The Peirce Decomposition and the Two Halves over $\mathbb{R}$

Introduction

The transpose sends the corner $A_{kl}$ to the corner $A_{lk}$, so each corner pair splits into a symmetric and an antisymmetric part of equal dimension, and the two halves of $M_{n}(\mathbb{R})$ are thereby graded by the frame.

The article shows that a symmetric frame makes the transpose permute the corners, computes both halves of a corner pair and of a diagonal corner, proves the total identity for the symmetric dimensions, and checks it on the rank vectors $(2,2)$, $(3,1)$ and $(2,1,1)$ at $n=4$. It assumes The Peirce Basis and the Corners over $\mathbb{R}$ for the corners and the frame, and it defers the Jordan reading of the symmetric half to The Peirce Decomposition of $\operatorname{Sym}_n$ as a Jordan Algebra.

The Frame and the Transpose

A frame is called symmetric when every idempotent is symmetric. This article treats frames with

$$ P_{k}^{\mathsf T}=P_{k},\qquad k=1,\dots,n, $$

together with $P_{k}P_{l}=\delta_{kl}P_{k}$ and $\sum_{k}P_{k}=\mathrm{I}$. A symmetric frame is an orthogonal decomposition $\mathbb{R}^{n}=V_{1}\oplus\cdots\oplus V_{n}$ into pairwise orthogonal subspaces, with $P_{k}$ the orthogonal projection onto $V_{k}$ and $\operatorname{rk}P_{k}=r_{k}$.

The transpose acts on the corners of a symmetric frame by exchanging the two indices. Because $P_{k}^{\mathsf T}=P_{k}$,

$$ (P_{k}XP_{l})^{\mathsf T}=P_{l}^{\mathsf T}X^{\mathsf T}P_{k}^{\mathsf T}=P_{l}X^{\mathsf T}P_{k}\in A_{lk}. $$

So $A_{kl}^{\mathsf T}\subseteq A_{lk}$, and applying the same computation to $A_{lk}$ gives equality of dimensions and the identity

$$ A_{kl}^{\mathsf T}=A_{lk}. $$

For a non-symmetric frame the statement changes. The transpose of a corner of $P$ is a corner of the transposed frame, not of $P$, so the exchange of indices needs the frame to be fixed by the transpose.

The transpose is $\mathbb{R}$-linear, so it acts inside the real algebra. It is not conjugate linear and it fixes every real scalar. Its fixed space is $\operatorname{Sym}_n$ and its space of eigenvalue $-1$ is $\operatorname{Skew}_n$, of dimensions $n(n+1)/2$ and $n(n-1)/2$.

The Transpose on the Corner Pair

A corner pair $A_{kl}+A_{lk}$ is invariant under the transpose. The transpose carries $A_{kl}$ to $A_{lk}$ and $A_{lk}$ to $A_{kl}$, so it preserves their sum. On that sum it is an involution, and the sum splits into its two eigenspaces.

Proposition. For $k\ne l$,

$$ \operatorname{Sym}(A_{kl}+A_{lk})=\{B+B^{\mathsf T}:B\in A_{kl}\},\qquad \operatorname{Skew}(A_{kl}+A_{lk})=\{B-B^{\mathsf T}:B\in A_{kl}\}, $$

and both have dimension $r_{k}r_{l}$.

Proof. An element of the sum is $X=B+C$ with $B\in A_{kl}$ and $C\in A_{lk}$. Then $X^{\mathsf T}=C^{\mathsf T}+B^{\mathsf T}$, so $X$ is symmetric exactly when $C=B^{\mathsf T}$, and antisymmetric exactly when $C=-B^{\mathsf T}$. The map $B\mapsto B+B^{\mathsf T}$ is injective on $A_{kl}$, because $B+B^{\mathsf T}=0$ with $B\in A_{kl}$ forces $B=-B^{\mathsf T}\in A_{kl}\cap A_{lk}=0$; likewise for $B\mapsto B-B^{\mathsf T}$. Both images therefore have the dimension $r_{k}r_{l}$ of $A_{kl}$, and $r_{k}r_{l}+r_{k}r_{l}=2r_{k}r_{l}=\dim(A_{kl}+A_{lk})$.

The two halves of a corner pair are the symmetric and the antisymmetric maps between two orthogonal subspaces. A map $B:V_{l}\to V_{k}$ has a transpose $B^{\mathsf T}:V_{k}\to V_{l}$, and the pair $(B,B^{\mathsf T})$ is symmetric while $(B,-B^{\mathsf T})$ is antisymmetric. The symmetric part is the graph of the transpose, the antisymmetric part the graph of its negative.

The Diagonal Corner

A diagonal corner is invariant under the transpose, and it splits into its symmetric and antisymmetric parts. For $k=l$ the corner $A_{kk}$ is carried to itself, and

$$ A_{kk}=\operatorname{Sym}(A_{kk})\oplus\operatorname{Skew}(A_{kk}), $$

with

$$ \dim\operatorname{Sym}(A_{kk})=\frac{r_{k}(r_{k}+1)}{2},\qquad \dim\operatorname{Skew}(A_{kk})=\frac{r_{k}(r_{k}-1)}{2}. $$

The two dimensions add to $r_{k}^{2}=\dim A_{kk}$, and the split is the one of The Matrix Algebra $M_n(\mathbb{R})$ applied to the smaller algebra $A_{kk}\cong M_{r_{k}}(\mathbb{R})$.

The two parts of a diagonal corner close under different products. The symmetric elements are closed under $X\circ Y=\tfrac12(XY+YX)$ and the antisymmetric ones under $[X,Y]=XY-YX$; no single product closes on the whole corner.

The trace form is positive definite on the symmetric part and negative definite on the antisymmetric part. For a symmetric $S$ in the corner, $\operatorname{tr}(S^{2})=\sum S_{ij}^{2}>0$, and for an antisymmetric $K$, $\operatorname{tr}(K^{2})=-\sum K_{ij}^{2}<0$. The corner therefore contributes $r_{k}(r_{k}+1)/2$ positive and $r_{k}(r_{k}-1)/2$ negative directions to the inertia.

The Total Identity

The symmetric dimensions of all the corners add to $n(n+1)/2$.

$$ \sum_{k=1}^{n}\frac{r_{k}(r_{k}+1)}{2}+\sum_{k

Proof. The second sum is half of $n^{2}-\sum_{k}r_{k}^{2}$, since $\big(\sum_{k}r_{k}\big)^{2}=\sum_{k}r_{k}^{2}+2\sum_{k

$$ \frac{\sum_{k}r_{k}^{2}+n}{2}+\frac{n^{2}-\sum_{k}r_{k}^{2}}{2}=\frac{n^{2}+n}{2}. $$

The antisymmetric dimensions add to $n(n-1)/2$. The same computation with $r_{k}(r_{k}-1)/2$ gives

$$ \frac{\sum_{k}r_{k}^{2}-n}{2}+\frac{n^{2}-\sum_{k}r_{k}^{2}}{2}=\frac{n^{2}-n}{2}. $$

The two totals are the dimensions of the two halves. Summing them gives $n(n+1)/2+n(n-1)/2=n^{2}$, so the corner pairs account for the whole algebra, and the identity depends on the frame only through its rank vector.

The Grading of the Symmetric Half

The symmetric half is the direct sum of its corner contributions.

$$ \operatorname{Sym}_n=\bigoplus_{k=1}^{n}\operatorname{Sym}(A_{kk})\ \oplus\ \bigoplus_{k

Proof. Every $X\in\operatorname{Sym}_n$ splits as $X=\sum_{k,l}X_{kl}$ with $X_{kl}\in A_{kl}$, and the symmetry gives $X_{kl}^{\mathsf T}=X_{lk}$. So $X$ is determined by the diagonal blocks $X_{kk}\in A_{kk}$ and by the pairs $(X_{kl},X_{lk})$ with $k

The graded pieces are the eigenspaces of the frame's Peirce operators cut by the transpose. The piece labelled $k$ is $\operatorname{Sym}(A_{kk})$, the fixed part of the corner $A_{kk}$; the piece labelled $k

At a complete frame every piece is a line. With all $r_{k}=1$ the diagonal pieces have dimension $1$ and the pair pieces too, so the symmetric half is a direct sum of $n+n(n-1)/2=n(n+1)/2$ lines. This is the finest grading the frame gives.

The grading refines the two-block Peirce decomposition. Taking the coarse frame $P_{1}=E$, $P_{2}=\mathrm{I}-E$ with ranks $r$ and $n-r$ gives three pieces: $\operatorname{Sym}(A_{11})$ of dimension $r(r+1)/2$, $\operatorname{Sym}(A_{22})$ of dimension $(n-r)(n-r+1)/2$ and $\operatorname{Sym}(A_{12}+A_{21})$ of dimension $r(n-r)$. These are the three Peirce spaces of the Jordan algebra $\operatorname{Sym}_n$ relative to $E$, in the order $1$, $0$, $\tfrac12$.

The Grading of the Antisymmetric Half

The antisymmetric half is graded the same way.

$$ \operatorname{Skew}_n=\bigoplus_{k=1}^{n}\operatorname{Skew}(A_{kk})\ \oplus\ \bigoplus_{k

The proof is the same computation with $X_{lk}=-X_{kl}^{\mathsf T}$, and the dimensions are $r_{k}(r_{k}-1)/2$ and $r_{k}r_{l}$. At a complete frame the diagonal pieces vanish, and the antisymmetric half is a direct sum of $n(n-1)/2$ lines, one for each pair $k

The antisymmetric half is a Lie algebra graded by its corners. $\operatorname{Skew}_n=\mathfrak{so}(n)$ is closed under the commutator, and the bracket of two corner pieces lands in the piece named by the surviving indices, which is the root-space pattern of $\mathfrak{so}(n)$ read in the frame.

The Two Halves over $\mathbb{R}$ and over $\mathbb{C}$

Over $\mathbb{R}$ the two halves are different algebras. $\operatorname{Sym}_n$ is closed under the anticommutator and is a formally real Jordan algebra; $\operatorname{Skew}_n$ is closed under the commutator and is the Lie algebra $\mathfrak{so}(n)$ of the orthogonal group. The dimensions differ, $n(n+1)/2$ against $n(n-1)/2$, and the trace form has opposite signs on the two halves.

Over $\mathbb{C}$ the same two eigenspaces exist, and the form on them is isotropic. Complexifying gives $\operatorname{Sym}_n(\mathbb{C})$ and $\operatorname{Skew}_n(\mathbb{C})$, of complex dimensions $n(n+1)/2$ and $n(n-1)/2$, and the corner exchange $A_{kl}^{\mathsf T}=A_{lk}$ holds for a complex symmetric frame word for word. What the complex picture loses is the sign: over $\mathbb{C}$ the form $\operatorname{tr}(XY)$ vanishes on nonzero elements of both halves. A symmetric example at $n=2$ is

$$ X=\begin{pmatrix}1&i\\ i&-1\end{pmatrix},\qquad X^{2}=0, $$

and an antisymmetric one at $n=3$ is built from $Y_{12}=1$, $Y_{13}=i$, whose square has zero trace. So over $\mathbb{C}$ neither half is definite and the sign that separates them over $\mathbb{R}$ has no meaning.

The inertia of the trace form is the real invariant. Over $\mathbb{R}$ the form $\langle X,Y\rangle=\operatorname{tr}(XY)$ has inertia $\big(n(n+1)/2,\,n(n-1)/2\big)$, positive on $\operatorname{Sym}_n$ and negative on $\operatorname{Skew}_n$. A complex bilinear form has no inertia, so the definite split computed here is a real statement.

The grading is a real statement about the transpose. The eigenvalues $\pm1$ are real numbers, so the pieces of the grading are real subspaces of real dimensions $r_{k}r_{l}$ and $r_{k}(r_{k}\pm1)/2$, and the two halves are the two real eigenspaces of one $\mathbb{R}$-linear involution. The Jordan algebra $\operatorname{Sym}_n$ is formally real, while $\operatorname{Sym}_n(\mathbb{C})$ contains the nilpotent $X$ above; that difference, not the existence of the splitting, is what the base field decides.

Checks at $n=4$

The identity holds for the rank vectors $(2,2)$, $(3,1)$ and $(2,1,1)$. For each vector, the symmetric total is

ranks diagonal terms pair terms total
$(2,2)$ $3+3$ $4$ $10$
$(3,1)$ $6+1$ $3$ $10$
$(2,1,1)$ $3+1+1$ $2+2+1$ $10$

and $10=4\cdot5/2$ in all three cases. The antisymmetric totals are $1+1+4=6$, $3+0+3=6$ and $1+0+0+2+2+1=6$, which is $4\cdot3/2$.

The corner dimensions are the predicted ones. With the rank vector $(2,1,1)$ the corner pairs give symmetric and antisymmetric dimensions $2,2$ for $(0,1)$ and $(0,2)$, and $1,1$ for $(1,2)$; the diagonal corner of rank $2$ gives $3,1$ and the two corners of rank $1$ give $1,0$. The numbers agree with $r_{k}r_{l}$ and $r_{k}(r_{k}\pm1)/2$.

Summary

A symmetric frame $P_{k}=P_{k}^{\mathsf T}$ makes the transpose exchange the corners, $A_{kl}^{\mathsf T}=A_{lk}$, so each corner pair $A_{kl}+A_{lk}$ with $k\ne l$ splits into a symmetric part and an antisymmetric part, both of dimension $r_{k}r_{l}$, and each diagonal corner $A_{kk}$ splits into $\operatorname{Sym}(A_{kk})$ of dimension $r_{k}(r_{k}+1)/2$ and $\operatorname{Skew}(A_{kk})$ of dimension $r_{k}(r_{k}-1)/2$. The symmetric dimensions satisfy $\sum_{k}r_{k}(r_{k}+1)/2+\sum_{k

Summary of Notation

Symbol Meaning
$P_{k}$ A symmetric idempotent of the frame, $P_{k}^{\mathsf T}=P_{k}$
$r_{k}=\operatorname{rk}P_{k}$ The rank of $P_{k}$, with $\sum_{k}r_{k}=n$
$A_{kl}=P_{k}M_{n}(\mathbb{R})P_{l}$ The corner, with $A_{kl}^{\mathsf T}=A_{lk}$
$\operatorname{Sym}(A_{kl}+A_{lk})$ The symmetric part of a corner pair, of dimension $r_{k}r_{l}$
$\operatorname{Skew}(A_{kl}+A_{lk})$ The antisymmetric part of a corner pair, of dimension $r_{k}r_{l}$
$\operatorname{Sym}(A_{kk})$ The symmetric part of a diagonal corner, of dimension $r_{k}(r_{k}+1)/2$
$\operatorname{Skew}(A_{kk})$ The antisymmetric part of a diagonal corner, of dimension $r_{k}(r_{k}-1)/2$
$\{k,l\}$ The label of a graded piece, with $k\le l$

Further Reading

  • Roger A. Horn and Charles R. Johnson, Matrix Analysis (Cambridge, 2013), for symmetric and antisymmetric matrices and the transpose.
  • Nathan Jacobson, Structure and Representations of Jordan Algebras (American Mathematical Society, 1968), for Peirce decompositions and the relation between a Jordan algebra and its corners.
  • Sigurdur Helgason, Differential Geometry, Lie Groups and Symmetric Spaces (American Mathematical Society, 2001), for the graded structure of $\mathfrak{so}(n)$ and its root spaces.
  • The Peirce Basis and the Corners over $\mathbb{R}$, for the corners and the frame that the grading uses.
  • The Peirce Decomposition of $\operatorname{Sym}_n$ as a Jordan Algebra, for the symmetric half read as a Jordan algebra.
  • The Matrix Algebra $M_n(\mathbb{R})$, for the trace form, its inertia and the two halves.