Skip to content
Tutorials Center
My learning

Topics

Lesson 47 of 62 ยท Eigenvalues and Decompositions

Jordan Normal Form and Schur Decomposition

Handle defective matrices with generalized eigenvectors and the Jordan form, then use the always-available Schur decomposition A = U T U*, with SymPy and SciPy examples.

Diagonalization is a lovely idea: in the right basis, a matrix is just a list of scaling factors. But it has a catch. It needs n independent eigenvectors, and some matrices simply do not have them. The matrix [[2, 1], [0, 2]] has the eigenvalue 2 twice but only a single eigenvector direction. It is called defective, and no change of basis can make it diagonal.

Two decompositions handle every square matrix. The Jordan normal form is the closest you can get to diagonal using a change of basis: a diagonal matrix with a few extra 1s just above it. It is a perfect theoretical tool for understanding powers, exponentials and differential equations. The Schur decomposition is the numerical workhorse: it triangularizes any matrix using a stable unitary change of basis. This lesson explains both and checks them with SymPy and SciPy. It builds on eigenvalues and eigenvectors, diagonalization and Cayley-Hamilton.

Why some matrices are defective

An eigenvalue has two multiplicities:

  • The algebraic multiplicity is how many times it appears as a root of the characteristic polynomial.
  • The geometric multiplicity is the dimension of its eigenspace, dim null(A - lambda*I).

Always, 1 <= geometric <= algebraic. A matrix is diagonalizable exactly when the two agree for every eigenvalue. When geometric is smaller than algebraic, we are short of eigenvectors.

For B = [[2, 1], [0, 2]]: the root 2 has algebraic multiplicity 2. But B - 2*I = [[0, 1], [0, 0]] has a one-dimensional null space (spanned by (1, 0)), so the geometric multiplicity is 1. The matrix is defective.

Generalized eigenvectors and Jordan blocks

To fill the gap, we use generalized eigenvectors. A vector v is a generalized eigenvector of rank m for lambda if (A - lambda*I)**m @ v = 0 but (A - lambda*I)**(m-1) @ v != 0. Ordinary eigenvectors have rank 1. They link into chains:

(A - lambda*I) @ v_2 = v_1, where v_1 is a true eigenvector.

In the basis v_1, v_2, the matrix acts as A @ v_1 = lambda*v_1 and A @ v_2 = lambda*v_2 + v_1, which is exactly the Jordan block

J = [[lambda, 1], [0, lambda]].

Jordan normal form theorem. Every square matrix over the complex numbers is similar to a block diagonal matrix J = P**(-1) * A * P made of Jordan blocks J_k(lambda): lambda on the diagonal, 1 on the superdiagonal. The number of blocks for an eigenvalue equals its geometric multiplicity, and their sizes add up to its algebraic multiplicity. The form is unique up to the order of blocks.

flowchart LR
    A["square matrix A"] --> B{"enough independent<br/>eigenvectors?"}
    B -- "yes" --> C["diagonalizable:<br/>J is diagonal"]
    B -- "no" --> D["add generalized<br/>eigenvectors (chains)"]
    D --> E["Jordan form:<br/>diagonal plus some 1s"]

Jordan form is diagonalization with chains of generalized eigenvectors added where eigenvectors run out.

Python Jordan form with SymPy

Press Run (or Ctrl + Enter) to see the result here.

Program input ยท what you'd type while the program runs

Type the answers before you press Run, one per line. Each time the program reads a line (with input()), it gets the next line from this box.

Notice that P contains the eigenvector in one column and the generalized eigenvector in the next, in exactly the order that creates the 1 above the diagonal.

What Jordan form tells you

Powers. For a Jordan block J = lambda*I + N, where N is the shift matrix with ones on the superdiagonal and N**m = 0 for a block of size m, the binomial theorem gives a finite sum. For a 2-by-2 block:

J**k = [[lambda**k, k*lambda**(k-1)], [0, lambda**k]].

A new factor k appears: the entries of A**k can grow like k * lambda**(k-1), polynomial growth on top of the exponential. In a stable system (|lambda| < 1) it still decays, but a defective matrix can show a large temporary bump before it does.

Exponentials. e**(J*t) = e**(lambda*t) * [[1, t], [0, 1]]: a factor t times the exponential. This is exactly the origin of terms like t*e**(lambda*t) in solutions of ODEs with repeated roots, such as the critically damped oscillator in second-order ODEs.

Diagonalizability. A is diagonalizable if and only if every Jordan block is 1-by-1.

Python Powers and exponentials of a Jordan block

Press Run (or Ctrl + Enter) to see the result here.

Program input ยท what you'd type while the program runs

Type the answers before you press Run, one per line. Each time the program reads a line (with input()), it gets the next line from this box.

The last lines show why Jordan form is a theoretical tool rather than a numerical one. A perturbation of size eps moves the eigenvalues by about sqrt(eps), far larger. Floating-point rounding on a defective matrix therefore changes the computed Jordan structure completely, and no stable algorithm for it exists. In practice, rounding makes a "nearly defective" matrix look diagonalizable with a hopelessly ill-conditioned eigenvector matrix.

The Schur decomposition

The Schur decomposition gives up diagonal and settles for triangular, and in return it exists for every square matrix and can be computed stably.

Schur's theorem. For any square matrix A (complex), there is a unitary matrix U (U* U = I, where U* is the conjugate transpose) and an upper triangular matrix T with

A = U @ T @ U*.

The eigenvalues of A are the diagonal entries of T, and the first column of U is an eigenvector. The change of basis is by an orthonormal basis, so it does not amplify errors, unlike the possibly wild P in Jordan form. For a real matrix with complex eigenvalues, the real Schur form uses a real orthogonal Q and a quasi-triangular T with 2-by-2 blocks on the diagonal for each complex-conjugate pair.

Special cases reveal structure:

  • A is normal (A A* = A* A, including symmetric, Hermitian, orthogonal and unitary matrices) if and only if T is diagonal. That recovers the spectral theorem.
  • If A is not normal, the size of the part of T above the diagonal measures how far it is from normal, and how much transient growth to expect.

The QR algorithm that computes eigenvalues in every numerical library is, at heart, an algorithm for finding the Schur form.

Python Schur decomposition in SciPy

Press Run (or Ctrl + Enter) to see the result here.

Program input ยท what you'd type while the program runs

Type the answers before you press Run, one per line. Each time the program reads a line (with input()), it gets the next line from this box.

Which one to use

Jordan form Schur form
Exists for every square matrix yes yes
Change of basis any invertible P unitary U, perfectly conditioned
Result block diagonal, almost diagonal upper triangular
Numerically stable no, discontinuous in A yes
Best for theory, ODEs, powers, proofs computing: eigenvalues, matrix functions, solving Sylvester equations

SciPy's funm and sqrtm use Schur-based methods, and expm uses a Pade approximation that also works on defective matrices. When you need an answer on a computer, use Schur (or the library function built on it). When you need to understand why a matrix behaves as it does, think in Jordan blocks.

Common mistakes

  • Concluding a matrix is diagonalizable because its eigenvalues are real. Repeated eigenvalues with too few eigenvectors still give a defective matrix.
  • Computing Jordan forms of floating-point matrices and trusting the structure. Use exact arithmetic (SymPy) or avoid it.
  • Writing A = U T U**(-1) with a non-unitary U and calling it Schur: Schur requires a unitary (orthogonal) U.
  • Forgetting that real matrices may need complex U for the plain Schur form; use the real Schur form for real blocks.

Check your understanding

Quick check ยท 5 questions

of 5 answered You scored / 5 ยท perfect!

  1. 1. What does it mean for a matrix to be defective?

    A defective matrix has an eigenvalue whose geometric multiplicity is smaller than its algebraic multiplicity, so there are not enough eigenvectors.
  2. 2. What is the Jordan form of [[2, 1], [0, 2]]?

    The matrix has eigenvalue 2 with only one eigenvector, so it is already a single 2 x 2 Jordan block.
  3. 3. How does the 2 x 2 Jordan block J with eigenvalue lambda behave under powers?

    Writing J = lambda I + N with N^2 = 0, the binomial theorem gives lambda^k I + k lambda^(k-1) N.
  4. 4. What is special about the Schur decomposition A = U T U*?

    Schur's theorem applies to every square matrix; unitary changes of basis do not amplify rounding errors, unlike Jordan's possibly ill-conditioned basis.
  5. 5. For which matrices is the Schur factor T diagonal?

    A is normal (A A* = A* A) exactly when T is diagonal; this includes symmetric, Hermitian and orthogonal matrices.

Summary

  • A defective matrix lacks a full set of eigenvectors: its geometric multiplicity is below its algebraic multiplicity.
  • Generalized eigenvectors form chains and give the Jordan normal form: A = P J P**(-1) with Jordan blocks lambda*I + N.
  • Jordan blocks explain k*lambda**(k-1) factors in powers and t*e**(lambda*t) in exponentials; A is diagonalizable iff all blocks are 1-by-1.
  • The Jordan form is unstable numerically: perturbations of size eps move repeated eigenvalues by about sqrt(eps).
  • The Schur decomposition A = U T U* (unitary U, upper triangular T) exists for every square matrix, is stable to compute and is diagonal exactly for normal matrices.

Next: speed up the solution of huge systems in Iterative Solvers: Jacobi, Gauss-Seidel and Conjugate Gradient.