Ok, this means that I can have N number of samples, but we have to consider only those bounded region decided by the range of K. This makes it clear.
Now coming to the convention. Delayed samples are always taken to the right axis x(n-k). But in the convolution sum evaluation we take the image and flip it such that it falls under positive axis. This is not clear because as per definition , convolution sum= x
*h(n-k). h(n-k) should be in positive axis only. Why flip it and then multiply accumulate?