Now I saw the code:

The situation is like this:

If `lag' is non-negative, a recalculation starts at most `lag' milleseconds
after each change.  That means that if `lag' is zero, all auto expressions are
recalculated immediately after each change.
If `lag' is negative, the recalculation is started |lag| milleseconds after last
change.  In other words, gnumeric waits for |lag| milleseconds after each
change; if no other change appears during that period, a recalculation is started.

Another proposed formulation:

If `lag' were 0, gnumeric would recalculate all auto expressions immediately
after  each change.  Non-zero values allow gnumeric to acumulate more changes
before it does the recalculation.
If `lag' is positive, then whenever a change appears, gnumeric waits that many
milleseconds and then recalculates; if more changes appears during that period,
it's processed together with this one.
If `lag' is negative, then recalculation happens only after a quiet period of
|lag| milleseconds.

In other words:
If `lag' is non-zero, gnumeric accumulates changes. If the oldest (for lag
positive) or newest (for lag negative) change is |lag| milleseconds old, it
recalculates all the auto expressions.

Please choose the best parts of my proposals.

I have changed the schema to include essentially your last description. Please
note that the internal settings have been removed from the preference dialog, so
this description is only visible for those using the gconf-editor to finetune
their settings.

Thanks!  It's always good to have a comprehensive description (I got there when
I grepped the source tree).