basin 0.8.0

//! Limited-memory BFGS / L-BFGS-B state.
//!
//! Carries the current iterate plus the limited-memory history
//! `(s_k, y_k)` capped at `m_capacity` pairs, and the compact-form
//! Gram matrices `SᵀY` and `SᵀS`. Mirrors the Fortran v3.0 storage
//! (`ws`, `wy`, `sy`, `ss`, `theta` in `references/lbfgsb-v3.0/`)
//! but keeps the history chronologically (oldest at index 0) rather
//! than in a ring buffer with a `head` pointer. The numerical result
//! is identical because cauchy / subsm only require "column j of W
//! in oldest-to-newest order" and never depend on the ring-buffer
//! index modulo `m_capacity`.

use crate::core::math::Dot;
use crate::core::problem::EvalCounts;
use crate::core::state::{CountsMirror, GradientState, State};

/// Solver state for L-BFGS-B (and the unbounded L-BFGS solver, when
/// it lands).
///
/// `theta` initializes to `1.0`; after the first accepted update
/// it becomes `(y · y) / (s · y)`, matching the Fortran convention
/// at `mainlb`'s `matupd` call site.
pub struct LbfgsState<V> {
    pub(crate) param: V,
    pub(crate) cost: Option<f64>,
    pub(crate) gradient: Option<V>,

    /// History capacity. Fortran's `m`; recommended `[3, 20]`.
    pub(crate) m_capacity: usize,
    /// `s_k = x_{k+1} − x_k`, chronological, oldest first.
    pub(crate) ws: Vec<V>,
    /// `y_k = g_{k+1} − g_k`, same length and order as `ws`.
    pub(crate) wy: Vec<V>,
    /// `SᵀY` as row-major `m_capacity²` storage; only the leading
    /// `col × col` block (`col = ws.len()`) is live.
    pub(crate) sy: Vec<f64>,
    /// `SᵀS`, same row-major layout as `sy`.
    pub(crate) ss: Vec<f64>,
    /// Compact-form scaling. `1.0` until first accepted update,
    /// thereafter `(y · y) / (s · y)`.
    pub(crate) theta: f64,

    pub(crate) iter: u64,
    pub(crate) cost_evals: u64,
    pub(crate) gradient_evals: u64,

    /// Working buffers and persistent solver-side scalars for the
    /// L-BFGS-B iteration (Fortran's `mainlb` scratch arrays plus the
    /// pieces of `isave`/`dsave` that survive across iterations).
    /// Initialized lazily by `LBFGSB::init`; absent when [`LbfgsState`]
    /// is used by other solvers (e.g. a future unbounded L-BFGS).
    pub(crate) work: Option<LbfgsbWork>,
}

/// Mutable working storage threaded through the L-BFGS-B iteration.
///
/// Allocates once in [`crate::solver::LBFGSB::init`] and is reused
/// across every [`crate::core::solver::Solver::next_iter`] call.
/// Mirrors the layout Fortran `mainlb` carves out of the user-
/// supplied scratch arrays (`ws`, `wy`, `sy`, `ss`, `wt`, `wn`, `snd`,
/// `z`, `r`, `d`, `t`, `xp`, `wa`, `index`, `iwhere`, `indx2`) plus
/// the iteration-persistent scalars that live in `isave` / `dsave`
/// between coroutine returns.
///
/// Stored on [`LbfgsState`] rather than the solver struct so that
/// [`crate::core::solver::Solver`] implementations stay
/// configuration-only (mirroring [`crate::solver::BFGS`]).
pub(crate) struct LbfgsbWork {
    // ---- Compact-form matrices ----
    /// `2m × 2m` row-major; stores the `L·E·Lᵀ` factor of the
    /// indefinite middle matrix `K`. Output of `formk`, consumed by
    /// `subsm`.
    pub(crate) wn: Vec<f64>,
    /// `2m × 2m` row-major; the lower-triangular `N` Gram cache that
    /// `formk` maintains incrementally across outer iterations.
    pub(crate) wn1: Vec<f64>,
    /// `m × m` row-major; Cholesky factor of `T = θ SᵀS + LD⁻¹Lᵀ`,
    /// produced by `formt`, consumed by `bmv` inside `cauchy`.
    pub(crate) wt: Vec<f64>,

    // ---- n-sized working vectors ----
    /// Cauchy point / subspace Newton point (Fortran `z`).
    pub(crate) z: Vec<f64>,
    /// Reduced gradient at the Cauchy point (Fortran `r`).
    pub(crate) r: Vec<f64>,
    /// Search direction `d = z − x` (Fortran `d`).
    pub(crate) d: Vec<f64>,
    /// Cauchy breakpoint buffer / line-search previous iterate
    /// (Fortran `t`).
    pub(crate) t_buf: Vec<f64>,
    /// Subspace projected-Newton safeguard slot (Fortran `xp`).
    pub(crate) xp: Vec<f64>,

    // ---- 2m-sized cauchy / subsm scratch ----
    /// `Wᵀ d` accumulator inside `cauchy` (Fortran `wa(1..2m)`).
    pub(crate) wa_p: Vec<f64>,
    /// `Wᵀ (xcp − x)` accumulator (Fortran `wa(2m+1..4m)`); fed to
    /// `subsm` via `cmprlb`.
    pub(crate) wa_c: Vec<f64>,
    /// Breakpoint `W` row inside `cauchy` (Fortran `wa(4m+1..6m)`).
    pub(crate) wa_wbp: Vec<f64>,
    /// Middle-matrix solve scratch (Fortran `wa(6m+1..8m)`); reused
    /// by `subsm` as `wv`.
    pub(crate) wa_v: Vec<f64>,

    // ---- Integer working arrays ----
    /// Cauchy-point variable classification (`FREE_NOT_MOVED`, etc.).
    pub(crate) iwhere: Vec<i8>,
    /// Free + active partition at the GCP, free first.
    pub(crate) index: Vec<usize>,
    /// Entering + leaving variables since the previous GCP. Doubles
    /// as `iorder` inside `cauchy` (the breakpoint heap).
    pub(crate) indx2: Vec<usize>,

    // ---- Iteration-persistent scalars / flags ----
    /// True iff at least one variable has a finite bound.
    pub(crate) cnstnd: bool,
    /// True iff every variable is two-sided (both bounds finite).
    pub(crate) boxed: bool,
    /// True iff the limited-memory history was updated in the
    /// previous outer iteration.
    pub(crate) updatd: bool,
    /// Total number of accepted BFGS updates (Fortran `iupdat`).
    pub(crate) iupdat: u32,
    /// `‖d‖` from the last line search; used on subsequent calls to
    /// set the initial step.
    pub(crate) dnorm: f64,
    /// `−gᵀd` from the previous line search (Fortran `gdold`); needed
    /// for the curvature-skip threshold.
    pub(crate) gdold: f64,
    /// Number of free variables at the GCP (`nfree`). Carried across
    /// iterations because `freev` uses the *previous* `index` to
    /// detect leaving variables.
    pub(crate) nfree: usize,
}

impl LbfgsbWork {
    /// Pre-allocate every buffer to its required size given the
    /// problem dimension `n` and history capacity `m`.
    pub(crate) fn new(n: usize, m: usize) -> Self {
        let two_m = 2 * m;
        Self {
            wn: vec![0.0; two_m * two_m],
            wn1: vec![0.0; two_m * two_m],
            wt: vec![0.0; m * m],
            z: vec![0.0; n],
            r: vec![0.0; n],
            d: vec![0.0; n],
            t_buf: vec![0.0; n],
            xp: vec![0.0; n],
            wa_p: vec![0.0; two_m],
            wa_c: vec![0.0; two_m],
            wa_wbp: vec![0.0; two_m],
            wa_v: vec![0.0; two_m],
            iwhere: vec![0; n],
            index: (0..n).collect(),
            indx2: vec![0; n],
            cnstnd: false,
            boxed: true,
            updatd: false,
            iupdat: 0,
            dnorm: 0.0,
            gdold: 0.0,
            nfree: n,
        }
    }

    /// Reset the limited-memory state for an iteration restart
    /// (matches Fortran `col = 0; head = 1; theta = 1; iupdat = 0;
    /// updatd = false`).
    pub(crate) fn reset_history(&mut self) {
        self.iupdat = 0;
        self.updatd = false;
    }
}

impl<V> LbfgsState<V> {
    /// Build state at the given starting point with capacity for
    /// `m_capacity` history pairs. Use `m_capacity = 10` as a
    /// reasonable default; Fortran recommends `[3, 20]`.
    ///
    /// # Panics
    ///
    /// Panics if `m_capacity == 0`.
    pub fn new(param: V, m_capacity: usize) -> Self {
        assert!(m_capacity >= 1, "m_capacity must be ≥ 1");
        let mm = m_capacity * m_capacity;
        Self {
            param,
            cost: None,
            gradient: None,
            m_capacity,
            ws: Vec::with_capacity(m_capacity),
            wy: Vec::with_capacity(m_capacity),
            sy: vec![0.0; mm],
            ss: vec![0.0; mm],
            theta: 1.0,
            iter: 0,
            cost_evals: 0,
            gradient_evals: 0,
            work: None,
        }
    }

    /// Current history length (`col` in Fortran). In `[0, m_capacity]`.
    /// Used by tests; the solver inlines `state.ws.len()` directly.
    #[allow(dead_code)]
    pub(crate) fn col(&self) -> usize {
        self.ws.len()
    }

    /// Append a `(s, y)` pair to the history and update `sy`, `ss`,
    /// `theta`. When the history is at capacity, the oldest pair is
    /// dropped (left shift on `ws`, `wy`, and the leading block of
    /// `sy`, `ss`).
    ///
    /// Returns `false` if `s·y ≤ 0` or any product is non-finite —
    /// the curvature condition is the caller's responsibility, this
    /// is just a final safeguard. The state is left unchanged in
    /// that case.
    pub(crate) fn append_pair(&mut self, s: V, y: V) -> bool
    where
        V: Dot,
    {
        let sy_dot = s.dot(&y);
        let yy_dot = y.dot(&y);
        if !(sy_dot > 0.0 && sy_dot.is_finite() && yy_dot.is_finite()) {
            return false;
        }

        let m = self.m_capacity;

        // Drop oldest when at capacity.
        if self.ws.len() == m {
            self.ws.remove(0);
            self.wy.remove(0);
            // Shift the leading `(m-1) × (m-1)` block of sy and ss
            // up-and-left by one row+column. Use a forward sweep —
            // each (i, j) only reads from (i+1, j+1) which we haven't
            // written yet.
            for i in 0..m - 1 {
                for j in 0..m - 1 {
                    self.sy[i * m + j] = self.sy[(i + 1) * m + (j + 1)];
                    self.ss[i * m + j] = self.ss[(i + 1) * m + (j + 1)];
                }
            }
            // Zero the now-vacated last row and last column.
            for i in 0..m {
                self.sy[i * m + (m - 1)] = 0.0;
                self.sy[(m - 1) * m + i] = 0.0;
                self.ss[i * m + (m - 1)] = 0.0;
                self.ss[(m - 1) * m + i] = 0.0;
            }
        }

        // Fill the new last row/column of sy = SᵀY and ss = SᵀS.
        // `new_idx` is where (s, y) will sit after we push, so the
        // existing history occupies indices `0..new_idx`.
        let new_idx = self.ws.len();
        for i in 0..new_idx {
            let s_i_y_new = self.ws[i].dot(&y);
            let s_new_y_i = s.dot(&self.wy[i]);
            let s_i_s_new = self.ws[i].dot(&s);
            // sy[i, new] = sᵢ · y_new, sy[new, i] = s_new · yᵢ.
            self.sy[i * m + new_idx] = s_i_y_new;
            self.sy[new_idx * m + i] = s_new_y_i;
            // ss is symmetric.
            self.ss[i * m + new_idx] = s_i_s_new;
            self.ss[new_idx * m + i] = s_i_s_new;
        }
        self.sy[new_idx * m + new_idx] = sy_dot;
        self.ss[new_idx * m + new_idx] = s.dot(&s);

        self.theta = yy_dot / sy_dot;

        // Push last so the dot products above saw the pre-extension
        // history.
        self.ws.push(s);
        self.wy.push(y);
        true
    }
}

impl<V> State for LbfgsState<V> {
    type Param = V;
    type Float = f64;

    fn iter(&self) -> u64 {
        self.iter
    }
    fn increment_iter(&mut self) {
        self.iter += 1;
    }
    fn cost_evals(&self) -> u64 {
        self.cost_evals
    }
    fn param(&self) -> &V {
        &self.param
    }
    /// # Panics
    ///
    /// Panics if read before [`Solver::init`](crate::core::solver::Solver::init)
    /// has populated the cached cost — see [`BasicState::cost`] for
    /// the full safety argument; same contract.
    ///
    /// [`BasicState::cost`]: crate::core::state::BasicState::cost
    fn cost(&self) -> f64 {
        self.cost
            .expect("LbfgsState::cost read before Solver::init populated it")
    }
}

impl<V> GradientState for LbfgsState<V> {
    fn gradient(&self) -> Option<&V> {
        self.gradient.as_ref()
    }
    fn gradient_evals(&self) -> u64 {
        self.gradient_evals
    }
}

impl<V> CountsMirror for LbfgsState<V> {
    fn mirror(&mut self, delta: &EvalCounts) {
        self.cost_evals = delta.cost_evals + delta.residual_evals;
        self.gradient_evals = delta.gradient_evals + delta.jacobian_evals + delta.hessian_evals;
    }
}

#[cfg(test)]
// Explicit `i * m + j` indexing (including `0 * m + 0`) mirrors the
// Fortran source's 2-D layout for `sy` / `ss` — load-bearing for
// readability when cross-checking against `lbfgsb.f`.
#[allow(clippy::identity_op, clippy::erasing_op)]
mod tests {
    use super::*;

    #[test]
    fn new_state_is_empty() {
        let s = LbfgsState::<Vec<f64>>::new(vec![0.0; 4], 5);
        assert_eq!(s.col(), 0);
        assert_eq!(s.m_capacity, 5);
        assert_eq!(s.theta, 1.0);
        assert!(s.cost.is_none());
        assert!(s.gradient.is_none());
        assert_eq!(s.ws.len(), 0);
        assert_eq!(s.wy.len(), 0);
        assert_eq!(s.sy.len(), 25);
        assert_eq!(s.ss.len(), 25);
    }

    #[test]
    fn first_append_sets_theta_and_diagonal() {
        let mut state = LbfgsState::<Vec<f64>>::new(vec![0.0, 0.0], 3);
        let s = vec![1.0, 2.0]; // ‖s‖² = 5
        let y = vec![3.0, 4.0]; // ‖y‖² = 25, s·y = 1·3 + 2·4 = 11
        let ok = state.append_pair(s, y);
        assert!(ok);
        assert_eq!(state.col(), 1);
        assert_eq!(state.theta, 25.0 / 11.0);
        assert_eq!(state.sy[0], 11.0); // sy[0,0] = s·y
        assert_eq!(state.ss[0], 5.0); // ss[0,0] = s·s
    }

    #[test]
    fn second_append_fills_off_diagonal_gram_blocks() {
        let mut state = LbfgsState::<Vec<f64>>::new(vec![0.0; 2], 3);
        let s1 = vec![1.0, 0.0];
        let y1 = vec![2.0, 0.0];
        let s2 = vec![0.0, 3.0];
        let y2 = vec![0.0, 4.0];
        state.append_pair(s1, y1);
        state.append_pair(s2, y2);

        // m = 3, so indexing is i * 3 + j.
        // sy[0,0] = s1·y1 = 2, sy[0,1] = s1·y2 = 0,
        // sy[1,0] = s2·y1 = 0, sy[1,1] = s2·y2 = 12.
        assert_eq!(state.sy[0 * 3 + 0], 2.0);
        assert_eq!(state.sy[0 * 3 + 1], 0.0);
        assert_eq!(state.sy[1 * 3 + 0], 0.0);
        assert_eq!(state.sy[1 * 3 + 1], 12.0);

        // ss is symmetric: ss[0,0]=1, ss[0,1]=ss[1,0]=0, ss[1,1]=9.
        assert_eq!(state.ss[0 * 3 + 0], 1.0);
        assert_eq!(state.ss[0 * 3 + 1], 0.0);
        assert_eq!(state.ss[1 * 3 + 0], 0.0);
        assert_eq!(state.ss[1 * 3 + 1], 9.0);

        // theta from the most recent update: ‖y2‖²/(s2·y2) = 16/12.
        assert_eq!(state.theta, 16.0 / 12.0);
    }

    #[test]
    fn appending_beyond_capacity_drops_oldest() {
        let mut state = LbfgsState::<Vec<f64>>::new(vec![0.0], 2);
        // Three appends with distinct identifiable pairs; m_capacity=2
        // means the third should evict the first.
        let s1 = vec![1.0];
        let y1 = vec![2.0];
        let s2 = vec![3.0];
        let y2 = vec![4.0];
        let s3 = vec![5.0];
        let y3 = vec![6.0];
        state.append_pair(s1.clone(), y1.clone());
        state.append_pair(s2.clone(), y2.clone());
        state.append_pair(s3.clone(), y3.clone());

        assert_eq!(state.col(), 2);
        // After eviction, history is [s2, s3] / [y2, y3].
        assert_eq!(state.ws[0], s2);
        assert_eq!(state.ws[1], s3);
        assert_eq!(state.wy[0], y2);
        assert_eq!(state.wy[1], y3);

        // Gram blocks should reflect the post-eviction history.
        // sy[0,0] = s2·y2 = 12, sy[0,1] = s2·y3 = 18,
        // sy[1,0] = s3·y2 = 20, sy[1,1] = s3·y3 = 30.
        let m = 2;
        assert_eq!(state.sy[0 * m + 0], 12.0);
        assert_eq!(state.sy[0 * m + 1], 18.0);
        assert_eq!(state.sy[1 * m + 0], 20.0);
        assert_eq!(state.sy[1 * m + 1], 30.0);
    }

    #[test]
    fn curvature_failure_leaves_state_untouched() {
        let mut state = LbfgsState::<Vec<f64>>::new(vec![0.0, 0.0], 3);
        // s · y = -1 (negative curvature) — must be rejected.
        let s = vec![1.0, 0.0];
        let y = vec![-1.0, 0.0];
        let ok = state.append_pair(s, y);
        assert!(!ok);
        assert_eq!(state.col(), 0);
        assert_eq!(state.theta, 1.0);
    }

    #[test]
    fn state_implements_state_and_gradient_state_traits() {
        // Sanity check that the trait impls are reachable through the
        // generic State / GradientState bounds.
        let s: LbfgsState<Vec<f64>> = LbfgsState::new(vec![1.0, 2.0], 5);
        // Param round-trip via the State trait.
        let p: &Vec<f64> = State::param(&s);
        assert_eq!(p, &vec![1.0, 2.0]);
        // GradientState exposes the None gradient pre-init.
        assert!(GradientState::gradient(&s).is_none());
        assert_eq!(GradientState::gradient_evals(&s), 0);
        assert_eq!(State::iter(&s), 0);
        assert_eq!(State::cost_evals(&s), 0);
    }
}