Cryptosystems defined over the real numbers

Question

In a computational setting it is clear that we cannot do computation over the real numbers, as we couldn't generate, much less store them, so we have to instead deal with numbers that approximate them. If a cryptosystem has some component defined over the real numbers, or uses "real valued" functions such as $\sin$, then does using such systems over the discretized space compromise the theoretical security of the system?

I am particularly thinking of chaos based cryptosystems for example, where encryption consists of iterating a map over an interval of the real line. Are such systems inherently less secure because they cannot be actually implemented over the reals, and instead are approximations? Can such approximations be used to cryptanalyze the systems?

Answer 1

The complex roots of unity form a finite group under multiplication

$$z^k = \exp\bigg(\frac{2k{\pi}i}{n}\bigg) = \cos\frac{2k{\pi}}{n} + i\sin\frac{2k{\pi}}{n},$$

where $z$ is a complex generator of the order-$n$ multiplicative group. This is a cyclic group and is isomorphic to any other cyclic group of the same order, for example the integers modulo 2 under addition, which is equivalent to Boolean XOR

$$a \oplus b \iff z^a z^b = z^{a+b}$$

This equivalence gives rise to the group Fourier transform, in which vector convolution in the complex domain correspond to point-wise multiplication in the additive group. This fact is used by Joan Daemen in his thesis to derive the DC/LC properties for the AES S-BOX.

The obvious practical problem with computing in this group is the error associated with floating-point arithmetic. As the number of arithmetic operations increases, so does the error. This kind of round-off error may be more appropriately called quantization noise in the signal processing context.