In characteristic not $2$, the Theorem of Cartan-Dieudonné states:

  1. [Grove, Theorem 6.6]: Let $q$ be a nondegenerate symmetric quadratic form of dimension $n$ in characteristic not $2$. Then every element in the orthogonal group $O(q)$ can be written as a product of at most $n$ reflections.

In a nutshell: I am looking for a similar statement but over a field in characteristic $2$ and in particular in odd dimension. At the end I will make the statement that I would like to have more precise.

As I am new to quadratic forms, I got confused by different notions of non-degeneracy, so I want to fix the terminology; I take the one from [Grove]. To a quadratic form $q$ on $V$ one can associate a bilinear form $b$ on $V$ via $b(v,w)=q(v+w)-q(v)-q(w)$. In characteristic $2$, several quadratic forms can give the same bilinear form (e.g. $y^2+xz$ and $xz$). Let $\mathrm{rad}_b(V)=\{w\in V\mid b(v,w)=0 ~\forall v\in V\}$. If $\mathrm{rad}_b(V)\neq\{0\}$, then $q$ is called

  • degenerate if the characteristic is not $2$,
  • defective if the characteristic is $2$.

In characteristic $2$ and dimension odd, all quadratic forms are defective! In characteristic $2$, we say that $q$ is regular if $q(v)\neq0$ for all non-zero $v\in\mathrm{rad}(V)$. (In [Connors], regular is called nondegenerate).

Example: The two quadratic forms $q_1=y^2+xz$ and $q_2=xz$ on $V=k^3$, where $k$ is a field of characteristic $2$, give the same bilinear form and they both have $\mathrm{rad}(V)=\langle(0,1,0)\rangle$. So they are defective. While $q_1$ is regular, $q_2$ is not.

From now on, assume that $k$ is a field in characteristic $2$. Here transvections play the role of reflections. I have found the following statements that go in the right direction:

  1. [Grove, Theorem 14.16]: Assume that $q$ is regular and non-defective (see the assumption on p. 129 which I believe is still taken in Theorem 14.16). Then $O(q)$ is generated by transvections, except for one example over the field $\mathbb{F}_2$ where $V$ is $4$-dimensional.

The above theorem can not be applied in odd dimension. The next statement is exactly what I want, but it is stated only for dimension $3$:

  1. [Connors, Proposition 3.2]: Let $q$ be a defective, anisotropic regular quadratic form in dimension $3$. Then every element of $O(q)$ is the product of at most $2$ transvections.

I am looking for a reference of the following statement (if true): Let $V$ be a vector space over any field in characteristic $2$, and let $q$ be a regular, anisotropic quadratic form on $V$ of odd dimension $n$ (i.e. $q$ defective). Then every element in the orthogonal group $O(q)$ can be written as a product of (at most $n$) transvections.

[Grove] : Larry C. Grove, "Classical groups and geometric algebra", 2002

[Connors] : Edward A. Connors, ''The Structure of $O'(V)/DO(V)$ in the Defective Case'', 1973


I suggest reading the result 6.2.17 in the book written by A.J.Hahn and O.T.O'Meara called "The Classical Groups and K-Theory". The result is a general version of the Cartan-Dieudonné-Scherk theorem (which implies the Cartan-Diuedonné theorem). The result is stated in the language of "Quadratic spaces over form rings". The case of the defective orthogonal groups is a particular case of the general "unitary groups" $U_n(V)$. More precisely, in the book you find

6.2.17 Theorem. Suppose $U_n(V)$ is not a symplectic group. If $J=\mathrm{id}_R$, assume that $R$ is not $F_2$ and if $J\neq \mathrm{id}_R$, assume $R$ is not $F_4$. Let $\sigma$ in $U_n(V)$ be non-trivial and let $S$ be its residual space.

(i) If $\sigma$ is not totally isotropic, $\sigma$ is a product of $\mathrm{dim} S$ symmetries, but no fewer.

(ii) If $\sigma$ is totally isotropic, $\sigma$ is a product of $\mathrm{dim} S+2$ symmetries, but no fewer.

When $J=\mathrm{id}_R$ you can get the defective orthogonal groups and so any $\sigma$ will be a product of "symmetries" (i.e. transvections in the case of the defective orthogonal group).

  • $\begingroup$ Thanks for stating the result! I accepted this answer, but I'd still be interested in a reference that uses the language of quadratic form over vector space. $\endgroup$ – JNS Mar 25 at 9:26

For a reference, you could use Theorem I.5.1 from C. Chevalley, The Algebraic Theory of Spinors, Columbia University Press, New York, 1954. (This also appears in volume 2 of his collected works.) The result is the following (rephrased):

The orthogonal group $O(V,q)$ is generated by reflections unless the underlying field has only 2 elements, $\dim V = 4$ and $q$ is hyperbolic.

In fact, if you only care about anisotropic quadratic forms, then the proof of Cartan-Dieudonné, in any characteristic, is much easier than the general case, and very elegant. (I learned this proof from Richard Weiss but I can't remember the orginal source, unfortunately).

Let $f$ be the bilinear form corresponding to $q$. For any $c \in V$ with $c \neq 0$, let $\pi_c$ be the corresponding reflection $$ \pi_c(v) := v - \frac{f(v,c)}{q(c)} c .$$ Now let $\varphi \in O(V,q)$ be arbitrary, and let $F = \operatorname{Fix}_V(\varphi)$. We will show by induction on the codimension of $F$ that $\varphi$ is generated by reflections. If $\operatorname{codim} F = 0$, then $F = V$ so $\varphi = 1$. Now assume that $\operatorname{codim} F > 0$ and let $v$ be an element of $V \setminus F$, so $w := \varphi(v) - v \neq 0$. We claim that $\pi_w$ fixes $F$ and maps $v$ to $\varphi(v)$.

First, notice that by applying $q$ on $\varphi(v) = v + w$, we get $q(w) = - f(v,w)$, so indeed $\pi_w(v) = v + w = \varphi(v)$. Next, if $u \in F$, then $$ f(u,w) = f(u,v) - f(u,\varphi(v)) = f(u,v) - f(\varphi(u),\varphi(v)) = 0, $$ so $\pi_w$ fixes $u$. This proves our claim. We conclude that $\varphi^{-1} \circ \pi_w$ fixes the space $\langle F, u \rangle$, which has codimension one less than $F$, so we can apply induction, and we are done.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.