algorithm - SIFT - Having troubles understanding certain steps -

algorithm - SIFT - Having troubles understanding certain steps -

currently working on sift-algorithm , trying understand it.

i have problem understanding fitting process done after histogram of gradients has been determined. paper says take 2 bins next peak , fitting process. x-axis , y-axis , around point evolve taylor series?

in histogram of gradients magnitude weights? magnitude of gradient of pixel (i,j)?

in 3rd step supposed create feature descriptor. paper states have take 16blocks each 4x4 around keypoint , weight them magnitude weight , gaussian weight. how big should standard deviation of gaussian step? afaik, need cover 256pixels gauss kernel has big enough. also, happens if our keypoint on image border , cant fit 16blocks on image plane?

algorithm sift