This is indeed an interesting question. Define a *quasi-primary* state to be one that is annihilated by $L_1$ i.e., it is a highest-weight vector (state) of the $sl(2)$ subalgebra of the Virasoro algebra. Consider a generic Virasoro highest weight vector, $|h\rangle$, of weight $h$. The Verma module is constructed by acting on $|h\rangle$ by all combinations $L_{-n}$ for $n=1,2,3,\ldots$. By generic I mean none of the descendants are also highest weight vectors (aka null vectors). This is done for simplicity. It follows that the Verma module is irreducible.

Now we wish to decompose this Verma module into irreps of $sl(2)$. The first quasi-primary appears at level $0$ is $|h\rangle$ and its descendants are $L_{-1}^n|h\rangle$ for $n\geq1$. There is no quasi-primary at level 1. At level 2, there is $L_{-2}|h\rangle$ in addition to $(L_1)^2|h\rangle$. But it is not quasi-primary. But a simple calculation shows that $|\phi\rangle:=\left(L_{-2}-\tfrac32 (L_{-1})^2\right)|h\rangle$ is a quasi-primary. This along with its descendants $(L_{-1})^n |\phi\rangle$ is the second irrep of $sl(2)$. One can continue in this fashion at each level and look for quasi-primaries i.e., states annihilated by $L_1$. The statement in the reference mentioned by Trimok, if I understood it correctly, states that at level $(n+1)$, there are descendants that appear by the action of $L_{-1}$ on states at level $n$ and the remaining are necessarily quasi-primaries.

Recall that at level $n$, there are $p(n)$ states where $p(n)$ is the number of partitions of $n$. So it follows that there must be $(p(n+1)-p(n))$ quasi-primaries at level (n+1). I suspect that the proof is not too hard but I have not worked it out. Each quasi-primary is the highest weight vector for an infinite dimensional irrep of sl(2).

This post imported from StackExchange Physics at 2014-03-31 10:14 (UCT), posted by SE-user suresh