Minimal polynomial (linear algebra)

In linear algebra, the minimal polynomial μ_A of an n × n matrix Template:Mvar over a field F is the monic polynomial Template:Mvar over F of least degree such that P(A) = 0. Any other polynomial Template:Mvar with Q(A) = 0 is a (polynomial) multiple of μ_A.

The following three statements are equivalent:

1. Template:Mvar is a root of μ_A,
2. Template:Mvar is a root of the characteristic polynomial χ_A of Template:Mvar,
3. Template:Mvar is an eigenvalue of matrix Template:Mvar.

The multiplicity of a root Template:Mvar of μ_A is the largest power Template:Mvar such that ker((A − λI_n)^m) strictly contains ker((A − λI_n)^m−1). In other words, increasing the exponent up to Template:Mvar will give ever larger kernels, but further increasing the exponent beyond Template:Mvar will just give the same kernel.

If the field F is not algebraically closed, then the minimal and characteristic polynomials need not factor according to their roots (in F) alone, in other words they may have irreducible polynomial factors of degree greater than 1. For irreducible polynomials Template:Mvar one has similar equivalences:

1. Template:Mvar divides μ_A,
2. Template:Mvar divides χ_A,
3. the kernel of P(A) has dimension at least 1.
4. the kernel of P(A) has dimension at least deg(P).

Like the characteristic polynomial, the minimal polynomial does not depend on the base field. In other words, considering the matrix as one with coefficients in a larger field does not change the minimal polynomial. The reason for this differs from the case with the characteristic polynomial (where it is immediate from the definition of determinants), namely by the fact that the minimal polynomial is determined by the relations of linear dependence between the powers of Template:Mvar: extending the base field will not introduce any new such relations (nor of course will it remove existing ones).

The minimal polynomial is often the same as the characteristic polynomial, but not always. For example, if Template:Mvar is a multiple aI_n of the identity matrix, then its minimal polynomial is X − a since the kernel of aI_n − A = 0 is already the entire space; on the other hand its characteristic polynomial is (X − a)ⁿ (the only eigenvalue is Template:Mvar, and the degree of the characteristic polynomial is always equal to the dimension of the space). The minimal polynomial always divides the characteristic polynomial, which is one way of formulating the Cayley–Hamilton theorem (for the case of matrices over a field).

Formal definition

Given an endomorphism Template:Mvar on a finite-dimensional vector space Template:Mvar over a field F, let I_T be the set defined as

${{I}}_T=\{p\in \mathbf{𝐅}[t]\mid p(T)=0\},$

where F[t] is the space of all polynomials over the field F. I_T is a proper ideal of F[t]. Since F is a field, F[t] is a principal ideal domain, thus any ideal is generated by a single polynomial, which is unique up to a unit in F. A particular choice among the generators can be made, since precisely one of the generators is monic. The minimal polynomial is thus defined to be the monic polynomial that generates I_T. It is the monic polynomial of least degree in I_T.

Applications

An endomorphism Template:Mvar of a finite-dimensional vector space over a field F is diagonalizable if and only if its minimal polynomial factors completely over F into distinct linear factors. The fact that there is only one factor X − λ for every eigenvalue Template:Mvar means that the generalized eigenspace for Template:Mvar is the same as the eigenspace for Template:Mvar: every Jordan block has size 1. More generally, if Template:Mvar satisfies a polynomial equation P(φ) = 0 where Template:Mvar factors into distinct linear factors over F, then it will be diagonalizable: its minimal polynomial is a divisor of Template:Mvar and therefore also factors into distinct linear factors. In particular one has:

• P = X^k − 1: finite order endomorphisms of complex vector spaces are diagonalizable. For the special case k = 2 of involutions, this is even true for endomorphisms of vector spaces over any field of characteristic other than 2, since X² − 1 = (X − 1)(X + 1) is a factorization into distinct factors over such a field. This is a part of representation theory of cyclic groups.
• P = X² − X = X(X − 1): endomorphisms satisfying φ² = φ are called projections, and are always diagonalizable (moreover their only eigenvalues are 0 and 1).
• By contrast if μ_φ = X^k with k ≥ 2 then Template:Mvar (a nilpotent endomorphism) is not necessarily diagonalizable, since X^k has a repeated root 0.

These cases can also be proved directly, but the minimal polynomial gives a unified perspective and proof.

Computation

For a nonzero vector Template:Mvar in Template:Mvar define:

${{I}}_{T,v}=\{p\in \mathbf{𝐅}[t]|p(T)(v)=0\}.$

This definition satisfies the properties of a proper ideal. Let μ_T,v be the monic polynomial which generates it.

Example

Define Template:Mvar to be the endomorphism of R³ with matrix, on the canonical basis,

$\left(\begin{array}{ccc}1& -1& -1\\ 1& -2& 1\\ 0& 1& -3\end{array}\right).$

Taking the first canonical basis vector e₁ and its repeated images by Template:Mvar one obtains

$e_1=\left[\begin{array}{c}1\\ 0\\ 0\end{array}\right],T\cdot e_1=\left[\begin{array}{c}1\\ 1\\ 0\end{array}\right].T^2\cdot e_1=\left[\begin{array}{c}0\\ -1\\ 1\end{array}\right]\text{ and}{T}^3\cdot e_1=\left[\begin{array}{c}0\\ 3\\ -4\end{array}\right]$

of which the first three are easily seen to be linearly independent, and therefore span all of R³. The last one then necessarily is a linear combination of the first three, in fact

T³ ⋅ e₁ = −4T² ⋅ e₁ − T ⋅ e₁ + e₁,

so that:

μ_T,e1 = X³ + 4X² + X − I.

This is in fact also the minimal polynomial μ_T and the characteristic polynomial χ_T: indeed μ_T,e1 divides μ_T which divides χ_T, and since the first and last are of degree 3 and all are monic, they must all be the same. Another reason is that in general if any polynomial in Template:Mvar annihilates a vector Template:Mvar, then it also annihilates T⋅v (just apply Template:Mvar to the equation that says that it annihilates Template:Mvar), and therefore by iteration it annihilates the entire space generated by the iterated images by Template:Mvar of Template:Mvar; in the current case we have seen that for v = e₁ that space is all of R³, so μ_T,e1(T) = 0. Indeed one verifies for the full matrix that T³ + 4T² + T − I₃ is the zero matrix:

$\left[\begin{array}{ccc}0& 1& -3\\ 3& -13& 23\\ -4& 19& -36\end{array}\right]+4\left[\begin{array}{ccc}0& 0& 1\\ -1& 4& -6\\ 1& -5& 10\end{array}\right]+\left[\begin{array}{ccc}1& -1& -1\\ 1& -2& 1\\ 0& 1& -3\end{array}\right]+\left[\begin{array}{ccc}-1& 0& 0\\ 0& -1& 0\\ 0& 0& -1\end{array}\right]=\left[\begin{array}{ccc}0& 0& 0\\ 0& 0& 0\\ 0& 0& 0\end{array}\right]$

See also

• Annihilating polynomial

References

• Template:Lang Algebra