Isospeed convolutions

In [PLSK11] and [Lie17], it was observed that one can use pseudoinverses to switch from a (min,+) computation to a (max,+) one, and back (under some non-restrictive hypotheses). Furthermore, doing so would often result in significant performance improvements.

The isomorphism between (min,+) and (max,+) convolution.

In [ZNS23a] we explored this further, finding the main properties that cause these performance and devising a new algorithm, Isospeed, which exploits them to gain significant performance improvements over both the state-of-art (min,+) and (max,+) convolution algorithms.

These improvements are implemented in Nancy transparently to the user, so that no action or configuration in necessary to get these improvements.

For more details, check out the paper [ZNS23a] or [Zippo23], Chapter 14.