pounce_algorithm/sqp/line_search.rs
1//! l1-merit backtracking line search for SQP. The classic
2//! Han-Powell scheme:
3//!
4//! ```text
5//! φ(x; ν) = f(x) + ν · violation(x)
6//! violation(x) = Σ_i max(bl_i − c_i, 0) + max(c_i − bu_i, 0)
7//! + Σ_j max(xl_j − x_j, 0) + max(x_j − xu_j, 0)
8//! ```
9//!
10//! Step `p` is a descent direction of `φ(·; ν)` whenever
11//! `ν ≥ ‖λ_g‖_∞` (Nocedal-Wright §18.4). We adapt `ν` at every
12//! iteration as `ν ← max(ν, ‖λ_g_qp‖_∞ + buffer)` so the QP-
13//! derived multipliers are always dominated.
14//!
15//! Backtracking is plain Armijo:
16//!
17//! ```text
18//! φ(x + αp) ≤ φ(x) + η · α · D_p φ(x; ν)
19//! ```
20//!
21//! with predicted derivative
22//!
23//! ```text
24//! D_p φ ≈ ∇f(x)ᵀ p − ν · violation(x)
25//! ```
26//!
27//! (correct under the standard assumption that the QP step
28//! reduces the linearized constraint violation to zero).
29//!
30//! Phase 5b commit 5 deliverable. The filter alternative
31//! (Fletcher-Leyffer 2002) is opt-in via
32//! `SqpGlobalization::Filter` and lands as a follow-up.
33
34use crate::sqp::options::SqpOptions;
35use crate::sqp::problem::SqpProblemSpec;
36use pounce_common::types::{Number, NLP_LOWER_BOUND_INF, NLP_UPPER_BOUND_INF};
37
38/// l1 constraint + bound violation `‖max(bl − c, 0) + max(c − bu, 0)‖_1`
39/// (plus the same for variable bounds). Infinite bounds are treated
40/// as never violated.
41pub fn l1_violation(
42 x: &[Number],
43 c_vals: &[Number],
44 bl: &[Number],
45 bu: &[Number],
46 xl: &[Number],
47 xu: &[Number],
48) -> Number {
49 let mut v = 0.0_f64;
50 for (i, &ci) in c_vals.iter().enumerate() {
51 if bl[i] > NLP_LOWER_BOUND_INF {
52 v += (bl[i] - ci).max(0.0);
53 }
54 if bu[i] < NLP_UPPER_BOUND_INF {
55 v += (ci - bu[i]).max(0.0);
56 }
57 }
58 for (j, &xj) in x.iter().enumerate() {
59 if xl[j] > NLP_LOWER_BOUND_INF {
60 v += (xl[j] - xj).max(0.0);
61 }
62 if xu[j] < NLP_UPPER_BOUND_INF {
63 v += (xj - xu[j]).max(0.0);
64 }
65 }
66 v
67}
68
69pub struct LineSearchResult {
70 pub alpha: Number,
71 pub nu: Number,
72 pub x_new: Vec<Number>,
73 pub f_new: Number,
74 pub c_new: Vec<Number>,
75 pub success: bool,
76}
77
78/// Adapt `ν` against the QP multiplier magnitude and run Armijo
79/// backtracking on the l1 merit function. Returns the accepted
80/// step length, the updated ν, and the resulting trial state
81/// (`x_new`, `f_new`, `c_new`) so the caller doesn't have to
82/// re-evaluate.
83#[allow(clippy::too_many_arguments)]
84pub fn l1_merit_line_search<N: SqpProblemSpec>(
85 nlp: &mut N,
86 x: &[Number],
87 p: &[Number],
88 qp_lambda_g: &[Number],
89 grad_f: &[Number],
90 f_curr: Number,
91 c_curr: &[Number],
92 bl: &[Number],
93 bu: &[Number],
94 xl: &[Number],
95 xu: &[Number],
96 current_nu: Number,
97 opts: &SqpOptions,
98) -> LineSearchResult {
99 // ν adaptation (Han-Powell): dominate the QP multipliers by
100 // an additive safety margin, then clamp at l1_penalty_max so
101 // a pathological |λ_qp| spike doesn't blow the merit into a
102 // regime where Armijo always fails. Nocedal-Wright §18.4
103 // recommends `ν ≥ ‖λ‖_∞`; we use `+ l1_penalty_safety` to
104 // give the test a comfortable inequality.
105 let lambda_inf = qp_lambda_g.iter().map(|l| l.abs()).fold(0.0_f64, f64::max);
106 let nu = current_nu
107 .max(lambda_inf + opts.l1_penalty_safety)
108 .min(opts.l1_penalty_max);
109
110 let viol_curr = l1_violation(x, c_curr, bl, bu, xl, xu);
111 let phi_curr = f_curr + nu * viol_curr;
112
113 let grad_p: Number = grad_f.iter().zip(p.iter()).map(|(g, pi)| g * pi).sum();
114 // Predicted decrease: linear-objective contribution minus
115 // the violation we expect the QP to eliminate.
116 let predicted = grad_p - nu * viol_curr;
117 let eta = 1e-4_f64;
118
119 let mut alpha = 1.0_f64;
120 let mut x_trial = vec![0.0; x.len()];
121 let mut last_f = f_curr;
122 let mut last_c = c_curr.to_vec();
123 while alpha > opts.bt_min_alpha {
124 for (xt, (&xi, &pi)) in x_trial.iter_mut().zip(x.iter().zip(p.iter())) {
125 *xt = xi + alpha * pi;
126 }
127 let f_trial = nlp.eval_f(&x_trial);
128 let c_trial = nlp.eval_c(&x_trial);
129 let viol_trial = l1_violation(&x_trial, &c_trial, bl, bu, xl, xu);
130 let phi_trial = f_trial + nu * viol_trial;
131 last_f = f_trial;
132 last_c.clone_from(&c_trial);
133
134 let target = phi_curr + eta * alpha * predicted;
135 // Standard Armijo sufficient-decrease (Nocedal-Wright
136 // §3.1). The earlier `|| phi_trial < phi_curr` fallback
137 // (PR #50 review C3) accepted *any* descent and
138 // effectively bypassed the inequality on nonconvex
139 // problems where `predicted ≥ 0` makes the Armijo target
140 // monotone-non-decreasing. We now gate the fallback on
141 // `predicted >= 0` only — i.e. fall back to "any merit
142 // decrease wins" only when the predicted derivative is
143 // not a descent direction, which is the case the original
144 // fallback was intended to cover (cf. Wächter-Biegler 2006
145 // §3.3 backtracking rule).
146 let armijo_ok = if predicted < 0.0 {
147 phi_trial <= target
148 } else {
149 phi_trial < phi_curr
150 };
151 #[cfg(test)]
152 if opts.print_level >= 2 {
153 tracing::debug!(target: "pounce::sqp",
154 " ls trial α={alpha:.3e} phi_t={phi_trial:.4e} \
155 phi_c={phi_curr:.4e} target={target:.4e} pred={predicted:.3e} \
156 grad_p={grad_p:.3e} viol_c={viol_curr:.3e} viol_t={viol_trial:.3e} \
157 ok={armijo_ok}"
158 );
159 }
160 if armijo_ok {
161 return LineSearchResult {
162 alpha,
163 nu,
164 x_new: x_trial,
165 f_new: f_trial,
166 c_new: c_trial,
167 success: true,
168 };
169 }
170 alpha *= opts.bt_reduction;
171 }
172
173 LineSearchResult {
174 alpha,
175 nu,
176 x_new: x_trial,
177 f_new: last_f,
178 c_new: last_c,
179 success: false,
180 }
181}