Question

I know there is something wrong with the following reasoning but I’m not sure what it is.

The FFT:

1. given two polynomials

   A = a_0 + a_1 x + a_2 x^2 + ... + a_n x^n

   and

   B = b_0 + b_1 x + b_2 x^2 + ... + b_n x^n

   you can compute the coefficients of the product

   AB = \sum _k = 0 ^ 2n ( \sum _ j = 0 ^ k (a_j b_{k-j}))x^k

   in O(n log n ) time.

2. So given two vectors (a_0, ..., a_n) and (b_0, ..., b_n) we can calculate
   the vector v_i = \sum j = 0 ^ k ( a_j b_{k-j}) in O(n log n) time (by embedding the vectors in zeros.)

3. Given the above, we should be able to calculate the dot product of A =(a_0, ..., a_n) and B =(b_0, ..., b_n) which is A.B = \sum_j=0 ^ n a_j b_j in O(n log n) time by preprocessing one of the vectors say B to be B' = (b_n, b_{n-1}, ..., b_1, b_0) then computing the convolution as in 2. in O(n log n) time.

If the above reasoning is correct, then that means we can implement Matrix Multiplication of two nxn matrices in O(n^2 log n ) time by computing the dot product in O(n log n) time O(n) times.

However, the best run-time for matrix multiplication we know is about O(n^2.4) so this seems unlikely to be true, which probably means of of the steps 1,2, or 3 is incorrect.

Answer 1

There are n^2 entries in the product, not n and so this algorithm would be O(n^2 * n * log n) = O(n^3 log n).

And the best algorithm for computing dot products is O(n), not O(n log n). That’s why the naive algorithm for matrix multiplication is O(n^3). (It’s n^2 dot products that can be done in O(n) time).