TITLE: Local normal forms of covariantly constant  selfadjoint (1,1)-tensors 
QUESTION [7 upvotes]: Consider the pair $(g, L)$, where $g$ is a pseudo-Riemannian (i.e., 
nondegenerate and of arbitrary signature) metric and $L$ is an (1,1) $g$-selfadjoint tensor  field. 
Does there exist a local description of such pairs; for example in the form 
"in a certain coordinate system $g$ and $L$ are given by formulas ...". 
For me, the state of art is as follows: for Riemannian metrics, the answer was known to  classics: The existence of such $L$ implies  the local 
decomposition of $g$ into a direct product: $g= g_0+ ... + g_k$, where 
$g_0$  is flat (we can think therefore that $g_0= dx_1^2 +...+dx_m^2$ in a certain coordinate system) and each $g_i$ has  irreducible holonomy group. For such  metrics $g= g_0+ ... + g_k$, the tensor   $L$ also can be decomposed   into the product $L= L_0+...+L_k$; moreover, $L_0$ is given by  arbitrary symmetric matrix with constant entries  and other $L_i$ are proportional to identity with constant coefficients  (the coefficients depend on the component). 
For the pseudo-Riemannian metrics one can do the same splitting if $L$ has different eigenvalues; so the interesting case is when $L$ has one real eigenvalue or two conjugated complex eigenvalues. And  this case looks quite open for me; only the special case when 
$L^2= 0$ or $L^2= -1$ are known.
Does  anybody know more?

REPLY [5 votes]: Well, I don't have the complete answer, but then I don't think that a 'complete' answer is going to be simple.  For example, you haven't ruled out the case $L=0$, which amounts to giving a `normal form' for all pseudo-Riemannian metrics.  What is true, of course, is that the algebraic type of $(g,L)$ is the same at all points (assuming that $M$ is connected) and, as Willie Wong points out, you will at least have to classify those algebraic types as a starting point.
Of course, you can get complete answers in some algebraic types.  For example, when $L$ has a single real eigenvalue $r$ (which must be constant), you can assume that this eigenvalue is $0$ by replacing $L$ by $L-rI$, and the `least degenerate' case would presumably then be when $L^n=0$ but $L^{n-1}\not=0$. (I'm assuming that the dimension of the manifold is $n$.)  Then it is not difficult to show that there is a (local) frame field $e_1,\ldots,e_n$ such that $e_i = L^{i-1}e_1$ for $i = 2,\ldots,n$ and, moreover, such that $g(e_i,e_{n+1-i}) = 1$ with all other $g(e_i,e_j) = 0$. (You may have to replace $g$ by $-g$ in order to arrange this, but that's a trivial matter.)  Moreover, this frame field is unique up to replacing $e_i$ by $-e_i$ for all $i$.  Since $L$ is $g$-parallel, it follows that this frame field is also $g$-parallel.  In particular, the connection is flat and one can (locally) choose coordinates $x^i$ such that $dx^i(e_j) = \delta^i_j$.  In these coordinates, $g$ and $L$ have constant coefficients.
Added after Vladimir's comment:  Yes, it appears that the case of more than one Jordan block is more interesting.  I did a back-of-the-envelope calculation for the case of a $5$-dimensional manifold $M$ with a pair $(g,L)$ where $L$ is nilpotent with $2$ Jordan blocks, one of size $3$ and one of size $2$.  The result is that there are $4$ algebraic types possible ($2$ if you allow the replacement of $g$ by $-g$), and each of them exists as a $1$-parameter family of inequivalent types.  One member in each family is 'flat', i.e., $g$ and $L$ have constant coefficients in the appropriate coordinate system, but the others aren't flat, even though they are homogeneous (with a $7$-parameter family of $L$-preserving isometries).  Hmm.