Constructing symmetric invariants of the exceptional simple Lie algebra as restrictions

Let $\mathfrak{g}$ be a complex simple Lie algebra with maximal torus $\mathfrak{h}$, Weyl group $W$. The adjoint representation $\operatorname{ad} : \mathfrak{g} \rightarrow \mathfrak{gl(g)}$ extends by derivations to a representation of $\mathfrak{g}$ in $S(\mathfrak{g})$. We identify $S(\mathfrak{g})$ and $\mathbb{C}[\mathfrak{g}]$ as $\mathfrak{g}$-modules via the Killing form. The most famous description of $S(\mathfrak{g})^{\mathfrak{g}}$ goes as follows: the Chevalley restriction theorem states that the restriction map $\mathbb{C}[\mathfrak{g}]^\mathfrak{g} \rightarrow \mathbb{C}[\mathfrak{h}]^W$ is an isomorphism and the Chevalley-Sheppard-Todd theorem implies that the Weyl group invariants are a polynomial ring on $\operatorname{rank}(\mathfrak{g})$ generators.

I am interested in another approach to constructing the invariant rings via restriction. Take an $n$-dimensional vector space $V$ and set $\mathfrak{g} = \mathfrak{gl}(V)$. The characteristic polynomial $p(t, x)$ of an element $x \in \mathfrak{g}$ can be written as $t^n + \sum_{i=0}^{n-1} p_i(x) t^i$, and the coefficients $p_i$ form a complete set of algebraically independent generators for $S(\mathfrak{g})^{\mathfrak{g}}$. Now suppose that $\mathfrak{k} \subseteq \mathfrak{g}$ is a classical Lie algebra of type $\mathfrak{so}(V)$ or $\mathfrak{sp}(V)$. We can consider the restrictions $p_{i}\vert_{\mathfrak{k}}$ and in this case we have $p_{2i}\vert_{\mathfrak{k}} = 0$ for all $i$ and the remaining invariants form a complete set of algebraically independent generators for $S(\mathfrak{k})^\mathfrak{k}$, unless of course $\mathfrak{k}$ has type ${\sf D}$, in which case $p_{n-1} = P^2$ for some $P \in S(\mathfrak{g})$ known as the Pfaffian. In this case $p_{1}\vert_{\mathfrak{k}}, p_{3}\vert_{\mathfrak{k}},...,p_{n-3}\vert_{\mathfrak{k}}, P$ form a basic set of generators for the symmetric invariants. This is all very classical and probably goes back to Weyl.

What I would like to know is whether we can play the same game with exceptional Lie algebras. This is my question: for $\mathfrak{g}$ simple and exceptional, does there always exist a simple finite dimensional representation $\mathfrak{g} \rightarrow \mathfrak{gl}(V)$ such that some of the restrictions $p_0\vert_{\mathfrak{g}},...,p_{n-1}\vert_{\mathfrak{g}}$ are zero, whilst the others form a complete set of algebraically independent generators for $S(\mathfrak{g})^{\mathfrak{g}}$, perhaps after Pfaffing around with them a bit?

• Though the idea seems plausible enough, a big obstacle to treating the exceptional Lie algebras is that the smallest nontrivial irreducible representation of $E_8$ has dimension $n=248$. This is the adjoint representation, which may or may not be amenable to the sort of treatment you outline. On the plus side, various people have worked out structure constants for a Chevalley basis of $E_8$. On the minus side, it's still a very big representation. Maybe your approach works efficiently for type $G_2$? Here the smallest (faithful) representation occurs in dimension $n=7$. – Jim Humphreys Nov 28 '15 at 0:51
• @JimHumphreys thanks for the comment. The $G_2$ suggestion seems possible: $\mathfrak{g}_2 \subseteq \mathfrak{so}_7 \subseteq \mathfrak{gl}_7$ so we get $p_{2i}\vert_{\mathfrak{g}_2} = 0$. The remaining invariants have degrees 2, 4, 6. The basic invariants for $\mathfrak{g}_2$ have degrees 2, 6, so it would suffice to check that $p_{3}\vert_{\mathfrak{g}_2} = 0$ and $p_{1}\vert_{\mathfrak{g}_2}, p_{5}\vert_{\mathfrak{g}_2} = 0$ are algebraically independent. This'll require $\mathfrak{g}_2$-ology. Any chance of a general approach? Maybe $\mathfrak{g} \rightarrow \mathfrak{gl}(L(\omega_1))$? – Lewis Topley Nov 28 '15 at 11:58

For type $G_2$ with the minimal faithful representation we obtain $p_1(sh_1+th_2)=-2(t^2-3st+3s^2)$ and $$p_5(sh_1+th_2)=-4s^6+12s^5t-13s^4t^2+6s^3t^3-s^2t^4=-s^2(t-s)^2(t-2s)^2$$ from which you can already see the main problem - these polynomials become somewhat unwieldy as the degree increases. In this case we know that these two polynomials (in $k[s,t]$) are algebraically independent since otherwise $p_5$ has to be a multiple of $p_1^3$.
This doesn't quite answer your question, as in this case we have $p_3=(t^2-3st+3s^2)^2\neq 0$. My feeling is that for the exceptional types it is extremely unlikely that you will find a representation satisfying the condition you require (that all except $r={\rm rank}({\mathfrak g})$ of the polynomials $p_i$ are zero). On the other hand, I would expect that if you pick any faithful representation $(\rho,V)$ of a simple Lie algebra ${\mathfrak g}$ then the restrictions of $r={\rm rank}({\mathfrak g})$ of the basic invariants on $\mathfrak{gl}(V)$ to $\rho({\mathfrak g})$ are algebraically independent. (I probably should be able to see why this is true straight away.) Usually we should be able to take the same degrees as the degrees of the basic invariants on ${\mathfrak g}$. I would point out that the issue with the Pfaffian for $\mathfrak{so}_{2n}$ arises because ${\rm GL}_{2n}$ contains the full orthogonal group, which includes all of the outer automorphisms of $\mathfrak{so}_{2n}$. So in fact when we restrict ${\rm GL}_{2n}$-invariants on $\mathfrak{gl}_{2n}$ to $\mathfrak{so}_{2n}$ then we obtain elements of $k[\mathfrak{so}_{2n}]^{{\rm O}_{2n}}$, and this is a polynomial ring with homogeneous generators of degrees $2,4,\ldots, 2n$. I expect that what happens is: if $V$ is ${\mathfrak g}$-isomorphic to its twist via any outer automorphism then we get the same situation as you have for $\mathfrak{so}_{2n}\hookrightarrow\mathfrak{gl}_{2n}$; otherwise the restriction of the invariants on $\mathfrak{gl}(V)$ to ${\mathfrak g}$ gives us an exact copy of the invariants on ${\mathfrak g}$ (plus extra complications in type $D_4$). This happens, for example, for either 27-dimensional faithful representation for $E_6$.
For the higher rank cases GAP can't compute the characteristic polynomial directly as the memory required (e.g. for a 248 x 248 matrix involving 8 or 9 variables) is too great for it to handle. Magma might be able to do it, but it's not clear what value you can get out of running the computation - for example, if we take the adjoint representation in type $E_8$ and we compute ${\rm Trace}(\rho(x)^i)$ for $i=2,8,12,\ldots$ and $x=sh_1+\ldots +zh_8$ then already for $i=8$ this produces 3 pages of unreadable degree 8 polynomial in $k[s,t,\ldots ,z]$. We can then prove that this is independent of ${\rm Trace}(\rho(x)^2)$, but you still have a long way to go to prove by this route that the polynomials ${\rm Trace}(\rho(x)^i)$ are algebraically independent.
• I can see now it's quite unlikely to find a general result of the form first suggested. I like your suggestion that the restriction map $k[\mathfrak{gl}(V)]^{GL(V)} \rightarrow k[\mathfrak{g}]^{G}$ might often surject. Let $\mathfrak{h}$ be a max torus in $\mathfrak{g}$ and extend to a maximal torus $\mathfrak{h}'$ in $\mathfrak{gl}(V)$. Let $\mathfrak{g}' = \mathfrak{g} + \mathfrak{h}'$. Let $W$ and $W'$ be the Weyl groups of $GL(V)$ and $G'$ normalising $\mathfrak{h}'$. Note that $W' \cong N_G(\mathfrak{h})/H$. Applying Chevalley restriction to $GL(V)$, $G'$ and $G$. The question is now... – Lewis Topley Nov 30 '15 at 14:57
• (cntd) when does the composition $k[\mathfrak{h}']^W \hookrightarrow k[\mathfrak{h}']^{W'} \twoheadrightarrow k[\mathfrak{h}]^{W'}$ surject? This seems tricky since the Weyl groups generally don't embed in any reasonable way. – Lewis Topley Nov 30 '15 at 14:58
• The composition surjects precisely when: two elements of ${\mathfrak h}$ are $W$-conjugate if and only if they are $W'$-conjugate. – Paul Levy Dec 1 '15 at 16:23