Linear Algebra: Jordan Normal Form

When a matrix cannot be diagonalized, it signals a fundamental complication in its structure. In engineering contexts—from control theory to structural dynamics—this is not a rare exception but a frequent challenge. The Jordan normal form (or Jordan canonical form) provides the next best thing: a nearly diagonal canonical form that reveals the underlying algebraic and geometric structure of any linear operator, even when eigenvectors are insufficient.

Generalized Eigenvectors and the Core Idea

For a diagonalizable matrix $A$, you can find a basis of eigenvectors. This corresponds to the geometric multiplicity of each eigenvalue equaling its algebraic multiplicity. When this fails, the matrix is defective. The key insight is to extend the concept of an eigenvector. A generalized eigenvector of rank $k$ for eigenvalue $\lambda$ is a nonzero vector $\mathbf{v}$ such that $(A-\lambda I{)}^k\mathbf{v}=0$, but $(A-\lambda I{)}^{k-1}\mathbf{v}\ne 0$. An eigenvector is a generalized eigenvector of rank 1.

This definition directly addresses the deficiency. While you may not have enough true eigenvectors to form a basis, you are guaranteed a complete basis of generalized eigenvectors for the complex numbers. This collection of vectors provides the columns for the transformation matrix that brings $A$ into its Jordan form.

Jordan Blocks, Chains, and Nilpotent Matrices

The building blocks of the Jordan form are Jordan blocks. A Jordan block of size $m$ for eigenvalue $\lambda$ is an $m\times m$ matrix of the form: $$J_m(\lambda )=\left[\begin{array}{cccc}\lambda & 1& & \\ & \lambda & \ddots & \\ & & \ddots & 1\\ & & & \lambda \end{array}\right].$$ It has $\lambda$ on the diagonal, 1s on the superdiagonal, and zeros elsewhere.

Each Jordan block corresponds to a Jordan chain. A Jordan chain is an ordered set of vectors $\{{\mathbf{v}}_1,{\mathbf{v}}_2,...,{\mathbf{v}}_m\}$ such that:

• ${\mathbf{v}}_1$ is an eigenvector: $(A-\lambda I){\mathbf{v}}_1=0$.
• ${\mathbf{v}}_2$ is a generalized eigenvector of rank 2: $(A-\lambda I){\mathbf{v}}_2={\mathbf{v}}_1$.
• In general, $(A-\lambda I){\mathbf{v}}_k={\mathbf{v}}_{k-1}$.

The matrix $(A-\lambda I)$, when restricted to the subspace spanned by a Jordan chain for $\lambda$, acts as a nilpotent matrix. A matrix $N$ is nilpotent if $N^k=0$ for some positive integer $k$. On a chain of length $m$, $(A-\lambda I{)}^m$ sends every vector to zero, demonstrating nilpotency. The Jordan block $J_m(\lambda )$ is simply $\lambda I+N$, where $N$ is the nilpotent matrix with 1s on the superdiagonal.

Computing the Jordan Normal Form: A Step-by-Step Process

Finding the Jordan form $J$ and transformation matrix $P$ such that $P^{-1}AP=J$ is a systematic, though sometimes tedious, procedure. Let's outline it with a conceptual example.

Step 1: Find Eigenvalues. Compute the characteristic polynomial and find all eigenvalues ${\lambda }_i$, with their algebraic multiplicities.

Step 2: For Each Eigenvalue, Analyze Generalized Eigenspaces. For a fixed eigenvalue $\lambda$, consider the sequence of nullspaces: $N_k=\text{Null}((A-\lambda I{)}^k)$. These form a nested chain, $N_1\subseteq N_2\subseteq ...\subseteq N_d$, where $d$ is the index of nilpotency—the size of the largest Jordan block for $\lambda$. The dimension of $N_1$ is the geometric multiplicity (number of eigenvectors/ number of Jordan blocks). The differences in dimensions between these spaces tell you the sizes of the Jordan blocks.

Step 3: Construct Jordan Chains. Start from the top. Find vectors in $N_d$ that are not in $N_{d-1}$. These are the highest-rank generalized eigenvectors. For each such vector ${\mathbf{v}}_d$, build its chain backwards: ${\mathbf{v}}_{d-1}=(A-\lambda I){\mathbf{v}}_d$, ${\mathbf{v}}_{d-2}=(A-\lambda I){\mathbf{v}}_{d-1}$, ... until you reach an eigenvector ${\mathbf{v}}_1$. This forms one Jordan chain corresponding to a $d\times d$ block. Repeat until you have exhausted the basis for the generalized eigenspace for $\lambda$.

Example: Suppose for $\lambda =2$, $dim(N_1)=1$ and $dim(N_2)=2$. This means there is only 1 eigenvector (so 1 Jordan block total) and the jump from $N_1$ to $N_2$ tells us the block size is 2. You would find a vector ${\mathbf{v}}_2\in N_2$ not in $N_1$, and set ${\mathbf{v}}_1=(A-2I){\mathbf{v}}_2$.

Step 4: Assemble. Collect all Jordan chains for all eigenvalues. The vectors form the columns of $P$, and the corresponding Jordan blocks, arranged in order, form $J$.

The Matrix Exponential via Jordan Form

This is a prime engineering application. Solving systems of linear differential equations $\dot{\mathbf{x}}=A\mathbf{x}$ requires computing the matrix exponential $e^{At}$. For a diagonal matrix $D$, this is trivial: $e^{Dt}$ has $e^{d_{ii}t}$ on its diagonal. For a non-diagonalizable $A$, the Jordan form is the savior.

First, if $A=PJ{P}^{-1}$, then $e^{At}=P{e}^{Jt}{P}^{-1}$. The exponential of a Jordan block is easily computed. For a $3\times 3$ block $J_3(\lambda )$: $$J_3(\lambda )=\lambda I+N,\text{where }N=\left[\begin{array}{ccc}0& 1& 0\\ 0& 0& 1\\ 0& 0& 0\end{array}\right].$$ Since $\lambda I$ and $N$ commute, $e^{J_3(\lambda )t}=e^{\lambda It}{e}^{Nt}=e^{\lambda t}{e}^{Nt}$. The nilpotent matrix $N$ has $N^3=0$, so its exponential is a finite sum: $$e^{Nt}=I+Nt+\frac{N^2{t}^2}{2!}=\left[\begin{array}{ccc}1& t& t^2/2\\ 0& 1& t\\ 0& 0& 1\end{array}\right].$$ Thus, $$e^{J_3(\lambda )t}=e^{\lambda t}\left[\begin{array}{ccc}1& t& t^2/2\\ 0& 1& t\\ 0& 0& 1\end{array}\right].$$ The solution to the differential system inherits these polynomial-in-$t$ terms multiplied by $e^{\lambda t}$, which correspond directly to the algebraic structure of the generalized eigenvectors.

Understanding the Algebraic Structure

The Jordan form provides a complete invariant for similarity of matrices over algebraically closed fields (like the complex numbers). It tells you everything about the algebraic structure of the linear operator:

• The eigenvalues reveal the scaling factors.
• The number of Jordan blocks for an eigenvalue equals its geometric multiplicity (the dimension of the eigenspace).
• The sizes of the Jordan blocks describe the nilpotent action of $(A-\lambda I)$ on the generalized eigenspace. The largest block size is the index of nilpotency of $(A-\lambda I)$.
• The form makes the minimal polynomial explicit: it is ${\prod }_i(x-{\lambda }_i{)}^{s_i}$, where $s_i$ is the size of the largest Jordan block for ${\lambda }_i$.

This structure is why the Jordan form is "canonical." While the transformation matrix $P$ is not unique, the Jordan matrix $J$ itself is unique up to the order of the blocks. It is the simplest representation that fully captures the operator's action when diagonalization is impossible.

Common Pitfalls

1. Confusing Algebraic and Geometric Multiplicity: The algebraic multiplicity of $\lambda$ (its multiplicity as a root of the characteristic polynomial) is the total number of times $\lambda$ appears on the diagonal of $J$. The geometric multiplicity (dimension of the eigenspace) is the number of Jordan blocks for $\lambda$. Mistaking these will lead to an incorrect block count.

1. Incorrect Chain Construction: A frequent error is to start building a Jordan chain from a true eigenvector by applying $(A-\lambda I)$ repeatedly. This just gives zero. You must start from a highest-rank generalized eigenvector (in the nullspace of $(A-\lambda I{)}^k$ but not of $(A-\lambda I{)}^{k-1}$) and work backwards to the eigenvector.

1. Assuming Real Jordan Form for Real Matrices: Over the real numbers, if eigenvalues are complex, the standard Jordan form requires complex numbers. For a real matrix, you can use a Real Jordan Form that uses $2\times 2$ blocks corresponding to complex conjugate eigenvalue pairs, avoiding complex entries in $P$. Applying the complex Jordan form algorithm to a real matrix with complex eigenvalues will yield a complex $P$, which may not be desirable for real-number applications.

1. Overlooking Nilpotency in Computations: When computing functions like $e^{At}$, failing to recognize that $(A-\lambda I)$ is nilpotent on the subspace of a Jordan chain leads to attempting infinite series. Recognizing nilpotency truncates the series to a simple polynomial, making computation feasible.

Summary

• The Jordan normal form is the canonical nearly-diagonal matrix similar to any square matrix, especially when it is not diagonalizable.
• It is constructed using a basis of generalized eigenvectors, organized into Jordan chains that correspond to Jordan blocks in the matrix.
• Each Jordan block for eigenvalue $\lambda$ combines a scaling ($\lambda I$) and a nilpotent action (the superdiagonal 1s).
• Computing the Jordan form involves analyzing the nested nullspaces of $(A-\lambda I{)}^k$ to determine the number and sizes of Jordan blocks for each eigenvalue.
• Its prime engineering utility is simplifying the computation of functions of matrices, most notably the matrix exponential $e^{At}$ for solving systems of differential equations, where polynomial terms arise from the nilpotent structure.
• The Jordan form completely reveals the operator's structure, including its eigenvalues, geometric multiplicities, and the intricacies of its non-diagonalizable behavior.