. What fundamental problem does forward mapping (pushing each source pixel to its transformed location) have?
When a transform enlarges or rotates an image, destination locations spread out and some destination pixels receive no source pixel at all.
. Why does inverse mapping avoid the holes problem that forward mapping has?
Since every destination pixel is visited and mapped back via M_inv, every output pixel is guaranteed to get a value.
. Why does inverse mapping need interpolation at all?
The mapped-back coordinate is generally fractional, so a value must be estimated from the surrounding source pixels.
. What is the key trade-off between nearest-neighbor and bilinear interpolation?
Bilinear blends the 4 neighboring pixels, removing the blocky aliasing nearest-neighbor produces at the cost of a softer image.