Let $G= SU(n)$ and let $\mathbb{T}$ be the maximal torus in $G$ given by diagonal matrices. We have $$ H^*(G,\mathbb{Q}) \cong \Lambda_{\mathbb{Q}}[x_3, x_5, \dots, x_{2n-1}] \ . $$ Now consider the Weyl map $$ p \colon G/\mathbb{T} \times \mathbb{T} \to G \quad , \quad ([g],z) \mapsto gzg^{-1}\ . $$ The induced map in rational cohomology $$ p^* \colon H^*(G,\mathbb{Q}) \to H^*(G/\mathbb{T}, \mathbb{Q}) \otimes H^*(\mathbb{T},\mathbb{Q}) $$ is injective. In fact, if we restrict the codomain to fixed points with respect to a certain action of the Weyl group $W$ it becomes an isomorphism (see for example Reeder, On the Cohomology of Compact Lie Groups). The cohomology ring $H^*(\mathbb{T},\mathbb{Q})$ is isomorphic to another exterior algebra and there are also explicit descriptions of $H^*(G/\mathbb{T},\mathbb{Q})$ (see Reeder's paper again for a reference).

Is there a formula describing $p^*(x_{2i+1})$ in terms of any set of natural generators for the codomain?


First, let me fix generators for $H^*(SU(n))$ and $H^*(SU(n)/\mathbb T)$: For the first, consider the vector bundle on $\Sigma SU(n)$ with clutching map $\operatorname{id}_{SU(n)}$, i.e. with classifying map $f_n\colon\Sigma SU(n)\simeq \Sigma\Omega BSU(n)\to BSU(n)$, and let $\Sigma x_{2i-1} = f^*c_i$. For the second, let $\pi_k\colon \mathbb T\subset U(1)^k\to U(1)$ be the projection to the $k$-th coordinate and $L_k = SU(n)\times_{T,\pi_k} \mathbb C$ be the associated line bundle, and set $y_k = c_1(L_k)$; since $\bigoplus_{k=1}^n \pi_k$ is the restriction of the defining representation of $SU(n)$ to $\mathbb T$, the sum $E := \bigoplus_{k=1}^n L_k$ is trivial, so that all symmetric polynomials in the $y_k$ vanish since they can be expressed via Chern classes of $E$. The induced map $Q[y_1,\dots,y_n]/Q[y_1,\dots,y_n]^{S_n}\to H^*(SU(n)/\mathbb T)$ is an isomorphism; a basis for this cokernel is given by the monomials $y_1^{\alpha_1}\dots y_n^{\alpha_n}$ with $0\le \alpha_k < k$.

Now think of $p:G/\mathbb T\times\mathbb T\to SU(n)$ as an automorphism $\phi$ of $E\boxtimes\underline{\mathbb C}$; it splits as $\phi = \phi_1\oplus\dots\oplus \phi_n$, where $\phi_k = \operatorname{id}_{L_k}\boxtimes \pi_k$ is the automorphism of $L_k\boxtimes \underline{\mathbb C}$ obtained as the external tensor product of the identity of $L_k$ with the projection $\pi_k:\mathbb T\subset U(1)^k\to U(1)$.

Consider the clutching construction $(E\boxtimes\underline{\mathbb C})^\phi\to G/\mathbb T\times\mathbb T\times S^1$ which is obtained by gluing the two ends of $E\boxtimes\underline{\mathbb C}$ together using $\phi$. Denote by $z_k = \pi_k^*([U(1)]), u = [S^1]$ the obvious cohomology classes in degree $1$, and itdentify them and the $y_k$ with their image in the product. An easy argument shows that $c_1(\mathbb C^{\pi_k}) = uz_k\in H^2(\mathbb T\times S^1)$. By naturality, we have \begin{align*} (E\boxtimes\underline{\mathbb C})^\phi &\cong \bigoplus_{k=1}^n(L_k\boxtimes\underline{\mathbb C})^{\phi_k}\\ c_1\big((L_k\boxtimes\underline{\mathbb C})^{\phi_k}\big) &= c_1(L_k) + c_1(\underline{\mathbb C}^{\pi_k}) = y_k + uz_k\\ c\big((E\boxtimes\underline{\mathbb C})^\phi\big) &= \prod_{k=1}^n c\big((L_k\boxtimes\underline{\mathbb C})^{\phi_k}\big)\\ &= \prod_{k=1}^n (1 + y_k + uz_k)\\ &= \underbrace{c(E)}_{= 0} + u\sum_{k=1}^n z_k \sum_{S\subset \{1,\dots,n\}\smallsetminus\{k\}}\prod_{l\in S} y_l \end{align*}

It is no surprise that this expression vanishes for $u=0$ since we can use the chosen trivialization of $E\boxtimes \underline{\mathbb C}$ to descend $(E\boxtimes \underline{\mathbb C})^\phi$ to $\Sigma(G/\mathbb T\times \mathbb T)$, and the resulting vector bundle has classifying map $f_n\circ \Sigma p$. Chasing through the definitions, we see that $$ p^* x_{2i-1} = \sum_{k=1}^n\Big[\sum_{\substack{S\subset \{1,\dots,n\}\smallsetminus\{k\}\\|S| = i - 1}}\prod_{l\in S} y_l\Big]\otimes z_k\ . $$


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.