A set of linear or nonlinear transformations applied to a numerical problem—commonly a linear system or eigenproblem—to change its algebraic properties (spectral distribution, conditioning) so that iterative solvers converge faster or more stably.

Principle

Introduce a preconditioner M (or transform variables) such that the transformed operator M^{-1}A (or M^{-1}AM^{-T}, or equivalent) has a more favorable spectrum or conditioning number, reducing iteration counts or improving numerical stability of iterative methods.

Demonstration

Solving Ax = b with Conjugate Gradient using an incomplete Cholesky preconditioner: constructing an approximate factorization M ≈ A reduces the condition number of M^{-1}A and typically cuts the number of CG iterations dramatically compared with no preconditioner.

Misapplication

Using an expensive or poorly matched preconditioner whose setup cost and storage outweigh iteration savings, or applying a non-symmetric preconditioner to a method that requires symmetry, causing breakdown or incorrect convergence.

Consequence

Appropriate preconditioning reduces wall-clock time and memory usage by lowering iteration counts and improving numerical robustness; it may introduce setup cost, approximation error, or additional implementation complexity that must be justified by net gain.

Reversal

No preconditioning (identity preconditioner) leaves the original spectrum and conditioning, which for ill-conditioned problems yields slow or stalled convergence; conversely, over-aggressive preconditioning can introduce bias or instability.

Boundary

Primarily applies to iterative solvers for linear systems, eigenproblems, and optimization subproblems; it does not replace direct factorization where exact solves are required, and its benefit depends on problem structure and solver compatibility.

Semantic Tension

Tension exists with scaling and deflation techniques: scaling changes variable magnitudes to improve conditioning, deflation removes problematic invariant subspaces, while preconditioning encompasses and sometimes combines these ideas but is distinct in implementation and goals.

Synthesis

Preconditioning systematically alters the algebraic form of a numerical problem to produce a transformed operator with improved spectral properties, thereby making iterative solvers faster and more reliable when the preconditioner is well matched to the problem.