The Peirce Basis and the Corners over $\mathbb{R}$
Introduction
A real frame cuts $M_{n}(\mathbb{R})$ into $n^{2}$ real corners, the corners multiply by the matrix-unit table, and the residual freedom is conjugation by an invertible real matrix, so the frame variety has dimension $n^{2}-n$.
The article states the corners, proves that they sum directly to the algebra, states the products $E_{ij}E_{kl}=\delta_{jk}E_{il}$, defines a system of matrix units over $\mathbb{R}$, reads the corners as joint eigenspaces of two commuting families of operators, and computes the real gauge. Frames of equal ranks are conjugate by $\mathrm{GL}_{n}(\mathbb{R})$, the stabiliser is the diagonal subgroup $D$, and the frame variety is $\mathrm{GL}_{n}(\mathbb{R})/D$, of dimension $n^{2}-n$. It assumes The Peirce Basis in $n$ Dimensions, which runs the same list over $\mathbb{C}$, and defers the symmetric half of the corners to The Peirce Decomposition of $\operatorname{Sym}_n$ as a Jordan Algebra.
The Frame
A frame is an ordered family of orthogonal idempotents summing to the identity.
$$ P_{k}P_{l}=\delta_{kl}P_{k},\qquad \sum_{k=1}^{n}P_{k}=\mathrm{I}. $$
The rank $r_{k}=\operatorname{rk}P_{k}$ is part of the data and $\sum_{k}r_{k}=n$. The frame is complete when every $r_{k}=1$, and the standard frame is $F_{k}=E_{kk}$.
The subspaces $V_{k}=\operatorname{im}P_{k}$ are independent and sum to $\mathbb{R}^{n}$. The idempotent $P_{k}$ is the identity on $V_{k}$ and kills the other $V_{l}$, so $\mathbb{R}^{n}=\bigoplus_{k}V_{k}$ with $\dim V_{k}=r_{k}$. A frame is the same datum as a decomposition of $\mathbb{R}^{n}$ into ordered real subspaces of dimensions $r_{k}$.
The Corners
A corner is the space of maps from one subspace of the frame to another.
$$ A_{kl}=P_{k}M_{n}(\mathbb{R})P_{l}=\{P_{k}XP_{l}:X\in M_{n}(\mathbb{R})\}. $$
Proposition. $\dim A_{kl}=r_{k}r_{l}$, the corners are independent, and
$$ M_{n}(\mathbb{R})=\bigoplus_{k,l=1}^{n}A_{kl},\qquad \sum_{k,l}r_{k}r_{l}=\Big(\sum_{k}r_{k}\Big)^{2}=n^{2}. $$
Proof. As a space of maps, $A_{kl}=\operatorname{Hom}(V_{l},V_{k})$, so its dimension is $r_{k}r_{l}$. Expanding the identity on both sides of $X$,
$$ X=\Big(\sum_{k}P_{k}\Big)X\Big(\sum_{l}P_{l}\Big)=\sum_{k,l}P_{k}XP_{l}, $$
gives the sum. For independence, multiply $\sum_{k,l}X_{kl}=0$ on the left by $P_{i}$ and on the right by $P_{j}$; every term dies except $X_{ij}$.
The diagonal corner is the endomorphism algebra of its subspace. $A_{kk}=\operatorname{End}(V_{k})$ has dimension $r_{k}^{2}$, and the off-diagonal corner is the space of maps $V_{l}\to V_{k}$, of dimension $r_{k}r_{l}$. At a complete frame all $n^{2}$ corners have dimension one.
The Multiplication Table
The product of the corner $(k,l)$ with the corner $(p,q)$ lands in the corner $(k,q)$, and only when $l=p$. From $P_{l}P_{p}=\delta_{lp}P_{l}$,
$$ A_{kl}A_{pq}\subseteq\delta_{lp}A_{kq}. $$
One nonzero element per corner gives the table of the matrix units. For a complete frame and $E_{kl}\in A_{kl}$ nonzero, the rule is
$$ E_{ij}E_{kl}=\delta_{jk}E_{il}, $$
which is the table of the elementary matrices read in the frame. It forces $E_{kk}=P_{k}$: the diagonal element is $c_{k}P_{k}$, and the rule with all indices equal gives $c_{k}^{2}=c_{k}$, hence $c_{k}=1$.
A complete real frame realises the table with real rank-one matrices. Let $v_{k}\in V_{k}$ be nonzero and let $\ell_{k}$ be the real form with $\ell_{k}(v_{l})=\delta_{kl}$. Then $E_{kl}=v_{k}\ell_{l}$ is real, lies in $A_{kl}$, and
$$ E_{kl}E_{pq}=v_{k}\ell_{l}(v_{p})\ell_{q}=\delta_{lp}E_{kq},\qquad E_{kk}=P_{k}. $$
The $v_{k}$ are a real basis of $\mathbb{R}^{n}$ and the $\ell_{k}$ its real dual basis.
Systems of Matrix Units
A system of $n\times n$ matrix units over $\mathbb{R}$ is a family $E_{ij}\in M_{n}(\mathbb{R})$ with that table and diagonal sum the identity.
$$ E_{ij}E_{kl}=\delta_{jk}E_{il},\qquad \sum_{k=1}^{n}E_{kk}=\mathrm{I}. $$
Proposition. Such a family is a basis of the algebra, its diagonal members are the idempotents of a frame, and $E_{ij}\mapsto$ elementary array is an isomorphism of algebras.
Proof. The table gives $E_{kk}^{2}=E_{kk}$ and $E_{kk}E_{ll}=0$ for $k\ne l$, so the $E_{kk}$ are a frame, and the second condition makes it sum to the identity. The products span the algebra by the table and $\mathrm{I}=\sum_{k}E_{kk}$; by the corner decomposition of that frame each corner is the line spanned by $E_{kl}$, so the $n^{2}$ elements are independent. The assignment $E_{ij}\mapsto$ elementary array is linear and bijective, and both sides obey the same table, so it is multiplicative.
Every system of units is conjugate to the standard one. The $E_{kk}$ have rank one and are mutually orthogonal, so $\mathbb{R}^{n}=\bigoplus_{k}\operatorname{im}E_{kk}$. For $v_{k}\in\operatorname{im}E_{kk}$ nonzero, the real matrix $S$ with columns $v_{k}$ is invertible, and $S^{-1}E_{ij}S$ sends $v_{j}$ to $v_{i}$ and kills the other $v_{m}$, hence is the elementary array. The conjugating matrix is unique up to the diagonal subgroup.
The Two Peirce Operators
The corners are the joint eigenspaces of two commuting families of operators.
$$ L_{k}(X)=P_{k}X,\qquad R_{k}(X)=XP_{k}. $$
Proposition. The $2n$ operators commute, and
$$ A_{kl}=\{X:L_{k}X=X,\ R_{l}X=X,\ L_{i}X=R_{j}X=0\ \text{for}\ i\ne k,\ j\ne l\}. $$
Proof. $L_{k}L_{i}=\delta_{ki}L_{k}$ and $L_{k}R_{j}=R_{j}L_{k}$, so the family commutes. The displayed conditions say $P_{k}XP_{l}=X$, which leaves exactly the $r_{k}r_{l}$ entries that define the corner.
The eigenvalues are the real numbers $0$ and $1$. Each $L_{k}$ and $R_{k}$ is $\mathbb{R}$-linear with spectrum $\{0,1\}$, so the joint eigenspaces of a real frame are real subspaces, and no complex number enters the cut.
The Real Gauge
Two frames with the same rank vector differ by conjugation by an invertible real matrix. There is $S\in\mathrm{GL}_{n}(\mathbb{R})$ with
$$ P'_{k}=SP_{k}S^{-1},\qquad k=1,\dots,n . $$
Proof. Take a real basis adapted to $P$, with the first $r_{1}$ vectors in $V_{1}$, the next $r_{2}$ in $V_{2}$, and so on, and a real basis $v'_{m}$ adapted to $P'$ in the same way. The matrix $S$ with $Sv_{m}=v'_{m}$ is real and invertible, and it carries $V_{k}$ isomorphically to $V'_{k}$; such a map conjugates the projection onto $V_{k}$ into the projection onto $V'_{k}$.
The group acting is $\mathrm{GL}_{n}(\mathbb{R})$, not $\mathrm{GL}_{n}(\mathbb{C})$. Both bases above are real, so $S$ is real. A complex $S$ that carries the standard frame to a real frame is real up to a diagonal factor: the $k$-th idempotent is sent to the projection onto the line spanned by the $k$-th column, so that column is a complex multiple of a real vector, and $S=A\operatorname{diag}(u_{1},\dots,u_{n})$ with $A$ real and all $u_{k}\ne0$. Diagonal matrices act trivially on the standard frame, so the complex matrices producing a given real frame are the real ones times the diagonal ones.
The stabiliser is the diagonal subgroup. If $S$ commutes with every $E_{kk}$, comparing entries in $SE_{kk}=E_{kk}S$ gives $S_{ij}=0$ for $i\ne j$. Hence
$$ D=\{\operatorname{diag}(t_{1},\dots,t_{n}):t_{k}\in\mathbb{R}^{\times}\}, $$
of dimension $n$.
The gauge acts on the units by ratios. Conjugation sends $E_{kl}=v_{k}\ell_{l}$ to $(Sv_{k})(\ell_{l}S^{-1})$; for diagonal $S$ with entries $t_{k}$,
$$ E_{kl}\longmapsto t_{k}t_{l}^{-1}E_{kl}, $$
and the idempotents $E_{kk}$ are fixed.
The Frame Variety
The complete frames form the homogeneous space $\mathrm{GL}_{n}(\mathbb{R})/D$, of dimension $n^{2}-n$. The map $S\mapsto(SE_{11}S^{-1},\dots,SE_{nn}S^{-1})$ is surjective by the conjugation theorem and has fibres the right cosets of $D$, so
$$ \dim(\text{frames})=n^{2}-n,\qquad n^{2}=\underbrace{(n^{2}-n)}_{\text{frames}}+\underbrace{n}_{\dim D}. $$
At $n=3$ the dimension is $6$. The numbers are $9-3=6$. A frame is a decomposition of $\mathbb{R}^{3}$ into three lines, which costs $3\cdot 2=6$ parameters. Over $\mathbb{C}$ the same quotient has complex dimension $n^{2}-n$, hence real dimension $2(n^{2}-n)$: the real variety has half that, and the factor two is the passage from $\mathbb{R}$ to $\mathbb{C}$ in the group.
The tangent space confirms the count. With $S=\mathrm{I}+\epsilon X$ the frame moves by $P_{k}(\epsilon)=\mathrm{I}+\epsilon[X,P_{k}]$, so the tangent space at the standard frame is spanned by the commutators $[X,E_{kk}]$. That span has dimension $6$ at $n=3$ and $12$ at $n=4$, and the kernel of $X\mapsto([X,E_{kk}])_{k}$ is the diagonal matrices.
The Real Content: Simultaneous Diagonalisation
A frame of real idempotents exists exactly when the idempotents are simultaneously diagonalisable over $\mathbb{R}$. For commuting real idempotents $P_{k}$ with $\sum_{k}P_{k}=\mathrm{I}$ and ranks $r_{k}$ there is $S\in\mathrm{GL}_{n}(\mathbb{R})$ with
$$ S^{-1}P_{k}S=\operatorname{diag}(\underbrace{1,\dots,1}_{r_{k}},\underbrace{0,\dots,0}_{n-r_{k}}) $$
in block position $k$. This is the adapted-basis argument again: the images $V_{k}$ are independent and sum to $\mathbb{R}^{n}$, and a real basis chosen block by block diagonalises all the $P_{k}$ at once.
The diagonalising matrix is real because the eigenvalues are real. The real content is not commutativity, which holds over $\mathbb{C}$ too, but the fact that the eigenvalues are the real numbers $0$ and $1$. For distinct real $\lambda_{k}$ the combination $H=\sum_{k}\lambda_{k}P_{k}$ has eigenvalues $\lambda_{k}$; every eigenvector of $H$ is an eigenvector of every $P_{k}$, and the eigenvectors of a real matrix with distinct real eigenvalues are real.
The ranks are arbitrary and they are the only invariant. For every composition $n=r_{1}+\cdots+r_{n}$ with $r_{k}\ge1$, a real invertible $S$ and the diagonal blocks $\Delta_{k}$ give the frame $P_{k}=S\Delta_{k}S^{-1}$. Two frames are conjugate by $\mathrm{GL}_{n}(\mathbb{R})$ exactly when their rank vectors agree.
The frame comes from a real basis. The unit $E_{kl}=v_{k}\ell_{l}$ is a real matrix only if $v_{k}$ and $\ell_{l}$ are real, and a real frame supplies them, since each $\operatorname{im}P_{k}$ is a real line. No real frame needs complex vectors in its Peirce basis.
A frame need not be symmetric. The corner $A_{kl}=\operatorname{Hom}(V_{l},V_{k})$ requires real subspaces only. The frames fixed by the transpose, with $P_{k}^{\mathsf T}=P_{k}$, are treated in The Peirce Decomposition and the Two Halves over $\mathbb{R}$.
Summary
A frame of $M_{n}(\mathbb{R})$ is an ordered family of orthogonal real idempotents summing to the identity, equivalently a decomposition of $\mathbb{R}^{n}$ into ordered real subspaces $V_{k}$ of ranks $r_{k}$. The corners $A_{kl}=P_{k}M_{n}(\mathbb{R})P_{l}=\operatorname{Hom}(V_{l},V_{k})$ have dimension $r_{k}r_{l}$, sum directly to the algebra, and multiply by $A_{kl}A_{pq}\subseteq\delta_{lp}A_{kq}$; a system of matrix units is one real element $E_{kl}$ per corner with $E_{ij}E_{kl}=\delta_{jk}E_{il}$ and $\sum_{k}E_{kk}=\mathrm{I}$, and it is a basis of the algebra. The corners are the joint eigenspaces of the Peirce operators $L_{k}(X)=P_{k}X$ and $R_{k}(X)=XP_{k}$, with eigenvalues $0$ and $1$. Frames of equal ranks are conjugate by $\mathrm{GL}_{n}(\mathbb{R})$, with stabiliser the diagonal group $D$ of dimension $n$, so the frame variety is $\mathrm{GL}_{n}(\mathbb{R})/D$ of real dimension $n^{2}-n$; at $n=3$ that is $9-3=6$, half the real dimension $12$ of the complex frame space. The real content is that a frame of real idempotents exists exactly when the idempotents are simultaneously diagonalisable over $\mathbb{R}$, which holds because the eigenvalues are real; then the ranks are arbitrary and the units come from a real basis.
Summary of Notation
| Symbol | Meaning |
|---|---|
| $P_{k}$ | The $k$-th idempotent of the frame |
| $V_{k}=\operatorname{im}P_{k}$ | The real subspace of the $k$-th idempotent |
| $r_{k}=\operatorname{rk}P_{k}$ | The rank, with $\sum_{k}r_{k}=n$ |
| $A_{kl}=P_{k}M_{n}(\mathbb{R})P_{l}$ | The corner, isomorphic to $\operatorname{Hom}(V_{l},V_{k})$ |
| $E_{kl}$ | The chosen element of $A_{kl}$ |
| $L_{k},R_{k}$ | The Peirce operators $X\mapsto P_{k}X$ and $X\mapsto XP_{k}$ |
| $D$ | The diagonal subgroup of $\mathrm{GL}_{n}(\mathbb{R})$ |
| $\mathrm{GL}_{n}(\mathbb{R})/D$ | The frame variety, of dimension $n^{2}-n$ |
Further Reading
- Richard S. Pierce, Associative Algebras (Springer, 1982), for Peirce corners, systems of matrix units and their uniqueness up to conjugation.
- Roger A. Horn and Charles R. Johnson, Matrix Analysis (Cambridge, 2013), for idempotents, adapted bases and simultaneous diagonalisation over $\mathbb{R}$.
- The Peirce Basis in $n$ Dimensions, for the complex version of the same table and the gauge on the units.
- The Matrix Algebra $M_n(\mathbb{R})$, for the matrix units and the transpose over $\mathbb{R}$.
- The Peirce Decomposition of $\operatorname{Sym}_n$ as a Jordan Algebra, for the symmetric half of the corners.
- The Frame, the Simple Nilpotents and the Real Gauge of $M_n(\mathbb{R})$, for the nilpotents attached to a frame.