The Moore–Penrose Inverse in an Involutive Algebra
Introduction
The inverse of an element asks for $a^{-1}$ with $aa^{-1}=a^{-1}a=1$, and exists only for the invertibles. The Moore–Penrose inverse weakens the request. It asks for an element that inverts $a$ only on the part of the algebra where $a$ acts, and that is self-adjoint in the two products it forms with $a$. The definition uses nothing but the product and the involution, so it belongs to the algebra and not to a matrix representation.
The article proves four things. The four equations determine $a^{+}$ when they have a solution. The two products $aa^{+}$ and $a^{+}a$ are the projections that generate the right ideal $aA$ and the left ideal $Aa$. A solution exists exactly when those two ideals are generated by projections and $a$ is injective on the second, a criterion that can fail. And the inverse obeys the identities of the matrix case, $(a^{+})^{+}=a$, $(a^{*})^{+}=(a^{+})^{*}$ and $(aa^{*})^{+}=(a^{+})^{*}a^{+}$, with the same proofs.
The article is purely algebraic: no topology, no norm, no positivity and no matrix. Existence is not automatic in an arbitrary involutive algebra, and the criterion says exactly when it holds. The matrix algebras are the model case in which it always does; they are treated in The Moore–Penrose Inverse of a Matrix, where the two projections are the projections onto the column and the row space, and where the rank factorisation gives the inverse in closed form. The biquaternion algebra is the case needed by the polar decomposition, and the inverse of its Hermitian factor is computed from a quadratic identity in The Polar Decomposition of a Biquaternion, §The Moore–Penrose Inverse and §The Moore–Penrose Inverse of the Hermitian Factor.
Throughout, $A$ is a unital associative algebra over a field $F$, with an involution $a\mapsto a^{*}$: an anti-automorphism of order two, $(ab)^{*}=b^{*}a^{*}$ and $a^{**}=a$. An element $p$ is a projection when $p^{2}=p=p^{*}$. The right ideal generated by $a$ is $aA=\{ax:x\in A\}$ and the left ideal is $Aa=\{xa:x\in A\}$; the two are computed in Ideals and Quotients of Algebras and Left and Right Multiplication. The star here is the involution on the elements, as in Involutive Algebras; no operator adjoint is used in this article.
The Four Conditions
Definition. A Moore–Penrose inverse of $a\in A$ is an element $a^{+}\in A$ satisfying the four Penrose equations
$$ \boxed{\; aa^{+}a=a,\qquad a^{+}aa^{+}=a^{+},\qquad \bigl(aa^{+}\bigr)^{*}=aa^{+},\qquad \bigl(a^{+}a\bigr)^{*}=a^{+}a . \;} $$
The first two say that $a^{+}$ inverts $a$ on the part of the algebra where $a$ is not zero. The last two say that the two products are self-adjoint, hence that they are projections, and the four together say that $a^{+}$ is the inverse of $a$ read between those two projections.
Lemma (the two projections). If $a^{+}$ exists, then $p=aa^{+}$ and $q=a^{+}a$ are projections, and
$$ aA=pA,\qquad Aa=Aq,\qquad a=pa=aq,\qquad a^{+}=qa^{+}p . $$
Proof. The first equation of the definition gives $a=aa^{+}a=pa$, and the second gives $p^{2}=aa^{+}aa^{+}=a(a^{+}aa^{+})=aa^{+}=p$; the third gives $p^{*}=p$. Hence $p$ is a projection and $a=pa\in pA$, so $aA\subseteq pA$; conversely $p=aa^{+}\in aA$, so $pA\subseteq aA$ and the two right ideals are equal. The same computation with $a$ and $a^{+}$ exchanged gives $q^{2}=q=q^{*}$ and $Aa=Aq$. Then $a^{+}=a^{+}aa^{+}=qa^{+}$ and $a^{+}=a^{+}aa^{+}=a^{+}p$, so $a^{+}=qa^{+}p$.
Remark (what the inverse reads). Let $\ker a=\{x:ax=0\}$ and $\ker a^{+}=\{x:a^{+}x=0\}$, right ideals. The lemma gives
$$ \ker a=(1-q)A,\qquad \ker a^{+}=(1-p)A, $$
because $ax=0$ is equivalent to $qx=a^{+}ax=0$, and symmetrically. The element $a$ maps $qA$ bijectively onto $pA$, with inverse $a^{+}$, and $a^{+}$ is zero on the complementary ideal $(1-p)A$. The four conditions are therefore a statement about one isomorphism of right ideals and two complementary projections, and nothing more.
Uniqueness
Lemma (the projection of an ideal). Let $p$ and $q$ be projections. If $pA=qA$, then $p=q$.
Proof. From $pA=qA$ one has $p\in qA$, so $p=qx$ for some $x$, and $qp=q^{2}x=qx=p$. Symmetrically $q\in pA$ gives $pq=q$. Starring the first identity, $p=p^{*}=(qp)^{*}=p^{*}q^{*}=pq=q$.
Theorem (uniqueness). An element of $A$ has at most one Moore–Penrose inverse.
Proof. Let $a^{+}$ and $b$ both satisfy the four conditions for $a$. By the lemma of §The Four Conditions, the products $p=aa^{+}$ and $q=a^{+}a$ generate $aA$ and $Aa$, and so do $ab$ and $ba$. The lemma above gives $ab=p$ and $ba=q$. Therefore
$$ b=bab=b\bigl(aa^{+}a\bigr)b=(ba)a^{+}(ab)=qa^{+}p=a^{+}, $$
the last equality by the lemma of §The Four Conditions.
The uniqueness is algebraic and uses neither a basis nor a norm. It also shows that the two projections are part of the data of $a$ alone: they are the projections that generate its two ideals, and no choice is made.
Existence
Existence is a condition on the two ideals, and it can fail.
Theorem (the criterion). The element $a$ has a Moore–Penrose inverse exactly when there are projections $p$ and $q$ with
$$ aA=pA,\qquad Aa=Aq, $$
and $a$ injective on $qA$. In that case $p=aa^{+}$ and $q=a^{+}a$.
Proof. If $a^{+}$ exists, the lemma of §The Four Conditions gives the two ideal identities and the projections, and $a^{+}a=q$ gives the injectivity: if $x=qx\in qA$ and $ax=0$, then $x=qx=a^{+}ax=0$.
Conversely, suppose the two projections and the injectivity are given. Since $a=aq$, the restriction $a:qA\to pA$ is surjective, hence bijective; let $f:pA\to qA$ be its inverse. For $x\in pA$ and $r\in A$ one has $a\bigl(f(x)r\bigr)=xr$, and $f(x)r\in qA$, so $f(xr)=f(x)r$ by the uniqueness of the preimage. Thus $f$ is a map of right $A$-modules, and $f(x)=f(px)=f(p)x$; put $b=f(p)\in qA$. Then $f$ is left multiplication by $b$, and $b=f(p)=f(p^{2})=f(p)p=bp$, so $b=qbp$. Extended by zero on $(1-p)A$, the same element $b$ equals $f$ on $pA$ and vanishes on $(1-p)A$, because $b(1-p)=qbp(1-p)=0$.
Now $ab=p$: on $pA$, $abx=af(x)=x$, and on $(1-p)A$, $bx=0$ while $a(1-p)=a-ap=0$. Similarly $ba=q$: on $qA$, $bax=x$, and on $(1-q)A$, $ax=0$ while $b(1-q)=0$. Hence
$$ aba=pa=a,\qquad bab=qb=b, $$
and the two products $ab=p$ and $ba=q$ are projections, so all four Penrose equations hold.
The proof also shows what the inverse is: the inverse of the isomorphism $a:qA\to pA$, extended by zero. Nothing else enters, and no positivity is used.
Corollary (the invertibles). If $a$ is invertible, then $a^{+}=a^{-1}$, with $p=q=1$.
Proof. Take $p=q=1$; the restriction $a:A\to A$ is bijective.
Corollary (the projections). If $p$ is a projection, then $p^{+}=p$.
Proof. Take $p=q$ in the criterion; $p$ is the identity on $pA$.
Corollary (matrix algebras). In $M_{n}(\mathbb{C})$ with the conjugate transpose, every element has a Moore–Penrose inverse.
Proof. Let $p$ be the orthogonal projection onto the column space of $a$ and $q$ the orthogonal projection onto the row space. Both are projections of the algebra, $aA=pA$ and $Aa=Aq$, and $\ker a$ is the orthogonal complement of the row space, so $a$ is injective on $qA$. The criterion applies; the closed form is in The Moore–Penrose Inverse of a Matrix, which also treats the rectangular case, where the inverse leaves the algebra of square matrices.
Remark (existence is not automatic). Let $A=\mathbb{C}[x]/(x^{2})$ with the identity involution, and $a=x$. Then $a^{2}=0$, and for every $b=u+vx$,
$$ aba=x(u+vx)x=ux^{2}+vx^{3}=0\neq x, $$
so the first Penrose equation has no solution. The ideal $aA=(x)$ is not generated by a projection: the only projections are $0$ and $1$. The criterion fails at the nilpotent, and the element has no Moore–Penrose inverse in the algebra. The failure is not a defect of the definition but a property of the algebra: the involution is degenerate, and the ideal generated by $x$ has no self-adjoint idempotent.
The Algebra of the Inverse
Assume throughout that $a^{+}$ exists. The inverse obeys the identities of the matrix case, and the proofs are the uniqueness theorem applied to the four equations.
Proposition (the elementary identities).
$$ \boxed{\;(a^{+})^{+}=a,\qquad (a^{*})^{+}=(a^{+})^{*},\qquad (aa^{*})^{+}=(a^{+})^{*}a^{+},\qquad (a^{*}a)^{+}=a^{+}(a^{+})^{*},\;} $$
and
$$ a^{+}=(a^{*}a)^{+}a^{*}=a^{*}(aa^{*})^{+}. $$
Proof. The four equations are symmetric under the exchange of $a$ and $a^{+}$: the first and the second are exchanged, and the third and the fourth are unchanged. Hence $a$ satisfies the equations that define $(a^{+})^{+}$, and uniqueness gives $(a^{+})^{+}=a$.
For the second identity, the candidate $(a^{+})^{*}$ satisfies the four equations for $a^{*}$:
$$ a^{*}(a^{+})^{*}a^{*}=(aa^{+}a)^{*}=a^{*},\qquad (a^{+})^{*}a^{*}(a^{+})^{*}=(a^{+}aa^{+})^{*}=(a^{+})^{*}, $$
while $a^{*}(a^{+})^{*}=(a^{+}a)^{*}=a^{+}a$ and $(a^{+})^{*}a^{*}=(aa^{+})^{*}=aa^{+}$ are self-adjoint. Uniqueness gives $(a^{*})^{+}=(a^{+})^{*}$.
For the third, put $b=(a^{+})^{*}a^{+}$ and $x=aa^{*}$, and use $a^{+}a=a^{*}(a^{+})^{*}$ and $(a^{+})^{*}a^{*}=aa^{+}$. Then
$$ xbx=a\bigl(a^{*}(a^{+})^{*}\bigr)a^{+}aa^{*}=a\,(a^{+}a)\,a^{+}aa^{*}=aa^{+}aa^{*}=aa^{*}=x, $$
$$ bxb=(a^{+})^{*}\bigl(a^{+}aa^{*}(a^{+})^{*}\bigr)a^{+}=(a^{+})^{*}(a^{+}a)(a^{+}a)^{*}a^{+}=(a^{+})^{*}a^{+}=b, $$
and $xb=aa^{*}(a^{+})^{*}a^{+}=a(a^{+}a)a^{+}=aa^{+}$ and $bx=(a^{+})^{*}(a^{+}a)a^{*}=((a^{+})^{*}a^{*})^{2}=(aa^{+})^{2}=aa^{+}$ are projections. Hence $b=x^{+}$. The fourth identity follows from the third applied to $a^{*}$, with the second used to remove the stars. Finally $(a^{*}a)^{+}a^{*}=a^{+}(a^{+})^{*}a^{*}=a^{+}aa^{+}=a^{+}$, and symmetrically $a^{*}(aa^{*})^{+}=a^{+}$.
Proposition (unitary factors). If $u$ is unitary, $u^{*}u=uu^{*}=1$, then
$$ (ua)^{+}=a^{+}u^{*},\qquad (au)^{+}=u^{*}a^{+}. $$
Proof. For the first, the candidate $a^{+}u^{*}$ satisfies the four equations for $ua$: the two cancellation equations are
$$ ua\,a^{+}u^{*}\,ua=uaa^{+}a=ua,\qquad a^{+}u^{*}\,ua\,a^{+}u^{*}=a^{+}aa^{+}u^{*}=a^{+}u^{*}, $$
and the two products are $uaa^{+}u^{*}$, which is the conjugate of the projection $aa^{+}$ by the unitary $u$, and $a^{+}a$, which is a projection. The second identity is the first applied to $a^{*}$ and conjugated.
Remark (the product rule fails). The identity $(ab)^{+}=b^{+}a^{+}$ does not hold in general. In $M_{2}(\mathbb{C})$, take
$$ a=\begin{pmatrix}1&1\end{pmatrix},\qquad b=\begin{pmatrix}1\\0\end{pmatrix}. $$
Then $ab=1$, so $(ab)^{+}=1$, while $b^{+}=\begin{pmatrix}1&0\end{pmatrix}$ and $a^{+}=\tfrac12\begin{pmatrix}1\\1\end{pmatrix}$, hence $b^{+}a^{+}=\tfrac12$. The product rule fails by the factor two. It holds when one factor is unitary, by the proposition above, but in general the inverse of a product is not the product of the inverses.
Examples
The strata of the algebra behave differently, and the examples below show the difference.
The invertibles. For invertible $a$ the inverse is $a^{-1}$, and $p=q=1$. The Moore–Penrose inverse agrees with the algebra inverse exactly here, and on no other stratum.
The projections. For a projection $p$ the inverse is $p$ itself, and $p^{+}=p$ is the identity on the ideal it generates. The four equations reduce to $p^{2}=p=p^{*}$.
The zero divisors. Let $A=M_{2}(\mathbb{C})$ and $a=E_{12}$. Then
$$ E_{12}^{+}=E_{21},\qquad E_{12}E_{12}^{+}=E_{11},\qquad E_{12}^{+}E_{12}=E_{22}, $$
as the four equations check at once. The element is a zero divisor, its inverse does not exist, and its Moore–Penrose inverse does. The two projections are different, $E_{11}\neq E_{22}$, so $aa^{+}\neq a^{+}a$; the two projections of a zero divisor need not coincide.
The rank-one elements. For nonzero columns $u\in\mathbb{C}^{m}$ and $v\in\mathbb{C}^{n}$,
$$ (uv^{*})^{+}=\frac{vu^{*}}{\lVert u\rVert^{2}\lVert v\rVert^{2}}. $$
The verification is the four equations, with $uv^{*}vu^{*}=\lVert v\rVert^{2}uv^{*}$ and the two products the projections onto the lines of $u$ and $v$.
A table. The examples collect as follows, with $A=M_{2}(\mathbb{C})$.
| element | $a^{+}$ | $aa^{+}$ | $a^{+}a$ |
|---|---|---|---|
| invertible $a$ | $a^{-1}$ | $1$ | $1$ |
| projection $p$ | $p$ | $p$ | $p$ |
| $E_{12}$ | $E_{21}$ | $E_{11}$ | $E_{22}$ |
| $uv^{*}$ | $vu^{*}/(\lVert u\rVert^{2}\lVert v\rVert^{2})$ | $uu^{*}/\lVert u\rVert^{2}$ | $vv^{*}/\lVert v\rVert^{2}$ |
The failure. In $A=\mathbb{C}[x]/(x^{2})$ with the identity involution the element $x$ has no Moore–Penrose inverse, by the remark of §Existence. The same algebra shows that the criterion is not a formality: the ideal $(x)$ is a one-dimensional ideal generated by an element of square zero, and no projection generates it.
The Group Inverse
A second generalized inverse is defined by replacing the two self-adjointness conditions with a commutation condition.
Definition. A group inverse of $a$ is an element $a^{\#}$ with
$$ aa^{\#}a=a,\qquad a^{\#}aa^{\#}=a^{\#},\qquad aa^{\#}=a^{\#}a . $$
In a finite-dimensional algebra it exists exactly when $a$ and $a^{2}$ generate the same two ideals, $aA=a^{2}A$ and $Aa=Aa^{2}$, that is when $a$ has index one; it is then unique. The Drazin inverse $a^{D}$ relaxes the first equation to $a^{k}a^{D}a=a^{k}$ for some $k\ge1$ and keeps the other two; at index one $k=1$ and $a^{D}=a^{\#}$.
Remark (which inverse answers which question). The group inverse inverts $a$ on the ideal $aA$ and commutes with $a$. The Moore–Penrose inverse inverts $a$ on the projection $qA$ and is self-adjoint in its two products. When both exist they agree exactly when the two projections coincide,
$$ a^{+}=a^{\#}\iff aa^{+}=a^{+}a . $$
For a matrix the condition is the equality of the column space of $a$ and the column space of $a^{*}$, the EP condition; it holds for every Hermitian matrix of index one and fails for $E_{12}$, which has a Moore–Penrose inverse and no group inverse, since $E_{12}^{2}=0$ while $E_{12}A\neq0$. The two inverses are not competitors: one is a spectral inverse on the image, the other a projection inverse on the two ideals.
Summary
The Moore–Penrose inverse of an element of an involutive algebra is the unique solution of the four Penrose equations, when a solution exists.
| item | result |
|---|---|
| definition | $aa^{+}a=a$, $a^{+}aa^{+}=a^{+}$, $(aa^{+})^{*}=aa^{+}$, $(a^{+}a)^{*}=a^{+}a$ |
| projections | $aa^{+}$ generates $aA$, $a^{+}a$ generates $Aa$ |
| kernel | $\ker a=(1-a^{+}a)A$, $\ker a^{+}=(1-aa^{+})A$ |
| uniqueness | at most one inverse, by the uniqueness of the projection of an ideal |
| existence | iff $aA$ and $Aa$ are generated by projections and $a$ is injective on the second |
| units | $a^{+}=a^{-1}$ |
| projections | $p^{+}=p$ |
| identities | $(a^{+})^{+}=a$, $(a^{*})^{+}=(a^{+})^{*}$, $(aa^{*})^{+}=(a^{+})^{*}a^{+}$ |
| factorisation | $a^{+}=(a^{*}a)^{+}a^{*}=a^{*}(aa^{*})^{+}$ |
| product rule | fails in general; $(ua)^{+}=a^{+}u^{*}$ for unitary $u$ |
| group inverse | agrees with $a^{+}$ exactly when $aa^{+}=a^{+}a$ |
The three statements to carry away are these. The four equations are the definition, and they determine the inverse uniquely whenever they have a solution. The solution is the inverse of $a$ between the two ideals, extended by zero, and existence is exactly the existence of the two projections and the injectivity between them. The matrix case is the case in which the criterion is always satisfied, and the biquaternion case is the case in which the inverse of the Hermitian factor is computed explicitly.
Summary of Notation
| symbol | meaning |
|---|---|
| $A$ | a unital associative algebra over a field $F$ |
| $a^{*}$ | the involution of the algebra |
| $a^{+}$ | the Moore–Penrose inverse of $a$ |
| $p,q$ | projections, $p=aa^{+}$, $q=a^{+}a$ |
| $aA$, $Aa$ | the right and the left ideal generated by $a$ |
| $\ker a$ | $\{x:ax=0\}$ |
| $a^{\#}$ | the group inverse |
| $a^{D}$ | the Drazin inverse |
| $E_{ij}$ | the matrix unit |
Further Reading
- Roger Penrose, "A generalized inverse for matrices", Proceedings of the Cambridge Philosophical Society 51 (1955) 406–413. The four equations, their uniqueness, and the matrix case.
- Eliakim H. Moore, "On the reciprocal of the general algebraic matrix", Bulletin of the American Mathematical Society 26 (1920) 394–395. The earliest formulation, as a reciprocal for singular matrices.
- Adi Ben-Israel and Thomas N. E. Greville, Generalized Inverses: Theory and Applications, 2nd edition (Springer, 2003). The algebraic theory, the group inverse and the Drazin inverse.
- Stephen L. Campbell and Carl D. Meyer, Generalized Inverses of Linear Transformations (Pitman, 1979; Dover, 2009). The Drazin inverse and the singular case.
- Max-Albert Knus, Alexander Merkurjev, Markus Rost and Jean-Pierre Tignol, The Book of Involutions, Colloquium Publications 44 (American Mathematical Society, 1998). Algebras with involution, the setting in which the four equations are read here.