Positive dimensionless projected-gradient termination threshold in the solver's normalized chart.
Sparse symmetric Hessian rows; each entry stores a column index and its numeric coefficient.
OptionalinitialValues?: readonly number[]Optional feasible starting values in solver coordinates; omission uses the declared solver initialization.
Lower componentwise control bound in the normalized solver chart.
Explicit positive numerical iteration ceiling; exhaustion remains a reported termination, not convergence.
Assembled normalized right-hand-side values in the same unknown-slot order as the Hessian.
Upper componentwise control bound in the normalized solver chart.
The same normalized quadratic objective before iterative updates.
Actual numerical iterations attempted, including iterations without an accepted update.
Largest Hessian diagonal used in scaling and diagnostics.
Smallest positive Hessian diagonal; null means no positive entry was observed.
Current normalized quadratic objective value under the declared row weights.
Row-sum spectral upper estimate for the original Hessian, not a condition number.
Infinity norm of the normalized projected-gradient mapping used for termination.
Actual acceleration-momentum resets after objective increase.
Row-sum spectral upper estimate after diagonal congruence, not a condition number.
Owned scalar solution values in the operation's declared coefficient ordering and units.
Minimize sum_axis(0.5x' H x - rhs' x) on the supplied original control box. H is an exactly symmetric sparse positive-semidefinite Gram matrix supplied by the caller. This helper validates finite entries, distinct sparse addresses, reciprocal symmetry and nonnegative diagonals; it does not certify PSD or rank. An exactly zero diagonal requires an exactly zero row and three zero RHS entries. Such an unseen coordinate uses scale one and retains its feasible initial value. Dii=1/sqrt(Hii), x=Dy, Hs=DHD and bs=Db give the same real-arithmetic objective and box after scaling. FISTA projects in y, with objective-based momentum restarts. Numeric scaling is binary64; failed finite/feasible transformations throw rather than adding a regularizer or perturbing H. Active bounds retain original endpoints. Termination uses the original infinity norm of G_L(x)=L(x-proj(x-gradient/L)). An equivalent component formula avoids cancellation of two nearly equal x values. The best finite feasible original objective is returned, with its own gradient metric; exhausting the attempt budget is not convergence. Iterations counts all projected proposals, including rejected momentum attempts. Empty systems return empty. The attempt budget is a positive safe integer and gradient tolerance is finite and nonnegative. Reported gradients/objectives are binary64 evaluations, not exact KKT, uniqueness, eigenvalue or geometric accuracy certificates. Optional initialValues are copied original normalized coordinates, interleaved by three axes per control. They must be finite and feasible without clipping; omission starts at projected zero. initialObjective records their original Q, and every invocation starts with fresh momentum and retains that feasible best.