AutoMovie API
    Preparing search index...

    Function polishAutoMovieQuadraticProgram

    • Prepare a correction near an unchanged native primal and signed dual. When the independent original-row owner has already admitted feasibility and complementarity, retain that primal and cyclically minimize stationarity's squared Euclidean norm over each signed-dual coordinate. Its box comes from the same original equality, null-bound and complementarity predicate; native multipliers need no initial global clipping. Original full-row KKT alone admits the result, and Euclidean decrease guarantees no infinity-norm success or convergence within the caller's finite work budget.

      Wright (2015), sections 1.1-1.2, treats box constraints by coordinate indicators and exact scalar minimization: https://arxiv.org/pdf/1502.04759 A different native residual signature instead selects joint correction. Neither problem phase names nor particular original rows select the method. Each started dual sweep or working solve consumes that one unchanged budget.

      The refinement owner supplies exact coefficient groups and their interval intersections. Only identical sparse A rows share an equation: proportional, differently signed or differently scaled rows remain separate. Their active endpoint is the strongest original lower/upper owner, so nested copies cannot demand incompatible endpoint equations for the same affine value.

      The sum of original signed multipliers is the group multiplier because the coefficients are identical. Corrections preserve that equation and restore its signed multiplier to the original endpoint owner. A point intersection allows a free multiplier: its sign chooses the lower or upper owner, including point intersections formed by separate one-sided inequalities. Every original row stays in independent final admission; grouping is no deleted QP constraint.

      Joint correction uses a changing independent working basis. Its native start must be feasible on every original row within the unchanged caller threshold; initial inequalities must also be practically active at that threshold. This is an explicit finite-arithmetic premise, not an exact feasible-point theorem. Wrong-sign working multipliers leave the basis; every original row can block a primal step and supply its exact endpoint owner. Structural zero equalities compress exactly; other unresolved dependence refuses without row deletion. QR solves the unchanged PSD objective on the working null space. Flat descent needs a finite original-row blocker; rank uncertainty preserves native state. Each attempted working solve shares the same finite budget as dual sweeps. Gill and Wong (2014), sections 2-3, describe independent working bases, sign-aware changes and blocking constraints: https://www.ccom.ucsd.edu/~peg/papers/genqp.pdf No shifted/penalized feasibility phase, extra native solve or convergence guarantee is supplied here. A trial is adopted only after original full KKT. Before a flat step, a separate working-normal dual estimate is paired with its own affine origin and measured by that same original full-row owner. Nonzero projected scalars introduce no new stopping epsilon. A failed normal candidate neither replaces the live flat state nor supplies optimum sign departures; finite objective and native status remain with the final refiner.

      Parameters

      • input: IAutoMovieSolveQuadraticProgram.IInput
      • native: IAutoMovieSolveQuadraticProgram.IResult
      • options: IAutoMovieQuadraticRefinementOptions
      • groups: readonly IAutoMovieQuadraticRefinementInterval[]

        Refiner-owned exact coefficient groups with original endpoint owners and complete original ordinals; inputs have passed its intersection check.

      • readMerit: (primal: readonly number[], dual: readonly number[]) => number

        Original full-row KKT maximum, owned by the refinement caller.

      • readGradient: (primal: readonly number[], dual: readonly number[]) => number[]

        Original stationarity measurement, owned by that caller.

      • readComplementarity: (row: number, value: number, multiplier: number) => number

        Original signed/null/equality predicate, owned by that caller.

      • fixedPrimalEligible: boolean

        Original native feasibility and complementarity already pass.

      Returns IAutoMovieQuadraticPolishingResult

      When options.maximumRefinements is not a nonnegative safe integer, the budget both correction loops count against.