Title: A Tighter Lower Bound For Fourier Transform in the Well Conditioned Model

Abstract: The Fourier Transform is one of the most important linear transformations used in science and technology. Cooley and Tukey's Fast Fourier Transform (FFT) from 1964 is a method for computing this transformation in time $O(n\log n)$. In spite of its importance, no nontrivial (super-linear) lower bounds were known without making very strong assumptions about the computation. Morgenstern's result from 1974 provides an $\Omega(n\log n)$ lower bound for the unnormalized Fourier Transform (of determinant $n^{n/2}$), assuming the linear computational model using numbers of at most constant modulus. Ailon shows in 2013 an $\Omega(n\log n)$ for computing the normalized Fourier Transform (of determinant $1$) assuming only unitary operations on two coordinates are allowed at each step, and no extra memory is allowed. In 2014 Ailon then improved the result to show that $\Omega((n\log n)/R)$ operations are required to compute any scaling of the Fourier transform in the well conditioned model, in which the linear transformation defined at any intermediate point has condition number at most $R$. The last result also essentially removed the memory restriction.
In this work we improve the last result to obtain a lower bound of $\Omega((n\log n)/(1 + \log R))$ in the same setting as in [Ailon 2014]. The main additional tool is bounding the quasi-entropy change upon rotation of $2$ rows of a matrix (and its inverse) by embedding in a $4$ dimensional space on which an SVD argument can be used. The implication is that, for example, a linear time Fourier transform algorithm in this model, if exists, would require computation polynomially ill conditioned, instead of just logarithmically as in [Ailon 2014]. Hence, yet more numerical instability would be introduced.