---- Case 1: fp > fx ----------------------------------------------------- Reference model: phi(a) = 0.5 * a^2 - a, minimum at a* = 1. phi(0) = 0, p

Strictly convex, well-conditioned quadratic. Starting at a point where the full Newton step exceeds the default TR radius, the solver grows the radiu

---- Case 1: convex quadratic `phi(a) = a^2 - 2a` ------------------------ Starting at `a = 0` with `phi'(0) = -2`, the exact minimizer is `a = 1`. HZ

---- A.1: QuadraticEqualityPenalty is 0 at feasibility -------------------- c(x) = 0 + 1*x. At x = 0, c = 0, so P_eq(c)(x=0) = 0.5 * 0^2 = 0.

Overload that accepts a pre-configured stopping `Progress` (mirrors the other solvers' form) alongside the trust-region config. The two arguments are

HZ, stages U0-U3. Given a third point `c` with value/slope already in `history[ic]`, shrink the bracket `[ia, ib]` to two indices of a bracket around

Pareto comparison for the "best iterate so far" tracker. The order is: 1. An iterate with max_violation <= feas_tol is strictly better than any itera

Exposes the stored interface for mode-downgrading converting construction into a weaker-mode `FunctionExpr`. Returns a deep-copied interface pointer

TODO(patwie): cpplint prefers pointers here, but this would make the code unreadable. As these are all changing values a configuration structure would

Main driver. `*x` carries the initial iterate on entry and the accepted iterate on exit; `*f`, `*g` are filled with the accepted function value and g

Returns true if the two vectors are nearly equal (using the infinity norm).

Main demo: Solve the constrained problem with multiple constraints

Main demo: solve the constrained problem

----------------------------------------------------------------- Example Usage

Create an initial simplex given the starting point x using adaptive perturbations.

Simplified operator(): only function value and gradient.

Call operator.