[1]), |
| 198 |
f64::from_bits(C[2]), |
| 199 |
f64::from_bits(C[3]), |
| 200 |
f64::from_bits(C[4]), |
| 201 |
f64::from_bits(C[5]), |
| 202 |
f64::from_bits(C[6]), |
| 203 |
f64::from_bits(C[7]), |
| 204 |
f64::from_bits(C[8]), |
| 205 |
f64::from_bits(C[9]), |
| 206 |
f64::from_bits(C[10]), |
| 207 |
f64::from_bits(C[11]), |
| 208 |
); |
| 209 |
p * recip |
| 210 |
} |
| 211 |
|
| 212 |
/** |
| 213 |
Note expansion generation below: this is negative series expressed in Sage as positive, |
| 214 |
so before any real evaluation `x=1/x` should be applied |
| 215 |
|
| 216 |
Generated by SageMath: |
| 217 |
```python |
| 218 |
def binomial_like(n, m): |
| 219 |
prod = QQ(1) |
| 220 |
z = QQ(4)*(n**2) |
| 221 |
for k in range(1,m + 1): |
| 222 |
prod *= (z - (2*k - 1)**2) |
| 223 |
return prod / (QQ(2)**(2*m) * (ZZ(m).factorial())) |
| 224 |
|
| 225 |
R = LaurentSeriesRing(RealField(300), 'x',default_prec=300) |
| 226 |
x = R.gen() |
| 227 |
|
| 228 |
def Pn_asymptotic(n, y, terms=10): |
| 229 |
# now y = 1/x |
| 230 |
return sum( (-1)**m * binomial_like(n, 2*m) / (QQ(2)**(2*m)) * y**(QQ(2)*m) for m in range(terms) ) |
| 231 |
|
| 232 |
def Qn_asymptotic(n, y, terms=10): |
| 233 |
return sum( (-1)**m * binomial_like(n, 2*m + 1) / (QQ(2)**(2*m + 1)) * y**(QQ(2)*m + 1) for m in range(terms) ) |
| 234 |
|
| 235 |
P = Pn_asymptotic(1, x, 50) |
| 236 |
Q = Qn_asymptotic(1, x, 50) |
| 237 |
|
| 238 |
def sqrt_series(s): |
| 239 |
val = S.valuation() |
| 240 |
lc = S[val] # Leading coefficient |
| 241 |
b = lc.sqrt() * x**(val // 2) |
| 242 |
|
| 243 |
for _ in range(5): |
| 244 |
b = (b + S / b) / 2 |
| 245 |
b = b |
| 246 |
return b |
| 247 |
|
| 248 |
S = (P**2 + Q**2).truncate(50) |
| 249 |
|
| 250 |
b_series = sqrt_series(S).truncate(30) |
| 251 |
# see the beta series |
| 252 |
print(b_series) |
| 253 |
``` |
| 254 |
|
| 255 |
See notes/bessel_asympt.ipynb for generation |
| 256 |
**/ |
| 257 |
#[inline] |
| 258 |
pub(crate) fn j1f_asympt_beta(x: f64) -> f64 { |
| 259 |
const C: [u64; 10] = [ |
| 260 |
0x3ff0000000000000, |
| 261 |
0x3fc8000000000000, |
| 262 |
0xbfc8c00000000000, |
| 263 |
0x3fe9c50000000000, |
| 264 |
0xc01ef5b680000000, |
| 265 |
0x40609860dd400000, |
| 266 |
0xc0abae9b7a06e000, |
| 267 |
0x41008711d41c1428, |
| 268 |
0xc15ab70164c8be6e, |
| 269 |
0x41bc1055e24f297f, |
| 270 |
]; |
| 271 |
let recip = 1. / x; |
| 272 |
let x2 = recip * recip; |
| 273 |
f_polyeval10( |
| 274 |
x2, |
| 275 |
f64::from_bits(C[0]), |
| 276 |
f64::from_bits(C[1]), |
| 277 |
f64::from_bits(C[2]), |
| 278 |
f64::from_bits(C[3]), |
| 279 |
f64::from_bits(C[4]), |
| 280 |
f64::from_bits(C[5]), |
| 281 |
f64::from_bits(C[6]), |
| 282 |
f64::from_bits(C[7]), |
| 283 |
f64::from_bits(C[8]), |
| 284 |
f64::from_bits(C[9]), |
| 285 |
) |
| 286 |
} |
| 287 |
|
| 288 |
/** |
| 289 |
Generated in Sollya: |
| 290 |
```python |
| 291 |
pretty = proc(u) { |
| 292 |
return ~(floor(u*1000)/1000); |
| 293 |
}; |
| 294 |
|
| 295 |
bessel_j1 = library("./cmake-build-release/libbessel_sollya.dylib"); |
| 296 |
|
| 297 |
f = bessel_j1(x)/x; |
| 298 |
d = [0, 0.921]; |
| 299 |
w = 1; |
| 300 |
pf = fpminimax(f, [|0,2,4,6,8,10,12|], [|D...|], d, absolute, floating); |
| 301 |
|
| 302 |
w = 1; |
| 303 |
or_f = bessel_j1(x); |
| 304 |
pf1 = pf * x; |
| 305 |
err_p = -log2(dirtyinfnorm(pf1*w-or_f, d)); |
| 306 |
print ("relative error:", pretty(err_p)); |
| 307 |
|
| 308 |
for i from 0 to degree(pf) by 2 do { |
| 309 |
print("'", coeff(pf, i), "',"); |
| 310 |
}; |
| 311 |
``` |
| 312 |
See ./notes/bessel_sollya/bessel_j1f_at_zero.sollya |
| 313 |
**/ |
| 314 |
#[inline] |
| 315 |
fn poly_near_zero(x: f32) -> f32 { |
| 316 |
let dx = x as f64; |
| 317 |
let x2 = dx * dx; |
| 318 |
let p = f_polyeval7( |
| 319 |
x2, |
| 320 |
f64::from_bits(0x3fe0000000000000), |
| 321 |
f64::from_bits(0xbfaffffffffffffc), |
| 322 |
f64::from_bits(0x3f65555555554089), |
| 323 |
f64::from_bits(0xbf0c71c71c2a74ae), |
| 324 |
f64::from_bits(0x3ea6c16bbd1dc5c1), |
| 325 |
f64::from_bits(0xbe384562afb69e7d), |
| 326 |
f64::from_bits(0x3dc248d0d0221cd0), |
| 327 |
); |
| 328 |
(p * dx) as f32 |
| 329 |
} |
| 330 |
|
| 331 |
/// This method on small range searches for nearest zero or extremum. |
| 332 |
/// Then picks stored series expansion at the point end evaluates the poly at the point. |
| 333 |
#[inline] |
| 334 |
fn small_argument_path(x: f32) -> f32 { |
| 335 |
static SIGN: [f64; 2] = [1., -1.]; |
| 336 |
let sign_scale = SIGN[x.is_sign_negative() as usize]; |
| 337 |
let x_abs = f32::from_bits(x.to_bits() & 0x7fff_ffff) as f64; |
| 338 |
|
| 339 |
// let avg_step = 74.60109 / 47.0; |
| 340 |
// let inv_step = 1.0 / avg_step; |
| 341 |
|
| 342 |
const INV_STEP: f64 = 0.6300176043004198; |
| 343 |
|
| 344 |
let fx = x_abs * INV_STEP; |
| 345 |
const J1_ZEROS_COUNT: f64 = (J1_ZEROS.len() - 1) as f64; |
| 346 |
let idx0 = unsafe { fx.min(J1_ZEROS_COUNT).to_int_unchecked::<usize>() }; |
| 347 |
let idx1 = unsafe { |
| 348 |
fx.cpu_ceil() |
| 349 |
.min(J1_ZEROS_COUNT) |
| 350 |
.to_int_unchecked::<usize>() |
| 351 |
}; |
| 352 |
|
| 353 |
let found_zero0 = DoubleDouble::from_bit_pair(J1_ZEROS[idx0]); |
| 354 |
let found_zero1 = DoubleDouble::from_bit_pair(J1_ZEROS[idx1]); |
| 355 |
|
| 356 |
let dist0 = (found_zero0.hi - x_abs).abs(); |
| 357 |
let dist1 = (found_zero1.hi - x_abs).abs(); |
| 358 |
|
| 359 |
let (found_zero, idx, dist) = if dist0 < dist1 { |
| 360 |
(found_zero0, idx0, dist0) |
| 361 |
} else { |
| 362 |
(found_zero1, idx1, dist1) |
| 363 |
}; |
| 364 |
|
| 365 |
if idx == 0 { |
| 366 |
return poly_near_zero(x); |
| 367 |
} |
| 368 |
|
| 369 |
// We hit exact zero, value, better to return it directly |
| 370 |
if dist == 0. { |
| 371 |
return (f64::from_bits(J1_ZEROS_VALUE[idx]) * sign_scale) as f32; |
| 372 |
} |
| 373 |
|
| 374 |
let c = &J1F_COEFFS[idx - 1]; |
| 375 |
|
| 376 |
let r = (x_abs - found_zero.hi) - found_zero.lo; |
| 377 |
|
| 378 |
let p = f_polyeval14( |
| 379 |
r, |
| 380 |
f64::from_bits(c[0]), |
| 381 |
f64::from_bits(c[1]), |
| 382 |
f64::from_bits(c[2]), |
| 383 |
f64::from_bits(c[3]), |
| 384 |
f64::from_bits(c[4]), |
| 385 |
f64::from_bits(c[5]), |
| 386 |
f64::from_bits(c[6]), |
| 387 |
f64::from_bits(c[7]), |
| 388 |
f64::from_bits(c[8]), |
| 389 |
f64::from_bits(c[9]), |
| 390 |
f64::from_bits(c[10]), |
| 391 |
f64::from_bits(c[11]), |
| 392 |
f64::from_bits(c[12]), |
| 393 |
f64::from_bits(c[13]), |
| 394 |
); |
| 395 |
|
| 396 |
(p * sign_scale) as f32 |
| 397 |
} |
| 398 |
|
| 399 |
#[cfg(test)] |
| 400 |
mod tests { |
| 401 |
use super::*; |
| 402 |
|
| 403 |
#[test] |
| 404 |
fn test_f_j1f() { |
| 405 |
assert_eq!( |
| 406 |
f_j1f(77.743162408196766932633181568235159), |
| 407 |
0.09049267898021947 |
| 408 |
); |
| 409 |
assert_eq!( |
| 410 |
f_j1f(-0.000000000000000000000000000000000000008827127), |
| 411 |
-0.0000000000000000000000000000000000000044135635 |
| 412 |
); |
| 413 |
assert_eq!( |
| 414 |
f_j1f(0.000000000000000000000000000000000000008827127), |
| 415 |
0.0000000000000000000000000000000000000044135635 |
| 416 |
); |
| 417 |
assert_eq!(f_j1f(5.4), -0.3453447907795863); |
| 418 |
assert_eq!( |
| 419 |
f_j1f(84.027189586293545175976760219782591), |
| 420 |
0.0870430264022591 |
| 421 |
); |
| 422 |
assert_eq!(f_j1f(f32::INFINITY), 0.); |
| 423 |
assert_eq!(f_j1f(f32::NEG_INFINITY), 0.); |
| 424 |
assert!(f_j1f(f32::NAN).is_nan()); |
| 425 |
assert_eq!(f_j1f(-1.7014118e38), 0.000000000000000000006856925); |
| 426 |
} |
| 427 |
} |