Wed, 07 Dec 2022 06:59:19 +0200
Added tag v1.0.0 for changeset 23821a01d013
0 | 1 | |
2 | //! Things for constructing new kernels from component kernels and traits for analysing them | |
3 | use serde::Serialize; | |
4 | use numeric_literals::replace_float_literals; | |
5 | ||
6 | use alg_tools::types::*; | |
7 | use alg_tools::norms::*; | |
8 | use alg_tools::loc::Loc; | |
9 | use alg_tools::sets::Cube; | |
10 | use alg_tools::bisection_tree::{ | |
11 | Support, | |
12 | Bounds, | |
13 | LocalAnalysis, | |
14 | GlobalAnalysis, | |
15 | Bounded, | |
16 | }; | |
17 | use alg_tools::mapping::Apply; | |
18 | use alg_tools::maputil::{array_init, map2}; | |
19 | use alg_tools::sets::SetOrd; | |
20 | ||
21 | use crate::fourier::Fourier; | |
22 | ||
23 | /// Representation of the product of two kernels. | |
24 | /// | |
25 | /// The kernels typically implement [`Support`] and [`Mapping`][alg_tools::mapping::Mapping]. | |
26 | /// | |
27 | /// The implementation [`Support`] only uses the [`Support::support_hint`] of the first parameter! | |
28 | #[derive(Copy,Clone,Serialize,Debug)] | |
29 | pub struct SupportProductFirst<A, B>( | |
30 | /// First kernel | |
31 | pub A, | |
32 | /// Second kernel | |
33 | pub B | |
34 | ); | |
35 | ||
36 | impl<A, B, F : Float, const N : usize> Apply<Loc<F, N>> | |
37 | for SupportProductFirst<A, B> | |
38 | where A : for<'a> Apply<&'a Loc<F, N>, Output=F>, | |
39 | B : for<'a> Apply<&'a Loc<F, N>, Output=F> { | |
40 | type Output = F; | |
41 | #[inline] | |
42 | fn apply(&self, x : Loc<F, N>) -> Self::Output { | |
43 | self.0.apply(&x) * self.1.apply(&x) | |
44 | } | |
45 | } | |
46 | ||
47 | impl<'a, A, B, F : Float, const N : usize> Apply<&'a Loc<F, N>> | |
48 | for SupportProductFirst<A, B> | |
49 | where A : Apply<&'a Loc<F, N>, Output=F>, | |
50 | B : Apply<&'a Loc<F, N>, Output=F> { | |
51 | type Output = F; | |
52 | #[inline] | |
53 | fn apply(&self, x : &'a Loc<F, N>) -> Self::Output { | |
54 | self.0.apply(x) * self.1.apply(x) | |
55 | } | |
56 | } | |
57 | ||
58 | impl<'a, A, B, F : Float, const N : usize> Support<F, N> | |
59 | for SupportProductFirst<A, B> | |
60 | where A : Support<F, N>, | |
61 | B : Support<F, N> { | |
62 | #[inline] | |
63 | fn support_hint(&self) -> Cube<F, N> { | |
64 | self.0.support_hint() | |
65 | } | |
66 | ||
67 | #[inline] | |
68 | fn in_support(&self, x : &Loc<F, N>) -> bool { | |
69 | self.0.in_support(x) | |
70 | } | |
71 | ||
72 | #[inline] | |
73 | fn bisection_hint(&self, cube : &Cube<F, N>) -> [Option<F>; N] { | |
74 | self.0.bisection_hint(cube) | |
75 | } | |
76 | } | |
77 | ||
78 | impl<'a, A, B, F : Float> GlobalAnalysis<F, Bounds<F>> | |
79 | for SupportProductFirst<A, B> | |
80 | where A : GlobalAnalysis<F, Bounds<F>>, | |
81 | B : GlobalAnalysis<F, Bounds<F>> { | |
82 | #[inline] | |
83 | fn global_analysis(&self) -> Bounds<F> { | |
84 | self.0.global_analysis() * self.1.global_analysis() | |
85 | } | |
86 | } | |
87 | ||
88 | impl<'a, A, B, F : Float, const N : usize> LocalAnalysis<F, Bounds<F>, N> | |
89 | for SupportProductFirst<A, B> | |
90 | where A : LocalAnalysis<F, Bounds<F>, N>, | |
91 | B : LocalAnalysis<F, Bounds<F>, N> { | |
92 | #[inline] | |
93 | fn local_analysis(&self, cube : &Cube<F, N>) -> Bounds<F> { | |
94 | self.0.local_analysis(cube) * self.1.local_analysis(cube) | |
95 | } | |
96 | } | |
97 | ||
98 | /// Representation of the sum of two kernels | |
99 | /// | |
100 | /// The kernels typically implement [`Support`] and [`Mapping`][alg_tools::mapping::Mapping]. | |
101 | /// | |
102 | /// The implementation [`Support`] only uses the [`Support::support_hint`] of the first parameter! | |
103 | #[derive(Copy,Clone,Serialize,Debug)] | |
104 | pub struct SupportSum<A, B>( | |
105 | /// First kernel | |
106 | pub A, | |
107 | /// Second kernel | |
108 | pub B | |
109 | ); | |
110 | ||
111 | impl<'a, A, B, F : Float, const N : usize> Apply<&'a Loc<F, N>> | |
112 | for SupportSum<A, B> | |
113 | where A : Apply<&'a Loc<F, N>, Output=F>, | |
114 | B : Apply<&'a Loc<F, N>, Output=F> { | |
115 | type Output = F; | |
116 | #[inline] | |
117 | fn apply(&self, x : &'a Loc<F, N>) -> Self::Output { | |
118 | self.0.apply(x) + self.1.apply(x) | |
119 | } | |
120 | } | |
121 | ||
122 | impl<A, B, F : Float, const N : usize> Apply<Loc<F, N>> | |
123 | for SupportSum<A, B> | |
124 | where A : for<'a> Apply<&'a Loc<F, N>, Output=F>, | |
125 | B : for<'a> Apply<&'a Loc<F, N>, Output=F> { | |
126 | type Output = F; | |
127 | #[inline] | |
128 | fn apply(&self, x : Loc<F, N>) -> Self::Output { | |
129 | self.0.apply(&x) + self.1.apply(&x) | |
130 | } | |
131 | } | |
132 | ||
133 | impl<'a, A, B, F : Float, const N : usize> Support<F, N> | |
134 | for SupportSum<A, B> | |
135 | where A : Support<F, N>, | |
136 | B : Support<F, N>, | |
137 | Cube<F, N> : SetOrd { | |
138 | #[inline] | |
139 | fn support_hint(&self) -> Cube<F, N> { | |
140 | self.0.support_hint().common(&self.1.support_hint()) | |
141 | } | |
142 | ||
143 | #[inline] | |
144 | fn in_support(&self, x : &Loc<F, N>) -> bool { | |
145 | self.0.in_support(x) || self.1.in_support(x) | |
146 | } | |
147 | ||
148 | #[inline] | |
149 | fn bisection_hint(&self, cube : &Cube<F, N>) -> [Option<F>; N] { | |
150 | map2(self.0.bisection_hint(cube), | |
151 | self.1.bisection_hint(cube), | |
152 | |a, b| a.or(b)) | |
153 | } | |
154 | } | |
155 | ||
156 | impl<'a, A, B, F : Float> GlobalAnalysis<F, Bounds<F>> | |
157 | for SupportSum<A, B> | |
158 | where A : GlobalAnalysis<F, Bounds<F>>, | |
159 | B : GlobalAnalysis<F, Bounds<F>> { | |
160 | #[inline] | |
161 | fn global_analysis(&self) -> Bounds<F> { | |
162 | self.0.global_analysis() + self.1.global_analysis() | |
163 | } | |
164 | } | |
165 | ||
166 | impl<'a, A, B, F : Float, const N : usize> LocalAnalysis<F, Bounds<F>, N> | |
167 | for SupportSum<A, B> | |
168 | where A : LocalAnalysis<F, Bounds<F>, N>, | |
169 | B : LocalAnalysis<F, Bounds<F>, N>, | |
170 | Cube<F, N> : SetOrd { | |
171 | #[inline] | |
172 | fn local_analysis(&self, cube : &Cube<F, N>) -> Bounds<F> { | |
173 | self.0.local_analysis(cube) + self.1.local_analysis(cube) | |
174 | } | |
175 | } | |
176 | ||
177 | /// Representation of the convolution of two kernels. | |
178 | /// | |
179 | /// The kernels typically implement [`Support`]s and [`Mapping`][alg_tools::mapping::Mapping]. | |
180 | // | |
181 | /// Trait implementations have to be on a case-by-case basis. | |
182 | #[derive(Copy,Clone,Serialize,Debug,Eq,PartialEq)] | |
183 | pub struct Convolution<A, B>( | |
184 | /// First kernel | |
185 | pub A, | |
186 | /// Second kernel | |
187 | pub B | |
188 | ); | |
189 | ||
190 | /// Representation of the autoconvolution of a kernel. | |
191 | /// | |
192 | /// The kernel typically implements [`Support`] and [`Mapping`][alg_tools::mapping::Mapping]. | |
193 | /// | |
194 | /// Trait implementations have to be on a case-by-case basis. | |
195 | #[derive(Copy,Clone,Serialize,Debug,Eq,PartialEq)] | |
196 | pub struct AutoConvolution<A>( | |
197 | /// The kernel to be autoconvolved | |
198 | pub A | |
199 | ); | |
200 | ||
201 | /// Representation a multi-dimensional product of a one-dimensional kernel. | |
202 | /// | |
203 | /// For $G: ℝ → ℝ$, this is the function $F(x\_1, …, x\_n) := \prod_{i=1}^n G(x\_i)$. | |
204 | /// The kernel $G$ typically implements [`Support`] and [`Mapping`][alg_tools::mapping::Mapping] | |
205 | /// on [`Loc<F, 1>`]. Then the product implements them on [`Loc<F, N>`]. | |
206 | #[derive(Copy,Clone,Serialize,Debug,Eq,PartialEq)] | |
207 | struct UniformProduct<G, const N : usize>( | |
208 | /// The one-dimensional kernel | |
209 | G | |
210 | ); | |
211 | ||
212 | impl<'a, G, F : Float, const N : usize> Apply<&'a Loc<F, N>> | |
213 | for UniformProduct<G, N> | |
214 | where G : Apply<Loc<F, 1>, Output=F> { | |
215 | type Output = F; | |
216 | #[inline] | |
217 | fn apply(&self, x : &'a Loc<F, N>) -> F { | |
218 | x.iter().map(|&y| self.0.apply(Loc([y]))).product() | |
219 | } | |
220 | } | |
221 | ||
222 | impl<G, F : Float, const N : usize> Apply<Loc<F, N>> | |
223 | for UniformProduct<G, N> | |
224 | where G : Apply<Loc<F, 1>, Output=F> { | |
225 | type Output = F; | |
226 | #[inline] | |
227 | fn apply(&self, x : Loc<F, N>) -> F { | |
228 | x.into_iter().map(|y| self.0.apply(Loc([y]))).product() | |
229 | } | |
230 | } | |
231 | ||
232 | impl<G, F : Float, const N : usize> Support<F, N> | |
233 | for UniformProduct<G, N> | |
234 | where G : Support<F, 1> { | |
235 | #[inline] | |
236 | fn support_hint(&self) -> Cube<F, N> { | |
237 | let [a] : [[F; 2]; 1] = self.0.support_hint().into(); | |
238 | array_init(|| a.clone()).into() | |
239 | } | |
240 | ||
241 | #[inline] | |
242 | fn in_support(&self, x : &Loc<F, N>) -> bool { | |
243 | x.iter().all(|&y| self.0.in_support(&Loc([y]))) | |
244 | } | |
245 | ||
246 | #[inline] | |
247 | fn bisection_hint(&self, cube : &Cube<F, N>) -> [Option<F>; N] { | |
248 | cube.map(|a, b| { | |
249 | let [h] = self.0.bisection_hint(&[[a, b]].into()); | |
250 | h | |
251 | }) | |
252 | } | |
253 | } | |
254 | ||
255 | impl<G, F : Float, const N : usize> GlobalAnalysis<F, Bounds<F>> | |
256 | for UniformProduct<G, N> | |
257 | where G : GlobalAnalysis<F, Bounds<F>> { | |
258 | #[inline] | |
259 | fn global_analysis(&self) -> Bounds<F> { | |
260 | let g = self.0.global_analysis(); | |
261 | (0..N).map(|_| g).product() | |
262 | } | |
263 | } | |
264 | ||
265 | impl<G, F : Float, const N : usize> LocalAnalysis<F, Bounds<F>, N> | |
266 | for UniformProduct<G, N> | |
267 | where G : LocalAnalysis<F, Bounds<F>, 1> { | |
268 | #[inline] | |
269 | fn local_analysis(&self, cube : &Cube<F, N>) -> Bounds<F> { | |
270 | cube.iter_coords().map( | |
271 | |&[a, b]| self.0.local_analysis(&([[a, b]].into())) | |
272 | ).product() | |
273 | } | |
274 | } | |
275 | ||
276 | macro_rules! product_lpnorm { | |
277 | ($lp:ident) => { | |
278 | impl<G, F : Float, const N : usize> Norm<F, $lp> | |
279 | for UniformProduct<G, N> | |
280 | where G : Norm<F, $lp> { | |
281 | #[inline] | |
282 | fn norm(&self, lp : $lp) -> F { | |
283 | self.0.norm(lp).powi(N as i32) | |
284 | } | |
285 | } | |
286 | } | |
287 | } | |
288 | ||
289 | product_lpnorm!(L1); | |
290 | product_lpnorm!(L2); | |
291 | product_lpnorm!(Linfinity); | |
292 | ||
293 | ||
294 | /// Trait for bounding one kernel with respect to another. | |
295 | /// | |
296 | /// The type `F` is the scalar field, and `T` another kernel to which `Self` is compared. | |
297 | pub trait BoundedBy<F : Num, T> { | |
298 | /// Calclate a bounding factor $c$ such that the Fourier transforms $ℱ\[v\] ≤ c ℱ\[u\]$ for | |
299 | /// $v$ `self` and $u$ `other`. | |
300 | /// | |
301 | /// If no such factors exits, `None` is returned. | |
302 | fn bounding_factor(&self, other : &T) -> Option<F>; | |
303 | } | |
304 | ||
305 | /// This [`BoundedBy`] implementation bounds $(uv) * (uv)$ by $(ψ * ψ) u$. | |
306 | #[replace_float_literals(F::cast_from(literal))] | |
307 | impl<F, C, BaseP> | |
308 | BoundedBy<F, SupportProductFirst<AutoConvolution<C>, BaseP>> | |
309 | for AutoConvolution<SupportProductFirst<C, BaseP>> | |
310 | where F : Float, | |
311 | C : Clone + PartialEq, | |
312 | BaseP : Fourier<F> + PartialOrd, // TODO: replace by BoundedBy, | |
313 | <BaseP as Fourier<F>>::Transformed : Bounded<F> + Norm<F, L1> { | |
314 | ||
315 | fn bounding_factor(&self, kernel : &SupportProductFirst<AutoConvolution<C>, BaseP>) -> Option<F> { | |
316 | let SupportProductFirst(AutoConvolution(ref cutoff2), base_spread2) = kernel; | |
317 | let AutoConvolution(SupportProductFirst(ref cutoff, ref base_spread)) = self; | |
318 | let v̂ = base_spread.fourier(); | |
319 | ||
320 | // Verify that the cut-off and ideal physical model (base spread) are the same. | |
321 | if cutoff == cutoff2 | |
322 | && base_spread <= base_spread2 | |
323 | && v̂.bounds().lower() >= 0.0 { | |
324 | // Calculate the factor between the convolution approximation | |
325 | // `AutoConvolution<SupportProductFirst<C, BaseP>>` of $A_*A$ and the | |
326 | // kernel of the seminorm. This depends on the physical model P being | |
327 | // `SupportProductFirst<C, BaseP>` with the kernel `K` being | |
328 | // a `SupportSum` involving `SupportProductFirst<AutoConvolution<C>, BaseP>`. | |
329 | Some(v̂.norm(L1)) | |
330 | } else { | |
331 | // We cannot compare | |
332 | None | |
333 | } | |
334 | } | |
335 | } | |
336 | ||
337 | impl<F : Float, A, B, C> BoundedBy<F, SupportSum<B, C>> for A | |
338 | where A : BoundedBy<F, B>, | |
339 | C : Bounded<F> { | |
340 | ||
341 | #[replace_float_literals(F::cast_from(literal))] | |
342 | fn bounding_factor(&self, SupportSum(ref kernel1, kernel2) : &SupportSum<B, C>) -> Option<F> { | |
343 | if kernel2.bounds().lower() >= 0.0 { | |
344 | self.bounding_factor(kernel1) | |
345 | } else { | |
346 | None | |
347 | } | |
348 | } | |
349 | } | |
350 | ||
351 | /// Generates on $[a, b]$ [`Support::support_hint`] for a symmetric interval $[-r, r]$. | |
352 | /// | |
353 | /// It will attempt to place the subdivision point at $-r$ or $r$. | |
354 | /// If neither of these points lies within $[a, b]$, `None` is returned. | |
355 | #[inline] | |
356 | pub(super) fn symmetric_interval_hint<F : Float>(r : F, a : F, b : F) -> Option<F> { | |
357 | if a < -r && -r < b { | |
358 | Some(-r) | |
359 | } else if a < r && r < b { | |
360 | Some(r) | |
361 | } else { | |
362 | None | |
363 | } | |
364 | } | |
365 | ||
366 | /// Generates on $[a, b]$ [`Support::support_hint`] for a function with monotone derivative, | |
367 | /// support on $[-r, r]$ and peak at $0. | |
368 | /// | |
369 | /// It will attempt to place the subdivision point at $-r$, $r$, or $0$, depending on which | |
370 | /// gives the longer length for the shorter of the two subintervals. If none of these points | |
371 | /// lies within $[a, b]$, or the resulting interval would be shorter than $0.3r$, `None` is | |
372 | /// returned. | |
373 | #[replace_float_literals(F::cast_from(literal))] | |
374 | #[inline] | |
375 | pub(super) fn symmetric_peak_hint<F : Float>(r : F, a : F, b : F) -> Option<F> { | |
376 | let stage1 = if a < -r { | |
377 | if b <= -r { | |
378 | None | |
379 | } else if a + r < -b { | |
380 | Some(-r) | |
381 | } else { | |
382 | Some(0.0) | |
383 | } | |
384 | } else if a < 0.0 { | |
385 | if b <= 0.0 { | |
386 | None | |
387 | } else if a < r - b { | |
388 | Some(0.0) | |
389 | } else { | |
390 | Some(r) | |
391 | } | |
392 | } else if a < r && b > r { | |
393 | Some(r) | |
394 | } else { | |
395 | None | |
396 | }; | |
397 | ||
398 | // Ignore stage1 hint if either side of subdivision would be just a small fraction of the | |
399 | // interval | |
400 | match stage1 { | |
401 | Some(h) if (h - a).min(b-h) >= 0.3 * r => Some(h), | |
402 | _ => None | |
403 | } | |
404 | } |