1 Cayley–Hamilton theorem for square matrices

1.1 Statement and notation via the characteristic polynomial

Let \(A\) be an \(n\times n\) matrix over a field (or commutative ring where the usual polynomial operations are well-defined). Its characteristic polynomial is \[ \chi_A(\lambda)=\det(\lambda I-A), \] a monic polynomial of degree \(n\). The Cayley–Hamilton theorem states that substituting the matrix \(A\) into its own characteristic polynomial yields the zero matrix: \[ \chi_A(A)=0. \] Here, \(\chi_A(A)\) is interpreted as a polynomial in \(A\) obtained by replacing \(\lambda\) with \(A\). Because matrix powers of \(A\) commute with each other (all are powers of the same matrix), the polynomial evaluation is unambiguous.

1.2 Interpreting \( \chi_A(A)=0 \) as an identity in matrix powers

Write the characteristic polynomial in the form \[ \chi_A(\lambda)=\lambda^n+c_{n-1}\lambda^{n-1}+\cdots +c_1\lambda+c_0. \] Then Cayley–Hamilton becomes \[ A^n+c_{n-1}A^{n-1}+\cdots+c_1A+c_0 I=0. \] This is an identity among matrix powers: it lets one rewrite the highest power \(A^n\) as a linear combination of lower powers of \(A\). Repeatedly applying this reduction turns any polynomial in \(A\) into an equivalent polynomial whose degree is \(<n\).

The minimal polynomial \(\mu_A(\lambda)\) of \(A\) is the monic polynomial of smallest degree such that \(\mu_A(A)=0\). It always divides \(\chi_A(\lambda)\). Degree bounds matter for inverse formulas: any expression of \(A^{-1}\) as a polynomial in \(A\) can be reduced modulo \(\mu_A\), implying that the effective degree needed is governed by \(\deg(\mu_A)\), not necessarily by \(n\).

2 From characteristic polynomial to an inverse identity

2.1 Expressing the characteristic polynomial coefficients in terms of traces/determinants (conceptual overview)

The coefficients of \(\chi_A(\lambda)\) encode invariants of \(A\). In general, they can be computed from principal minors or expressed using symmetric functions of eigenvalues. When \(A\) is diagonalizable over an algebraic closure, these coefficients become elementary symmetric polynomials in the eigenvalues. In characteristic zero fields or under suitable assumptions, traces of exterior powers (or power sums) relate to these symmetric functions, yielding conceptual formulas connecting coefficients to traces and determinant-like quantities.

2.2 Using invertibility: nonzero constant term criteria

The constant term of \(\chi_A\) is \[ c_0 = \chi_A(0)=\det(0\cdot I-A)=\det(-A)=(-1)^n\det(A). \] Thus \(c_0\neq 0\) exactly when \(\det(A)\neq 0\), i.e., precisely when \(A\) is invertible. This fact is central to deriving a polynomial expression for \(A^{-1}\): it ensures that the Cayley–Hamilton identity contains an invertible scalar multiple of \(I\), allowing division by that scalar.

2.3 Deriving \(A^{-1}\) by rearranging the Cayley–Hamilton equation

2.3.1 Solving for \(A^{-1}\) in the general polynomial form

Starting from \[ A^n+c_{n-1}A^{n-1}+\cdots+c_1A+c_0 I=0, \] isolate the constant term: \[ c_0 I = -\bigl(A^n+c_{n-1}A^{n-1}+\cdots+c_1A\bigr). \] Assuming \(c_0\neq 0\), multiply both sides by \(c_0^{-1}\): \[ I = -c_0^{-1}\bigl(A^n+c_{n-1}A^{n-1}+\cdots+c_1A\bigr). \] To obtain \(A^{-1}\), factor out one power of \(A\) from the right-hand side: \[ A^n+c_{n-1}A^{n-1}+\cdots+c_1A = A\bigl(A^{n-1}+c_{n-1}A^{n-2}+\cdots+c_1 I\bigr). \] So \[ I = -c_0^{-1}\,A\bigl(A^{n-1}+c_{n-1}A^{n-2}+\cdots+c_1 I\bigr). \] Multiplying on the right by \(A^{-1}\) (or equivalently on the left by \(A^{-1}\), using commutation with its own powers) gives \[ A^{-1} = -c_0^{-1}\bigl(A^{n-1}+c_{n-1}A^{n-2}+\cdots+c_1 I\bigr). \] This has the advertised form \(A^{-1}=p(A)\) with \(p\) a polynomial of degree at most \(n-1\) whose coefficients depend only on the characteristic polynomial coefficients.

3 Polynomial representation of the inverse

3.1 Existence and form of the polynomial \(p(A)\)

For an invertible \(n\times n\) matrix, the Cayley–Hamilton theorem guarantees a polynomial representation of the inverse: \[ A^{-1}=p(A), \] where one valid choice is \[ p(\lambda) = -\frac{1}{c_0}\bigl(\lambda^{n-1}+c_{n-1}\lambda^{n-2}+\cdots+c_1\bigr), \] with \(c_0\) the constant term of \(\chi_A\). This yields a concrete constructive method: compute \(\chi_A\), extract coefficients, then evaluate the resulting polynomial at \(A\).

3.2 Relationship between \(p(A)\), \chi_A, and the minimal polynomial

Since \(\chi_A(A)=0\), any polynomial \(q\) satisfies \(\chi_A(A)\,q(A)=0\). The polynomial \(p\) above arises by rewriting \(\chi_A\) so that \(A^{-1}\) appears as the factor that turns a reduced expression into the identity. In general, there exists some polynomial \(r\) such that \[ \lambda\cdot p(\lambda)-1 = r(\lambda)\chi_A(\lambda), \] and substituting \(\lambda=A\) gives \(A p(A)=I\). Replacing \(\chi_A\) by the minimal polynomial \(\mu_A\) does not change the core idea, because \(\mu_A\) also annihilates \(A\).

3.3 Uniqueness aspects under different polynomial bases

3.3.1 Reduced polynomial degrees and bounds

The polynomial \(p\) representing \(A^{-1}\) is not unique as a polynomial in the formal variable \(\lambda\), because one can add multiples of \(\chi_A(\lambda)\) or \(\mu_A(\lambda)\) that vanish upon evaluation at \(A\). However, there is a canonical “reduced” representative: reduce modulo the minimal polynomial so that the degree is \(<\deg(\mu_A)\). Under that reduction, the inverse polynomial becomes unique.

4 Worked examples

4.1 \(2\times 2\) matrices: explicit polynomial inverse formula

Let \(A\) be \(2\times 2\). Write the characteristic polynomial as \[ \chi_A(\lambda)=\lambda^2-t\lambda+d, \] where \(t=\operatorname{tr}(A)\) and \(d=\det(A)\). Cayley–Hamilton gives \[ A^2-tA+dI=0 \quad \Rightarrow \quad A^2-tA=-dI. \] Assume \(d\neq 0\). Then \[ I=-d^{-1}(A^2-tA) = -d^{-1}A(A-tI). \] Thus \[ A^{-1} = -d^{-1}(A-tI)=\frac{1}{d}(tI-A). \] This is the familiar explicit inverse formula for \(2\times 2\) matrices, expressed as a degree-one polynomial in \(A\).

4.2 \(3\times 3\) matrices: expressing \(A^{-1}\) using characteristic coefficients

Let \[ \chi_A(\lambda)=\lambda^3+c_2\lambda^2+c_1\lambda+c_0, \] with \(c_0\neq 0\) for invertible \(A\). Cayley–Hamilton yields \[ A^3+c_2A^2+c_1A+c_0 I=0. \] Rearrange: \[ c_0 I =-(A^3+c_2A^2+c_1A)= -A(A^2+c_2A+c_1 I). \] Since \(c_0\neq 0\), \[ A^{-1}= -c_0^{-1}(A^2+c_2A+c_1 I). \] So the inverse is given by a quadratic polynomial in \(A\), with coefficients determined by \(c_2,c_1,c_0\) from \(\chi_A\).

4.3 Example with a diagonalizable matrix (spectral viewpoint)

Suppose \(A\) is diagonalizable over a field where its eigenvalues \(\lambda_1,\dots,\lambda_n\) lie, so that \(A=S\operatorname{diag}(\lambda_i)S^{-1}\). Then any polynomial \(p(A)\) satisfies \[ p(A)=S\operatorname{diag}(p(\lambda_i))S^{-1}. \] If \(A\) is invertible, none of the \(\lambda_i\) equals \(0\). The polynomial derived from Cayley–Hamilton can be interpreted as enforcing \[ p(\lambda_i)=\lambda_i^{-1}\quad \text{for each eigenvalue } \lambda_i, \] because substituting \(A\) into the relation yields \(A p(A)=I\). This matches the general principle that, on the spectrum, the matrix inverse acts like scalar inversion.

4.4 Example with a non-diagonalizable matrix (Jordan form viewpoint)

If \(A\) is not diagonalizable, one may represent it in Jordan form \(J\) over a suitable extension field. Then \(A=SJS^{-1}\), and \[ p(A)=S\,p(J)\,S^{-1}. \] For Jordan blocks, evaluating a polynomial at \(J\) produces not only diagonal values \(p(\lambda)\) but also controlled contributions on the superdiagonal determined by derivatives of \(p\) at \(\lambda\). The polynomial constructed from Cayley–Hamilton still gives the correct inverse because the matrix identity \(A p(A)=I\) holds in the ring of polynomial expressions in \(A\), independent of diagonalizability. In effect, the Cayley–Hamilton construction ensures the necessary derivative conditions automatically for the nilpotent parts of Jordan blocks.

5 Computational and algorithmic considerations

5.1 How to compute the coefficients of the characteristic polynomial in practice

Computing \(\chi_A\) exactly is often done via symbolic algorithms (e.g., determinant-based computation of \(\det(\lambda I-A)\)), fraction-free or modular methods, or structured approaches for special matrix classes. For numerical work, one may approximate eigenvalues and reconstruct coefficients from them; however, reconstructing coefficients can be sensitive to rounding. The polynomial-in-\(A\) inverse method benefits most from contexts where the characteristic polynomial coefficients are available exactly or with controlled error.

5.2 Complexity and stability considerations for polynomial evaluation

Once the coefficients are known, evaluating \(p(A)\) typically uses Horner’s method or repeated squaring-and-multiplication strategies to reduce the number of matrix multiplications. The dominant cost is matrix multiplication, so the practical complexity depends on matrix size and multiplication algorithm. Numerically, evaluating higher-degree polynomials in \(A\) can amplify errors, particularly if \(A\) has eigenvalues spread over widely different magnitudes. Stability can improve when the degree is reduced using the minimal polynomial (when known or computable) and when evaluation is performed with numerically robust multiplication schemes.

5.3 Verification: checking that \(A\,p(A)=I\) (and \(p(A)\,A=I\))

5.3.1 Symbolic verification vs numeric verification workflows

In symbolic settings, one can verify the identity by substituting the derived polynomial into the Cayley–Hamilton relation and simplifying, guaranteeing exactness. In numeric workflows, verification is done by computing the residual \[ R = A\,p(A)-I \] and measuring its norm. Because polynomial evaluation and arithmetic are subject to floating-point rounding, the residual provides an error estimate; small residual norms indicate that the computed polynomial inverse behaves like an inverse within numerical tolerance.

6 Special cases and common structures

6.1 Inverse as a low-degree polynomial when the minimal polynomial has small degree

If the minimal polynomial \(\mu_A\) has degree \(m\ll n\), then the inverse can be represented by a polynomial of degree at most \(m-1\). This occurs, for instance, when \(A\) satisfies a short recurrence relation beyond the characteristic one. In such cases, the polynomial-in-\(A\) method becomes more efficient: fewer powers of \(A\) are needed, and evaluation tends to be more stable due to reduced degree.

6.2 Applications to matrices satisfying additional polynomial relations

Some matrices satisfy extra identities such as \(A^k\) being expressible in terms of lower powers due to geometric or algebraic constraints (e.g., projection-like behavior, idempotence, or limited distinct eigenvalues). When an additional annihilating polynomial is known, one can replace \(\chi_A\) by that smaller annihilator to derive a tighter inverse polynomial bound. The resulting formula is still rooted in the same mechanism: invertibility supplies a nonzero scalar constant term enabling division.

6.3 Normal, symmetric, and Hermitian matrices (high-level consequences)

For normal matrices (including Hermitian), the spectral decomposition is well-behaved. Polynomial expressions in \(A\) interact naturally with eigenvalues: if the characteristic polynomial is known, the constructed \(p\) approximates or matches scalar inversion on each spectral value. While the Cayley–Hamilton-based formula does not rely on diagonalization, the normal case often makes computation and interpretation easier—especially when eigenvalues are determined stably, or when minimal polynomials have small degree due to clustered spectra.

7 Alternative derivations and connections (non-controversial overview)

7.1 Using adjugate and determinant identities as a comparison

A classical identity expresses the inverse using the adjugate matrix: \[ A^{-1}=\frac{\operatorname{adj}(A)}{\det(A)}. \] Because the adjugate can be written in terms of cofactors, and cofactors are linked to coefficients of \(\chi_A\), there is a conceptual bridge between adjugate-based formulas and Cayley–Hamilton-based polynomial formulas. The Cayley–Hamilton approach can be viewed as collapsing the adjugate expression into a polynomial in \(A\) that necessarily exists whenever \(\det(A)\neq 0\).

7.2 Relation to Lagrange interpolation on eigenvalues (conceptual)

For diagonalizable matrices, one can construct a polynomial \(p\) such that \(p(\lambda_i)=1/\lambda_i\) on the eigenvalues \(\lambda_i\). Interpolation yields a polynomial meeting these conditions, and Cayley–Hamilton ensures that the same polynomial (after reduction) gives the correct matrix inverse. Thus, the polynomial \(p\) can be interpreted as a “spectral interpolant” of reciprocal values, with the matrix identity enforcing consistency beyond merely eigenvalues.

7.3 Connections to functional calculus for matrices

Matrix polynomial expressions belong to the general framework of functional calculus: for suitable scalar functions \(f\), one can define \(f(A)\) through polynomials or limits of polynomials. The inverse function \(f(\lambda)=1/\lambda\) is analytic away from \(0\), and for invertible \(A\) it can be approximated or represented by polynomials. Cayley–Hamilton provides an exact representation in this case, turning the general functional-calculus philosophy into a concrete closed-form identity.

8 Generalizations and limitations

8.1 What fails when \(A\) is singular (constant term zero)

If \(A\) is singular, then \(\det(A)=0\) and the constant term of \(\chi_A\) is zero. Cayley–Hamilton then gives an identity of the form \[ A(\text{polynomial in }A)=0, \] but there is no way to divide by \(c_0\) to isolate \(A^{-1}\). In this situation, \(A^{-1}\) does not exist, so no polynomial \(p\) can satisfy \(A p(A)=I\). The obstruction is precisely the vanishing of the constant term in the characteristic polynomial.

8.2 Field dependence and polynomial coefficient behavior

The coefficients of \(\chi_A\) depend on the base field, but the Cayley–Hamilton theorem holds over any commutative field for which determinants and polynomial evaluation are defined. If computations are performed over extensions (e.g., to access eigenvalues), the polynomial identity still lives in the original coefficient field because \(\chi_A\) has coefficients determined by \(A\) intrinsically. However, practical computation may differ in efficiency depending on the chosen field and available algorithms.

8.3 Extending the idea to rational functions of matrices

8.3.1 Bounds on polynomial degree needed to represent \(A^{-1}\)

The inverse is a special rational function: \(1/\lambda\). For general rational functions \(g(\lambda)/h(\lambda)\), one may attempt to express \(h(A)^{-1}\) when \(h(A)\) is invertible and then reduce the resulting expression to a polynomial in \(A\) using annihilating polynomials. Degree bounds depend on which polynomial annihilates \(A\) and how \(h\) factors relative to the spectrum. In the specific case \(1/\lambda\), the Cayley–Hamilton construction provides a uniform bound of degree at most \(n-1\), and potentially smaller when reduced modulo the minimal polynomial.