Scaling

Default value: equilibrate
rosepy parameter name: scaling

See the linked article for information on How to change and view the solver parameters.

The solver supports a variety of matrix scaling techniques to improve numerical stability and solver performance. These techniques rescale constraint coefficients and bounds to reduce the condition number of the matrix and mitigate the effects of rounding errors.

Use the scaling setting to specify the method. The available options are:

Equilibrate

rosepy parameter value: equilibrate

This is the default option. Scales each row and column by the largest absolute element found in that row or column. This quick heuristic ensures no single element dominates the matrix, offering a simple form of balancing.

Arithmetic mean

rosepy parameter value: arithmetic

Applies arithmetic mean scaling, which rescales each row and column based on the arithmetic mean of the absolute values of the coefficients. This method is simple and effective for moderately scaled models.

Geometric mean

rosepy parameter value: geometric

Applies geometric mean scaling, which rescales each row and column based on the geometric mean of the absolute value of the coefficients. This method often provides better balance than arithmetic scaling as it is less sensitive to extremes.

De Buchet (p=1)

rosepy parameter value: de_buchet

Implements de Buchet scaling (p=1). A variant of norm-based scaling that optimizes for the 1-norm, aiming to balance row and column magnitudes while preserving sparsity.

De Buchet (p=2)

rosepy parameter value: de_buchet2

Uses de Buchet scaling (p=2), which instead minimizes the 2-norm (Euclidean norm). It is more aggressive than de_buchet and may result in better conditioning at the cost of denser scaling matrices.

Euclidean (L2) norm

rosepy parameter value: l2

Applies L2-norm scaling, where rows and columns are scaled based on their Euclidean (L2) norms. This method emphasizes the overall energy of the coefficients and is useful for certain numerically sensitive problems.

Hamming-Curtis-Reid

rosepy parameter value: hcr

Applies the Hamming-Curtis-Reid (HCR) scaling method, which is designed to maintain sparsity and improve numerical robustness in structured problems such as network flows.

None

rosepy parameter value: none

No scaling is applied. The model is solved using the original coefficients. Suitable when the model is already well-conditioned or for debugging purposes.