Skip to content

Drazin inverse

A unique generalized inverse for a square matrix defined through its index, commutation, and power equations.

Version
v1 · 2026-09-28 · History
Domain-specific #
7628
Domain group
Formal Sciences
Origin domain
Mathematics
Subdomain
Matrix Theory → Mathematics

Core Idea

The Drazin inverse of a square matrix \(A\) is the unique matrix \(A^D\) that inverts the stable invertible part of \(A\) while annihilating its nilpotent part.[1] Let the index \(k\) of \(A\) be the least nonnegative integer for which \(\operatorname{rank}(A^{k+1})=\operatorname{rank}(A^k)\).[2] Then \(A^D\) is characterized by

\[ A^{k+1}A^D=A^k,\qquad A^DAA^D=A^D,\qquad AA^D=A^DA. \]

These three relations—index-dependent recovery, reflexivity, and commutation—are jointly constitutive.[3]

In a basis where \(A\) is block diagonal with an invertible block \(B\) and a nilpotent block \(N\), the operation is transparent: \(A^D\) replaces \(B\) by \(B^{-1}\) and \(N\) by zero.[4] Thus an invertible matrix has its ordinary inverse as its Drazin inverse, a nilpotent matrix has Drazin inverse zero, and the construction is preserved under conjugation.[5]

The Drazin inverse is not generally a classical generalized inverse satisfying \(AA^DA=A\). Its identity instead depends on the matrix index and the three displayed equations. A pseudoinverse or other matrix assigned by a different optimality or adjoint condition does not qualify merely because it agrees on a special case.

How would you explain it like I'm…

Undo What Can Be Undone

Imagine a machine that does two jobs: one part mixes things up in a way you can undo, and another part squashes things until they disappear. The Drazin inverse is a helper that undoes the first part and just lets the disappearing part stay gone. It doesn't pretend it can bring the squashed things back.

The Partial Undo Machine

Some math machines, called matrices, change lists of numbers in a way that has two parts mixed together. One part can be run backwards. The other part is special: if you run it over and over, after enough runs it always turns everything into zero. The Drazin inverse is a partner machine that runs the first part backwards and simply turns the second part into zero right away. So it is not a full undo button, because whatever the second part handled can't be brought back.

Invert the Core, Zero the Rest

A square matrix can be split, in a suitable choice of coordinates, into an invertible part and a nilpotent part, meaning a part that becomes the zero matrix after being multiplied by itself enough times. Ordinary inverses only exist when the whole matrix is invertible. The Drazin inverse A^D works for any square matrix: it inverts the invertible part and sends the nilpotent part to zero. So if A is invertible, A^D is just the usual inverse, and if A is nilpotent, A^D is zero. Unlike the familiar pseudoinverse, it is not built to satisfy A A^D A = A, and in general it does not; instead it is defined by rules that depend on the matrix's index and by the requirement that it commutes with A.

 

The Drazin inverse of a square matrix A is the unique matrix A^D that inverts the stable, invertible part of A and annihilates its nilpotent part. Define the index k of A as the smallest nonnegative integer with rank(A^(k+1)) = rank(A^k). Then A^D is characterized by three relations holding together: A^(k+1) A^D = A^k (index-dependent recovery), A^D A A^D = A^D (reflexivity), and A A^D = A^D A (commutation). In a basis where A is block diagonal with an invertible block B and a nilpotent block N, A^D simply replaces B by B^(-1) and N by zero. Consequences: an invertible matrix's Drazin inverse is its ordinary inverse, a nilpotent matrix's is zero, and the construction respects similarity (conjugation). Unlike classical generalized inverses, it generally does not satisfy A A^D A = A, and the Moore–Penrose pseudoinverse or other inverses chosen by optimality or adjoint conditions are different objects even where they happen to coincide.

Structural Signature

Sig role-phrases:

  • the square matrix — A supplies the possibly singular linear operator whose stable and nilpotent actions must be separated.
  • the matrix index — the least k ≥ 0 satisfying rank(A^(k+1)) = rank(A^k) records when the range sequence stabilizes.
  • the candidate inverse — a square matrix A^D of the same size is tested against the index-dependent algebraic contract.
  • the power-recovery identity — A^(k+1)A^D = A^k recovers the stabilized action selected by the index.
  • the reflexive identity — A^D A A^D = A^D constrains the candidate on its retained range.
  • the commutation identity — AA^D = A^D A preserves the operator's invariant decomposition.
  • the uniqueness guarantee — the three identities at the least index determine one Drazin inverse.
  • the block construction — an invertible block B is replaced by B^−1 while a nilpotent block N is replaced by zero.
  • the similarity guarantee — conjugating A and A^D by the same invertible change of basis preserves the Drazin relation.
  • the limiting branches — an invertible matrix returns its ordinary inverse, a nilpotent matrix returns zero, and low-index cases recover the corresponding group-inverse regime.
  • the generalized-inverse boundary — agreement on a special case or satisfaction of least-squares, adjoint, or classical AXA = A conditions does not replace the Drazin contract.
  • the computational limitation — algebraic uniqueness does not guarantee that a particular decomposition or numerical algorithm is well-conditioned.

What It Is Not

  • Not an ordinary inverse assigned despite singularity. An ordinary inverse requires (A{-1}A=AA=I), which a singular matrix cannot satisfy. The Drazin inverse instead recovers the stable invertible action and sends the nilpotent action to zero under its index-dependent contract.
  • Not just any generalized inverse. A matrix qualifies only when it satisfies the power-recovery, reflexive, and commutation identities at the matrix's least index. Producing some inverse-like matrix for a singular (A) is not enough.
  • Not the Moore–Penrose pseudoinverse. The Moore–Penrose construction is selected by adjoint and least-squares conditions; the Drazin inverse is selected by commutation and index-dependent algebraic identities. Their agreement on a special matrix does not make the definitions equivalent.
  • Not required to satisfy (AA^DA=A) in every case. That classical generalized-inverse equation can fail for a legitimate Drazin inverse of index greater than one; imposing it universally substitutes a stronger, different contract.
  • Not defined by an arbitrary exponent that makes one equation work. The exponent (k) is the least point at which the ranks of consecutive powers stabilize, and the other two defining identities must hold as well.
  • Not a lossless reversal of all of (A). The nilpotent block is annihilated rather than reconstructed, so transient action in that block is deliberately excluded from what (A^D) retains.
  • Not a particular decomposition or numerical algorithm. Jordan or block form supplies a construction and explanation, but the Drazin identity is basis-independent and remains governed by the three equations; computational conditioning is a separate issue.

Scope of Application

The Drazin inverse applies to square matrices whose rank-stabilization index and three defining identities are meaningful, with extensions to compatible algebraic settings only when the same index-dependent contract is explicitly defined.

  • Singular square matrices — the construction separates stable invertible action from nilpotent action when no ordinary inverse exists for the full matrix.
  • Invertible matrices — index-zero cases recover the ordinary matrix inverse, providing a limiting regime of the same definition rather than a separate operation.
  • Nilpotent matrices — matrices whose powers eventually vanish have Drazin inverse zero, with the nilpotency depth reflected in the matrix index.
  • Block and Jordan decompositions — an invertible block can be replaced by its inverse and a nilpotent block by zero, then transported back through a similarity transformation.
  • Projection and low-index cases — a projection is its own Drazin inverse, while the appropriate index-zero or index-one regime yields the associated group-inverse specialization.
  • Matrices over perfect fields — a Jordan–Chevalley decomposition can define the operation by inverting the stable component on its supported subspace and annihilating the nilpotent component.
  • Iterative computation — hyper-power sequences and other algorithms can approximate the Drazin inverse when their commutation and convergence conditions hold, without altering the algebraic identity being computed.
  • Categorical generalizations — morphisms in categories equipped with a suitably defined Drazin structure can instantiate an extension of the matrix contract; an arbitrary generalized inverse or inverse-like morphism does not qualify without those identities.

Clarity

Naming the Drazin inverse identifies a particular algebraic contract, not a generic license to assign an inverse-like matrix to a singular matrix. The matrix index fixes the power equation, while reflexivity and commutation fix how the candidate interacts with the original matrix. All three conditions matter: agreement with an ordinary inverse on invertible matrices or with zero on a nilpotent example does not by itself identify the Drazin inverse.

The term lets a matrix theorist ask: What is the index of this matrix, and does the proposed inverse satisfy each defining equation at that index? A block decomposition then makes the answer legible: invert the stable invertible block and annihilate the nilpotent block. This also sharpens the distinction from the Moore–Penrose inverse, whose adjoint and least-squares conditions answer a different problem, and from a classical generalized inverse, which normally requires the equation \(AXA=A\) that the Drazin inverse need not satisfy.

Manages Complexity

A singular matrix can mix invertible dynamics with chains that eventually collapse to zero, making an ordinary inverse impossible and entrywise case analysis opaque. The Drazin construction compresses this behavior into the matrix index k and three algebraic identities. The rank-stabilization condition locates the transient nilpotent depth; commutation preserves the matrix's invariant decomposition; and the power and reflexivity equations specify recovery on the stable component.

In block or Jordan form, the outcome branches are immediate: invertible blocks are replaced by their inverses, nilpotent blocks by zero, an invertible matrix returns its ordinary inverse, and a nilpotent matrix returns zero. Conjugation invariance allows the calculation to be performed in a convenient basis and transported back. Index zero and one cases expose the ordinary-inverse and group-inverse regimes without changing the general contract.

The compression does not preserve the transient action inside the annihilated nilpotent block, guarantee numerical stability of a chosen algorithm, or make the result satisfy every generalized-inverse convention. It isolates the stable invertible action selected by the index and defining equations, while basis computation, conditioning, and comparison with Moore–Penrose or other inverses remain separate questions.

Abstract Reasoning

An index diagnostic runs from the rank sequence rank(A), rank(A^2), … to the least k at which rank(A^(k+1)) = rank(A^k). That stabilization identifies how many powers are needed before the nilpotent transient stops shrinking the range, and it fixes the power equation a proposed Drazin inverse must satisfy. Choosing k merely because one displayed equality happens to hold is insufficient; minimality is part of the diagnosis.

A decomposition-to-inverse move runs from a similarity basis separating an invertible block B and a nilpotent block N to A^D: replace B by B^−1, replace N by zero, and conjugate back. This predicts the limiting cases—ordinary inverse for invertible A and zero for nilpotent A—and explains why basis change preserves the construction. The resulting candidate is certified only after checking commutation, reflexivity, and A^(k+1) A^D = A^k together.

An intervention-and-boundary move asks how the result changes when the stable and nilpotent parts change. Altering an eigenvalue inside the invertible block changes its reciprocal contribution; lengthening a nilpotent chain can raise the index without creating an invertible contribution. Replacing the Drazin contract with adjoint or least-squares conditions selects a different generalized inverse, and requiring A A^D A = A in every case wrongly excludes legitimate higher-index matrices. Thus agreement on an invertible or projection example does not establish identity outside that regime.

Knowledge Transfer

Within matrix theory, Drazin inversion transfers literally across matrix sizes, similarity bases, fields supporting the needed decomposition, and index regimes. The rank-stabilization definition of k, the three defining identities, and the block operation—ordinary inversion on the invertible part and zero on the nilpotent part—carry together. Conjugation lets a computation in Jordan or block form be transported back, while the invertible, nilpotent, projection, and group-inverse cases provide reusable diagnostics rather than unrelated exceptions.

The evidence also supports (B) a shared abstract mechanism under inversion beyond matrices: suitably defined categorical morphisms can have Drazin inverses that preserve the same stable-versus-nilpotent separation. What transfers is the index-dependent algebraic contract, not merely the idea of an inverse-like object. Matrix rank, Jordan blocks, and a chosen numerical algorithm remain home-bound to the matrix realization. Treating a Moore–Penrose inverse, least-squares solution, or any singular-matrix workaround as “the Drazin inverse” is only (A) analogy unless commutation, reflexivity, and the power equation hold at the least index. Transfer stops when a different generalized-inverse convention replaces those identities or when agreement occurs only on a special case.

Examples

Canonical

Take A = [[2,0,0],[0,0,1],[0,0,0]]. Its upper-left block [2] is invertible and its lower-right block [[0,1],[0,0]] is nilpotent of order two. The ranks of A, A², and A³ are respectively 2, 1, and 1, so the least stabilization index is k = 2.[6] Replacing the invertible block by [1/2] and the nilpotent block by zero gives A^D = diag(1/2,0,0).[7] Direct calculation yields A³A^D = A², A^D A A^D = A^D, and AA^D = A^D A; the three equations certify the result.[8]

Mapped back: the displayed A is the square matrix, k = 2 is the matrix index, and diag(1/2,0,0) is the candidate inverse. The three calculations instantiate the power-recovery identity, the reflexive identity, and the commutation identity, so the uniqueness guarantee identifies diag(1/2,0,0) as the Drazin inverse. Separating [2] from the nilpotent shift uses the block construction, which also shows why the transient nilpotent action is annihilated.

Applied / In Practice

As an in-domain diagnostic, let P = [[1,0],[0,0]], a singular projection. Because P² = P, consecutive ranks stabilize at k = 1.[9] Testing the natural candidate P^D = P gives P²P = P, PPP = P, and PP = PP, so the Drazin contract holds even though no ordinary inverse exists.[10] This small case is useful for checking symbolic or numerical implementations: returning the projection itself is required, whereas returning an arbitrary generalized inverse that merely satisfies PXP = P does not establish the full Drazin identity.

Mapped back: P supplies the square matrix, rank stabilization at one supplies the matrix index, and P itself is the candidate inverse. Idempotence verifies the power-recovery identity and the reflexive identity, while multiplication verifies the commutation identity. The result occupies the limiting branches for a projection and enforces the generalized-inverse boundary: a single classical PXP = P check cannot replace the three-condition contract.

Structural Tensions

T1: Stable recovery versus nilpotent information loss. The Drazin inverse exactly inverts the stable invertible block while sending the nilpotent block to zero, making singular dynamics tractable. That recovery is necessarily selective: transient action inside the nilpotent chains is not reversed. Diagnostic: decompose the operator into stable and nilpotent parts and state explicitly which behavior the proposed inverse preserves and which it annihilates.

T2: Minimal index versus equation satisfaction. The power-recovery identity depends on an exponent k, but the Drazin index is the least rank-stabilization point, not any larger exponent that happens to make an equality true. Minimality captures the nilpotent depth; ignoring it weakens the classification. Diagnostic: compute consecutive ranks to locate the first stabilization before testing the three defining identities.

T3: Algebraic uniqueness versus numerical conditioning. The index, power recovery, reflexivity, and commutation determine one mathematical matrix, yet a Jordan decomposition or iterative algorithm can be unstable or inaccurate on a concrete input. Uniqueness does not certify a computation; numerical difficulty does not alter the algebraic object. Diagnostic: separate residual checks for the defining equations from conditioning or convergence evidence for the chosen computational method.

T4: Shared special cases versus distinct generalized inverses. Ordinary inverses, group inverses, projections, and some pseudoinverses can agree with the Drazin inverse in restricted regimes, while their general defining contracts differ. Agreement on a convenient example aids verification but can encourage false equivalence. Diagnostic: test commutation, reflexivity, and index-dependent power recovery rather than identifying the construction from a special-case numerical match.

T5: Drazin-inverse autonomy versus reduction to Inversion. Every qualifying Drazin inverse is a strict algebraic specialization of the exact parent Prime Inversion (Inversion): a rule reverses the stable part of a square operator while preserving its invariant decomposition and producing an inverse-like result. Reduction preserves that reversal structure, but loses the least rank-stabilization index, power-recovery, reflexive and commutation identities, and the deliberate zeroing of the nilpotent block. Treating it as wholly autonomous hides its inversion contract; treating any generalized inverse as Drazin erases the defining equations.
Diagnostic: Is there merely an inverse-like reversal, or does the operator satisfy the Drazin index-dependent identities and stable-versus-nilpotent block construction?

Structural–Framed Character

Drazin Inverse is structural-leaning. Its identity is fixed by an exact algebraic contract that survives basis change, while its domain accent specifies square operators, rank stabilization, and the stable-versus-nilpotent decomposition.

Its evaluative_weight is low: a matrix either satisfies the defining identities or does not, independently of whether computing the inverse is useful or well-conditioned. Its human_practice_bound is low because notation and algorithms are practices but the algebraic relation does not depend on a practitioner once the matrix and field are fixed. Its institutional_origin is low; naming and publication history stabilize terminology without constituting the object. Its vocab_travels is moderate because a compatible Drazin notion can extend beyond matrices only when the index-dependent contract is carried intact, not whenever something is merely inverse-like. Its import_vs_recognize balance strongly favors recognition: the least stabilization index and three equations determine the unique candidate rather than importing an optional interpretive frame.

The smallest reviewed portable skeleton is Inversion (Inversion). The original operator, reversal operation, preserved stable action, inverse-like result, and collapse when stable-part reversal is removed fill that Prime's signature; the Drazin construction specializes it by annihilating the nilpotent part rather than recovering every action. That portable reach belongs to the Inversion Prime. The least index, power-recovery, reflexive, and commutation equations remain the matrix-theoretic accent owned by Drazin Inverse.

Its character: structural-leaning because an exact, basis-independent inversion relation governs identity while the singular-operator contract sharply restricts its literal domain.

Structural Core vs. Domain Accent

Drazin Inverse remains domain-specific rather than a Prime because its portable reversal structure is constituted by an index-dependent matrix contract for separating stable invertible action from nilpotent action.

What is skeletal (could lift toward a cross-domain prime). The complete thin skeleton is an original structure, a reversal operation, an invariant retained through reversal, an inverse-like result, and a failure test showing when the reversal contract is absent. Drazin Inverse strictly instantiates Inversion: the original structure is a square operator, the operation reverses its stable invertible block, the invariant decomposition is preserved, and the result recovers the reversible action while making the nilpotent action zero. Remove stable-part reversal and what remains is not a Drazin inverse even if some generalized-inverse equation holds.

What is domain-bound. The matrix-theoretic accent comprises the least rank-stabilization index, the power-recovery equation, reflexivity, commutation, the stable-versus-nilpotent invariant decomposition, similarity equivariance, and the limiting invertible, nilpotent, projection, and group-inverse regimes. Those conditions determine one matrix and distinguish it from Moore–Penrose and other generalized inverses.

Why this does not clear the prime bar. The complete signature of a square operator, least rank-stabilization index, three Drazin identities, and nilpotent-block annihilation does not recur literally across three unrelated domains—Bayesian reversal, software dependency inversion, and rhetorical inversion. Those unrelated domains can preserve an original relation, reversal operation, retained equivalence, and inverse-like result, but they do not thereby possess a Drazin inverse; its portable reach therefore belongs to Inversion. Remove the matrix accent and the residue is structure-preserving reversal, not this candidate. Preserve the specialist vocabulary of index, stable block, and nilpotent block but remove reversal of the stable action, and the residue is a decomposition of an operator rather than Drazin inversion.

This entry is a kind of Inversion.

Instantiates — Inversion (Inversion). The original structure is the square operator (A), including its stable invertible action and nilpotent action. The Drazin operation uses the least rank-stabilization index and the power-recovery, reflexive, and commutation identities to produce (A^D); on an invariant block decomposition it replaces the invertible block by its ordinary inverse and the nilpotent block by zero. What is preserved is the stable range, the operator's invariant decomposition, and similarity-equivariant algebraic structure, rather than every transient action of (A). The resulting inverse recovers the reversible part of a singular operator and makes its limiting invertible, nilpotent, projection, and group-inverse regimes legible. A positive test identifies the original operator, the index-dependent reversal rule, the preserved stable structure, the inverse-like result, and the algebraic payoff. A collapse test leaves only a generic generalized inverse chosen by least-squares or adjoint conditions, an arbitrary matrix satisfying one equation, or zeroing without stable-part inversion. Replacing matrix-specific terms by those typed roles preserves Inversion's full signature, whereas deleting stable-block reversal destroys the Drazin identity even if commutation survives. The Drazin inverse is therefore a strict specialization of Inversion whose domain accent is singular square matrices, rank stabilization, and deliberate annihilation of the nilpotent component.

Relationships to Other Abstractions

Local relationship map for Drazin inverseParents appear above the current abstraction, mutual partners to the right, and children below. Node labels state whether each abstraction is prime or domain-specific; colors identify relation types.Drazin inverseDOMAINPrime abstraction: Inversion — is a kind ofInversionPRIME

Current abstraction Drazin inverse Domain-specific

Parents (1) — more general patterns this builds on

  • Drazin inverse is a kind of Inversion Prime

    The original structure is the square operator (A), including its stable invertible action and nilpotent action.

Hierarchy paths (3) — routes to 3 parentless roots

Neighborhood in Abstraction Space

Drazin inverse sits in a sparse region of the domain-specific corpus (65th percentile for distinctiveness): few abstractions share its structure, so a faithful description tends to retrieve it precisely.

Family — Unclustered & Miscellaneous (2551 abstractions)

Nearest neighbors

Computed from structural-signature embeddings · 2026-10-08

Not to Be Confused With

  • Ordinary matrix inverse. The ordinary inverse reverses the whole action of an invertible matrix and satisfies (A{-1}A=AA=I); the Drazin inverse remains defined for singular matrices by inverting only the stable invertible block. Tell: test invertibility and the identity equations rather than treating singular reversal as ordinary inversion.
  • Generalized inverse. A generalized inverse is the broader class of inverse-like matrices selected by various algebraic conditions, while the Drazin inverse is uniquely fixed by power recovery, reflexivity, and commutation at the matrix's least index. Tell: verify all three Drazin identities and the rank-stabilization index rather than one generic inverse equation.
  • Moore–Penrose pseudoinverse. The Moore–Penrose pseudoinverse is selected by adjoint symmetry and least-squares conditions; the Drazin inverse is selected by commutation and index-dependent algebra. Tell: inspect the defining equations used to choose the matrix, especially adjoint conditions versus (A{k+1}AD=A^k).
  • Group inverse. The group inverse is the index-one specialization of the Drazin inverse, not a coextensive name for every Drazin case. Tell: compute the least rank-stabilization index; index one permits the group-inverse label, while higher index remains Drazin-only.
  • Jordan decomposition. A Jordan or invertible–nilpotent block decomposition is a construction used to calculate and explain (A^D), not the inverse itself. Tell: distinguish the change of basis and block structure from the resulting matrix that replaces the invertible block by its inverse and the nilpotent block by zero.

References

[1] Thomas Butler and Marc C. Wilson, On the Drazin Inverse of the Rate Matrix, Linear Algebra and its Applications 434 (2011) (accessed 2026-09-13). registry ↩

[2] Unverified encyclopedia synthesis; no authoritative source located for the claim as written. ↩

[3] Unverified encyclopedia synthesis; no authoritative source located for the claim as written. ↩

[4] Unverified encyclopedia synthesis; no authoritative source located for the claim as written. ↩

[5] Unverified encyclopedia synthesis; no authoritative source located for the claim as written. ↩

[6] Unverified encyclopedia synthesis; no authoritative source located for the claim as written. ↩

[7] Unverified encyclopedia synthesis; no authoritative source located for the claim as written. ↩

[8] Unverified encyclopedia synthesis; no authoritative source located for the claim as written. ↩

[9] Unverified encyclopedia synthesis; no authoritative source located for the claim as written. ↩

[10] Unverified encyclopedia synthesis; no authoritative source located for the claim as written. ↩