The Continuous Quadratic Form and the Polar Form

Introduction

A quadratic form on a vector space over a field is a polynomial datum: its value at a vector is an element of the base, and its polar form is the symmetric bilinear pairing obtained from it by polarisation. This article reads the same object in a topological module. The question is not what a quadratic form is — that is settled in Quadratic Forms and Polarisation — but which forms belong to the layer, what the topology on the set of forms is, and what the continuity of the form implies for the polar form, for the radical, for the isometries and for the classification.

Two facts organise the answer. The first is that the polar form is a continuous bilinear map as soon as the quadratic form is continuous, by the polarisation identity, so a single continuous function of one variable carries the whole symmetric bilinear structure with it; the two henceforth move together, and it is the quadratic form that is taken as the primitive datum of the layer. The second is that over a normed space the polarisation is an isometry between the continuous quadratic forms and the continuous symmetric bilinear forms, both carrying their natural uniform norms, so that the algebraic passage between the two readings is also a metric one; the constant of equivalence of the two norms is one, and the passage loses nothing.

The article defines the continuous quadratic form and its polar form, constructs the topological vector space of continuous forms with its uniform norm and proves the polarisation is an isometric isomorphism, treats the radical and compares it with the closure of zero of the Clifford algebra, and states what continuity gives for the isometries and for the classification: the orthogonal group is closed, the non-degenerate forms are open in finite dimension, and the signature is locally constant there. It owns no Clifford algebra of its own, no alternating form and no completion. The Clifford algebra generated by the form is Topological Clifford Algebras; the alternating forms and the Poisson bracket are The Continuous Symplectic Form and the Poisson Bracket; the orthogonality, the reflection and the group of isometries are The Topological Orthogonal Group and the Spin Group; the classification modulo hyperbolic forms is The Topological Witt Group and the Brauer–Wall Group; the algebraic theory of the form is Bilinear Forms, Quadratic Forms and Polarisation and Witt's Theorems. Throughout, $F$ is a topological field of characteristic not $2$, $V$ a topological $F$-module with a continuous scalar action, and $m$ and $B$ are related by polarisation.

The Continuous Quadratic Form

Definition and the Polar Form

Definition. Let $F$ be a topological field and let $V$ be a topological $F$-module. A continuous quadratic form is a map $m : V \to F$ such that $\mathfrak{Q}(\lambda v) = \lambda^{2}q(v)$ for all $\lambda \in F$ and $v \in V$, such that the map $(x,y)\mapsto \mathfrak{Q}(x+y)-\mathfrak{Q}(x)-\mathfrak{Q}(y)$ is bilinear, and which is continuous for the given topologies. Its polar form is the symmetric bilinear form

$$ \boxed{\;B(x,y) = \tfrac{1}{2}\bigl(\mathfrak{Q}(x+y)-\mathfrak{Q}(x)-\mathfrak{Q}(y)\bigr)\;} , $$

the form of Quadratic Forms and Polarisation, §Quadratic Forms, and one has $\mathfrak{Q}(x)=B(x,x)$.

The three conditions are the algebraic definition of Quadratic Forms and Polarisation with the continuity added, and the coefficient $\tfrac12$ exists because the characteristic is not $2$. The polar form is symmetric, $B(x,y)=B(y,x)$, since the difference is symmetric in $x$ and $y$, and it recovers $m$ on the diagonal. Nothing is assumed about the form beyond this: it need not be non-degenerate, need not be definite, and no positivity is named.

Proposition (the polar form is continuous). Let $m$ be a continuous quadratic form. Then $B : V\times V\to F$ is continuous.

Proof. By the definition, $B$ is the composite of the map $(x,y)\mapsto (\mathfrak{Q}(x+y),\mathfrak{Q}(x),\mathfrak{Q}(y))$, which is continuous because $m$ is and the operations of the module are, with the continuous map $(u,v,w)\mapsto\tfrac12(u-v-w)$ of $F^{3}$ into $F$. A composite of continuous maps is continuous.

Remark (the converse is the definition). For a symmetric bilinear form $B$, the diagonal $\mathfrak{Q}(x)=B(x,x)$ is a quadratic form, and it is continuous whenever $B$ is; the statement of the proposition is that the continuity of $m$ alone implies that of $B$, so in the layer the quadratic form is the primitive datum and the bilinear form a derived one. In characteristic $2$ the polarisation identity is not available and a quadratic form is not determined by a symmetric bilinear form; the corpus excludes that case, as Quadratic Forms and Polarisation, §The Failure of Polarisation records.

Degeneracy, Rank and the Radical

Definition. The radical of a continuous quadratic form is the subspace $\operatorname{rad}(m)=\{v\in V : B(v,w)=0 \text{ for every } w\in V\}$ of Bilinear Forms, §Rank and the Radical; the form is non-degenerate when the radical is zero. The rank is the codimension of the radical when $V$ is finite-dimensional over a field, and the form is degenerate otherwise.

Proposition. Let $m$ be a continuous quadratic form on a Hausdorff module over a Hausdorff field. Then the radical is a closed subspace of $V$.

Proof. For each $w\in V$ the map $\varphi_{w}(v)=B(v,w)$ is continuous, by the continuity of $B$ and that of the pairing with the fixed vector $w$; its kernel is closed. The radical is the intersection $\bigcap_{w\in V}\ker\varphi_{w}$, and an intersection of closed sets is closed.

The proposition is the reason the radical is a topological object and not only an algebraic one, and it is what allows the quotient by the radical to be formed as a topological space in the finite-dimensional case; the quotient by the radical of a continuous form is treated in Topological Clifford Algebras, §The Closure of Zero and the Hausdorff Quotient, where it is compared with the closure of zero.

The Space of Continuous Forms

The Uniform Norm

Definition. Let $V$ be a normed space over $\mathbb{R}$ or $\mathbb{C}$ and let $m$ be a continuous quadratic form on $V$. Its uniform norm is

$$ \lVert m\rVert = \sup\{\lvert \mathfrak{Q}(v)\rvert : \lVert v\rVert\le 1\}, $$

and the operator norm of a continuous bilinear form is $\lVert B\rVert = \sup\{\lvert B(x,y)\rvert : \lVert x\rVert\le1,\ \lVert y\rVert\le1\}$. The set of continuous quadratic forms on $V$ is written $\mathfrak{Q}(V)$ and the set of continuous symmetric bilinear forms $\operatorname{sym}(V)$.

These are the norms of Normed and Banach Spaces, read on the forms; the supremum is finite exactly when the form is bounded, and by Bounded Operators on a Topological Vector Space a quadratic form on a normed space is continuous if and only if it is bounded. The norm of a form is a topological datum, and no norm is imposed on the algebra: this article carries no length on the elements beyond the one that $V$ already had.

Theorem (polarisation is an isometric isomorphism). Let $V$ be a normed space over $\mathbb{R}$. Recall that $B$ is the polar form of $m$ and $\mathfrak{Q}(x)=B(x,x)$. The map $m\mapsto B$ is a linear isomorphism $\mathfrak{Q}(V)\to\operatorname{sym}(V)$ and

$$ \boxed{\;\lVert m\rVert = \lVert B\rVert\;} . $$

Proof. The passage $m\mapsto B$ is additive and homogeneous: for $m_{1},m_{2}$ one has $B_{m_{1}+m_{2}}=B_{m_{1}}+B_{m_{2}}$ by the polarisation identity applied to $m_{1}+m_{2}$, and $B_{\lambda m}=\lambda B_{m}$ for the same reason. The inverse is $B\mapsto(x\mapsto B(x,x))$, which is additive and homogeneous as well, and the two are mutually inverse because $\mathfrak{Q}(x)=B(x,x)$ and $B$ is recovered from $m$ by the identity. For the norms, $\lvert \mathfrak{Q}(x)\rvert=\lvert B(x,x)\rvert\le\lVert B\rVert\lVert x\rVert^{2}$ gives $\lVert m\rVert\le\lVert B\rVert$. In the other direction use the parallelogram form of the identity,

$$ B(x,y) = \tfrac{1}{4}\bigl(\mathfrak{Q}(x+y)-\mathfrak{Q}(x-y)\bigr) , $$

and the parallelogram law of the norm: for $\lVert x\rVert\le1$ and $\lVert y\rVert\le1$,

$$ \lvert B(x,y)\rvert \le \tfrac14\lVert m\rVert\bigl(\lVert x+y\rVert^{2}+\lVert x-y\rVert^{2}\bigr) = \tfrac14\lVert m\rVert\cdot2\bigl(\lVert x\rVert^{2}+\lVert y\rVert^{2}\bigr) \le \lVert m\rVert , $$

using $\lVert u\pm v\rVert^{2}+\lVert u\mp v\rVert^{2}=2\lVert u\rVert^{2}+2\lVert v\rVert^{2}$. Taking the supremum over the unit ball gives $\lVert B\rVert\le\lVert m\rVert$.

Corollary (completeness). The spaces $\mathfrak{Q}(V)$ and $\operatorname{sym}(V)$ are isometric, and $\mathfrak{Q}(V)$ is complete when $V$ is a Banach space: a uniform limit of continuous quadratic forms is a quadratic form, the algebraic identities passing to the limit because the field is complete.

Proof. The isometry is the theorem, so the two are complete together. A bounded bilinear form on a Banach space $V$ is a continuous bilinear map, and the space of continuous bilinear maps $V\times V\to\mathbb{R}$ is complete for the operator norm by Bounded Operators on a Topological Vector Space, §The Space of Continuous Operators; the symmetric ones form a closed subspace, so $\operatorname{sym}(V)$ is complete, and so is $\mathfrak{Q}(V)$. A uniform limit of quadratic forms is quadratic because each of the three defining properties is a closed condition.

Remark (the complex case). Over $\mathbb{C}$ the same computation applies to symmetric complex bilinear forms and complex-valued quadratic forms, and the polarisation is again an isometry. What is not obtained this way is the positive definite Hermitian form, whose conjugate slot and positivity make it a sesquilinear object of the category Topology on Sesqualgebras with a degree-2 form; over $\mathbb{C}$ the quadratic form $\mathfrak{Q}(v)=\lVert v\rVert^{2}$ is not a complex quadratic form in the present sense, and the norm is owned by Topology on Linear Spaces. The two layers are distinguished by the slot, exactly as Topology on Sesqualgebras distinguishes the two products.

The Topology of the Space of Forms

Proposition. The evaluation map $\mathfrak{Q}(V)\times V\to\mathbb{R}$ is continuous, and for each $v$ the evaluation $m\mapsto \mathfrak{Q}(v)$ is continuous; the closed subspaces of $\mathfrak{Q}(V)$ defined by a finite system of linear conditions on the coefficients are the closed conditions of the layer, and positivity, definiteness and non-degeneracy are not among them.

Proof. The first statement is the definition of the uniform norm as a uniform convergence on the unit ball, and the second follows from the first by fixing $v$. For the last clause, positivity and definiteness are inequalities with a strict reversal of sign and are not given by linear conditions; whether they are closed or open is studied below, and the answer depends on the dimension.

Theorem (the non-degenerate forms and the signature). Let $V=\mathbb{R}^{n}$ with its Euclidean topology, let $\operatorname{sym}_{n}(\mathbb{R})$ be the space of symmetric matrices with the operator norm, and let $m$ be the quadratic form of the symmetric matrix $G$. Then the set of non-degenerate forms is open, and the signature is locally constant on it. Consequently the forms of a given signature form an open set, and the number of signatures is the number of connected components of the open set of non-degenerate forms.

Proof. The determinant $\det G$ is a polynomial, hence continuous, in the entries, so the set $\{\det G\neq0\}$ is open and its complement, the singular forms, is closed. On that open set the eigenvalues of $G$, which depend continuously on $G$ by The Spectra of Self-Adjoint Operators of the Form, never vanish, so each eigenvalue has a constant sign in a neighbourhood, and the signature $(n,m,n-n-m)$ is therefore locally constant. A locally constant integer-valued function on the open set is constant on each connected component, and different signatures define disjoint open sets.

Corollary (the definite forms). The positive definite forms are one of those components, open in $\operatorname{sym}_{n}(\mathbb{R})$, and the negative definite forms another; a form on the boundary of the positive cone has a zero eigenvalue and is degenerate.

Proof. Positive definiteness is the signature $(n,0,0)$, which is locally constant by the theorem, so it is open; the same for the negative signature. A form with a zero eigenvalue is degenerate by the criterion of the radical in the diagonal basis, and lies in the closed complement of the open set of non-degenerate forms.

The corollary is the topological form of Sylvester's law of inertia: the classification of Quadratic Forms and Polarisation, §Sylvester's Law of Inertia is a partition into finitely many open sets, and the partition is locally constant, so no continuous path in the space of non-degenerate forms can change the signature. This is the reason the topological Witt group of The Topological Witt Group and the Brauer–Wall Group is a discrete group of components.

The Radical and the Closure of Zero

Remark (the radical is not the closure of zero). The radical is a subspace of $V$ and the closure of zero is an ideal of the Clifford algebra; the two measure different defects and neither is contained in the other in general. For the zero form the radical is all of $V$, while the topological exterior algebra is Hausdorff and its closure of zero is zero; for the indiscrete topology on the algebra of a non-degenerate form the radical is zero while the closure of zero is the whole algebra. What the radical does control is the algebra: the ideal generated by the images of the radical in $\mathrm{Cl}(V,\mathfrak{Q})$ is the nilpotent factor of Degenerate Clifford Algebras and the Radical, §The Nilpotent Factor and the Rank, and the Clifford algebra is the anti-Hermitian tensor product of the Clifford algebra of the reduced space with the exterior algebra of the radical.

Proposition. Let $m$ be a continuous quadratic form on a Hausdorff module and let $N$ be the image of $\operatorname{rad}(m)$ in $\mathrm{Cl}(V,\mathfrak{Q})$. Then $N$ is contained in every maximal ideal of the algebra, and $N^{2}=0$; the algebra is Hausdorff when the form is non-degenerate and the filtration is Hausdorff.

Proof. A radical element $v$ satisfies $B(v,w)=0$ for every $w$, hence $vw=-wv$ for every $w$ and, taking $w=v$, $v^{2}=-v^{2}$, so $v^{2}=\mathfrak{Q}(v)=0$; the subspace generated by such elements has square zero and its elements are nilpotent, and a nilpotent ideal is contained in every maximal ideal of an associative algebra with unit, by the standard argument of Algebras. For the last statement, if the radical is zero then $V$ injects into the algebra and the filtration of the non-degenerate algebra is Hausdorff, so the intersection of the neighbourhoods of zero is zero.

Isometries and the Continuity of the Group

Definition. An isometry of $(V,\mathfrak{Q})$ is a linear map $T : V\to V$ with $\mathfrak{Q}(Tv)=\mathfrak{Q}(v)$ for every $v$, equivalently $B(Tx,Ty)=B(x,y)$ for all $x,y$, as in Bilinear Forms, §Isometries; the group of all isometries is $\mathrm{O}(V,\mathfrak{Q})$ and its determinant-one part is $\mathrm{SO}(V,\mathfrak{Q})$.

Proposition. Let $m$ be continuous and $T$ an isometry that is continuous. Then the transpose of $T$ with respect to the polar form exists in the sense that $B(Tx,y)=B(x,T' y)$ for a continuous $T'$, and the two conditions $m\circ T=m$ and $B\circ(T\times T)=B$ are equivalent.

Proof. The equivalence is algebraic and holds for every form: $\mathfrak{Q}(Tv)=B(Tv,Tv)$ if $B$ is preserved, and conversely the polarisation of $m\circ T=m$ gives the bilinear identity, both sides being the polar forms of quadratic forms that agree, by the injectivity in the theorem above. For the transpose, the map $y\mapsto B(Tx,y)$ is continuous, and if the form is non-degenerate and $V$ is a reflexive topological module the Riesz-type representation supplies $T'$; the general construction is that of the adjoint of an operator in Adjoints of Bounded Sesquilinear Operators, whose bilinear case is the transpose read here.

Remark (finite dimension). When $V$ is finite-dimensional and Hausdorff over a complete valued field, every linear map $V\to V$ is continuous, so the condition on $T$ is automatic and $\mathrm{O}(V,\mathfrak{Q})$ is exactly the algebraic isometry group of Bilinear Forms, with no further hypothesis. In infinite dimension a $m$-preserving linear map need not be continuous, and the continuous isometries are the ones that form the object of The Topological Orthogonal Group and the Spin Group; the set of them is closed in the space of continuous linear maps, since it is the intersection of the closed sets $\{T : \mathfrak{Q}(Tv)=\mathfrak{Q}(v)\}$ over $v$ in a spanning set, and it is a topological group.

Proposition (the non-degenerate forms are preserved by an isometry). Let $T$ be a continuous linear isometry of $(V,\mathfrak{Q})$. Then $T$ carries the radical into itself and induces an isometry of the reduced space $V/\operatorname{rad}(m)$ for the induced form.

Proof. If $v\in\operatorname{rad}(m)$ then $B(Tv,Tw)=B(v,w)=0$ for every $w$, so $Tv\in\operatorname{rad}(m)$; the induced map on the quotient is well defined and preserves the induced form, which is non-degenerate and is the form of the algebra's non-degenerate factor.

Summary

A continuous quadratic form on a topological module is a quadratic form in the algebraic sense that is also a continuous map; its polar form is then continuous, by the polarisation identity, so the quadratic form is the primitive datum of the layer and the symmetric bilinear form its derived one. The radical is closed when the base and the module are Hausdorff, and it is a subspace of $V$, not to be confused with the closure of zero of the Clifford algebra, which is an ideal; the radical controls the nilpotent factor of the algebra and the reduction to the non-degenerate case.

Over a normed space the uniform norm of a quadratic form and the operator norm of its polar form are the same number, so polarisation is an isometric isomorphism $\mathfrak{Q}(V)\to\operatorname{sym}(V)$; both spaces are complete when $V$ is a Banach space and the base is complete, and the evaluation is continuous. The non-degenerate forms on $\mathbb{R}^{n}$ are open, the signature is locally constant there, and the definite forms are the signature $(n,0,0)$ and $(0,n,0)$ components of that open set; the classification of Sylvester is therefore a partition into finitely many open and closed pieces, and no continuous path changes the signature.

An isometry is a linear map preserving $m$, and preserving $m$ is equivalent to preserving the polar form; in finite dimension every linear map is continuous and the algebraic isometry group is the whole story, while in infinite dimension the continuous isometries are a closed subgroup of the space of continuous linear maps and are the object of The Topological Orthogonal Group and the Spin Group. The Hermitian form of a complex base with a conjugate slot is a different object, sesquilinear and owned by Topology on Sesqualgebras with a degree-2 form, and the positive norm itself is owned by Topology on Linear Spaces.

Summary of Notation

symbol meaning
$F$ a topological field of characteristic not $2$
$V$ a topological $F$-module, a normed space where a norm is used
$m$ a continuous quadratic form, $\mathfrak{Q}(\lambda v)=\lambda^{2}q(v)$
$B(x,y)=\tfrac12(\mathfrak{Q}(x+y)-\mathfrak{Q}(x)-\mathfrak{Q}(y))$ the polar form, continuous, symmetric, with $\mathfrak{Q}(x)=B(x,x)$
$\operatorname{rad}(m)$ the radical, closed in a Hausdorff module
$\mathfrak{Q}(V)$, $\operatorname{sym}(V)$ the continuous quadratic forms and continuous symmetric bilinear forms
$\lVert m\rVert$, $\lVert B\rVert$ the uniform norm and the operator norm, with $\lVert m\rVert=\lVert B\rVert$
$m\mapsto B$ the isometric isomorphism of polarisation
$\mathrm{O}(V,\mathfrak{Q})$, $\mathrm{SO}(V,\mathfrak{Q})$ the isometry group and its determinant-one part
signature locally constant on the non-degenerate forms over $\mathbb{R}^{n}$

Further Reading

  • John B. Conway, A Course in Functional Analysis, 2nd edition (Springer, 1990), for bounded quadratic forms, polarisation and the completeness of the space of operators.
  • Nicolas Bourbaki, Topological Vector Spaces (Springer, 1987), for the uniform structure of a normed space and the topology of uniform convergence on bounded sets.
  • Winfried Scharlau, Quadratic and Hermitian Forms (Springer, 1985), for polarisation over a general base, the radical and the reduction of a degenerate form.
  • T. Y. Lam, Introduction to Quadratic Forms over Fields (American Mathematical Society, 2005), for the algebraic classification and Sylvester's law of inertia.
  • John Milnor and Dale Husemoller, Symmetric Bilinear Forms (Springer, 1973), for the space of forms, its invariants and the Witt ring.
  • Tosio Kato, Perturbation Theory for Linear Operators, 2nd edition (Springer, 1976), for the continuity of the eigenvalues of a symmetric matrix and the local constancy of the signature.