I'm interested in ring/finite field arithmetic. The typical setup for this is a space of integers modulo a large prime number such that there are operators for addition, multiplication and their respective inverses. This comes up a lot in various encryption approaches.
My belief is that you can not do many common operations from machine learning and analytics without taking high order polynomial approximations. For example $\log(1+\exp(x))$ may be approximated using Chebyshev polynomials and if we wished to bound the error by a small constant then we may have to take a very high order polynomials depending on the size of the ring. Another challenge is that you can not simply have approximations for $\exp()$ and $\log()$ and compose them, as the addition of 1 in the above due to the modulo nature of rings. This can be seen for example if we multiply a value by another and add one and then inverse-multiply the value. As values are strictly integers, you do not receive the result as if it were a Real or a rounded version of real numbers. Rather the value is assumed to have some constant times the ring size added on to it such that multiplication-inverse lands you back on an integer.
However, I was challenged on this recently and it was implied I was simplifying things (which I may well be doing so). It would be really helpful if you could confirm if I am correct or incorrect and if the latter then why (or where I should read more etc).
Very much appreciate you help in advance :)
For finite fields or at least Galios Fields of the form GF(p) (where p is a prime number), all numbers are "exact" integers, or for GF(p^k), "exact" polynomials. Exponentiation can be sped up using repeated squaring:
https://en.wikipedia.org/wiki/Exponentiation_by_squaring
Logarithms for small fields can be implemented with a table, but for large fields there is still research being done on this, and it's complicated and/or computationally time consuming.
http://www.dtc.umn.edu/~odlyzko/doc/arch/discrete.logs.pdf