Polarisation and the Hermitian Square
Introduction
A sesquilinear form is a function of two variables; its diagonal $\mathfrak{Q}(x) = h(x,x)$ is a function of one. The diagonal determines the form only up to a well-defined part of it, and the passage from the diagonal back to the form is polarisation. The exact content of the passage, and the exact information the diagonal loses, are the subject of this article.
Two objects organise it. The polarisation identity
$$ \mathfrak{Q}(x+y) - \mathfrak{Q}(x) - \mathfrak{Q}(y) = h(x,y) + h(y,x) $$
recovers from the diagonal the trace form $h(x,y) + \varsigma(h(x,y))$, and leaves the anti-Hermitian part of $h$ unseen; at the trivial base involution and with $2$ invertible the passage is a bijection, and in characteristic two the diagonal sees nothing at all. And the Hermitian square $x^{*}x$ of the algebra is the element whose functional image is the diagonal, $\mathfrak{Q}(x) = \varphi(x^{*}x)$, so that the quadratic form of the layer is carried by the involution of the algebra and not by the form alone.
This is the counterpart, in the sesquilinear layer, of Quadratic Forms and Polarisation, whose identity and whose characteristic-two degeneracy are the same. The form is Sesqualgebras with a Form, the matrix and quaternion quadratics are Quadratic Forms over Algebras and Norms, the trace form is Hermitian Algebras, and the real form that the trace form defines is Real Forms and the Signature. Throughout, $(A,*,h)$ is a sesqualgebra with a form over a base $(R,\varsigma)$, $\varphi = h(\cdot, 1)$ is the functional of the form, and $R^{\varsigma}$ is the fixed ring.
The Diagonal
Definition. The diagonal of the form is the map
$$ q : A \longrightarrow R^{\varsigma}, \qquad \mathfrak{Q}(x) = h(x,x) . $$
It is valued in the fixed ring because $\mathfrak{Q}(x) = \varsigma(\mathfrak{Q}(x))$ is the Hermitian property read at the two equal arguments. The association $h \mapsto q$ is $R^{\varsigma}$-linear and it is the only part of the form that a single variable can see.
Definition. The trace form of $h$ is
$$ h_{\mathrm{tr}}(x,y) = h(x,y) + h(y,x) = h(x,y) + \varsigma(h(x,y)), $$
the sum of the two orderings of the pair, and the anti-Hermitian part is $h_{\mathrm{sk}}(x,y) = h(x,y) - \varsigma(h(x,y))$. The two are the eigenspaces of the involution $h \mapsto h^{\dagger}$, $h^{\dagger}(x,y) = \varsigma(h(y,x))$, of The Sesquilinear Form and the Conjugation.
Proposition. The trace form is symmetric, $h_{\mathrm{tr}}(x,y) = h_{\mathrm{tr}}(y,x)$, and $R^{\varsigma}$-bilinear; the anti-Hermitian part is alternating, $h_{\mathrm{sk}}(x,x) = 0$. Every form is the sum $h = \tfrac{1}{2}h_{\mathrm{tr}} + \tfrac{1}{2}h_{\mathrm{sk}}$ when $2$ is invertible in $R^{\varsigma}$.
Proof. Symmetry and bilinearity are immediate from the definition and from $R^{\varsigma} = R^{\varsigma\cdot\varsigma}$. For the alternating property, $h_{\mathrm{sk}}(x,x) = h(x,x) - \varsigma(h(x,x)) = 0$ because the diagonal lies in $R^{\varsigma}$. The decomposition is the eigenvector decomposition of the involution $h \mapsto h^{\dagger}$ of the space of the sesquilinear forms, which needs $\tfrac{1}{2}$.
Note that $h_{\mathrm{tr}}$ is not $R$-sesquilinear when $\varsigma \neq \mathrm{id}$: it carries $\lambda$ to $\tfrac{1}{2}(\lambda + \varsigma(\lambda))$-weight, so it is a form over the fixed ring and not over $R$. This is why the two classical readings over $\mathbb{R}$ and over $\mathbb{C}$ differ, and the difference is worked in the examples below.
Polarisation
Proposition (the polarisation identity). For all $x, y \in A$,
$$ \boxed{\;\mathfrak{Q}(x+y) = \mathfrak{Q}(x) + \mathfrak{Q}(y) + h_{\mathrm{tr}}(x,y)\;} . $$
Proof. Expand by biadditivity: $\mathfrak{Q}(x+y) = h(x,x) + h(x,y) + h(y,x) + h(y,y) = \mathfrak{Q}(x) + \mathfrak{Q}(y) + h(x,y) + h(y,x)$, and the last two terms are $h_{\mathrm{tr}}(x,y)$.
Corollary (the diagonal recovers the trace form). If $2$ is invertible in $R^{\varsigma}$ then
$$ \boxed{\;h_{\mathrm{tr}}(x,y) = \mathfrak{Q}(x+y) - \mathfrak{Q}(x) - \mathfrak{Q}(y)\;}, $$
so two forms with the same diagonal have the same trace form and differ by an anti-Hermitian form.
Corollary (the classical case). When $\varsigma = \mathrm{id}$ and $2$ is invertible, the polarisation $h(x,y) = \tfrac{1}{2}\bigl(\mathfrak{Q}(x+y) - \mathfrak{Q}(x) - \mathfrak{Q}(y)\bigr)$ is a bijection between the symmetric bilinear forms and the quadratic functions $q$ with $\mathfrak{Q}(\lambda x) = \lambda^{2}q(x)$. This is the correspondence of Quadratic Forms and Polarisation, and the sesquilinear layer is its $\varsigma$-twisted version.
Remark (what the diagonal does not see). The anti-Hermitian part is invisible to the diagonal: $h$ and $h - h_{\mathrm{sk}}$ have the same diagonal, and $h_{\mathrm{sk}}(x,x) = 0$ for every $x$. Over $\mathbb{C}$ with $\varsigma$ the conjugation, the trace form $\tfrac{1}{2}h_{\mathrm{tr}} = \operatorname{Re} h$ is a real symmetric form and the antisymmetric part $\tfrac{1}{2}h_{\mathrm{sk}} = \mathrm{i}\,\operatorname{Im} h$ is real alternating; a Hermitian form is exactly a real symmetric form together with a real alternating form tied by the complex structure, and this is the decomposition $h = g + \mathrm{i}\,\omega$ of the geometry of a Kähler manifold.
The Hermitian Square
The diagonal of the layer is computed by the algebra, because the functional of the form is evaluated on the Hermitian square.
Proposition. The diagonal is the functional image of the Hermitian square:
$$ \boxed{\;\mathfrak{Q}(x) = \varphi(x^{*}x), \qquad \varphi = h(\cdot, 1)\;}, $$
and $x^{*}x$ is self-adjoint, $(x^{*}x)^{*} = x^{*}x$.
Proof. $\varphi(x^{*}x) = h(x^{*}x, 1) = h(x, x^{*}1) = h(x,x)$, by the compatibility $h(ab,c) = h(b,a^{*}c)$ read at $(a,b,c) = (x^{*},x,1)$ and by $1^{*} = 1$. Self-adjointness is the involution axiom, $(x^{*}x)^{*} = x^{*}x^{**} = x^{*}x$.
Corollary. The quadratic form of the algebra $\mathfrak{Q}(x) = \varphi(x^{*}x)$ satisfies $\mathfrak{Q}(x) = h(x,x)$, and its polarisation is the trace form, $\mathfrak{Q}(x+y) - \mathfrak{Q}(x) - \mathfrak{Q}(y) = h_{\mathrm{tr}}(x,y)$.
The two products $x^{*}x$ and $xx^{*}$ are the two Hermitian squares of the element, exchanged by the involution. The second gives the same diagonal for a cyclic functional, in particular for the trace form $\varphi = \tau$, where $\varphi(xx^{*}) = \varphi(x^{*}x)$ and the diagonal is also the functional image of $xx^{*}$; for a non-cyclic functional the two differ, and it is $x^{*}x$ that the diagonal reads. The square is the norm map of the layer; it is quadratic over $R^{\varsigma}$ in the sense $(\lambda x)^{*}(\lambda x) = \varsigma(\lambda)\lambda\,x^{*}x = |\lambda|^{2}x^{*}x$, and it is multiplicative for the elements of the unitary slice, $(xy)^{*}(xy) = y^{*}(x^{*}x)y$. When the norm map is multiplicative on the whole algebra the form is that of a composition algebra, and the norm form of the quaternion and octonion algebras is the classical instance: the polarisation of the norm is the trace form, and the whole theory is Hermitian Forms over Algebras and Norms.
The Failure When 2 Is Not Invertible
Proposition. In characteristic two with $\varsigma = \mathrm{id}$ the polarisation identity degenerates: $\mathfrak{Q}(x+y) = \mathfrak{Q}(x) + \mathfrak{Q}(y)$, the trace form vanishes, and the diagonal determines no part of the form at all.
Proof. With $\varsigma = \mathrm{id}$ the form is symmetric, $h_{\mathrm{tr}} = 2h = 0$, and the polarisation identity reduces to additivity of $q$. A symmetric bilinear form and its diagonal are then unrelated data: an alternating form has a vanishing diagonal and is not zero, and distinct symmetric forms share the diagonal.
The quadratic datum is therefore strictly finer than the bilinear one when $2$ is not invertible: a quadratic refinement of the form is a function $q$ with $\mathfrak{Q}(x+y) - \mathfrak{Q}(x) - \mathfrak{Q}(y) = h(x,y)$, and it is an extra structure. The characteristic-two examples, including the Hermitian-and-anti-Hermitian-not-alternating form over $\mathbb{F}_{2}$, are worked in The Sesquilinear Form and the Conjugation; the quadratic forms themselves are Quadratic Forms and Polarisation.
Examples
The Matrices
On $M_n(\mathbb{C})$ with $h(X,Y) = \tau(XY^{*})$ the diagonal is $\mathfrak{Q}(X) = \tau(XX^{*}) = \sum_{ij}|X_{ij}|^{2}$, the square of the Frobenius norm, and the polarisation recovers $\tau(XY^{*} + YX^{*}) = 2\operatorname{Re}\tau(XY^{*})$. The form is Hermitian, its trace form is the real symmetric form of the real and imaginary parts of the trace pairing, and its antisymmetric part is the alternating form $\mathrm{i}\operatorname{Im}\tau(XY^{*})$. The smallest instance shows the square and its trace at once:
$$ X=\begin{pmatrix}1&i\\0&2\end{pmatrix},\qquad XX^{*}=\begin{pmatrix}2&2i\\-2i&4\end{pmatrix},\qquad \mathfrak{Q}(X)=\tau(XX^{*})=6=\lvert1\rvert^{2}+\lvert i\rvert^{2}+\lvert0\rvert^{2}+\lvert2\rvert^{2}, $$
the matrix $XX^{*}$ Hermitian positive semi-definite, as a Hermitian square must be.
The Real and the Complex Interpretation
Over $\mathbb{R}$ with the trivial involution every Hermitian form is symmetric and the polarisation is a bijection when $2$ is invertible. Over $\mathbb{C}$ it is not: the diagonal of a Hermitian form sees $\operatorname{Re}h$ and never $\operatorname{Im}h$, and the pair $(\operatorname{Re}h, \operatorname{Im}h)$ is the pair (symmetric, alternating) of the real reading. The transfer between the two readings is Real Forms and the Signature.
The Biquaternion Form
On $\mathbb{B}$ with the dagger, $\mathfrak{Q}(\tilde Q) = \operatorname{Sc}(\tilde Q\tilde Q^{\dagger}) = \sum_{\mu}|Q_{\mu}|^{2}$, and the polarisation gives the real symmetric form $\operatorname{Re}\operatorname{Sc}(\tilde Q_1\tilde Q_2^{\dagger})$ together with the alternating form $\operatorname{Im}\operatorname{Sc}(\tilde Q_1\tilde Q_2^{\dagger})$. The indefinite companion $\operatorname{Sc}(\bar{\tilde Q}\tilde Q')$ and its signature are The Form on the Biquaternion Algebra as a General Plain Sesquilinear Form.
Summary
- The diagonal $\mathfrak{Q}(x) = h(x,x)$ is valued in the fixed ring $R^{\varsigma}$ and is the only part of the form visible to one variable.
- The polarisation identity reads $\mathfrak{Q}(x+y) - \mathfrak{Q}(x) - \mathfrak{Q}(y) = h_{\mathrm{tr}}(x,y)$, the trace form.
- The trace form is symmetric and $R^{\varsigma}$-bilinear, the anti-Hermitian part $h_{\mathrm{sk}}$ is alternating, and $h = \tfrac{1}{2}h_{\mathrm{tr}} + \tfrac{1}{2}h_{\mathrm{sk}}$ when $2$ is invertible: the diagonal recovers the trace form and never the antisymmetric part.
- Over $\mathbb{C}$ the decomposition is $\operatorname{Re}h + \mathrm{i}\operatorname{Im}h$, a real symmetric form and a real alternating form.
- The diagonal is the functional image of the Hermitian square, $\mathfrak{Q}(x) = \varphi(x^{*}x)$, and the quadratic form $\mathfrak{Q}(x) = \varphi(x^{*}x)$ has the trace form for its polarisation; for a cyclic functional the square $xx^{*}$ gives the same diagonal.
- In characteristic two with the trivial involution the polarisation degenerates, the diagonal determines nothing, and a quadratic refinement is extra structure.
- The classical bilinear original is Quadratic Forms and Polarisation; the norm forms of the algebras are Hermitian Forms over Algebras and Norms.
Summary of Notation
| symbol | meaning |
|---|---|
| $q$ | the diagonal $h(x,x)$ of the form |
| $h_{\mathrm{tr}}$ | the trace form $h(x,y) + \varsigma(h(x,y))$ |
| $h_{\mathrm{sk}}$ | the anti-Hermitian part $h(x,y) - \varsigma(h(x,y))$ |
| $h^{\dagger}$ | the involution $h^{\dagger}(x,y) = \varsigma(h(y,x))$ on the forms |
| $\varphi$ | the functional $h(\cdot,1)$ |
| $xx^{*}$ | the Hermitian square of $x$ |
| $\mathfrak{Q}(x) = \varphi(xx^{*})$ | the quadratic form of the algebra |
| $R^{\varsigma}$ | the fixed ring of the base involution |
Further Reading
- Winfried Scharlau, Quadratic and Hermitian Forms (Springer, 1985), for the polarisation of a form, the trace form and the characteristic-two degeneration.
- T. Y. Lam, Introduction to Quadratic Forms over Fields, Graduate Studies in Mathematics 67 (American Mathematical Society, 2005), for the quadratic refinements and the failure of polarisation in characteristic two.
- Max-Albert Knus, Quadratic and Hermitian Forms over Rings, Grundlehren der mathematischen Wissenschaften 294 (Springer, 1991), for the trace form of a Hermitian form over a ring with an involution.