1 Matrix Similarity and Basic Setup

1.1 Unitary/Orthogonal Similarity

A square matrix \(A\in\mathbb{C}^{n\times n}\) is said to be *unitarily similar* to a matrix \(B\) if there exists a unitary matrix \(U\) such that \[ B = U^*AU, \] where \(U^*\) denotes the conjugate transpose. In real spaces, the analogous relation uses an orthogonal matrix \(Q\) with \(Q^TQ=I\): \[ B = Q^TAQ. \] Unitary (or orthogonal) similarity preserves many structural features, including the spectrum and a wide class of numerical measures tied to conditioning.

1.2 Upper Triangular Matrices and Invariant Subspaces

An upper triangular matrix \(T\) has the property that its leading principal subspaces are invariant under its action. More generally, when a similarity transform produces a triangular matrix, it reflects the existence of an invariant subspace chain for the original operator: the span of the first \(k\) basis vectors in the transformed coordinates corresponds to a subspace of the original space that is preserved by \(A\).

This viewpoint is especially useful for non-normal matrices, where eigenvectors may fail to form a basis and orthogonally computed invariant subspaces provide a stable alternative description.

1.3 Relationship to Eigenvalues

If \(B=U^*AU\), then \(A\) and \(B\) share the same characteristic polynomial, hence the same eigenvalues (with algebraic multiplicity). When a unitary similarity yields an upper triangular matrix \(T\), the eigenvalues become immediate: they are precisely the diagonal entries of \(T\).

2 Schur Decomposition

2.1 Complex Schur Form

2.1.1 Existence Theorem for Unitary Similarity

For every complex square matrix \(A\), there exist a unitary matrix \(U\) and an upper triangular matrix \(T\) such that \[ A = U T U^*. \] Equivalently, \(U^*AU=T\). The matrix \(T\) is called the *complex Schur form* of \(A\), and the unitary matrix \(U\) collects orthonormal basis vectors aligned with Schur’s invariant subspace structure.

2.1.2 Interpretation of the Triangular Structure

The triangular part \(T\) organizes the action of \(A\) in an orthonormal basis where invariant subspaces are built hierarchically. Even when eigenvectors are poorly behaved, the Schur basis remains numerically stable because orthonormality is enforced by construction. The off-diagonal entries of \(T\) contain information about how invariant subspaces are coupled, while the diagonal retains the eigenvalue data.

2.2 Real Schur Form

2.2.1 Quasi-Upper-Triangular Blocks for Real Matrices

For a real matrix \(A\in\mathbb{R}^{n\times n}\), one can seek an orthogonal similarity \(A = Q T Q^T\) where \(T\) is not necessarily triangular but has a *quasi-upper-triangular* structure. The key distinction is that real arithmetic cannot represent complex eigenvalues as single real diagonal entries; instead, they appear via \(2\times 2\) blocks.

Thus, \(T\) can be arranged to be block upper triangular with blocks of size \(1\times 1\) (real eigenvalues) and \(2\times 2\) (pairs of complex conjugate eigenvalues).

2.2.2 Complex Conjugate Eigenvalues and 2×2 Blocks

A real \(2\times 2\) block in Schur form corresponds to a complex conjugate eigenpair. If such a block is denoted \[ \begin{bmatrix} a & b\\ c & d \end{bmatrix}, \] then its eigenvalues are complex conjugates whenever the block has a negative discriminant. The trace and determinant of the block determine these eigenvalues, and the overall quasi-triangular layout still yields a clear spectral interpretation.

3 Properties of Schur Form

3.1 Diagonal Entries and Spectrum

For complex Schur form, \(T\) is upper triangular, so the eigenvalues of \(A\) are exactly the diagonal entries of \(T\). In the real Schur setting, eigenvalues are obtained by reading off \(1\times 1\) blocks directly and extracting the eigenvalues of each \(2\times 2\) block.

This property provides a practical bridge between abstract spectral theory and computable data.

3.2 Trace and Determinant Connections

Similarity transformations preserve both trace and determinant: \[ \operatorname{tr}(A)=\operatorname{tr}(T),\qquad \det(A)=\det(T). \] For complex Schur form, \(\operatorname{tr}(T)\) is the sum of diagonal entries. Also, since \(T\) is triangular, \(\det(T)\) equals the product of its diagonal entries. In the real case, the determinant factors over blocks, so the determinant equals the product of determinants of the \(1\times 1\) and \(2\times 2\) blocks.

3.3 Norm/Conditioning Considerations

Schur decomposition uses unitary or orthogonal transformations, which are norm-preserving in the Euclidean (2-)norm. This contributes to numerical stability: the computed Schur vectors can be arranged to be orthonormal up to rounding errors. While conditioning of eigenvalues can still vary widely for defective or nearly defective matrices, the Schur representation itself tends to be more reliable than directly computing eigenvectors.

In practice, the triangular/quasi-triangular structure also limits the growth of intermediate quantities in many algorithms.

3.4 Behavior Under Matrix Functions

Because Schur form expresses \(A\) through a similarity transform, many matrix functions can be computed by applying the function to \(T\) and transforming back: \[ f(A) = U f(T) U^* \quad (\text{complex case}). \] For polynomials, this follows directly from algebraic identities. For analytic functions, the same idea extends using the Jordan-like structure encoded by Schur form. Since \(T\) is triangular (or quasi-triangular), \(f(T)\) can be computed with structured methods, often exploiting block operations to handle complex conjugate pairs in the real setting.

4 Computation: Numerical Algorithms

4.1 Schur Iteration and the QR Algorithm

A primary computational route to Schur form is based on iterative QR-type methods applied to a matrix reduced to a structured form. These iterations generate a sequence of matrices whose limit is (up to unitary factors) the Schur form.

4.1.1 Hessenberg Reduction as a Preprocessing Step

Before QR iterations, one commonly transforms \(A\) to upper Hessenberg form: \[ H = Q^* A Q, \] where \(H\) is nearly upper triangular (all entries below the first subdiagonal are zero) and \(Q\) is unitary/orthogonal. This preprocessing step preserves eigenvalues and substantially reduces computational cost, because QR iterations on Hessenberg matrices are much cheaper than on general dense matrices.

The final Schur vectors can be assembled from the accumulated transformations used during reduction and iteration.

4.1.2 Convergence Criteria and Practical Stopping Rules

QR-based Schur iteration converges when off-diagonal elements corresponding to decoupled invariant subspaces become sufficiently small. Practical stopping rules are typically based on the magnitude of selected subdiagonal entries, which indicate whether the matrix is effectively block upper triangular.

For eigenvalue computations, the process is often *deflated*: once a portion of the matrix becomes nearly invariant, it can be separated and refined independently, improving efficiency and robustness.

4.2 Computational Complexity and Stability

For dense matrices, QR methods with Hessenberg reduction have a typical computational complexity on the order of \(O(n^3)\). Despite this cubic scaling, Schur-based approaches remain widely used because the unitary structure avoids unstable similarity scalings and because structured subproblems can be exploited.

Modern implementations also incorporate careful handling of shifts, balancing, and deflation to mitigate issues arising from finite precision arithmetic.

4.3 Extracting Eigenvalues and Eigenvectors from Schur Data

Once a Schur decomposition \(A = U T U^*\) is available, eigenvalues follow immediately from \(T\). To obtain eigenvectors, one typically solves triangular systems associated with \(T\). For each eigenvalue, the corresponding invariant subspace can be extracted by working with the appropriate rows/columns of \(T\) and then mapping back through \(U\).

For non-normal matrices, the concept of eigenvectors can be delicate (e.g., non-orthogonality or defectivity). The Schur approach still yields a stable basis for invariant subspaces, and eigenvector extraction is performed via structured solves rather than by solving an ill-conditioned eigenproblem directly from scratch.

5.1 Normal Matrices and Diagonalization

If \(A\) is normal (e.g., \(A^*A=AA^*\)), then Schur form can be chosen diagonal: the off-diagonal entries in \(T\) vanish. In that case, Schur decomposition reduces to the spectral theorem, and \(U\) becomes a unitary matrix of eigenvectors. Hence, Schur form generalizes diagonalization to broader classes of matrices while remaining valid in all cases.

5.2 Hermitian and Unitary Matrices

For Hermitian matrices, eigenvalues are real and the Schur form becomes diagonal with real entries. For unitary matrices, eigenvalues lie on the complex unit circle, and Schur form reflects that geometric constraint: the triangular structure is consistent with the modulus-preserving nature of the operator.

These classes often admit well-conditioned eigencomputations, though the Schur representation still provides a uniform framework.

5.3 Relationship to Jordan Canonical Form

Jordan canonical form describes how a matrix acts on generalized eigenspaces and organizes it into Jordan blocks. Schur form and Jordan form are related by similarity, but they differ in two important ways: Schur form uses orthonormal bases (unitary/orthogonal similarity), while Jordan form typically uses non-orthogonal generalized eigenvector bases. As a result, Schur form is often more numerically stable, even when the Jordan structure is highly sensitive.

In particular, when \(A\) is defective or nearly defective, Schur form remains well-defined and computable, whereas Jordan form can be extremely unstable under perturbations.

5.4 Condensed Notation: Block Schur Form

It is common to present Schur form in a block format where the matrix is partitioned into invariant subspace components. In the real case, the \(2\times 2\) blocks form a natural block structure. More generally, block Schur form highlights that iterations may separate the problem into smaller parts once convergence reveals nearly invariant blocks.

This representation is convenient for interpreting algorithms and for deriving matrix function evaluations that operate blockwise.

6 Applications

6.1 Eigenvalue Problems in Numerical Linear Algebra

Schur form is a foundational tool for eigenvalue analysis because it converts the eigenproblem into structured triangular computations. It supports reliable extraction of eigenvalues, computation of invariant subspaces, and methods for assessing spectral separation—properties that underlie many numerical routines.

Because the decomposition relies on orthonormal transformations, it integrates smoothly with stability-focused algorithm design.

6.2 Stability Analysis in Control Theory Contexts

In linear control, stability frequently depends on the locations of eigenvalues in the complex plane (for example, whether eigenvalues lie in a stable region). Schur form provides a robust computational mechanism for finding those eigenvalues without requiring explicit diagonalization. Moreover, real Schur forms with \(2\times 2\) blocks allow direct handling of complex conjugate pairs common in oscillatory dynamics.

The decomposition also aids in analyzing invariant subspaces associated with modes of the system.

6.3 Solving Matrix Equations via Schur Form

Many matrix equations can be reformulated using Schur decompositions, turning a difficult coupled problem into simpler triangular or block triangular subproblems. Examples include certain variants of Lyapunov and Sylvester-type equations, where transformation to (quasi-)triangular form enables efficient back-substitution schemes.

In these settings, the Schur basis helps preserve numerical stability, especially for problems sensitive to perturbations.

6.4 Model Reduction and Spectral Transform Methods

Model reduction aims to approximate a high-dimensional system with a smaller one while retaining key dynamical behavior. Spectral transform methods use eigen- and invariant-subspace information to identify dominant modes. Schur form contributes by providing a stable representation of those modes, even when the underlying matrix is non-normal.

By working with triangular or block triangular structures, reduced-order models can be constructed and analyzed using computations that avoid unstable eigenvector calculations.