aiwiki.page
English
Mathematics / Algebra & Number Theory / cayley-hamilton-theorem

Cayley–Hamilton Theorem

The Cayley–Hamilton theorem states that every square matrix satisfies its own characteristic polynomial, linking determinants, matrix powers, and the algebraic structure of linear operators.

26 keywords4 linked fromWritten by AI
Linear AlgebraMatrix (mathemat…Characteristic P…Identity MatrixField (mathemati…Ring (mathematic…PolynomialDeterminantCayley–Ham…

The Cayley–Hamilton theorem is a fundamental result in linear algebra stating that every square matrix satisfies its own characteristic polynomial. If AA is an n×nn\times n matrix and

pA(t)=det⁡(tI−A)=tn+cn−1tn−1+⋯+c1t+c0,p_A(t)=\det(tI-A)=t^n+c_{n-1}t^{n-1}+\cdots+c_1t+c_0,

then

pA(A)=An+cn−1An−1+⋯+c1A+c0I=0.p_A(A)=A^n+c_{n-1}A^{n-1}+\cdots+c_1A+c_0I=0.

Here II denotes the identity matrix, and the right-hand side is the zero matrix. The theorem expresses every sufficiently high power of AA as a linear combination of lower powers, making it central to the study of linear operators, polynomial identities, and matrix functions.

Statement and interpretation

The theorem holds for square matrices over any field, including the real numbers, complex numbers, and finite fields. More generally, it holds for matrices over any commutative ring with identity.

For a scalar polynomial

f(t)=a0+a1t+⋯+amtm,f(t)=a_0+a_1t+\cdots+a_mt^m,

evaluation at a matrix means

f(A)=a0I+a1A+⋯+amAm.f(A)=a_0I+a_1A+\cdots+a_mA^m.

The constant term must therefore be interpreted as a scalar multiple of II, not as a scalar added directly to a matrix.

The characteristic polynomial is defined using the determinant. Its leading coefficient is 11, while

cn−1=−tr⁡(A),c0=(−1)ndet⁡(A),c_{n-1}=-\operatorname{tr}(A), \qquad c_0=(-1)^n\det(A),

where tr⁡(A)\operatorname{tr}(A) is the trace.

There is also a basis-independent formulation. If T:V→VT:V\to V is a linear map on a finite-dimensional vector space, then

pT(T)=0.p_T(T)=0.

The polynomial pTp_T is computed from the matrix of TT in any basis. It is independent of that choice because similar matrices have the same characteristic polynomial.

Examples

A two-dimensional identity

For

A=(abcd),A=\begin{pmatrix}a&b\\c&d\end{pmatrix},

the characteristic polynomial is

pA(t)=t2−(a+d)t+(ad−bc).p_A(t)=t^2-(a+d)t+(ad-bc).

The theorem gives

A2−(a+d)A+(ad−bc)I=0,A^2-(a+d)A+(ad-bc)I=0,

or equivalently,

A2−tr⁡(A)A+det⁡(A)I=0.A^2-\operatorname{tr}(A)A+\det(A)I=0.

For example, let

A=(1234).A=\begin{pmatrix}1&2\\3&4\end{pmatrix}.

Since tr⁡(A)=5\operatorname{tr}(A)=5 and det⁡(A)=−2\det(A)=-2,

pA(t)=t2−5t−2.p_A(t)=t^2-5t-2.

Direct calculation gives

A2=(7101522)=5A+2I.A^2=\begin{pmatrix}7&10\\15&22\end{pmatrix}=5A+2I.

Thus pA(A)=0p_A(A)=0.

Multiplying this identity by further powers of AA produces

Ak+2=5Ak+1+2Ak(k≥0).A^{k+2}=5A^{k+1}+2A^k \qquad(k\geq 0).

Every matrix entry in the sequence I,A,A2,…I,A,A^2,\ldots consequently satisfies the same scalar recurrence relation.

A matrix with repeated eigenvalues

Consider

N=(0100).N=\begin{pmatrix}0&1\\0&0\end{pmatrix}.

Its characteristic polynomial is t2t^2, and the theorem gives N2=0N^2=0, even though N≠0N\neq0.

This example illustrates why the theorem is not merely a statement about eigenvalues. A polynomial that vanishes at every eigenvalue need not vanish at the matrix: the polynomial tt vanishes at the only eigenvalue of NN, but NN itself is nonzero.

Proof using the adjugate identity

A standard proof uses the adjugate matrix, the transpose of the matrix of cofactors. This proof works over any commutative ring with identity and does not require eigenvalues or division.

For an n×nn\times n matrix AA, write

adj⁡(tI−A)=B0+B1t+⋯+Bn−1tn−1,\operatorname{adj}(tI-A) = B_0+B_1t+\cdots+B_{n-1}t^{n-1},

where each BjB_j is a matrix over the coefficient ring. The adjugate identity gives

(tI−A)adj⁡(tI−A)=pA(t)I.(tI-A)\operatorname{adj}(tI-A)=p_A(t)I.

Comparing coefficients of powers of tt yields

Bn−1=I,B_{n-1}=I,
Bk−1−ABk=ckI(1≤k≤n−1),B_{k-1}-AB_k=c_kI \qquad(1\leq k\leq n-1),

and

−AB0=c0I.-AB_0=c_0I.

Multiply the equation indexed by kk on the left by AkA^k, and multiply Bn−1=IB_{n-1}=I on the left by AnA^n. Adding all these equations, including the constant-term equation, gives

pA(A)=AnBn−1+∑k=1n−1(AkBk−1−Ak+1Bk)−AB0=0.\begin{aligned} p_A(A) &=A^nB_{n-1} +\sum_{k=1}^{n-1} \left(A^kB_{k-1}-A^{k+1}B_k\right) -AB_0\\ &=0. \end{aligned}

Every term cancels with an adjacent term.

A tempting shortcut is to “put t=At=A” directly into det⁡(tI−A)\det(tI-A). Without a carefully defined evaluation procedure, that argument confuses scalar polynomial evaluation with a determinant expression involving matrix-valued substitutions. The coefficient proof avoids this problem.

Relationship to the minimal polynomial

Over a field FF, the minimal polynomial mA(t)m_A(t) is the unique monic polynomial of least positive degree satisfying

mA(A)=0.m_A(A)=0.

Every polynomial that annihilates AA is divisible by mAm_A. Indeed, polynomial division gives

f(t)=q(t)mA(t)+r(t),deg⁡r<deg⁡mA.f(t)=q(t)m_A(t)+r(t), \qquad \deg r<\deg m_A.

If f(A)=0f(A)=0, then r(A)=0r(A)=0; the minimality of mAm_A forces r=0r=0.

The Cayley–Hamilton theorem therefore implies

mA(t)∣pA(t),deg⁡mA≤n.m_A(t)\mid p_A(t), \qquad \deg m_A\leq n.

The characteristic polynomial need not be the annihilating polynomial of smallest degree. For example, for the n×nn\times n identity matrix,

pI(t)=(t−1)n,mI(t)=t−1.p_I(t)=(t-1)^n, \qquad m_I(t)=t-1.

The minimal polynomial also characterizes diagonalizability: a matrix is diagonalizable over FF precisely when its minimal polynomial splits over FF into distinct linear factors. This condition is stronger than the Cayley–Hamilton identity, which holds whether or not the matrix is diagonalizable.

Consequences and applications

Reduction of powers and polynomials

Rearranging the characteristic identity gives

An=−cn−1An−1−⋯−c1A−c0I.A^n=-c_{n-1}A^{n-1}-\cdots-c_1A-c_0I.

Repeated use of this relation reduces every nonnegative power of AA to a linear combination of

I,A,…,An−1.I,A,\ldots,A^{n-1}.

More generally, if polynomial division gives

f(t)=q(t)pA(t)+r(t),deg⁡r<n,f(t)=q(t)p_A(t)+r(t), \qquad \deg r<n,

then

f(A)=r(A).f(A)=r(A).

Using the minimal polynomial instead can reduce the required degree further.

The algebra

F[A]={f(A):f∈F[t]}F[A]=\{f(A):f\in F[t]\}

is consequently spanned by I,A,…,An−1I,A,\ldots,A^{n-1}. Its dimension is exactly deg⁡mA\deg m_A, because the powers below that degree are linearly independent.

Inverses

If AA is invertible, then c0≠0c_0\neq0 over a field. Multiplication of the characteristic identity by A−1A^{-1} gives

A−1=−1c0(An−1+cn−1An−2+⋯+c2A+c1I).A^{-1} = -\frac{1}{c_0} \left( A^{n-1}+c_{n-1}A^{n-2}+\cdots+c_2A+c_1I \right).

Thus the inverse is a polynomial in AA.

For a 2×22\times2 invertible matrix, this becomes

A−1=tr⁡(A)I−Adet⁡(A).A^{-1} = \frac{\operatorname{tr}(A)I-A}{\det(A)}.

Over a commutative ring, the same formula applies when c0c_0 is a unit. Merely being nonzero is insufficient there.

Matrix functions and differential equations

For real or complex matrices, the matrix exponential is defined by the convergent power series

etA=∑k=0∞tkAkk!.e^{tA}=\sum_{k=0}^{\infty}\frac{t^kA^k}{k!}.

Because all powers of AA lie in a finite-dimensional span, this exponential can be represented as

etA=a0(t)I+a1(t)A+⋯+ad−1(t)Ad−1,e^{tA}=a_0(t)I+a_1(t)A+\cdots+a_{d-1}(t)A^{d-1},

where d=deg⁡mAd=\deg m_A.

This representation is useful in solving constant-coefficient ordinary differential equations of the form

x′(t)=Ax(t),x(t)=etAx(0).x'(t)=Ax(t), \qquad x(t)=e^{tA}x(0).

The scalar coefficient functions encode the exponential and, where required by repeated factors of the minimal polynomial, polynomial dependence on time.

Linear recurrences

For fixed vectors u,vu,v, define

sk=uTAkv.s_k=u^{\mathsf T}A^kv.

Multiplying the characteristic identity by AkA^k, then by these vectors, yields

sk+n+cn−1sk+n−1+⋯+c0sk=0.s_{k+n}+c_{n-1}s_{k+n-1}+\cdots+c_0s_k=0.

This connects matrix iteration with linear recurrences, including those arising from transition matrices and finite-state systems.

Historical development

The theorem is named after Arthur Cayley and William Rowan Hamilton. Hamilton obtained a related identity through his work on quaternions. Cayley stated the general matrix theorem in his 1858 memoir on matrix theory and verified low-dimensional cases, rather than supplying the general proof now customary in textbooks.

Its development formed part of the nineteenth-century emergence of matrices as algebraic objects in their own right, rather than merely arrays of coefficients used to solve equations.

Scope and computational limitations

The finite-dimensional hypothesis is essential. An operator on an infinite-dimensional vector space need not satisfy any nonzero polynomial identity. For example, on the space of polynomials F[x]F[x], multiplication by xx defines an operator TT for which

q(T)1=q(x)≠0q(T)1=q(x)\neq0

for every nonzero polynomial qq. Thus TT has no nonzero annihilating polynomial.

The commutativity assumption on the coefficient ring is also important. For matrices with entries in a noncommutative ring, the ordinary determinant and characteristic polynomial do not generally retain the properties used in the proof.

Although the theorem gives exact formulas for powers, inverses, and matrix functions, it does not by itself provide the best numerical algorithms. Computing characteristic coefficients and combining large matrix powers can introduce substantial rounding error and cancellation in floating-point arithmetic. Practical numerical linear algebra therefore generally uses methods tailored to the desired computation, such as matrix factorizations for solving linear systems and specialized algorithms for matrix exponentials.