Tue, 21 Mar 2023 20:31:01 +0200
Implement non-negativity constraints for the conditional gradient methods
0 | 1 | |
2 | //! Implementation of the standard mollifier | |
3 | ||
4 | use rgsl::hypergeometric::hyperg_U; | |
5 | use float_extras::f64::{tgamma as gamma}; | |
6 | use numeric_literals::replace_float_literals; | |
7 | use serde::Serialize; | |
8 | use alg_tools::types::*; | |
9 | use alg_tools::euclidean::Euclidean; | |
10 | use alg_tools::norms::*; | |
11 | use alg_tools::loc::Loc; | |
12 | use alg_tools::sets::Cube; | |
13 | use alg_tools::bisection_tree::{ | |
14 | Support, | |
15 | Constant, | |
16 | Bounds, | |
17 | LocalAnalysis, | |
18 | GlobalAnalysis | |
19 | }; | |
20 | use alg_tools::mapping::Apply; | |
21 | use alg_tools::maputil::array_init; | |
22 | ||
23 | /// Reresentation of the (unnormalised) standard mollifier. | |
24 | /// | |
25 | /// For the `width` parameter $ε>0$, this is | |
26 | /// <div>$$ | |
27 | /// f(x)=\begin{cases} | |
28 | /// e^{\frac{ε^2}{\|x\|_2^2-ε^2}}, & \|x\|_2 < ε, \\ | |
29 | /// 0, & \text{otherwise}. | |
30 | /// \end{cases} | |
31 | /// $$</div> | |
32 | #[derive(Copy,Clone,Serialize,Debug,Eq,PartialEq)] | |
33 | pub struct Mollifier<C : Constant, const N : usize> { | |
34 | /// The parameter $ε$ of the mollifier. | |
35 | pub width : C, | |
36 | } | |
37 | ||
38 | #[replace_float_literals(C::Type::cast_from(literal))] | |
39 | impl<'a, C : Constant, const N : usize> Apply<&'a Loc<C::Type, N>> for Mollifier<C, N> { | |
40 | type Output = C::Type; | |
41 | #[inline] | |
42 | fn apply(&self, x : &'a Loc<C::Type, N>) -> Self::Output { | |
43 | let ε = self.width.value(); | |
44 | let ε2 = ε*ε; | |
45 | let n2 = x.norm2_squared(); | |
46 | if n2 < ε2 { | |
47 | (n2 / (n2 - ε2)).exp() | |
48 | } else { | |
49 | 0.0 | |
50 | } | |
51 | } | |
52 | } | |
53 | ||
54 | impl<C : Constant, const N : usize> Apply<Loc<C::Type, N>> for Mollifier<C, N> { | |
55 | type Output = C::Type; | |
56 | #[inline] | |
57 | fn apply(&self, x : Loc<C::Type, N>) -> Self::Output { | |
58 | self.apply(&x) | |
59 | } | |
60 | } | |
61 | ||
62 | impl<'a, C : Constant, const N : usize> Support<C::Type, N> for Mollifier<C, N> { | |
63 | #[inline] | |
64 | fn support_hint(&self) -> Cube<C::Type,N> { | |
65 | let ε = self.width.value(); | |
66 | array_init(|| [-ε, ε]).into() | |
67 | } | |
68 | ||
69 | #[inline] | |
70 | fn in_support(&self, x : &Loc<C::Type,N>) -> bool { | |
71 | x.norm2() < self.width.value() | |
72 | } | |
73 | ||
74 | /*fn fully_in_support(&self, _cube : &Cube<C::Type,N>) -> bool { | |
75 | todo!("Not implemented, but not used at the moment") | |
76 | }*/ | |
77 | } | |
78 | ||
79 | #[replace_float_literals(C::Type::cast_from(literal))] | |
80 | impl<'a, C : Constant, const N : usize> GlobalAnalysis<C::Type, Bounds<C::Type>> | |
81 | for Mollifier<C, N> { | |
82 | #[inline] | |
83 | fn global_analysis(&self) -> Bounds<C::Type> { | |
84 | // The function is maximised/minimised where the 2-norm is minimised/maximised. | |
85 | Bounds(0.0, 1.0) | |
86 | } | |
87 | } | |
88 | ||
89 | impl<'a, C : Constant, const N : usize> LocalAnalysis<C::Type, Bounds<C::Type>, N> | |
90 | for Mollifier<C, N> { | |
91 | #[inline] | |
92 | fn local_analysis(&self, cube : &Cube<C::Type, N>) -> Bounds<C::Type> { | |
93 | // The function is maximised/minimised where the 2-norm is minimised/maximised. | |
94 | let lower = self.apply(cube.maxnorm_point()); | |
95 | let upper = self.apply(cube.minnorm_point()); | |
96 | Bounds(lower, upper) | |
97 | } | |
98 | } | |
99 | ||
100 | /// Calculate integral of the standard mollifier of width 1 in $ℝ^n$. | |
101 | /// | |
102 | /// This is based on the formula from | |
103 | /// [https://math.stackexchange.com/questions/4359683/integral-of-the-usual-mollifier-function-finding-its-necessary-constant](). | |
104 | /// | |
105 | /// If `rescaled` is `true`, return the integral of the scaled mollifier that has value one at the | |
106 | /// origin. | |
107 | #[inline] | |
108 | pub fn mollifier_norm1(n_ : usize, rescaled : bool) -> f64 { | |
109 | assert!(n_ > 0); | |
110 | let n = n_ as f64; | |
111 | let q = 2.0; | |
112 | let p = 2.0; | |
113 | let base = (2.0*gamma(1.0 + 1.0/p)).powi(n_ as i32) | |
114 | /*/ gamma(1.0 + n / p) | |
115 | * gamma(1.0 + n / q)*/ | |
116 | * hyperg_U(1.0 + n / q, 2.0, 1.0); | |
117 | if rescaled { base } else { base / f64::E } | |
118 | } | |
119 | ||
120 | impl<'a, C : Constant, const N : usize> Norm<C::Type, L1> | |
121 | for Mollifier<C, N> { | |
122 | #[inline] | |
123 | fn norm(&self, _ : L1) -> C::Type { | |
124 | let ε = self.width.value(); | |
125 | C::Type::cast_from(mollifier_norm1(N, true)) * ε.powi(N as i32) | |
126 | } | |
127 | } | |
128 | ||
129 | #[replace_float_literals(C::Type::cast_from(literal))] | |
130 | impl<'a, C : Constant, const N : usize> Norm<C::Type, Linfinity> | |
131 | for Mollifier<C, N> { | |
132 | #[inline] | |
133 | fn norm(&self, _ : Linfinity) -> C::Type { | |
134 | 1.0 | |
135 | } | |
136 | } |