# Conjugation of root subgroups by the Weyl group

Fix a field $$k$$ of characteristic zero, and let $$G$$ be a connected reductive algebraic $$k$$-group of isotropic rank $$\ge 1$$. Fix a maximal $$k$$-split torus $$S$$, and let $$\Phi_k$$ be the relative root system of $$G$$ with respect to $$S$$. Assume that $$\Phi_k$$ is reduced and irreducible.

By Theorem 2 of Petrov and Stavrova, for every root $$\alpha \in \Phi_k$$ there is an embedding $$X_\alpha:V_\alpha \to G$$ where $$V_\alpha$$ is a vector group (EDIT: vector group scheme), and the image of $$X_\alpha$$ is the root subgroup $$U_\alpha$$. Given $$v \in V_\alpha(k)$$, $$v \neq 0$$, associated to $$X_\alpha(v) \in U_\alpha(k)$$ is a "Weyl group element" $$w_\alpha(v)$$, using Lemma 1.3 of Deodhar. If $$G$$ is a Chevalley group, these are the elements denoted $$w_\alpha(t)$$ in Steinberg, and $$w_\alpha(1)$$ is literally a representative of a Weyl group element. Steinberg lists some relations, of interest here is (R7), which says that for any two roots $$\alpha,\beta \in \Phi_k$$ and $$v \in k$$, $$w_\alpha(1) \cdot X_\beta(v) \cdot w_\alpha(1)^{-1} = X_{w_\alpha \beta} \Big( c_{\alpha,\beta} v\Big)$$ This says that the action of the (lifts of) Weyl group elements on $$G(k)$$ by conjugation corresponds to the action of the Weyl group on $$\Phi_k$$; conjugating the root subgroup $$U_\beta$$ by a Weyl group representative $$w_\alpha(1)$$ takes it to the root subgroup $$U_{w_\alpha \beta}$$. The coefficient $$c_{\alpha,\beta}$$ is just a sign $$\pm 1$$ depending on $$\alpha$$ and $$\beta$$.

I have done a variety of explicit computations in quasi-split groups, and found in every case that $$w_\alpha(u) \cdot X_\beta(v) \cdot w_\alpha(u)^{-1} = X_{w_\alpha \beta} \Big( c_{\alpha,\beta}(u,v) \Big)$$ for some function $$c_{\alpha,\beta}:V_\alpha(k) \times V_\beta(k) \to V_{w_\alpha \beta}(k)$$. Steinberg's (R7) says that in the split case $$c_{\alpha,\beta}(1,v) = \pm v$$. In the preprint (reference 3, equation 3) we give a version of this for a class of quasi-split special unitary groups, where $$c_{\alpha,\beta}(1,v) = \pm v$$ or $$\pm \overline{v}$$, with the bar denoting a Galois automorphism of order 2. I have also done computations in a quasi-split special orthogonal group, where some more complicated functions $$c_{\alpha,\beta}$$ arise.

Question 1: Why does the conjugation on the LHS above always end up in $$U_{w_\alpha \beta}$$, even in non-split cases? I suspect this is an obvious consequence of a Bruhat decomposition, but I don't understand that as well as I should.

Question 2: Is there a known generalization of this relation to reductive isotropic groups, or perhaps just for quasi-split groups?

1. Deodhar, Vinay V., On central extensions of rational points of algebraic groups, Am. J. Math. 100, 303-386 (1978). ZBL0392.20027.2.
2. Petrov, V.; Stavrova, A., Elementary subgroups of isotropic reductive groups., St. Petersbg. Math. J. 20, No. 4, 625-644 (2009); translation from Algebra Anal. 20, No. 4, 160-188 (2008). ZBL1206.20053.
3. Rapinchuk, I.; Ruiter, J., On abstract homomorphisms of some special unitary groups, arXiv:2107.07351, preprint.
4. Steinberg, Robert, Lectures on Chevalley groups, University Lecture Series 66. Providence, RI: American Mathematical Society (AMS) (ISBN 978-1-4704-3105-1/pbk; 978-1-4704-3631-5/ebook). xi, 160 p. (2016). ZBL1361.20003.
• The relative root groups are generated by the absolute root groups with the right restriction, with a predictable modification for multipliable relative roots. The relative Weyl group acts on absolute roots and preserves the restriction. Are you asking for something more subtle than that? EDIT: Oh, I see that your $V_\alpha$ is not the root group. So your worry is precisely about multipliable relative roots, right? Sep 7, 2021 at 0:47
• Notice, by the way, that (as far as I can tell—the generality is more than that in which I am used to working) the embedding $V_\alpha \to G$ is an embedding just of schemes, not of groups. (You didn't say otherwise, but I misread at first.) Sep 7, 2021 at 0:51
• Wait, one more comment: I see now you assume the relative root system is reduced, so multipliable roots are not a problem. So once again I'm not sure where the subtlety is …. Sep 7, 2021 at 0:56
• I am aware that $V_\alpha \to G$ is an embedding of group schemes, though I did not mention it. I have not seen other people phrase it as such, but in all the cases I have studied, $V_\alpha$ is essentially the root space $\mathfrak{g}_\alpha$, and $X_\alpha:V_\alpha \to G$ is just the restriction of the exponential map $\operatorname{exp}:\mathfrak{g} \to G$. Sep 7, 2021 at 2:13
• The Deodhar reference was incorrect, but I have fixed it. One of his papers has the exact same title as one of his thesis; I meant to refer to the thesis. Sep 8, 2021 at 21:22

Let me use $$a$$ and $$b$$ for relative roots, so that I can later switch to $$\alpha$$ and $$\beta$$ for absolute roots.

If $$b$$ is a non-multipliable root, then, as you have said, $$V_b$$ is the $$b$$-root space, and $$X_b$$ is an exponential-type map. Specifically, it is the unique group homomorphism $$V_b \to G$$ whose derivative at the identity is the inclusion of the $$b$$-root subspace of $$\operatorname{Lie}(G)$$. It can be described ‘explicitly’, for small values of explicitly, as $$v \mapsto \prod X_\beta(v_\beta)$$, where $$\beta$$ runs over the absolute roots whose restriction to $$S$$ is $$b$$, and $$v = \sum v_\beta$$.

If $$a$$ is also non-multipliable, then we have that $$w_a(u)$$ is the commuting product $$\prod w_\alpha(u_\alpha)$$, where $$\alpha$$ runs over the absolute roots whose restriction to $$S$$ is $$a$$, and $$u = \sum u_\alpha$$. (Actually, now that I think about it, I might already be assuming $$G$$ quasi-split here.) In particular, $$w_a(u)w_a(1)^{-1}$$ equals $$\prod \alpha^\vee(u_\alpha)$$ (or maybe the inverse of this, depending how things are normalised; I didn't check).

You have already observed that $$\operatorname{Int}(w_a(1))\bigl(X_b(v)\bigr)$$ equals $$\prod X_{w_a\beta}\bigl(c_{a\beta}\operatorname{Ad}(w_a(1))v_\beta\bigr)$$, where $$c_{a\beta} = \prod c_{\alpha\beta}$$. Now suppose that $$G$$ is quasi-split. Then the set of absolute roots $$\beta$$ restricting to $$b$$ is a Galois orbit, and it is clear that $$\beta \mapsto c_{a\beta}$$ is constant on Galois orbits, so $$\operatorname{Int}(w_a(1))\bigl(X_b(v)\bigr)$$ equals $$X_{w_a b}\bigl(c_{a b}\operatorname{Ad}(w_a(1)v)\bigr)$$, where $$c_{a b}$$ is the common value of $$c_{a\beta}$$. Now just conjugate by $$\prod \alpha^\vee(u_\alpha)$$ to finish.

• The usual definition of the $w_\alpha(t)$ requires that $t \ne 0$, so in our setting we must restrict to an open subset of $V_a$. I tried to check whether this is consistent with your pointer to the definition of $w_a$, but I cannot find it in Deodhar. However, this is how it "should" be defined, so I will proceed that way and correct if necessary once I know where to look. Sep 8, 2021 at 0:45
• I am confused by your placement of parentheses in the expression $w_a(1) X_b(w_a(1)^{-1} v)$ both times. Why have you not written this as $w_a(1) X_b(v) w_a(1)^{-1}$? Sep 9, 2021 at 0:28
• Can you provide any reference on the fact that $X_b(v) = \prod_\beta X_\beta(v_\beta)$, or that $w_\alpha(u) = \prod w_\alpha(u_\alpha)$? I believe the first equation, but I am skeptical of the second. In particular, in concrete calculations in special orthogonal groups, I have found that the elements $w_\alpha(u)$ associated with multidimensional root groups are not monomial matrices. But if what you say is true, it is a product of the absolute versions $w_\alpha(u_\alpha)$ which are monomial matrices (at least for absolute type $A_n$, which I think these special orthogonal groups are). Sep 9, 2021 at 0:32
• Re, in both cases, because root groups commute (since we're in the non-multipliable case). $X_b$ is the unique homomorphism whose derivative is the inclusion, and $v \mapsto \prod X_\beta(v_\beta)$ has that derivative. $w_a(u)$ is the unique element of $U_{-a}u U_{-a} \cap N$. Since $U_{-a}u U_{-a}$ equals $\prod U_{-\alpha}u_\alpha U_{-\alpha}$ (now also using quasisplit), it contains $\prod w_\alpha(u_\alpha)$. Sep 9, 2021 at 0:55
• Re again, special orthogonal groups are of type $B_n$ (if ($2n + 1$)-dimensional orthogonal space) or $D_n$ (if $2n$-dimensional). Unitary groups are of type $A_n$. Sep 9, 2021 at 0:56