Pushforwards of Line Bundles and Stability I recently finished reading this paper, and was wondering about a couple of things relating to theorem 1, which says that for any curve X there is a curve Y and f:Y->X such that pushforward is a dominant rational map from the Jacobian of Y to the moduli of semistable vector bundles on X (with numerical invariants fixed to make things more definite.)  So I had two questions:
1) Given a morphism of curves f:Y->X, is there a good characterization of the line bundles L on Y with f_*(L) semistable (or not semistable, equivalently)?
2) Given a morphism of curves f:Y->X, is there a good characterization of which semistable bundles are in/not in the image of f_*?
 A: A later paper of Beauville (here, preprint from 2000) seems to address your first question, but more generally by taking direct images of vector bundles.  He makes the following conjecture:
If f: X'-> X is a finite morphism of smooth projective curves and E is a generic vector bundle on X', then f_*E is stable if g(X) \geq 2 and is semistable if g(x) = 1.
The problem is actually equivalent to that of a pushing down a line bundle L (from a different cover).
In the paper, he shows that the conjecture holds with some restrictions on L (e.g. when \chi(X) is small), although obviously one wants it to hold more generally.  He shows, as well, that the conjecture holds as worded above whenever f is an etale covering.
Does this help?
A: Re: first question.
For semistability, we need a homomorphism from a line bundle U -> X of a certain degree to the pushdown of a line bundle L, which is the same thing as having a section on Y of f ^ * U ^ * \otimes L.
This can be expressed in terms of special spaces of divisors, and you can find details worked out explicitly for an example in rank 2 (i.e. a holomorphic double covering f: Y -> X) in pages 103-105 of [NJH,87].
