A measure of how far from the main diagonal the significant nonzero entries of a discretized linear operator (matrix) are concentrated, commonly quantified as the maximum distance |i-j| for which the matrix A(i,j) is nonzero (or the half-bandwidth in symmetric conventions).

Principle

The graph and ordering of degrees of freedom determine bandwidth: local stencils or small-element connectivity produce small bandwidth, and bandwidth controls storage layout, fill-in behaviour during factorization, and the complexity of certain direct solvers.

Demonstration

A one-dimensional second-order finite-difference discretization with nearest-neighbor coupling yields a tridiagonal matrix with bandwidth 1 (half-bandwidth of 1), while a naive global ordering of a multidimensional mesh can increase bandwidth and thus the cost of LU factorization.

Misapplication

Using bandwidth as the sole predictor of solver cost without accounting for sparsity pattern irregularity, fill-in from factorization, or the effect of reordering and block structure; or optimizing for bandwidth at the expense of parallel communication patterns.

Consequence

Low bandwidth enables compact banded storage, efficient banded direct solvers, and reduced fill-in; recognizing and reducing bandwidth through reordering (e.g., reverse Cuthill–McKee) can dramatically lower factorization cost for certain problems.

Reversal

Full width or global coupling where significant entries exist far from the diagonal, yielding large bandwidth and necessitating dense-storage techniques or factorization with heavy fill-in.

Boundary

Refers specifically to geometric distance from diagonal in matrix index space and excludes other notions called 'bandwidth' (signal bandwidth, communication bandwidth) unless the context is linear algebraic discretizations.

Semantic Tension

Tension exists between bandwidth and more nuanced sparsity metrics (profile, envelope, and graph separator sizes): bandwidth is a simple scalar summary but may hide critical pattern features that determine algorithmic cost.

Synthesis

Bandwidth is the diagonal-distance measure of nonzero concentration in a discrete linear operator; it guides storage and solver choices and is reducible by reordering, but must be considered alongside full sparsity structure for accurate performance prediction.