Tue, 06 Dec 2022 14:12:20 +0200
v1.0.0-pre-arxiv (missing arXiv links)
0 | 1 | /*! |
2 | Forward models from discrete measures to observations. | |
3 | */ | |
4 | ||
5 | use numeric_literals::replace_float_literals; | |
6 | use nalgebra::base::{ | |
7 | DMatrix, | |
8 | DVector | |
9 | }; | |
10 | use std::iter::Zip; | |
11 | use std::ops::RangeFrom; | |
12 | use std::marker::PhantomData; | |
13 | ||
14 | pub use alg_tools::linops::*; | |
15 | use alg_tools::euclidean::Euclidean; | |
16 | use alg_tools::norms::{ | |
17 | L1, Linfinity, Norm | |
18 | }; | |
19 | use alg_tools::bisection_tree::*; | |
20 | use alg_tools::mapping::RealMapping; | |
21 | use alg_tools::lingrid::*; | |
22 | use alg_tools::iter::{MapX, Mappable}; | |
23 | use alg_tools::nalgebra_support::ToNalgebraRealField; | |
24 | use alg_tools::tabledump::write_csv; | |
25 | use alg_tools::error::DynError; | |
26 | ||
27 | use crate::types::*; | |
28 | use crate::measures::*; | |
29 | use crate::seminorms::{ | |
30 | Lipschitz, | |
31 | ConvolutionOp, | |
32 | SimpleConvolutionKernel, | |
33 | }; | |
34 | use crate::kernels::{ | |
35 | Convolution, | |
36 | AutoConvolution, | |
37 | BoundedBy, | |
38 | }; | |
39 | ||
40 | pub type RNDM<F, const N : usize> = DiscreteMeasure<Loc<F,N>, F>; | |
41 | ||
42 | /// `ForwardeModel`s are bounded preadjointable linear operators $A ∈ 𝕃(𝒵(Ω); E)$ | |
43 | /// where $𝒵(Ω) ⊂ ℳ(Ω)$ is the space of sums of delta measures, presented by | |
44 | /// [`DiscreteMeasure`], and $E$ is a [`Euclidean`] space. | |
45 | pub trait ForwardModel<Domain, F : Float + ToNalgebraRealField> | |
46 | : BoundedLinear<DiscreteMeasure<Domain, F>, Codomain=Self::Observable, FloatType=F> | |
47 | + GEMV<F, DiscreteMeasure<Domain, F>, Self::Observable> | |
48 | + Linear<DeltaMeasure<Domain, F>, Codomain=Self::Observable> | |
49 | + Preadjointable<DiscreteMeasure<Domain, F>, Self::Observable> { | |
50 | /// The codomain or value space (of “observables”) for this operator. | |
51 | /// It is assumed to be a [`Euclidean`] space, and therefore also (identified with) | |
52 | /// the domain of the preadjoint. | |
53 | type Observable : Euclidean<F, Output=Self::Observable> | |
54 | + AXPY<F> | |
55 | + Clone; | |
56 | ||
57 | /// Return A_*A and A_* b | |
58 | fn findim_quadratic_model( | |
59 | &self, | |
60 | μ : &DiscreteMeasure<Domain, F>, | |
61 | b : &Self::Observable | |
62 | ) -> (DMatrix<F::MixedType>, DVector<F::MixedType>); | |
63 | ||
64 | /// Write an observable into a file. | |
65 | fn write_observable(&self, b : &Self::Observable, prefix : String) -> DynError; | |
66 | ||
67 | /// Returns a zero observable | |
68 | fn zero_observable(&self) -> Self::Observable; | |
69 | ||
70 | /// Returns an empty (uninitialised) observable. | |
71 | /// | |
72 | /// This is used as a placeholder for temporary [`std::mem::replace`] move operations. | |
73 | fn empty_observable(&self) -> Self::Observable; | |
74 | } | |
75 | ||
76 | pub type ShiftedSensor<F, S, P, const N : usize> = Shift<Convolution<S, P>, F, N>; | |
77 | ||
78 | /// Trait for physical convolution models. Has blanket implementation for all cases. | |
79 | pub trait Spread<F : Float, const N : usize> | |
80 | : 'static + Clone + Support<F, N> + RealMapping<F, N> + Bounded<F> {} | |
81 | ||
82 | impl<F, T, const N : usize> Spread<F, N> for T | |
83 | where F : Float, | |
84 | T : 'static + Clone + Support<F, N> + Bounded<F> + RealMapping<F, N> {} | |
85 | ||
86 | /// Trait for compactly supported sensors. Has blanket implementation for all cases. | |
87 | pub trait Sensor<F : Float, const N : usize> : Spread<F, N> + Norm<F, L1> + Norm<F, Linfinity> {} | |
88 | ||
89 | impl<F, T, const N : usize> Sensor<F, N> for T | |
90 | where F : Float, | |
91 | T : Spread<F, N> + Norm<F, L1> + Norm<F, Linfinity> {} | |
92 | ||
93 | ||
94 | pub trait SensorGridBT<F, S, P, const N : usize> : | |
95 | Clone + BTImpl<F, N, Data=usize, Agg=Bounds<F>> | |
96 | where F : Float, | |
97 | S : Sensor<F, N>, | |
98 | P : Spread<F, N> {} | |
99 | ||
100 | impl<F, S, P, T, const N : usize> | |
101 | SensorGridBT<F, S, P, N> | |
102 | for T | |
103 | where T : Clone + BTImpl<F, N, Data=usize, Agg=Bounds<F>>, | |
104 | F : Float, | |
105 | S : Sensor<F, N>, | |
106 | P : Spread<F, N> {} | |
107 | ||
108 | // We need type alias bounds to access associated types | |
109 | #[allow(type_alias_bounds)] | |
110 | type SensorGridBTFN<F, S, P, BT : SensorGridBT<F, S, P, N>, const N : usize> | |
111 | = BTFN<F, SensorGridSupportGenerator<F, S, P, N>, BT, N>; | |
112 | ||
113 | /// Sensor grid forward model | |
114 | #[derive(Clone)] | |
115 | pub struct SensorGrid<F, S, P, BT, const N : usize> | |
116 | where F : Float, | |
117 | S : Sensor<F, N>, | |
118 | P : Spread<F, N>, | |
119 | Convolution<S, P> : Spread<F, N>, | |
120 | BT : SensorGridBT<F, S, P, N>, { | |
121 | domain : Cube<F, N>, | |
122 | sensor_count : [usize; N], | |
123 | sensor : S, | |
124 | spread : P, | |
125 | base_sensor : Convolution<S, P>, | |
126 | bt : BT, | |
127 | } | |
128 | ||
129 | impl<F, S, P, BT, const N : usize> SensorGrid<F, S, P, BT, N> | |
130 | where F : Float, | |
131 | BT : SensorGridBT<F, S, P, N>, | |
132 | S : Sensor<F, N>, | |
133 | P : Spread<F, N>, | |
134 | Convolution<S, P> : Spread<F, N>, | |
135 | ShiftedSensor<F, S, P, N> : LocalAnalysis<F, BT::Agg, N> { | |
136 | ||
137 | pub fn new( | |
138 | domain : Cube<F, N>, | |
139 | sensor_count : [usize; N], | |
140 | sensor : S, | |
141 | spread : P, | |
142 | depth : BT::Depth | |
143 | ) -> Self { | |
144 | let base_sensor = Convolution(sensor.clone(), spread.clone()); | |
145 | let bt = BT::new(domain, depth); | |
146 | let mut sensorgrid = SensorGrid { | |
147 | domain, | |
148 | sensor_count, | |
149 | sensor, | |
150 | spread, | |
151 | base_sensor, | |
152 | bt, | |
153 | }; | |
154 | ||
155 | for (x, id) in sensorgrid.grid().into_iter().zip(0usize..) { | |
156 | let s = sensorgrid.shifted_sensor(x); | |
157 | sensorgrid.bt.insert(id, &s); | |
158 | } | |
159 | ||
160 | sensorgrid | |
161 | } | |
162 | ||
163 | pub fn grid(&self) -> LinGrid<F, N> { | |
164 | lingrid_centered(&self.domain, &self.sensor_count) | |
165 | } | |
166 | ||
167 | pub fn n_sensors(&self) -> usize { | |
168 | self.sensor_count.iter().product() | |
169 | } | |
170 | ||
171 | #[inline] | |
172 | fn shifted_sensor(&self, x : Loc<F, N>) -> ShiftedSensor<F, S, P, N> { | |
173 | self.base_sensor.clone().shift(x) | |
174 | } | |
175 | ||
176 | #[inline] | |
177 | fn _zero_observable(&self) -> DVector<F> { | |
178 | DVector::zeros(self.n_sensors()) | |
179 | } | |
180 | } | |
181 | ||
182 | impl<F, S, P, BT, const N : usize> Apply<RNDM<F, N>> for SensorGrid<F, S, P, BT, N> | |
183 | where F : Float, | |
184 | BT : SensorGridBT<F, S, P, N>, | |
185 | S : Sensor<F, N>, | |
186 | P : Spread<F, N>, | |
187 | Convolution<S, P> : Spread<F, N>, | |
188 | ShiftedSensor<F, S, P, N> : LocalAnalysis<F, BT::Agg, N> { | |
189 | ||
190 | type Output = DVector<F>; | |
191 | ||
192 | #[inline] | |
193 | fn apply(&self, μ : RNDM<F, N>) -> DVector<F> { | |
194 | self.apply(&μ) | |
195 | } | |
196 | } | |
197 | ||
198 | impl<'a, F, S, P, BT, const N : usize> Apply<&'a RNDM<F, N>> for SensorGrid<F, S, P, BT, N> | |
199 | where F : Float, | |
200 | BT : SensorGridBT<F, S, P, N>, | |
201 | S : Sensor<F, N>, | |
202 | P : Spread<F, N>, | |
203 | Convolution<S, P> : Spread<F, N>, | |
204 | ShiftedSensor<F, S, P, N> : LocalAnalysis<F, BT::Agg, N> { | |
205 | ||
206 | type Output = DVector<F>; | |
207 | ||
208 | fn apply(&self, μ : &'a RNDM<F, N>) -> DVector<F> { | |
209 | let mut res = self._zero_observable(); | |
210 | self.apply_add(&mut res, μ); | |
211 | res | |
212 | } | |
213 | } | |
214 | ||
215 | impl<F, S, P, BT, const N : usize> Linear<RNDM<F, N>> for SensorGrid<F, S, P, BT, N> | |
216 | where F : Float, | |
217 | BT : SensorGridBT<F, S, P, N>, | |
218 | S : Sensor<F, N>, | |
219 | P : Spread<F, N>, | |
220 | Convolution<S, P> : Spread<F, N>, | |
221 | ShiftedSensor<F, S, P, N> : LocalAnalysis<F, BT::Agg, N> { | |
222 | type Codomain = DVector<F>; | |
223 | } | |
224 | ||
225 | ||
226 | #[replace_float_literals(F::cast_from(literal))] | |
227 | impl<F, S, P, BT, const N : usize> GEMV<F, RNDM<F, N>, DVector<F>> for SensorGrid<F, S, P, BT, N> | |
228 | where F : Float, | |
229 | BT : SensorGridBT<F, S, P, N>, | |
230 | S : Sensor<F, N>, | |
231 | P : Spread<F, N>, | |
232 | Convolution<S, P> : Spread<F, N>, | |
233 | ShiftedSensor<F, S, P, N> : LocalAnalysis<F, BT::Agg, N> { | |
234 | ||
235 | fn gemv(&self, y : &mut DVector<F>, α : F, μ : &RNDM<F, N>, β : F) { | |
236 | let grid = self.grid(); | |
237 | if β == 0.0 { | |
238 | y.fill(0.0) | |
239 | } else if β != 1.0 { | |
240 | *y *= β; // Need to multiply first, as we have to be able to add to y. | |
241 | } | |
242 | if α == 1.0 { | |
243 | self.apply_add(y, μ) | |
244 | } else { | |
245 | for δ in μ.iter_spikes() { | |
246 | for &d in self.bt.iter_at(&δ.x) { | |
247 | let sensor = self.shifted_sensor(grid.entry_linear_unchecked(d)); | |
248 | y[d] += sensor.apply(&δ.x) * (α * δ.α); | |
249 | } | |
250 | } | |
251 | } | |
252 | } | |
253 | ||
254 | fn apply_add(&self, y : &mut DVector<F>, μ : &RNDM<F, N>) { | |
255 | let grid = self.grid(); | |
256 | for δ in μ.iter_spikes() { | |
257 | for &d in self.bt.iter_at(&δ.x) { | |
258 | let sensor = self.shifted_sensor(grid.entry_linear_unchecked(d)); | |
259 | y[d] += sensor.apply(&δ.x) * δ.α; | |
260 | } | |
261 | } | |
262 | } | |
263 | ||
264 | } | |
265 | ||
266 | impl<F, S, P, BT, const N : usize> Apply<DeltaMeasure<Loc<F, N>, F>> | |
267 | for SensorGrid<F, S, P, BT, N> | |
268 | where F : Float, | |
269 | BT : SensorGridBT<F, S, P, N>, | |
270 | S : Sensor<F, N>, | |
271 | P : Spread<F, N>, | |
272 | Convolution<S, P> : Spread<F, N>, | |
273 | ShiftedSensor<F, S, P, N> : LocalAnalysis<F, BT::Agg, N> { | |
274 | ||
275 | type Output = DVector<F>; | |
276 | ||
277 | #[inline] | |
278 | fn apply(&self, δ : DeltaMeasure<Loc<F, N>, F>) -> DVector<F> { | |
279 | self.apply(&δ) | |
280 | } | |
281 | } | |
282 | ||
283 | impl<'a, F, S, P, BT, const N : usize> Apply<&'a DeltaMeasure<Loc<F, N>, F>> | |
284 | for SensorGrid<F, S, P, BT, N> | |
285 | where F : Float, | |
286 | BT : SensorGridBT<F, S, P, N>, | |
287 | S : Sensor<F, N>, | |
288 | P : Spread<F, N>, | |
289 | Convolution<S, P> : Spread<F, N>, | |
290 | ShiftedSensor<F, S, P, N> : LocalAnalysis<F, BT::Agg, N> { | |
291 | ||
292 | type Output = DVector<F>; | |
293 | ||
294 | fn apply(&self, δ : &DeltaMeasure<Loc<F, N>, F>) -> DVector<F> { | |
295 | let mut res = DVector::zeros(self.n_sensors()); | |
296 | let grid = self.grid(); | |
297 | for &d in self.bt.iter_at(&δ.x) { | |
298 | let sensor = self.shifted_sensor(grid.entry_linear_unchecked(d)); | |
299 | res[d] += sensor.apply(&δ.x) * δ.α; | |
300 | } | |
301 | res | |
302 | } | |
303 | } | |
304 | ||
305 | impl<F, S, P, BT, const N : usize> Linear<DeltaMeasure<Loc<F, N>, F>> for SensorGrid<F, S, P, BT, N> | |
306 | where F : Float, | |
307 | BT : SensorGridBT<F, S, P, N>, | |
308 | S : Sensor<F, N>, | |
309 | P : Spread<F, N>, | |
310 | Convolution<S, P> : Spread<F, N>, | |
311 | ShiftedSensor<F, S, P, N> : LocalAnalysis<F, BT::Agg, N> { | |
312 | type Codomain = DVector<F>; | |
313 | } | |
314 | ||
315 | impl<F, S, P, BT, const N : usize> BoundedLinear<RNDM<F, N>> for SensorGrid<F, S, P, BT, N> | |
316 | where F : Float, | |
317 | BT : SensorGridBT<F, S, P, N, Agg=Bounds<F>>, | |
318 | S : Sensor<F, N>, | |
319 | P : Spread<F, N>, | |
320 | Convolution<S, P> : Spread<F, N>, | |
321 | ShiftedSensor<F, S, P, N> : LocalAnalysis<F, BT::Agg, N> { | |
322 | type FloatType = F; | |
323 | ||
324 | /// An estimate on the operator norm in $𝕃(ℳ(Ω); ℝ^n)$ with $ℳ(Ω)$ equipped | |
325 | /// with the Radon norm, and $ℝ^n$ with the Euclidean norm. | |
326 | fn opnorm_bound(&self) -> F { | |
327 | // With {x_i}_{i=1}^n the grid centres and φ the kernel, we have | |
328 | // |Aμ|_2 = sup_{|z|_2 ≤ 1} ⟨z,Αμ⟩ = sup_{|z|_2 ≤ 1} ⟨A^*z|μ⟩ | |
329 | // ≤ sup_{|z|_2 ≤ 1} |A^*z|_∞ |μ|_ℳ | |
330 | // = sup_{|z|_2 ≤ 1} |∑ φ(· - x_i)z_i|_∞ |μ|_ℳ | |
331 | // ≤ sup_{|z|_2 ≤ 1} |φ|_∞ ∑ |z_i| |μ|_ℳ | |
332 | // ≤ sup_{|z|_2 ≤ 1} |φ|_∞ √n |z|_2 |μ|_ℳ | |
333 | // = |φ|_∞ √n |μ|_ℳ. | |
334 | // Hence | |
335 | let n = F::cast_from(self.n_sensors()); | |
336 | self.base_sensor.bounds().uniform() * n.sqrt() | |
337 | } | |
338 | } | |
339 | ||
340 | type SensorGridPreadjoint<'a, A, F, const N : usize> = PreadjointHelper<'a, A, RNDM<F,N>>; | |
341 | ||
342 | ||
343 | impl<F, S, P, BT, const N : usize> | |
344 | Preadjointable<RNDM<F, N>, DVector<F>> | |
345 | for SensorGrid<F, S, P, BT, N> | |
346 | where F : Float, | |
347 | BT : SensorGridBT<F, S, P, N>, | |
348 | S : Sensor<F, N>, | |
349 | P : Spread<F, N>, | |
350 | Convolution<S, P> : Spread<F, N>, | |
351 | ShiftedSensor<F, S, P, N> : LocalAnalysis<F, BT::Agg, N>, | |
352 | Weighted<ShiftedSensor<F, S, P, N>, F> : LocalAnalysis<F, BT::Agg, N> { | |
353 | type PreadjointCodomain = BTFN<F, SensorGridSupportGenerator<F, S, P, N>, BT, N>; | |
354 | type Preadjoint<'a> = SensorGridPreadjoint<'a, Self, F, N> where Self : 'a; | |
355 | ||
356 | fn preadjoint(&self) -> Self::Preadjoint<'_> { | |
357 | PreadjointHelper::new(self) | |
358 | } | |
359 | } | |
360 | ||
361 | #[derive(Clone,Debug)] | |
362 | pub struct SensorGridSupportGenerator<F, S, P, const N : usize> | |
363 | where F : Float, | |
364 | S : Sensor<F, N>, | |
365 | P : Spread<F, N> { | |
366 | base_sensor : Convolution<S, P>, | |
367 | grid : LinGrid<F, N>, | |
368 | weights : DVector<F> | |
369 | } | |
370 | ||
371 | impl<F, S, P, const N : usize> SensorGridSupportGenerator<F, S, P, N> | |
372 | where F : Float, | |
373 | S : Sensor<F, N>, | |
374 | P : Spread<F, N>, | |
375 | Convolution<S, P> : Spread<F, N> { | |
376 | ||
377 | #[inline] | |
378 | fn construct_sensor(&self, id : usize, w : F) -> Weighted<ShiftedSensor<F, S, P, N>, F> { | |
379 | let x = self.grid.entry_linear_unchecked(id); | |
380 | self.base_sensor.clone().shift(x).weigh(w) | |
381 | } | |
382 | ||
383 | #[inline] | |
384 | fn construct_sensor_and_id<'a>(&'a self, (id, w) : (usize, &'a F)) | |
385 | -> (usize, Weighted<ShiftedSensor<F, S, P, N>, F>) { | |
386 | (id.into(), self.construct_sensor(id, *w)) | |
387 | } | |
388 | } | |
389 | ||
390 | impl<F, S, P, const N : usize> SupportGenerator<F, N> | |
391 | for SensorGridSupportGenerator<F, S, P, N> | |
392 | where F : Float, | |
393 | S : Sensor<F, N>, | |
394 | P : Spread<F, N>, | |
395 | Convolution<S, P> : Spread<F, N> { | |
396 | type Id = usize; | |
397 | type SupportType = Weighted<ShiftedSensor<F, S, P, N>, F>; | |
398 | type AllDataIter<'a> = MapX<'a, Zip<RangeFrom<usize>, | |
399 | std::slice::Iter<'a, F>>, | |
400 | Self, | |
401 | (Self::Id, Self::SupportType)> | |
402 | where Self : 'a; | |
403 | ||
404 | #[inline] | |
405 | fn support_for(&self, d : Self::Id) -> Self::SupportType { | |
406 | self.construct_sensor(d, self.weights[d]) | |
407 | } | |
408 | ||
409 | #[inline] | |
410 | fn support_count(&self) -> usize { | |
411 | self.weights.len() | |
412 | } | |
413 | ||
414 | #[inline] | |
415 | fn all_data(&self) -> Self::AllDataIter<'_> { | |
416 | (0..).zip(self.weights.as_slice().iter()).mapX(self, Self::construct_sensor_and_id) | |
417 | } | |
418 | } | |
419 | ||
420 | /// Helper structure for constructing preadjoints of `S` where `S : Linear<X>`. | |
421 | /// [`Linear`] needs to be implemented for each instance, but [`Adjointable`] | |
422 | /// and [`BoundedLinear`] have blanket implementations. | |
423 | #[derive(Clone,Debug)] | |
424 | pub struct PreadjointHelper<'a, S : 'a, X> { | |
425 | forward_op : &'a S, | |
426 | _domain : PhantomData<X> | |
427 | } | |
428 | ||
429 | impl<'a, S : 'a, X> PreadjointHelper<'a, S, X> { | |
430 | pub fn new(forward_op : &'a S) -> Self { | |
431 | PreadjointHelper { forward_op, _domain: PhantomData } | |
432 | } | |
433 | } | |
434 | ||
435 | impl<'a, X, Ypre, S> Adjointable<Ypre, X> | |
436 | for PreadjointHelper<'a, S, X> | |
437 | where Self : Linear<Ypre>, | |
438 | S : Clone + Linear<X> { | |
439 | type AdjointCodomain = S::Codomain; | |
440 | type Adjoint<'b> = S where Self : 'b; | |
441 | fn adjoint(&self) -> Self::Adjoint<'_> { | |
442 | self.forward_op.clone() | |
443 | } | |
444 | } | |
445 | ||
446 | impl<'a, X, Ypre, S> BoundedLinear<Ypre> | |
447 | for PreadjointHelper<'a, S, X> | |
448 | where Self : Linear<Ypre>, | |
449 | S : 'a + Clone + BoundedLinear<X> { | |
450 | type FloatType = S::FloatType; | |
451 | fn opnorm_bound(&self) -> Self::FloatType { | |
452 | self.forward_op.opnorm_bound() | |
453 | } | |
454 | } | |
455 | ||
456 | ||
457 | impl<'a, 'b, F, S, P, BT, const N : usize> Apply<&'b DVector<F>> | |
458 | for PreadjointHelper<'a, SensorGrid<F, S, P, BT, N>, RNDM<F,N>> | |
459 | where F : Float, | |
460 | BT : SensorGridBT<F, S, P, N>, | |
461 | S : Sensor<F, N>, | |
462 | P : Spread<F, N>, | |
463 | Convolution<S, P> : Spread<F, N>, | |
464 | ShiftedSensor<F, S, P, N> : LocalAnalysis<F, BT::Agg, N>, | |
465 | Weighted<ShiftedSensor<F, S, P, N>, F> : LocalAnalysis<F, BT::Agg, N> { | |
466 | ||
467 | type Output = SensorGridBTFN<F, S, P, BT, N>; | |
468 | ||
469 | fn apply(&self, x : &'b DVector<F>) -> Self::Output { | |
470 | self.apply(x.clone()) | |
471 | } | |
472 | } | |
473 | ||
474 | impl<'a, F, S, P, BT, const N : usize> Apply<DVector<F>> | |
475 | for PreadjointHelper<'a, SensorGrid<F, S, P, BT, N>, RNDM<F,N>> | |
476 | where F : Float, | |
477 | BT : SensorGridBT<F, S, P, N>, | |
478 | S : Sensor<F, N>, | |
479 | P : Spread<F, N>, | |
480 | Convolution<S, P> : Spread<F, N>, | |
481 | ShiftedSensor<F, S, P, N> : LocalAnalysis<F, BT::Agg, N>, | |
482 | Weighted<ShiftedSensor<F, S, P, N>, F> : LocalAnalysis<F, BT::Agg, N> { | |
483 | ||
484 | type Output = SensorGridBTFN<F, S, P, BT, N>; | |
485 | ||
486 | fn apply(&self, x : DVector<F>) -> Self::Output { | |
487 | let fwd = &self.forward_op; | |
488 | let generator = SensorGridSupportGenerator{ | |
489 | base_sensor : fwd.base_sensor.clone(), | |
490 | grid : fwd.grid(), | |
491 | weights : x | |
492 | }; | |
493 | BTFN::new_refresh(&fwd.bt, generator) | |
494 | } | |
495 | } | |
496 | ||
497 | impl<'a, F, S, P, BT, const N : usize> Linear<DVector<F>> | |
498 | for PreadjointHelper<'a, SensorGrid<F, S, P, BT, N>, RNDM<F,N>> | |
499 | where F : Float, | |
500 | BT : SensorGridBT<F, S, P, N>, | |
501 | S : Sensor<F, N>, | |
502 | P : Spread<F, N>, | |
503 | Convolution<S, P> : Spread<F, N>, | |
504 | ShiftedSensor<F, S, P, N> : LocalAnalysis<F, BT::Agg, N>, | |
505 | Weighted<ShiftedSensor<F, S, P, N>, F> : LocalAnalysis<F, BT::Agg, N> { | |
506 | ||
507 | type Codomain = SensorGridBTFN<F, S, P, BT, N>; | |
508 | } | |
509 | ||
510 | impl<F, S, P, BT, const N : usize> ForwardModel<Loc<F, N>, F> | |
511 | for SensorGrid<F, S, P, BT, N> | |
512 | where F : Float + ToNalgebraRealField<MixedType=F> + nalgebra::RealField, | |
513 | BT : SensorGridBT<F, S, P, N>, | |
514 | S : Sensor<F, N>, | |
515 | P : Spread<F, N>, | |
516 | Convolution<S, P> : Spread<F, N>, | |
517 | ShiftedSensor<F, S, P, N> : LocalAnalysis<F, BT::Agg, N>, | |
518 | Weighted<ShiftedSensor<F, S, P, N>, F> : LocalAnalysis<F, BT::Agg, N> { | |
519 | type Observable = DVector<F>; | |
520 | ||
521 | fn findim_quadratic_model( | |
522 | &self, | |
523 | μ : &DiscreteMeasure<Loc<F, N>, F>, | |
524 | b : &Self::Observable | |
525 | ) -> (DMatrix<F::MixedType>, DVector<F::MixedType>) { | |
526 | assert_eq!(b.len(), self.n_sensors()); | |
527 | let mut mA = DMatrix::zeros(self.n_sensors(), μ.len()); | |
528 | let grid = self.grid(); | |
529 | for (mut mAcol, δ) in mA.column_iter_mut().zip(μ.iter_spikes()) { | |
530 | for &d in self.bt.iter_at(&δ.x) { | |
531 | let sensor = self.shifted_sensor(grid.entry_linear_unchecked(d)); | |
532 | mAcol[d] += sensor.apply(&δ.x); | |
533 | } | |
534 | } | |
535 | let mAt = mA.transpose(); | |
536 | (&mAt * mA, &mAt * b) | |
537 | } | |
538 | ||
539 | fn write_observable(&self, b : &Self::Observable, prefix : String) -> DynError { | |
540 | let it = self.grid().into_iter().zip(b.iter()).map(|(x, &v)| (x, v)); | |
541 | write_csv(it, prefix + ".txt") | |
542 | } | |
543 | ||
544 | #[inline] | |
545 | fn zero_observable(&self) -> Self::Observable { | |
546 | self._zero_observable() | |
547 | } | |
548 | ||
549 | #[inline] | |
550 | fn empty_observable(&self) -> Self::Observable { | |
551 | DVector::zeros(0) | |
552 | } | |
553 | ||
554 | } | |
555 | ||
556 | /// Implements the calculation a factor $L$ such that $A_*A ≤ L 𝒟$ for $A$ the forward model | |
557 | /// and $𝒟$ a seminorm of suitable form. | |
558 | /// | |
559 | /// **This assumes (but does not check) that the sensors are not overlapping.** | |
560 | #[replace_float_literals(F::cast_from(literal))] | |
561 | impl<F, BT, S, P, K, const N : usize> Lipschitz<ConvolutionOp<F, K, BT, N>> | |
562 | for SensorGrid<F, S, P, BT, N> | |
563 | where F : Float + nalgebra::RealField + ToNalgebraRealField, | |
564 | BT : SensorGridBT<F, S, P, N>, | |
565 | S : Sensor<F, N>, | |
566 | P : Spread<F, N>, | |
567 | Convolution<S, P> : Spread<F, N>, | |
568 | K : SimpleConvolutionKernel<F, N>, | |
569 | AutoConvolution<P> : BoundedBy<F, K> { | |
570 | ||
571 | type FloatType = F; | |
572 | ||
573 | fn lipschitz_factor(&self, seminorm : &ConvolutionOp<F, K, BT, N>) -> Option<F> { | |
574 | // Sensors should not take on negative values to allow | |
575 | // A_*A to be upper bounded by a simple convolution of `spread`. | |
576 | if self.sensor.bounds().lower() < 0.0 { | |
577 | return None | |
578 | } | |
579 | ||
580 | // Calculate the factor $L_1$ for betwee $ℱ[ψ * ψ] ≤ L_1 ℱ[ρ]$ for $ψ$ the base spread | |
581 | // and $ρ$ the kernel of the seminorm. | |
582 | let l1 = AutoConvolution(self.spread.clone()).bounding_factor(seminorm.kernel())?; | |
583 | ||
584 | // Calculate the factor for transitioning from $A_*A$ to `AutoConvolution<P>`, where A | |
585 | // consists of several `Convolution<S, P>` for the physical model `P` and the sensor `S`. | |
586 | let l0 = self.sensor.norm(Linfinity) * self.sensor.norm(L1); | |
587 | ||
588 | // The final transition factor is: | |
589 | Some(l0 * l1) | |
590 | } | |
591 | } | |
592 | ||
593 | macro_rules! make_sensorgridsupportgenerator_scalarop_rhs { | |
594 | ($trait:ident, $fn:ident, $trait_assign:ident, $fn_assign:ident) => { | |
595 | impl<F, S, P, const N : usize> | |
596 | std::ops::$trait_assign<F> | |
597 | for SensorGridSupportGenerator<F, S, P, N> | |
598 | where F : Float, | |
599 | S : Sensor<F, N>, | |
600 | P : Spread<F, N>, | |
601 | Convolution<S, P> : Spread<F, N> { | |
602 | fn $fn_assign(&mut self, t : F) { | |
603 | self.weights.$fn_assign(t); | |
604 | } | |
605 | } | |
606 | ||
607 | impl<F, S, P, const N : usize> | |
608 | std::ops::$trait<F> | |
609 | for SensorGridSupportGenerator<F, S, P, N> | |
610 | where F : Float, | |
611 | S : Sensor<F, N>, | |
612 | P : Spread<F, N>, | |
613 | Convolution<S, P> : Spread<F, N> { | |
614 | type Output = SensorGridSupportGenerator<F, S, P, N>; | |
615 | fn $fn(mut self, t : F) -> Self::Output { | |
616 | std::ops::$trait_assign::$fn_assign(&mut self.weights, t); | |
617 | self | |
618 | } | |
619 | } | |
620 | ||
621 | impl<'a, F, S, P, const N : usize> | |
622 | std::ops::$trait<F> | |
623 | for &'a SensorGridSupportGenerator<F, S, P, N> | |
624 | where F : Float, | |
625 | S : Sensor<F, N>, | |
626 | P : Spread<F, N>, | |
627 | Convolution<S, P> : Spread<F, N> { | |
628 | type Output = SensorGridSupportGenerator<F, S, P, N>; | |
629 | fn $fn(self, t : F) -> Self::Output { | |
630 | SensorGridSupportGenerator{ | |
631 | base_sensor : self.base_sensor.clone(), | |
632 | grid : self.grid, | |
633 | weights : (&self.weights).$fn(t) | |
634 | } | |
635 | } | |
636 | } | |
637 | } | |
638 | } | |
639 | ||
640 | make_sensorgridsupportgenerator_scalarop_rhs!(Mul, mul, MulAssign, mul_assign); | |
641 | make_sensorgridsupportgenerator_scalarop_rhs!(Div, div, DivAssign, div_assign); | |
642 | ||
643 | macro_rules! make_sensorgridsupportgenerator_unaryop { | |
644 | ($trait:ident, $fn:ident) => { | |
645 | impl<F, S, P, const N : usize> | |
646 | std::ops::$trait | |
647 | for SensorGridSupportGenerator<F, S, P, N> | |
648 | where F : Float, | |
649 | S : Sensor<F, N>, | |
650 | P : Spread<F, N>, | |
651 | Convolution<S, P> : Spread<F, N> { | |
652 | type Output = SensorGridSupportGenerator<F, S, P, N>; | |
653 | fn $fn(mut self) -> Self::Output { | |
654 | self.weights = self.weights.$fn(); | |
655 | self | |
656 | } | |
657 | } | |
658 | ||
659 | impl<'a, F, S, P, const N : usize> | |
660 | std::ops::$trait | |
661 | for &'a SensorGridSupportGenerator<F, S, P, N> | |
662 | where F : Float, | |
663 | S : Sensor<F, N>, | |
664 | P : Spread<F, N>, | |
665 | Convolution<S, P> : Spread<F, N> { | |
666 | type Output = SensorGridSupportGenerator<F, S, P, N>; | |
667 | fn $fn(self) -> Self::Output { | |
668 | SensorGridSupportGenerator{ | |
669 | base_sensor : self.base_sensor.clone(), | |
670 | grid : self.grid, | |
671 | weights : (&self.weights).$fn() | |
672 | } | |
673 | } | |
674 | } | |
675 | } | |
676 | } | |
677 | ||
678 | make_sensorgridsupportgenerator_unaryop!(Neg, neg); |