Numerical relativity, applied to collisions of black holes, starts wit
h initial data for black holes already in each other's strong field. F
or the initial data to be astrophysically meaningful, it must approxim
ately represent conditions that evolved from holes originally at large
separation. The initial hypersurface data typically used for computat
ion is based on mathematically simplifying prescriptions, such as conf
ormal flatness of the 3-geometry and longitudinality of the extrinsic
curvature. In the case of head-on collisions of equal-mass holes, ther
e is evidence that such prescriptions work reasonably well, but it is
not clear why, or whether, this success is more generally valid. Here
we study these questions by considering the ''particle limit'' for hea
d on collisions of nonspinning holes, i.e., the limit of an extreme ra
tio of hole masses. The mass of the small hole is considered to be a p
erturbation of the Schwarzschild spacetime of the larger hole, and Ein
stein's equations are linearized in this perturbation and described by
a single gauge-invariant spacetime function psi for each multipole. T
he resulting quadrupole equations have been solved by numerical evolut
ion for collisions starting from various initial separations, and the
evolution is studied on a sequence of hypersurfaces. In particular, we
extract hypersurface data, that is, psi and its time derivative, on s
urfaces of constant background Schwarzschild time. These evolved data
can then be compared with ''prescribed'' data, evolved data can be rep
laced by prescribed data on any hypersurface and evolved further forwa
rd in time, a gauge-invariant measure of devia tion from conformal fla
tness can be evaluated, and other comparisons can be made. The main fi
ndings of this study are (i) for holes of unequal mass the use of pres
cribed data on late hypersurfaces is not successful, (ii) the failure
is likely due to the inability of the prescribed data to represent the
near field of the smaller hole, (iii) the discrepancy in the extrinsi
c curvature is more important than in the 3-geometry, and (iv) the use
of the more general conformally flat longitudinal data does not notab
ly improve this picture. [S0556-28u (97)03320-1].