Can Schur decomposition be applied to a rectangular matrix?

The standard decomposition described here is for square matrices; singular value decomposition handles rectangular matrices.

What does unitary mean?

Its conjugate transpose is its inverse, so it preserves the complex Euclidean inner product.

Why use real Schur form?

It keeps real arithmetic for real input while representing complex eigenvalue pairs through two-by-two blocks.

AI Glossary / Mathematics and optimization / Advanced

Schur decomposition

Definition

Schur decomposition expresses a square complex matrix as A=QTQ*, where Q is unitary and T is upper triangular. The diagonal entries of T are A's eigenvalues. A real matrix also admits a real Schur form using an orthogonal Q and a block upper-triangular T.

What it means

A matrix describes a linear transformation in a chosen coordinate system. Schur decomposition changes to an orthonormal coordinate system in which that transformation has a triangular structure. The change preserves lengths and angles while making certain calculations easier to organize.

The important advantage is existence: every square complex matrix has a Schur decomposition, including matrices that cannot be diagonalized. Upper-triangular structure permits nonzero entries above the diagonal, so it asks less than a diagonal eigenvalue decomposition.

The star denotes conjugate transpose. For real matrices and real factors, it becomes ordinary transpose. The real Schur form may contain two-by-two diagonal blocks representing complex-conjugate eigenvalue pairs. In that case, not every eigenvalue can be read from a single diagonal entry.

How it works

A conceptual construction begins with a unit-length eigenvector, extends it to an orthonormal basis and expresses the matrix in that basis. The first column then has the eigenvalue at its top and zeros below. Applying the same idea to the remaining block produces a triangular form over the complex numbers.

Practical numerical algorithms use structured transformations and iterations rather than repeatedly carrying out that conceptual proof literally. The computed factors should approximately satisfy Q*Q=I and A=QTQ*, within numerical accuracy.

The columns of Q are called Schur vectors. In general, they are not all individual eigenvectors. Instead, the leading columns describe nested invariant subspaces when the Schur form is ordered appropriately. Off-diagonal entries in T retain interactions that a diagonal representation would eliminate.

For a normal matrix, a unitary diagonalization is available and the complex Schur form can be diagonal. General non-normal matrices need the triangular structure. The distinction is useful when studying eigenvalues, matrix functions and the behavior of linear dynamical models.

Three practical examples

A numerical analyst uses a Schur form as a starting point for computing a function of a square matrix.
A control-model researcher groups selected eigenvalues to examine an associated invariant subspace.
A software engineer checks a decomposition routine on a matrix that has too few independent eigenvectors for diagonalization.

Think of it this way

Think of rotating your coordinate grid so that the transformation's dependencies can be followed in a triangular order. The transformation itself has not changed; only its representation has.

The analogy does not mean the underlying system becomes independent in every coordinate. Nonzero entries above the diagonal can still couple coordinates, which is precisely why triangularization is more general than diagonalization.

Worked example

Checking a two-by-two decomposition

Let A have rows (1,0) and (−2,3). Choose Q=(1/√2) times the matrix with rows (1,−1) and (1,1), and let T have rows (1,2) and (0,3).

The columns of Q have length one and inner product zero, so QᵀQ=I. Multiplying QT gives (1/√2) times the matrix with rows (1,−1) and (1,5). Multiplying again by Qᵀ yields A. Thus A=QTQᵀ is a valid Schur decomposition, with eigenvalues 1 and 3 on T's diagonal.

The first Schur vector q₁=(1,1)/√2 is an eigenvector: Aq₁=q₁. The second q₂=(−1,1)/√2 satisfies Aq₂=2q₁+3q₂, so it is not an eigenvector. The off-diagonal entry two records that coupling.

For a separate example, the matrix with rows (1,2) and (0,1) is already triangular. It has a repeated eigenvalue one but only a one-dimensional eigenspace. Its Schur decomposition exists with Q=I even though a diagonalization does not.

Illustrative case study

A fictional simulation-software team assumes that every square matrix can be represented by a full basis of eigenvectors. An internal test with a defective matrix breaks that assumption and produces an unreliable calculation.

The numerical specialist changes the design to use an appropriate Schur-based method for the required matrix computation. The team validates factor reconstruction and orthogonality, then tests the final quantity rather than stopping at the decomposition output.

The revised documentation distinguishes eigenvectors from Schur vectors. It also explains that a stable representation does not remove every sensitivity in the original problem; poorly separated spectral features can still require careful interpretation.

Common misconceptions and mistakes

Schur decomposition is always diagonal. General matrices require upper-triangular or real block-triangular structure.
Every column of Q is an eigenvector. The worked example shows otherwise.
Real input guarantees only real eigenvalues. Real Schur blocks can encode complex-conjugate pairs.
Failure of diagonalization prevents Schur decomposition. Existence is one of its key advantages.
The decomposition is unique. Eigenvalue ordering and basis choices can produce different valid factors.

Business use

This is a specialist numerical tool with indirect business relevance in simulation, optimization and model analysis. It supports reliable calculations involving square matrices when a simplistic eigenvector treatment would be unsuitable.

When reviewing technical software, ask which mathematical quantity the decomposition supports and how accuracy is assessed. Reconstruction error alone is not a complete assessment of every downstream inference. Record whether real or complex output is used, and avoid interpreting different valid factors from two implementations as evidence that one must be wrong.

Check your understanding

A matrix has only one independent eigenvector but is two-by-two. Can it still have a Schur decomposition?

Answer: Yes. The triangular matrix with rows (1,2) and (0,1) in the example is already a valid Schur form, despite not being diagonalizable.

Related terms

Learn next

Read eigenvalue decomposition and invariant subspace. Verify Aq₂=2q₁+3q₂ in the example, then explain why reading T's diagonal does not mean its off-diagonal entry can be discarded.

Eigenvalue

Sources

SciPy, scipy.linalg.schur: https://docs.scipy.org/doc/scipy/reference/generated/scipy.linalg.schur.html. Official real and complex forms and numerical interface.

Aalto Dictionary of Machine Learning, Schur decomposition: https://aaltodictionaryofml.github.io/ADictML_Main.pdf. Existence and conceptual construction. The explicit matrix products and fictional software scenario are original illustrations.

Previous: Schema migration Next: Schwa