Understanding why a GAM can't suppress an intercept

  Dear List,

I've just tried to specify a GAM without an intercept -- I've got one
of the (rare) cases where it is appropriate for E(y) -> 0 as X ->0.
Naively running a GAM with the "-1" appended to the formula and the
calling "predict.gam", I see that the model isn't behaving as expected.

I don't understand why this would be.  Google turns up this old R help
thread: http://r.789695.n4.nabble.com/GAM-without-intercept-td4645786.html

Simon writes:

     *Smooth terms are constrained to sum to zero over the covariate
     values. **
     **This is an identifiability constraint designed to avoid
     confounding with **
     **the intercept (particularly important if you have more than one
     smooth). *
     If you remove the intercept from you model altogether (m2) then the
     smooth will still sum to zero over the covariate values, which in
     your
     case will mean that the smooth is quite a long way from the data.
     When
     you include the intercept (m1) then the intercept is effectively
     shifting the constrained curve up towards the data, and you get a
     nice fit.

Why?  I haven't read Simon's book in great detail, though I have read
Ruppert et al.'s Semiparametric Regression.  I don't see a reason why
a penalized spline model shouldn't equal the intercept (or zero) when
all of the regressors equals zero.

Is anyone able to help with a bit of intuition?  Or relevant passages
from a good description of why this would be the case?

Furthermore, why does the "-1" formula specification work if it
doesn't work "as intended" by for example lm?

Many thanks,
Andrew

	[[alternative HTML version deleted]]