KEVOS
ArticlesServicesCase studiesAboutContact
ArticlesServicesCase studiesAboutContact
← ArticlesMatrix Representation and Cost ModelEngineering · Engineering MathematicsLesson 729/884← PrevNext →
ArticlePublished 7 Aug 20262 min readBy Kevin Joginmatrix representationsparse matrixcost modelexact linear algebra
On this page

Ask about this page

KEVOS AIMatrix Representation and Cost Model

KEVOS knowledge first · trusted web sources when needed

Linear Algebra Algorithms

Matrix Representation and Cost Model

Dense and sparse matrix representations, the cost model for exact linear algebra, and why coefficient growth rather than operation count usually decides performance.

Engineering / MathematicsLinear Algebra Algorithms2 min readKV-MATH-0528

Exact linear algebra differs from numerical linear algebra in one decisive respect: the entries grow. Operation counts that are accurate over the reals badly understate the cost over the integers or the rationals.

Representations

Matrix representations in exact computation
RepresentationStorageSuits
Dense arrayO(m n)Small matrices, most number field work
Sparse by rowsProportional to non-zerosRelation matrices from sieving
StructuredImplicitMatrices with exploitable form, such as Hessenberg

Key point

The relation matrices arising in class group computation and the quadratic sieve are very sparse and very large. Dense storage is not merely slow there; it does not fit in memory.

The cost model

Cost is counted as bit operations, so an operation on entries of b bits costs at least b. Since Gaussian elimination causes entries to grow, the average b during a computation can far exceed the input and output sizes.

Naive cost = (operation count) x (cost per operation)The second factor is not constant; it grows during elimination.

Pitfall

Quoting O(n^3) for exact Gaussian elimination is misleading. Over the integers, entry growth can be exponential in n without fraction-free techniques, so the true bit complexity is far worse than the operation count suggests.

Three responses to growth

Fraction-free elimination

Arrange the elimination so every intermediate entry is a determinant of a submatrix, bounded by Hadamard's inequality.

Modular methods

Compute modulo several primes and reconstruct via the Chinese remainder theorem. Entries never grow.

Lattice reduction

Apply LLL to keep a basis short. This is the standard fix for HNF explosion.

Where the matrices come from

Number field arithmetic
Small dense matrices over the integers or rationals, of size the field degree. Growth is the problem, not size.
Polynomial factorisation
Matrices over a finite field of size the polynomial degree — see Berlekamp. No growth; entries stay in the field.
Relation collection
Enormous sparse matrices over the field of two elements or over the integers. Size is the problem, not growth.

Note

These three regimes call for genuinely different algorithms. Applying a dense integer method to a sieving relation matrix, or a sparse method to a degree-eight number field matrix, wastes most of the available performance.

Source. Henri Cohen, A Course in Computational Algebraic Number Theory, Springer GTM 138 — 2.1-2.2.1. Structural reference unverified: the source file was not available during authoring; chapter and section numbers are taken from the published edition and have not been checked against a physical copy.

Related pages

  • Algorithm Notation and Complexity Conventions
  • Gaussian Elimination and Linear Systems

Continue learning

Finite Field Arithmetic in PracticeArticle · Engineering MathematicsNEXT LESSON →Gaussian Elimination and Linear SystemsArticle · Engineering MathematicsRoot Finding over Finite FieldsArticle · Engineering MathematicsGaussian Elimination over Finite FieldsArticle · Engineering Mathematics
KEVOS · Engineering, manufacturing and project improvement
ArticlesServicesCase studiesAboutContact
© 2026 KEVOS®