1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
// Copyright 2018-2024 argmin developers
//
// Licensed under the Apache License, Version 2.0 <LICENSE-APACHE or
// http://apache.org/licenses/LICENSE-2.0> or the MIT license <LICENSE-MIT or
// http://opensource.org/licenses/MIT>, at your option. This file may not be
// copied, modified, or distributed except according to those terms.

use crate::core::{ArgminFloat, Error, Gradient, Hessian, IterState, Problem, Solver, KV};
use argmin_math::{ArgminDot, ArgminInv, ArgminScaledSub};
#[cfg(feature = "serde1")]
use serde::{Deserialize, Serialize};

/// # Newton's method
///
/// Newton's method iteratively finds the stationary points of a function f by using a second order
/// approximation of f at the current point.
///
/// The stepsize `gamma` can be adjusted with the [`with_gamma`](`Newton::with_gamma`) method. It
/// must be in `(0, 1])` and defaults to `1`.
///
/// ## Requirements on the optimization problem
///
/// The optimization problem is required to implement [`Gradient`] and [`Hessian`].
///
/// ## Reference
///
/// Jorge Nocedal and Stephen J. Wright (2006). Numerical Optimization.
/// Springer. ISBN 0-387-30303-0.
#[derive(Clone, Copy)]
#[cfg_attr(feature = "serde1", derive(Serialize, Deserialize))]
pub struct Newton<F> {
    /// gamma
    gamma: F,
}

impl<F> Newton<F>
where
    F: ArgminFloat,
{
    /// Construct a new instance of [`Newton`]
    ///
    /// # Example
    ///
    /// ```
    /// # use argmin::solver::newton::Newton;
    /// let newton: Newton<f64> = Newton::new();
    /// ```
    pub fn new() -> Self {
        Newton { gamma: float!(1.0) }
    }

    /// Set step size gamma
    ///
    /// Gamma must be in `(0, 1]` and defaults to `1`.
    ///
    /// # Example
    ///
    /// ```
    /// # use argmin::solver::newton::Newton;
    /// # use argmin::core::Error;
    /// # fn main() -> Result<(), Error> {
    /// let newton: Newton<f64> = Newton::new().with_gamma(0.4)?;
    /// # Ok(())
    /// # }
    /// ```
    pub fn with_gamma(mut self, gamma: F) -> Result<Self, Error> {
        if gamma <= float!(0.0) || gamma > float!(1.0) {
            return Err(argmin_error!(
                InvalidParameter,
                "Newton: gamma must be in  (0, 1]."
            ));
        }
        self.gamma = gamma;
        Ok(self)
    }
}

impl<F> Default for Newton<F>
where
    F: ArgminFloat,
{
    fn default() -> Newton<F> {
        Newton::new()
    }
}

impl<O, P, G, H, F> Solver<O, IterState<P, G, (), H, (), F>> for Newton<F>
where
    O: Gradient<Param = P, Gradient = G> + Hessian<Param = P, Hessian = H>,
    P: Clone + ArgminScaledSub<P, F, P>,
    H: ArgminInv<H> + ArgminDot<G, P>,
    F: ArgminFloat,
{
    fn name(&self) -> &str {
        "Newton method"
    }

    fn next_iter(
        &mut self,
        problem: &mut Problem<O>,
        mut state: IterState<P, G, (), H, (), F>,
    ) -> Result<(IterState<P, G, (), H, (), F>, Option<KV>), Error> {
        let param = state.take_param().ok_or_else(argmin_error_closure!(
            NotInitialized,
            concat!(
                "`Newton` requires an initial parameter vector. ",
                "Please provide an initial guess via `Executor`s `configure` method."
            )
        ))?;
        let grad = problem.gradient(&param)?;
        let hessian = problem.hessian(&param)?;
        let new_param = param.scaled_sub(&self.gamma, &hessian.inv()?.dot(&grad));
        Ok((state.param(new_param), None))
    }
}

#[cfg(test)]
mod tests {
    use super::*;
    use crate::core::ArgminError;
    #[cfg(feature = "_ndarrayl")]
    use crate::core::Executor;
    #[cfg(feature = "_ndarrayl")]
    use approx::assert_relative_eq;

    test_trait_impl!(newton_method, Newton<f64>);

    #[test]
    fn test_new() {
        let solver: Newton<f64> = Newton::new();
        assert_eq!(solver.gamma.to_ne_bytes(), 1.0f64.to_ne_bytes());
    }

    #[test]
    fn test_default() {
        let solver_new: Newton<f64> = Newton::new();
        let solver_def: Newton<f64> = Newton::default();
        assert_eq!(
            solver_new.gamma.to_ne_bytes(),
            solver_def.gamma.to_ne_bytes()
        );
    }

    #[test]
    fn test_with_gamma() {
        for new_gamma in [f64::EPSILON, 0.1, 0.5, 0.9, 1.0] {
            let solver: Newton<f64> = Newton::new().with_gamma(new_gamma).unwrap();
            assert_eq!(solver.gamma.to_ne_bytes(), new_gamma.to_ne_bytes());
        }

        for new_gamma in [1.0 + f64::EPSILON, 2.0, 0.0, -1.0] {
            let res = Newton::new().with_gamma(new_gamma);
            assert_error!(
                res,
                ArgminError,
                "Invalid parameter: \"Newton: gamma must be in  (0, 1].\""
            );
        }
    }

    #[cfg(feature = "_ndarrayl")]
    #[test]
    fn test_next_iter_param_not_initialized() {
        use crate::core::State;
        use ndarray::{Array, Array1, Array2};
        let mut newton: Newton<f64> = Newton::new();

        struct NewtonProblem {}

        impl Gradient for NewtonProblem {
            type Param = Array1<f64>;
            type Gradient = Array1<f64>;

            fn gradient(&self, _p: &Self::Param) -> Result<Self::Gradient, Error> {
                Ok(Array1::from_vec(vec![1.0, 2.0]))
            }
        }

        impl Hessian for NewtonProblem {
            type Param = Array1<f64>;
            type Hessian = Array2<f64>;

            fn hessian(&self, _p: &Self::Param) -> Result<Self::Hessian, Error> {
                Ok(Array::from_shape_vec((2, 2), vec![1.0f64, 0.0, 0.0, 1.0])?)
            }
        }

        let res = newton.next_iter(&mut Problem::new(NewtonProblem {}), IterState::new());
        assert_error!(
            res,
            ArgminError,
            concat!(
                "Not initialized: \"`Newton` requires an initial parameter vector. ",
                "Please provide an initial guess via `Executor`s `configure` method.\""
            )
        );
    }

    #[cfg(feature = "_ndarrayl")]
    #[test]
    fn test_solver() {
        use crate::core::State;
        use ndarray::{Array, Array1, Array2};
        struct Problem {}

        impl Gradient for Problem {
            type Param = Array1<f64>;
            type Gradient = Array1<f64>;

            fn gradient(&self, _p: &Self::Param) -> Result<Self::Gradient, Error> {
                Ok(Array1::from_vec(vec![1.0, 2.0]))
            }
        }

        impl Hessian for Problem {
            type Param = Array1<f64>;
            type Hessian = Array2<f64>;

            fn hessian(&self, _p: &Self::Param) -> Result<Self::Hessian, Error> {
                Ok(Array::from_shape_vec((2, 2), vec![1.0f64, 0.0, 0.0, 1.0])?)
            }
        }

        // Single iteration, starting from [0, 0], gamma = 1
        let problem = Problem {};
        let solver: Newton<f64> = Newton::new();
        let init_param = Array1::from_vec(vec![0.0, 0.0]);

        let param = Executor::new(problem, solver)
            .configure(|config| config.param(init_param).max_iters(1))
            .run()
            .unwrap()
            .state
            .get_best_param()
            .unwrap()
            .clone();
        assert_relative_eq!(param[0], -1.0, epsilon = f64::EPSILON);
        assert_relative_eq!(param[1], -2.0, epsilon = f64::EPSILON);

        // Two iterations, starting from [0, 0], gamma = 1
        let problem = Problem {};
        let solver: Newton<f64> = Newton::new();
        let init_param = Array1::from_vec(vec![0.0, 0.0]);

        let param = Executor::new(problem, solver)
            .configure(|config| config.param(init_param).max_iters(2))
            .run()
            .unwrap()
            .state
            .get_best_param()
            .unwrap()
            .clone();
        assert_relative_eq!(param[0], -2.0, epsilon = f64::EPSILON);
        assert_relative_eq!(param[1], -4.0, epsilon = f64::EPSILON);

        // Single iteration, starting from [0, 0], gamma = 0.5
        let problem = Problem {};
        let solver: Newton<f64> = Newton::new().with_gamma(0.5).unwrap();
        let init_param = Array1::from_vec(vec![0.0, 0.0]);

        let param = Executor::new(problem, solver)
            .configure(|config| config.param(init_param).max_iters(1))
            .run()
            .unwrap()
            .state
            .get_best_param()
            .unwrap()
            .clone();
        assert_relative_eq!(param[0], -0.5, epsilon = f64::EPSILON);
        assert_relative_eq!(param[1], -1.0, epsilon = f64::EPSILON);

        // Two iterations, starting from [0, 0], gamma = 0.5
        let problem = Problem {};
        let solver: Newton<f64> = Newton::new().with_gamma(0.5).unwrap();
        let init_param = Array1::from_vec(vec![0.0, 0.0]);

        let param = Executor::new(problem, solver)
            .configure(|config| config.param(init_param).max_iters(2))
            .run()
            .unwrap()
            .state
            .get_best_param()
            .unwrap()
            .clone();
        assert_relative_eq!(param[0], -1.0, epsilon = f64::EPSILON);
        assert_relative_eq!(param[1], -2.0, epsilon = f64::EPSILON);
    }
}