Divisor each factored row was normalized by, paired with order: the
Euclidean norm of its weights, or the raw signed weight for a free
equation fixing one variable to zero.
Groups of working-list positions of free equations fixing the same variable to zero; the first of each group is the one factored.
variables − rank orthonormal columns spanning their null space.
Working-list positions of the factored rows, in pivot order.
Minimum-norm point satisfying every factored working equation.
rank orthonormal columns spanning the working normals.
Number of independent working equations factored, the size of R.
Pivot magnitude every factored column had to exceed:
max(variables, rows) · ε · largest normalized column norm. A rank
reading, not a geometric tolerance.
Why the working set could not be factored reliably, or null.
Upper-triangular R, rank by rank, of the pivoted QR.
Number of variables in the program; the length of every basis column.
An orthogonal factorization of one working set of affine equations
aᵢ · x = boundᵢ, ascreateAutoMovieQuadraticConstraintBasisreturns it.Each working row is normalized and the normals are factored by column-pivoted Householder QR,
[a₁/d₁ … aₖ/dₖ] P = Q R. Rows are identified by their position in the working list the caller passed, not by their row number in the whole program. Only exact structure removes a row: a free equation fixing one variable to zero repeats an earlier one on the same variable, or an all-zero row with a zero endpoint. Any other numerical dependence setsrefusalinstead of dropping a row. All arrays are new and owned by the caller; on a refusal they may be partial and must not be used.