Minimal Polynomial (Linear Algebra)¶

The unique monic generator of all polynomial identities satisfied by a finite-dimensional linear operator, encoding the least annihilating relation and the largest primary-block exponents.

Version: v2 · 2026-09-06 · History
Domain-specific #: 2278
Origin domain: linear algebra
Subdomain: operator theory and canonical forms
Aliases: Operator minimal polynomial, Matrix minimal polynomial, Minimum polynomial of a matrix

Core Idea¶

For a linear operator \(T:V\to V\) on a nonzero finite-dimensional vector space over a field \(F\), the minimal polynomial \(m_T(x)\) is the unique monic polynomial of least degree such that

\[ m_T(T)=0. \]

The short definition hides the more useful structural fact. Polynomial evaluation makes \(V\) an \(F[x]\)-module by letting \(x\) act as \(T\), and the set

\[ I_T=\{p(x)\in F[x]:p(T)=0\} \]

is a nonzero ideal. Because \(F[x]\) is a principal ideal domain, \(I_T\) has one monic generator. That generator is \(m_T\). Consequently,

\[ p(T)=0 \quad\Longleftrightarrow\quad m_T(x)\mid p(x). \]

Thus the minimal polynomial is not just one convenient annihilating equation. It is the divisibility certificate for all polynomial identities of the operator. It compresses an ideal of relations into one canonical polynomial.

Existence follows in finite dimensions because the characteristic polynomial \(\chi_T(x)\) annihilates \(T\) by the Cayley–Hamilton theorem. Therefore \(m_T\mid\chi_T\) and \(\deg m_T\leq\dim V\). The converse information is deliberately weaker: \(m_T\) usually does not record all multiplicities present in \(\chi_T\), and it is not a complete similarity invariant. What it records is exactly the deepest polynomial obstruction in each primary component.

When \(\chi_T\) splits over \(F\), write the distinct eigenvalues as \(\lambda_1,\ldots,\lambda_s\). Then

\[ m_T(x)=\prod_{i=1}^{s}(x-\lambda_i)^{r_i}, \]

where \(r_i\) is the size of the largest Jordan block associated with \(\lambda_i\). The characteristic polynomial records the total algebraic multiplicity of each eigenvalue; the minimal polynomial records the longest generalized-eigenvector chain. Over a field where the characteristic polynomial need not split, the same role is expressed through powers of distinct irreducible factors and rational canonical form.

Several central tests become one-line consequences. The operator is diagonalizable over \(F\) exactly when \(m_T\) splits over \(F\) into distinct linear factors. A nilpotent operator has \(m_T=x^r\), where \(r\) is its nilpotency index. A nontrivial projection satisfies \(m_T=x(x-1)\). If \(T^k=I\), then \(m_T\mid x^k-1\); diagonalizability follows when that polynomial splits into distinct factors over the field in question.

The identity is basis-independent. If \(A\) and \(B=S^{-1}AS\) represent the same operator in different bases, then \(p(B)=S^{-1}p(A)S\), so they have exactly the same annihilating polynomials and the same minimal polynomial. This makes \(m_T\) an operator invariant rather than a feature of one matrix display.

The minimal polynomial recurs in the same roles—definition, computation, diagnosis, and intervention—across canonical-form theory, diagonalizability, matrix functions, cyclic subspaces, representations, and finite-field algorithms. Yet the mechanism remains pinned to polynomial evaluation on linear endomorphisms; stripping that substrate leaves only the broader primes Invariance and Compression.

Structural Signature¶

The recurring structure is:

finite-dimensional operator \(T\) over \(F\) → evaluation homomorphism \(F[x]\to\operatorname{End}_F(V)\) → nonzero annihilator ideal \(I_T\) → unique monic generator \(m_T\) → universal divisibility test for polynomial identities → factor exponents encoding primary depth → operator-structure and computation consequences

Sig role-phrases:

Finite-dimensional operator. A linear endomorphism \(T\) or any square matrix representing it supplies the object whose polynomial relations are sought.
Polynomial functional calculus. Each \(p(x)=a_0+a_1x+\cdots+a_dx^d\) is evaluated as \(p(T)=a_0I+a_1T+\cdots+a_dT^d\).
Annihilator ideal. All \(p\) with \(p(T)=0\) form an ideal \(I_T\subset F[x]\), not an arbitrary list of unrelated equations.
Monic least-degree generator. Principal-ideal structure and monic normalization select one unique polynomial \(m_T\).
Universal divisibility invariant. A polynomial annihilates \(T\) precisely when it is a multiple of \(m_T\).
Characteristic-polynomial envelope. Cayley–Hamilton guarantees existence and gives \(m_T\mid\chi_T\).
Primary-depth code. The exponent of each irreducible factor records the longest corresponding primary chain; over a splitting field this is the largest Jordan-block size for that eigenvalue.
Basis-change invariance. Similar matrices yield identical minimal polynomials, so the identity belongs to the operator's similarity class.
Reduction rule. Polynomial expressions in \(T\) can be reduced modulo \(m_T\) to degree below \(\deg m_T\).
Local-to-global computation. Minimal polynomials of cyclic vectors or invariant components divide \(m_T\), and their least common multiple can recover it.

The invariant is retained when the basis, matrix coordinates, or computation algorithm changes. It is lost if the object is not an endomorphism, polynomial evaluation is absent, or the chosen polynomial is merely an annihilator without the generator/divisibility property.

What It Is Not¶

Not any polynomial that annihilates a matrix. If \(p(T)=0\), then \(p\) may contain unnecessary factors or powers. It is minimal only when it is the unique monic generator of \(I_T\).
Not the characteristic polynomial. The characteristic polynomial has degree \(\dim V\) and records total algebraic multiplicities. The minimal polynomial divides it and can have smaller degree.
Not a complete record of Jordan form. It gives each eigenvalue's largest block size but not the number or full partition of blocks. Distinct similarity classes can share both characteristic and minimal polynomials.
Not a basis-dependent formula. Computing from one matrix does not make the result coordinate-specific; similarity preserves it exactly.
Not necessarily split into linear factors. Over a non-algebraically-closed field it may contain irreducible factors of degree greater than one. Diagonalizability over \(F\) requires splitting over \(F\), not merely square-freeness in some larger field.
Not the minimal polynomial of an algebraic field element by default. The field-theory object is the monic generator of the kernel of evaluation \(F[x]\to E\), \(p\mapsto p(\alpha)\). The constructions are closely related, but the domain object and downstream roles differ.
Not a numerical approximation. Exact minimal-polynomial computation is algebraic and can be unstable as a floating-point inference for nearly defective matrices^[1]. Numerical eigenvalue routines answer a different problem.
Not the Minimal Pairs prime. “Minimal” here means least-degree monic generator under divisibility, not a controlled contrast between two observations.

Scope of Application¶

The abstraction travels literally across areas that study finite-dimensional linear operators over fields.

Elementary and abstract linear algebra. It unifies annihilating equations, Cayley–Hamilton consequences, diagonalizability, projections, involutions, nilpotence, and cyclic operators.
Matrix canonical forms. It is the largest invariant factor in rational canonical form and encodes the maximum Jordan-chain length for each eigenvalue when Jordan form exists over the field.
Primary decomposition. Factorization into coprime primary powers separates invariant subspaces and supports projection operators obtained through Bézout identities.
Representation theory. A group element acting linearly has a minimal polynomial constrained by its order or defining relations; repeated factors reveal possible nonsemisimple behavior in the field's characteristic^[2].
Matrix functions and recurrences. Powers and polynomial expressions reduce modulo \(m_T\), turning long sequences of matrix powers into recurrences of order \(\deg m_T\).
Krylov and cyclic-subspace methods. Linear dependence among \(v,Tv,T^2v,\ldots\) yields a vector-relative order polynomial; local polynomials are combined to determine the operator invariant.
Computational algebra over finite fields. Randomized and deterministic algorithms compute and verify minimal polynomials for large matrices, canonical-form work, module structure, and matrix-group calculations^[3].
Linear systems. The state-transition operator's polynomial relation bounds the independent powers needed for algebraic manipulation and exposes cyclic structure, though control-specific notions of reachability and observability require additional data.

Infinite-dimensional operators can have no nonzero annihilating polynomial, and bounded-operator spectral theory uses broader analytic functional calculi. Those extensions do not automatically instantiate the finite-dimensional minimal-polynomial abstraction.

Clarity¶

A claimed minimal polynomial should pass five checks.

Coefficient field: State \(F\). Factorization and diagonalizability depend on the field even though the polynomial identity itself remains defined.
Operator object: Identify an endomorphism \(T:V\to V\), not merely a rectangular matrix or a polynomial unrelated to an operator.
Annihilation: Verify \(m(T)=0\).
Minimality or generation: Show no monic polynomial of smaller degree annihilates \(T\), or prove that every annihilator is divisible by \(m\).
Normalization: Require monic form; otherwise nonzero scalar multiples would destroy uniqueness.

Common computations verify minimality by more informative routes. If a vector \(v\) is cyclic—its Krylov sequence spans \(V\)—then the first linear dependence among \(v,Tv,\ldots\) gives a degree-\(\dim V\) polynomial that is both local and global; it also equals the characteristic polynomial. If canonical form is known, take the least common multiple of block polynomials. If a candidate divides \(\chi_T\), annihilates \(T\), and each of its primary exponents is forced by a kernel-chain or cyclic-vector witness, minimality follows.

The fastest boundary test against the characteristic polynomial is a scalar operator: for \(T=aI\) on an \(n\)-dimensional space, \(m_T=x-a\) while \(\chi_T=(x-a)^n\). The fastest boundary test against a mere annihilator is that \((x-a)^2\) also annihilates \(aI\), but is not minimal.

Manages Complexity¶

An operator satisfies infinitely many polynomial identities once it satisfies one: every multiple of an annihilating polynomial also annihilates it. Listing identities therefore creates redundancy. The principal-ideal formulation compresses the entire set to \((m_T)\). Questions of the form “does \(p(T)=0\)?” become polynomial divisibility tests.

The minimal polynomial also bounds the algebra generated by \(T\). Division gives \(p=q m_T+r\) with \(\deg r<\deg m_T\), hence

\[ p(T)=r(T). \]

Every polynomial in \(T\) is therefore represented by a combination of \(I,T,\ldots,T^{d-1}\), where \(d=\deg m_T\). When \(d\ll\dim V\), this can be a substantial algebraic compression. High powers obey the recurrence obtained from \(m_T(T)=0\), and interpolation or Hermite-type data on the roots can determine matrix functions in the quotient algebra \(F[x]/(m_T)\).

Factorization separates structure by primary component. The distinct irreducible factors identify spectral/primary types; their exponents measure the depth of nonsemisimplicity. A square-free split polynomial certifies diagonalizability without constructing an eigenbasis. A pure power \(x^r\) immediately certifies nilpotence and its exact index. These diagnoses replace many case-specific matrix calculations with one invariant.

Computationally, local order polynomials obtained from Krylov sequences avoid constructing every matrix power as a dense global object. Their least common multiple approaches or equals \(m_T\), and deterministic evaluation can verify a candidate^[3]. The invariant organizes both the theorem and the algorithm.

Abstract Reasoning¶

Divisibility inference. From \(p(T)=0\), infer \(m_T\mid p\). This can rule out proposed structures: if \(m_T\) has a repeated factor, no square-free polynomial can annihilate \(T\).

Diagonalizability inference. If \(m_T\) splits with no repeated root, primary components are eigenspaces and \(T\) is diagonalizable. Conversely, a diagonal representation is annihilated by the product of distinct factors \((x-\lambda)\), so the minimal polynomial is square-free and split.

Direct-sum inference. For \(T\oplus S\), a polynomial annihilates the direct sum exactly when it annihilates both summands. Therefore

\[ m_{T\oplus S}=\operatorname{lcm}(m_T,m_S). \]

This makes block assembly compositional.

Restriction and quotient inference. If a subspace is \(T\)-invariant, the minimal polynomials of the restriction and induced quotient divide \(m_T\). They expose partial structure without exceeding the global primary exponents.

Similarity inference. Since \(p(S^{-1}TS)=S^{-1}p(T)S\), changing basis preserves the annihilator ideal. A discrepancy between alleged minimal polynomials proves that two matrices cannot be similar; agreement alone does not prove similarity.

Field inference. Extending scalars does not create a lower-degree polynomial relation among \(I,T,T^2,\ldots\) when the matrix entries begin in \(F\), although it may split irreducible factors and change the language of eigenvalues. This separates the invariant polynomial from its field-dependent factor interpretation.

Functional reduction. To compute \(p(T)\), first reduce \(p\) modulo \(m_T\). This predicts the maximum number of independent powers needed and turns operator equations into quotient-ring arithmetic.

Knowledge Transfer¶

Within algebra, transfer is literal. The same annihilator-ideal logic applies to a matrix in coordinates, an abstract endomorphism, a linear representation element, an \(F[x]\)-module, or a direct sum of cyclic components. Translation among Jordan blocks, invariant factors, Krylov sequences, and polynomial functional calculus preserves the mechanism.

There is a particularly close transfer to the minimal polynomial of an algebraic element \(\alpha\) over \(F\). Evaluation \(F[x]\to F(\alpha)\) has a principal kernel generated by \(m_\alpha\). Moreover, multiplication by \(\alpha\) on the finite extension is a linear operator with the same annihilating relation. This is a structural bridge, but it does not erase the distinction: field theory studies algebraic elements and extensions, while this entry studies operators and their invariant/canonical-form consequences.

Beyond algebra, “the smallest rule that generates all valid constraints” is a portable analogy. Its legitimate cross-domain carriers are Invariance, Compression, and minimal sufficient representation. A software specification, legal constraint, or scientific model does not have a linear-algebra minimal polynomial unless it supplies an operator over a field and polynomial evaluation. Borrowing the name without those roles is metaphor.

The computation pattern also transfers component-wise to recurrence finding and Krylov algorithms. What travels is discovery of a shortest linear dependence and reduction by the resulting relation. The exact minimal polynomial remains an algebraic invariant, not a generic name for any concise recurrence.

Examples¶

Canonical: same characteristic polynomial, different minimal polynomial¶

Over \(\mathbb R\), compare

\[ A=\operatorname{diag}(2,2,3),\qquad B=\begin{pmatrix}2&1&0\\0&2&0\\0&0&3\end{pmatrix}. \]

Both have characteristic polynomial \((x-2)^2(x-3)\). Matrix \(A\) is diagonal, so \((A-2I)(A-3I)=0\) and neither linear factor alone annihilates it; thus \(m_A=(x-2)(x-3)\). Matrix \(B\) has a size-two Jordan block at eigenvalue \(2\). The product \((B-2I)(B-3I)\) leaves a nonzero nilpotent contribution, while \((B-2I)^2(B-3I)=0\); hence \(m_B=(x-2)^2(x-3)\).

Mapped back: both cases use the operator, polynomial functional calculus, monic least-degree generator, and characteristic-polynomial envelope. The changed exponent demonstrates the primary-depth code and why the minimal polynomial is invariant but not the characteristic polynomial.

Applied / in practice: compressing powers to a recurrence¶

Let

\[ Q=\begin{pmatrix}1&1\\1&0\end{pmatrix}. \]

Direct multiplication gives \(Q^2=Q+I\), so \(x^2-x-1\) annihilates \(Q\). No linear polynomial can annihilate this nonscalar matrix, hence \(m_Q=x^2-x-1\). Multiplying the relation by \(Q^{n-2}\) yields \(Q^n=Q^{n-1}+Q^{n-2}\). By induction,

\[ Q^n=F_nQ+F_{n-1}I \qquad(n\geq1), \]

where \(F_n\) is the Fibonacci sequence. The same quotient-ring reduction applies to any polynomial expression in \(Q\): divide by \(m_Q\), retain a linear remainder, and evaluate only \(I\) and \(Q\).

Mapped back: the annihilator ideal supplies the recurrence, the universal divisibility invariant proves all polynomial identities are multiples of the same relation, and the reduction rule compresses arbitrarily high powers to fewer than \(\deg m_Q\) independent powers.

Structural Tensions¶

T1: Minimal relation versus complete structure. The polynomial is maximally economical for annihilating relations but intentionally discards block counts and smaller block sizes. Diagnostic: Does the task ask which polynomial identities hold, or does it require the full similarity class?

T2: Coordinate-free invariant versus coordinate computation. The result is basis-independent, yet practical algorithms manipulate matrices, Krylov bases, or canonical forms. Poor coordinates can obscure a simple invariant without changing it. Diagnostic: Is observed complexity intrinsic to \(T\), or caused by the chosen representation and algorithm?

T3: Factorization clarity versus field dependence. The polynomial itself persists under scalar extension, but roots, Jordan interpretation, and diagonalizability over the base field depend on factorization there. Diagnostic: Is a conclusion about \(m_T\) as a polynomial, or about how it splits over a specified field?

T4: Exact algebra versus numerical sensitivity. Over exact fields the definition and divisibility tests are crisp. For floating-point matrices, tiny perturbations can change exact Jordan structure and the minimal polynomial dramatically. Diagnostic: Is the input exact algebraic data, or is a tolerance-dependent numerical surrogate being mistaken for an exact invariant?

T5: Global certainty versus local computation. A few vector-relative order polynomials may recover \(m_T\) efficiently, especially probabilistically over finite fields, but a local polynomial can miss a primary exponent present elsewhere. Diagnostic: What witness or verification proves that the least common multiple of sampled local relations is global?

T6: Autonomous operator invariant versus parent-prime reduction. Invariance and Compression describe why the object is useful across domains; the annihilator ideal, polynomial evaluation, and primary exponents make it specifically a minimal polynomial. Diagnostic: Does the reasoning require \(F[x]\) acting through an endomorphism, or only the generic idea of a compact invariant?

Structural–Framed Character¶

Minimal Polynomial (Linear Algebra) is structural-leaning. Its evaluative weight is neutral: “minimal” is an exact least-degree condition, not a preference for simplicity. The construction is not human-practice-bound; once a field, vector space, and operator are specified, the ideal and its monic generator follow independent of institutional convention. Its historical presentation is part of mathematics, but its definition is not created by a standard-setting body.

The vocabulary partly travels—annihilation, generator, divisibility, invariance, and compression have wider uses—but the operative meanings of \(F[x]\), \(p(T)\), invariant factors, and Jordan chains remain algebraic. Import versus recognition is therefore strict: one recognizes the same object in any finite-dimensional operator setting, while use outside that setting imports only an analogy.

The portable skeleton is invariant compression: replace a transformation-dependent family of relations by one canonical preserved generator. That skeleton belongs to Invariance and Compression. Its character: an unusually structural domain abstraction whose exact recognition still depends on linear operators and polynomial-module machinery.

Structural Core vs. Domain Accent¶

This section decides why the linear-algebra minimal polynomial is domain-specific rather than a prime.

What is skeletal. The thin portable structure is: collect all constraints satisfied by an object; prove they form a generated family; choose a normalized least generator; use it as an invariant and reduction rule. This recurs as a pattern in ideals, recurrence relations, specifications, coding, and quotient constructions. It expresses Invariance and Compression.

What is domain-bound. The defining data are a field \(F\), a finite-dimensional vector space, an endomorphism \(T\), polynomial evaluation \(F[x]\to\operatorname{End}(V)\), the annihilator ideal, monic normalization, and factor exponents tied to primary or Jordan structure. Divisibility by \(m_T\), Cayley–Hamilton, cyclic vectors, and reduction modulo \(m_T\) are not optional accents; they are the mechanism. Remove the operator or polynomial action and the object is no longer this minimal polynomial.

Why this does not clear the prime bar. A prime must be recognized by substantially the same roles across unrelated substrates. Here, literal recognition occurs broadly inside algebra, representation theory, computation, and linear systems because they all retain the operator-over-a-field substrate. Outside it, “minimal polynomial” either means the related field-theory construction or becomes metaphor for a concise rule. The cross-domain reach belongs to Invariance and Compression, while the named object remains an autonomous, reusable algebraic invariant.

The minimal polynomial instantiates Invariance. The preserved feature is \(m_T\), the transformations are changes of basis or similarity, and \(p(S^{-1}TS)=S^{-1}p(T)S\) proves preservation.

It is strongly related to Compression because one generator represents every annihilating polynomial and reduces all powers modulo a degree-\(d\) relation. Compression is explanatory but need not be a second DAG parent alongside Invariance.

Canonical Form is a close but rejected parent. A canonical form is a complete distinguished representative for an equivalence class. The minimal polynomial is canonical as an invariant but incomplete: equal minimal polynomials do not imply similar operators. Basis is involved in representation and cyclic computation, but the result is specifically basis-independent. The domain-specific Ring and Field (Algebraic) entries provide ambient algebra, while Kernel appears in local tests; none covers the operator-specific generator.

Relationships to Other Abstractions¶

Current abstraction Minimal Polynomial (Linear Algebra) Domain-specific

Parents (1) — more general patterns this builds on

Minimal Polynomial (Linear Algebra) is part of Invariance Prime

The minimal polynomial instantiates Invariance.

Hierarchy path (1) — routes to 1 parentless root

Minimal Polynomial (Linear Algebra) → Invariance

Neighborhood in Abstraction Space¶

Minimal Polynomial (Linear Algebra) sits in a sparse region of the domain-specific corpus (72^nd percentile for distinctiveness): few abstractions share its structure, so a faithful description tends to retrieve it precisely.

Family — Algebraic Geometry & Bundle Structure (14 abstractions)

Nearest neighbors

Computed from structural-signature embeddings · 2026-09-08

Not to Be Confused With¶

Characteristic polynomial. \(\chi_T=\det(xI-T)\) always has degree \(\dim V\) and records total algebraic multiplicities; \(m_T\) is the least annihilator and records maximum primary depths. Tell: Is the polynomial defined by a determinant or by the annihilator ideal?
An annihilating polynomial. Any multiple of \(m_T\) annihilates \(T\), so annihilation alone does not prove minimality. Tell: Does the polynomial divide every other annihilator?
Minimal polynomial of an algebraic element. Field theory takes the kernel of evaluation at \(\alpha\); linear algebra takes the kernel of evaluation at an operator. The regular representation links them, but they answer different object-level questions. Tell: Is the input an algebraic element/extension or an endomorphism?
Vector-relative minimal or order polynomial. \(m_{T,v}\) is the least polynomial with \(m_{T,v}(T)v=0\) and divides \(m_T\). Equality requires that the vector witness see every maximal primary exponent, as a cyclic vector does. Tell: Does the relation annihilate one generated subspace or all of \(V\)?
Invariant factors. Rational canonical form has a divisibility chain of invariant factors; the largest is \(m_T\), while their product is \(\chi_T\). The full list contains more similarity information. Tell: Is one largest factor being reported or the complete chain?
Jordan canonical form. Jordan form lists every block when the polynomial splits. The minimal polynomial retains only the largest block size for each eigenvalue. Tell: Can the data recover block counts and all sizes?
Canonical Form prime. A complete canonical form classifies equivalence by equality of representatives; \(m_T\) supplies only a necessary similarity test. Tell: Does equality of outputs imply equivalence, or only fail to rule it out?
Minimal Pairs prime. Minimal pairs isolate one contrasting feature in observations. No annihilator ideal or operator evaluation is involved. Tell: Is minimality about degree/divisibility or experimental contrast?

References¶

[1] Golub and Wilkinson. “Ill-Conditioned Eigensystems and the Computation of the Jordan Canonical Form”. SIAM Review, 1976. The classic account of why Jordan structure is discontinuous under arbitrarily small perturbation — in floating point one cannot even decide whether a matrix is defective — which is what makes the minimal polynomial's primary exponents an unstable numerical inference for near-defective matrices. registry ↩

[2] Serre, Jean-Pierre. Linear Representations of Finite Groups. Springer (Graduate Texts in Mathematics 42), 1977. Serre supplies both halves in their standard form: an element of finite order n is annihilated by x^n - 1, and complete reducibility holds away from the modular case, failing in characteristic p dividing the group order where Brauer's theory takes over; the repeated-factor reading of that failure is the article's own. registry ↩

[3] Neunhöffer and Praeger. “Computing Minimal Polynomials of Matrices”. LMS Journal of Computation and Mathematics, 2008. Gives an O(n^3) Monte Carlo algorithm and an accompanying deterministic verification procedure for minimal polynomials of matrices over finite fields, motivated by Frobenius normal form, F[x]-module structure and matrix group recognition. The paper's own mechanism: order polynomials of random vectors under the matrix action are combined by least common multiple to reach the minimal polynomial with high probability, and a deterministic algorithm then certifies or refutes the candidate. registry ↩a ↩b