This subject is somewhat outside of my area of knowledge, but if I had to guess, this is what resolves the paradox. An abstract mathematical chaotic system is different than a real-life one, and it is my guess that if we performed the experiment, we would find the orbits would be different.
QM has withstood every test thrown at it so far, and while it's always possible that a correction to it will be found, I think it will be a very rare event should it happen.
Is that really true? Losslessly? Do you mean the entire phase space, or the sequence generated on any given run? Because I was under the impression that what distinguishes a chaotic system is that the output cannot be predicted in any way other than running the simulation (the article uses the word 'random', but it might be better to say 'unpredictable'). Compressibility relies on exploiting regularities, and regularity is predictability, isn't it?
The program that simulates the chaotic system is shorter than the unbounded amount of output being generated, and counts as a compression of the output.
I guess in other words, the more interesting question is this: can we generate hard to compress pseudorandom sequences with simple algorithms - if the algorithm description by itself can not be used as "compression"?
If you add a rule such as not using Algorithm X, it's trivial to create an Algorithm Y that generates the same output and it's barely any longer than X. For example, by adding any meaningless instruction at the start of algorithm X.
Of course, discovering what the shortest expression for a given sequence is is not a computable problem. So we can't always know if we've found the best form of compression for a sequence, we can only find an upper bound.
Or does 'simple' mean like an extractor for a standard archive file? Did you know that RAR is turing complete?
Maybe 'simple' has to be defined by limiting how much execution time you have per output byte?
You're confusing theoretical compressibility with the actual compression ratios you get when you put a string through typical compression functions.
Here's a different example. Suppose you have ASCII text encrypted with AES with the AES decryption key appended to it. Run that through gzip and you'll probably get output which is larger than the input. But here's a different compression algorithm that will work much better: Decrypt the encrypted data with the key and then compress the plaintext with gzip.
It isn't that the original string is of the sort that can't be compressed, it's that you need a compression function which is suited to the input. Pi is like that.
Compare this to other strings, like the output of an environmental hardware random number generator, which will almost always be totally incompressible whatsoever.
In randomness theory, the definition of the most compressed form of a sequence S is, given a computer C, the shortest program with output S in C.
So the program that generates the output of the simulation of the pendulums is a compression of it's output. Since it's and infinite sequence that can be compressed into a finite program, it is not random from the perspective of information theory or randomness theory.
My guess: because the abstract double pendulum is a model that only approximates real double pendulums. Analogously, computers based on Turing Machines exist despite Turing Machines requiring an infinite amount of space.
---
Also, a minor nitpick: the post seems to equate 'not periodic' or 'chaotic' with 'is random', which I think is misleading. Pseudo-random, I guess, but the output of a simulation of a chaotic system is highly compressible and thus not random in the information theoretic sense.