Struct DampedBfgs 

pub struct DampedBfgs { /* private fields */ }

Implementations

impl DampedBfgs

pub fn new(n: usize) -> Self

pub fn has_prev(&self) -> bool

Have we recorded a previous (x, ∇L)? false until the first call to Self::update.

pub fn seed_scale(&mut self, gamma: Number)

Seed B = γI directly and mark the one-time sizing done, so the first Self::update applies its rank-2 correction on top of this scale instead of re-seeding from its own (s, y).

Used by the driver’s iteration-0 curvature probe: the internal sizing in Self::update cannot fire until a first (s, y) pair exists, i.e. not until iteration 1 — but iteration 0 already solves a QP against B, and with the identity seed that step overshoots by ~cond(∇²L) on an ill-conditioned problem. See the sizing comment in Self::update for why that is fatal.

gamma must be finite and strictly positive; anything else is ignored (leaving B = I) rather than corrupting the matrix.

pub fn reset_to_scale(&mut self)

Discard the accumulated rank-2 curvature and fall back to a scaled identity γI, where γ is the current mean diagonal (a scale the accumulated matrix has already vouched for). prev_x / prev_grad_lag are retained, so the next Self::update resumes accumulating from the reset base.

Used as a recovery step when the QP subproblem fails: a quasi-Newton matrix that has drifted ill-conditioned makes the step subproblem numerically unsolvable, and that is recoverable — far better than aborting an otherwise healthy solve. Off-diagonals are zeroed; the diagonal keeps the problem’s scale.

pub fn update(&mut self, x_new: &[Number], grad_lag_new: &[Number])

Apply the Powell-damped BFGS update from the previous (x_old, ∇L_old) to the supplied (x_new, ∇L_new). The first call just stores the pair; subsequent calls also modify B.

pub fn update_sy(&mut self, s: &[Number], y: &[Number])

Apply the Powell-damped rank-2 update from an explicit curvature pair (s, y).

Prefer this over Self::update when the caller can form y itself: the SQP driver must difference ∇L at a single, fixed multiplier (see the note in sqp_alg.rs), which the (x, ∇L) form of Self::update cannot express because it stores the previous ∇L as evaluated at the previous multiplier.

pub fn as_triplet(&self) -> Triplet

Produce the current B as a Triplet over the upper triangle (1-based), ready to feed into SqpQpData::build.