Google found me the following, from the IBM JVM FAQ (how's that for an NLA):
When does the Java heap shrink?
Heap shrinkage occurs when GC determines that there is a lot of free heap storage, and releasing some heap memory is beneficial for system performance. Heap shrinkage occurs after GC, but when all the threads are still suspended.
If IBM is doing it, it could be that Sun does something similar.
Edit: They do.
The heap will grow or shrink to a size that will support the chosen throughput goal. Some oscillations in the size of the heap during initialization and during a change in the application's behavior can be expected.
...
It is typical that the size of the heap will oscillate as the garbage collector tries to satisfy competing goals. This is true even if the application has reached a steady state. The pressure to achieve a throughput goal (which may require a larger heap) competes with the goals for a maximum pause time and a minimum footprint (which both may require a small heap).
I suggest you have a look at the rest of that document; it may have more information relevant to your problem.