Improvements are believed to be exponential when simulating physical quantum systems.
Quantum computing have be designed to solve a single problem : the simulation of quantum systems on a classical computer is very slow, lets build a quantum computer so that it will be fast.
It might seem like a fringe use case but it matters (a lot) to industrials and researchers in a wide variety of topics.
I can tell you that Shor's algorithm for factoring takes O((log n)^2 (log log n)(log log log n)) (we can round that up to O((log n)^4) if you want) to a the best known classical algorithm of
O(exp(1.9 * (log n)^(1/3) (log log n)^(2/3)))
But unless I tell you how fast each gate is, that tells you nothing about the constant time factors.
Also, all these quantum algorithms are constructed out of logical gates, but each logical gate might be built out of 10 physical gates, using some error correcting code.
Since you really need to know the physical characteristics of the particular architecture to know how much error correction you need, it's hard to even say how many gates something will use and how fast they are until we pick a particular architecture.
All we can really say is that given a big enough number, a quantum computer can factor it faster than a classic computer.
Iff quantum computing is only useful at breaking encryption, I don't see the point in funding quantum computing research.
Quantum computing is not expected to have large impacts on machine learning at this time after Ewin Tang's paper. There's an especially large amount of fluff in this area, though.
You can get immense speedups for a very particular set of problems, turning them from "can't possibly complete ever" to solvable; so it would be like an assistive device for these niche tasks, but it would be very important for them because otherwise they're not really practical to solve.
I understand that this is the frontier and no one has any definite answers, but I hear everything from "trivial" improvements to "insta-crack AES-256" (thousands of trillion trillion trillion trillions times faster).
Does anyone have a well informed upper/lower bound estimate?