Struct ScalingPreview 

pub struct ScalingPreview {
    pub max_gradient: Number,
    pub max_grad_f: Number,
    pub obj_scale: Number,
    pub c: RowScaleBlock,
    pub d: RowScaleBlock,
    pub quad_rows: Vec<QuadRowScale>,
    pub n_quad_rows: usize,
    pub n_quad_unscaled: usize,
    pub n_quad_zero_jac: usize,
    pub max_quad_mismatch: Number,
}

What nlp_scaling_method=gradient-based will do to this model at this x0 — the solver’s own arithmetic (gradient_obj_scale / gradient_row_scale), not a copy of it.

Why a preflight reports this

Gradient-based scaling is a point sample: it reads ∇f and the Jacobian once, at x0, and never looks again. That is a good estimator of a row’s magnitude when the row’s derivative at x0 is representative of its derivative elsewhere, and no estimator at all when it is not. The extreme case is a row whose Jacobian vanishes at x0 — a ½xᵀQx ≤ b written about the origin, started from x = 0. The sample reads zero, the row is left at factor 1.0, and however badly Q and b disagree in magnitude the scaler has no way to know it. That is how AMPL emits qcqp1000-2c — every variable free, no initial guess, and k rows of pure ½xᵀQᵢx ≤ bᵢ.

So the two halves of this report are complementary: the block below says what the sample decided, and Self::quad_rows says what the sample could not see. See dev-notes/quadratic-structure-exploitation.md §8 (gh #703).

Fields

max_gradient: Number

The nlp_scaling_max_gradient cutoff this preview assumed.

max_grad_f: Number

‖∇f(x0)‖_∞.

obj_scale: Number

The objective factor df the scaler will pick.

c: RowScaleBlock

Equality rows (g_l == g_u).

d: RowScaleBlock

Everything else — the inequality and range rows.

quad_rows: Vec<QuadRowScale>

Rows recognized as quadratic, worst mismatch first. .nl models only; empty for a builtin or a model with no quadratic row.

n_quad_rows: usize

Total quadratic rows found (quad_rows is capped at --max-list).

n_quad_unscaled: usize

Of those, how many the sample leaves at factor 1.0.

n_quad_zero_jac: usize

Of those, how many have an identically-zero Jacobian row at x0.

max_quad_mismatch: Number

Largest rhs / curvature over the quadratic rows, or 0 when there are none.

Trait Implementations

impl Clone for ScalingPreview

fn clone(&self) -> ScalingPreview

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for ScalingPreview

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl Default for ScalingPreview

fn default() -> ScalingPreview

Returns the “default value” for a type.