Regression is a good analogy of the problem here. If you found a line of best fit for some datapoints, how would you get back the original datapoints, from the line?
Now imagine terabytes worth of datapoints, and thousands of dimensions rather than two.