7
$\begingroup$

Assume that $G$ is a Lie group with Lie algebra $\mathfrak{g}$. We fix an invariant Riemannian metric on $G$ and fix its corresponding $LC$ connection.

Consider the natural right action of $G$ on its Lie algebra $\mathfrak{g} \simeq \{X \in \chi^{\infty}({G}) \mid R_{g}^{*} X=X\}$, the space of smooth vector fields which are invariant under right multiplications.

In fact the right action is defined as follows:

For $g\in G$ and $X\in \mathfrak{g}$ define $X.g=L_{g}^{*} X$ where $L_{g}$ is the left multiplication by $g$.

So there is a natural (component wise) action of $G$ on $\mathfrak{g} \times \mathfrak{g}$

A smooth function $f:G \to \mathbb{R}$ is called $G$-invariant if $f(g^{-1}hg)=f(h)$ for all $g,h \in G$. For example $Det: Gl(n, \mathbb{R}) \to \mathbb{R}$.

A $2$- linear map $T: \mathfrak{g} \times \mathfrak{g} \to \mathbb{R}$ is $G$-invariant if $T((v,w).g)=T(v,w)$

Example The $2$-linear map $tr(u)tr(v)-tr(uv)$ defined on $M_{n}(\mathbb{R}) \times M_{n}(\mathbb{R})$ is a $Gl(n, \mathbb{R})$ invariant map, with the natural (conjugate) action of $Gl(n, \mathbb{R})$ on its Lie algebra $M_{n}(\mathbb{R})$ as described above.

Recall that for a Riemannian manifold with the corresponding $LC$ connection $\nabla$, the Hessian of a function $f$ defined on the manifold, is a two linear map on the tangent space with the formula $Hess (f).(V, W)=\nabla ^ {\nabla f}_{V}.W$

Question: Let $G$ be a Lie group and $f:G \to \mathbb{R}$ be a $G$-invariant smooth function. Is its Hessian $Hess(f)$ a $G$-invariant $2$-linear map on the Lie algebra $\mathfrak{g}$ of $G$?

This question is motivated by the following post.

Is there an explicit formula for the hessian of "Determinant"?

$\endgroup$
8
$\begingroup$

Yes.

In what follows, I use standard notation for the derivative; see, e.g., $\S$2.3 of Banach Spaces and Differential Calculus (Chapter 2) of the book referenced below.

First Derivative of $f$.

As the OP stated, a function $f: G \to \mathbb{R}$ is $G$ invariant means that $$ f(A) = f(g A g^{-1}) \quad \forall A, g \in G \tag{$\star$} $$ If we differentiate this relation once we obtain $$ D f(A) \cdot B_1 = \frac{d}{d\lambda} \left. f( g (A+\lambda B_1) g^{-1} ) \right|_{\lambda=0} = D f( g A g^{-1} ) \cdot g B_1 g^{-1} \tag{$\star \star$} $$ where $B_1 \in T_A G$. In the special case of the determinant on $G = Gl(n,\mathbb{R})$, this is simply saying that: $$ D \det(A) \cdot B_1 = \det(A) \operatorname{trace}(A^{-1} B_1) = \det( g A g^{-1} ) \operatorname{trace}(g A^{-1} g^{-1} g B_1 g^{-1} ) $$ which follows from basic properties of the determinant and trace of a matrix.

Second Derivative of $f$.

Now differentiate ($\star \star$) to obtain: $$ D^2 f(A) \cdot (B_1, B_2) = D^2 f( g A g^{-1} ) \cdot ( g B_1 g^{-1}, g B_2 g^{-1} ) \tag{$\star \star \star$} $$ where $B_1, B_2 \in T_A G$. In the special case of the determinant on $G = Gl(n,\mathbb{R})$, \begin{align*} D^2 \det(A) \cdot (B_1, B_2) &= \frac{d}{d \lambda} \left. D \det(A+\lambda B_2) \cdot B_1 \right|_{\lambda=0} \\ &= \frac{d}{d \lambda} \left. \det (A + \lambda B_2) \operatorname{trace} ( (A+\lambda B_2)^{-1} B_1 ) \right|_{\lambda=0} \\ &= \det(A) \operatorname{trace} (A^{-1} B_2) \operatorname{trace}(A^{-1} B_1) - \det(A) \operatorname{trace}(A^{-1} B_2 A^{-1} B_1 ) \end{align*} which is clearly $G$-invariant in the sense of relation ($\star \star \star$).

Higher Derivatives of $f$.

This result for the first and second derivative seems to be true for higher derivatives of a $G$-invariant function. In other words, differentiation preserves $G$-invariance in the sense given above.

Reference

For background info see, e.g., An Introduction to Lie groups Chapter 5 of

Jerrold E. Marsden, Tudor Ratiu, and Ralph Abraham. Manifolds, Tensors, and Applications. Second Edition. Vol. 75. Springer Science & Business Media, 2012.

$\endgroup$
  • 1
    $\begingroup$ Dear Nawaf Thank you very much for your complete and interesting answer. (As I gave +1 to your answer, already) $\endgroup$ – Ali Taghavi Sep 21 '16 at 20:07

Your Answer

By clicking "Post Your Answer", you acknowledge that you have read our updated terms of service, privacy policy and cookie policy, and that your continued use of the website is subject to these policies.

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